Upload folder using huggingface_hub

README.md

@@ -0,0 +1,202 @@
+---
+base_model: meta-llama/Llama-3.2-11B-Vision-Instruct
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.15.2

adapter_config.json

@@ -0,0 +1,31 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "/data1/chenjiazun/LLM/Llama-3.2-11B-Vision-Instruct",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": [],
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": "(?!^(multi_modal_projector))^(language_model|vision_model).*\\.(down_proj|gate_proj|v_proj|q_proj|o_proj|k_proj|up_proj|fc1|fc2|v_proj|q_proj|o_proj|k_proj)$",
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d745b7006942c8219cab2941dcf8a814443536cf05114088016b8902b9b15291
+size 134506880

additional_config.json

@@ -0,0 +1 @@
+{"lora_dtype": null, "lorap_lr_ratio": null, "lorap_emb_lr": 1e-06}

args.json

@@ -0,0 +1,450 @@
+{
+  "model": "/data1/chenjiazun/LLM/Llama-3.2-11B-Vision-Instruct",
+  "model_type": "llama3_2_vision",
+  "model_revision": null,
+  "task_type": "causal_lm",
+  "torch_dtype": "bfloat16",
+  "attn_impl": null,
+  "num_labels": null,
+  "problem_type": null,
+  "rope_scaling": null,
+  "device_map": null,
+  "max_memory": {},
+  "local_repo_path": null,
+  "template": "llama3_2_vision",
+  "system": null,
+  "max_length": 8092,
+  "truncation_strategy": "left",
+  "max_pixels": null,
+  "agent_template": null,
+  "norm_bbox": null,
+  "response_prefix": null,
+  "padding_side": "right",
+  "loss_scale": "last_round",
+  "sequence_parallel_size": 1,
+  "use_chat_template": true,
+  "template_backend": "swift",
+  "dataset": [
+    "./new_task1_nothink_respone_onlycorrect.json"
+  ],
+  "val_dataset": [],
+  "split_dataset_ratio": 0.01,
+  "data_seed": 42,
+  "dataset_num_proc": 1,
+  "dataset_shuffle": true,
+  "val_dataset_shuffle": false,
+  "streaming": false,
+  "interleave_prob": null,
+  "stopping_strategy": "first_exhausted",
+  "shuffle_buffer_size": 1000,
+  "enable_cache": false,
+  "download_mode": "reuse_dataset_if_exists",
+  "columns": {},
+  "strict": false,
+  "remove_unused_columns": false,
+  "model_name": [
+    null,
+    null
+  ],
+  "model_author": [
+    null,
+    null
+  ],
+  "custom_dataset_info": [],
+  "quant_method": null,
+  "quant_bits": null,
+  "hqq_axis": null,
+  "bnb_4bit_compute_dtype": "bfloat16",
+  "bnb_4bit_quant_type": "nf4",
+  "bnb_4bit_use_double_quant": true,
+  "bnb_4bit_quant_storage": null,
+  "max_new_tokens": 64,
+  "temperature": 1.0,
+  "top_k": 50,
+  "top_p": 0.9,
+  "repetition_penalty": 1.1,
+  "num_beams": 1,
+  "stream": false,
+  "stop_words": [],
+  "logprobs": false,
+  "top_logprobs": null,
+  "ckpt_dir": null,
+  "load_dataset_config": null,
+  "lora_modules": [],
+  "tuner_backend": "peft",
+  "train_type": "lora",
+  "adapters": [],
+  "external_plugins": [
+    "/data/kdd/crag/cjz/ms-swift/examples/train/grpo/plugin/plugin.py"
+  ],
+  "seed": 42,
+  "model_kwargs": {},
+  "load_args": false,
+  "load_data_args": false,
+  "use_hf": false,
+  "hub_token": null,
+  "custom_register_path": [],
+  "ignore_args_error": false,
+  "use_swift_lora": false,
+  "output_dir": "/data2/kdd/crag/cjz/agents/output/response没kl全程reward2/v10-20250518-175827",
+  "overwrite_output_dir": false,
+  "do_train": false,
+  "do_eval": false,
+  "do_predict": false,
+  "eval_strategy": "steps",
+  "prediction_loss_only": false,
+  "per_device_train_batch_size": 8,
+  "per_device_eval_batch_size": 8,
+  "per_gpu_train_batch_size": null,
+  "per_gpu_eval_batch_size": null,
+  "gradient_accumulation_steps": 16,
+  "eval_accumulation_steps": null,
+  "eval_delay": 0,
+  "torch_empty_cache_steps": null,
+  "learning_rate": 0.0002,
+  "weight_decay": 0.1,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.95,
+  "adam_epsilon": 1e-08,
+  "max_grad_norm": 0.5,
+  "num_train_epochs": 4.0,
+  "max_steps": -1,
+  "lr_scheduler_type": "cosine",
+  "lr_scheduler_kwargs": null,
+  "warmup_ratio": 0.05,
+  "warmup_steps": 0,
+  "log_level": "passive",
+  "log_level_replica": "warning",
+  "log_on_each_node": true,
+  "logging_dir": "/data2/kdd/crag/cjz/agents/output/response没kl全程reward2/v10-20250518-175827/runs",
+  "logging_strategy": "steps",
+  "logging_first_step": true,
+  "logging_steps": 1,
+  "logging_nan_inf_filter": true,
+  "save_strategy": "steps",
+  "save_steps": 5.0,
+  "save_total_limit": null,
+  "save_safetensors": true,
+  "save_on_each_node": false,
+  "save_only_model": false,
+  "restore_callback_states_from_checkpoint": false,
+  "no_cuda": false,
+  "use_cpu": false,
+  "use_mps_device": false,
+  "jit_mode_eval": false,
+  "use_ipex": false,
+  "bf16": true,
+  "fp16": false,
+  "fp16_opt_level": "O1",
+  "half_precision_backend": "auto",
+  "bf16_full_eval": false,
+  "fp16_full_eval": false,
+  "tf32": null,
+  "local_rank": 0,
+  "ddp_backend": null,
+  "tpu_num_cores": null,
+  "tpu_metrics_debug": false,
+  "debug": null,
+  "dataloader_drop_last": true,
+  "eval_steps": 500.0,
+  "dataloader_num_workers": 4,
+  "dataloader_prefetch_factor": null,
+  "past_index": -1,
+  "run_name": null,
+  "disable_tqdm": null,
+  "label_names": null,
+  "load_best_model_at_end": false,
+  "metric_for_best_model": "reward",
+  "greater_is_better": true,
+  "ignore_data_skip": false,
+  "fsdp": "",
+  "fsdp_min_num_params": 0,
+  "fsdp_config": null,
+  "tp_size": 0,
+  "fsdp_transformer_layer_cls_to_wrap": null,
+  "accelerator_config": {
+    "dispatch_batches": false
+  },
+  "deepspeed": {
+    "fp16": {
+      "enabled": "auto",
+      "loss_scale": 0,
+      "loss_scale_window": 1000,
+      "initial_scale_power": 16,
+      "hysteresis": 2,
+      "min_loss_scale": 1
+    },
+    "bf16": {
+      "enabled": "auto"
+    },
+    "zero_optimization": {
+      "stage": 3,
+      "offload_optimizer": {
+        "device": "none",
+        "pin_memory": true
+      },
+      "offload_param": {
+        "device": "none",
+        "pin_memory": true
+      },
+      "overlap_comm": false,
+      "contiguous_gradients": true,
+      "sub_group_size": 1000000000.0,
+      "reduce_bucket_size": "auto",
+      "zero_quantized_weights": false,
+      "zero_quantized_gradients": false,
+      "stage3_prefetch_bucket_size": "auto",
+      "stage3_param_persistence_threshold": "auto",
+      "stage3_max_live_parameters": 1000000000.0,
+      "stage3_max_reuse_distance": 1000000000.0,
+      "stage3_gather_16bit_weights_on_model_save": true
+    },
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "steps_per_print": 2000,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+  },
+  "label_smoothing_factor": 0.0,
+  "optim": "adamw_torch",
+  "optim_args": null,
+  "adafactor": false,
+  "group_by_length": false,
+  "length_column_name": "length",
+  "report_to": [
+    "tensorboard"
+  ],
+  "ddp_find_unused_parameters": null,
+  "ddp_bucket_cap_mb": null,
+  "ddp_broadcast_buffers": null,
+  "dataloader_pin_memory": true,
+  "dataloader_persistent_workers": false,
+  "skip_memory_metrics": true,
+  "use_legacy_prediction_loop": false,
+  "push_to_hub": false,
+  "resume_from_checkpoint": "/data2/kdd/crag/cjz/agents/output/response没kl全程reward2/v9-20250518-172344/checkpoint-15",
+  "hub_model_id": null,
+  "hub_strategy": "every_save",
+  "hub_private_repo": null,
+  "hub_always_push": false,
+  "gradient_checkpointing": true,
+  "gradient_checkpointing_kwargs": null,
+  "include_inputs_for_metrics": false,
+  "include_for_metrics": [],
+  "eval_do_concat_batches": true,
+  "fp16_backend": "auto",
+  "push_to_hub_model_id": null,
+  "push_to_hub_organization": null,
+  "push_to_hub_token": null,
+  "mp_parameters": "",
+  "auto_find_batch_size": false,
+  "full_determinism": false,
+  "torchdynamo": null,
+  "ray_scope": "last",
+  "ddp_timeout": 1800,
+  "torch_compile": false,
+  "torch_compile_backend": null,
+  "torch_compile_mode": null,
+  "include_tokens_per_second": false,
+  "include_num_input_tokens_seen": false,
+  "neftune_noise_alpha": null,
+  "optim_target_modules": null,
+  "batch_eval_metrics": false,
+  "eval_on_start": false,
+  "use_liger_kernel": false,
+  "eval_use_gather_object": false,
+  "average_tokens_across_devices": false,
+  "sortish_sampler": false,
+  "predict_with_generate": false,
+  "generation_max_length": null,
+  "generation_num_beams": null,
+  "generation_config": null,
+  "check_model": true,
+  "acc_strategy": "token",
+  "train_dataloader_shuffle": true,
+  "metric_warmup_step": 0,
+  "fsdp_num": 1,
+  "acc_steps": 1,
+  "eval_use_evalscope": false,
+  "eval_datasets": [],
+  "eval_limit": null,
+  "eval_datasets_args": null,
+  "eval_generation_config": null,
+  "freeze_parameters": [
+    "multi_modal_projector"
+  ],
+  "freeze_parameters_ratio": 0.0,
+  "trainable_parameters": [],
+  "freeze_llm": false,
+  "freeze_vit": false,
+  "freeze_aligner": true,
+  "target_modules": [
+    "all-linear"
+  ],
+  "target_regex": null,
+  "modules_to_save": [],
+  "lora_rank": 16,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "lora_bias": "none",
+  "lora_dtype": null,
+  "lorap_lr_ratio": null,
+  "use_rslora": false,
+  "use_dora": false,
+  "lora_ga_batch_size": 2,
+  "lora_ga_iters": 2,
+  "lora_ga_max_length": 1024,
+  "lora_ga_direction": "ArB2r",
+  "lora_ga_scale": "stable",
+  "lora_ga_stable_gamma": 16,
+  "init_weights": true,
+  "fourier_n_frequency": 2000,
+  "fourier_scaling": 300.0,
+  "boft_block_size": 4,
+  "boft_block_num": 0,
+  "boft_n_butterfly_factor": 1,
+  "boft_dropout": 0.0,
+  "vera_rank": 256,
+  "vera_projection_prng_key": 0,
+  "vera_dropout": 0.0,
+  "vera_d_initial": 0.1,
+  "adapter_act": "gelu",
+  "adapter_length": 128,
+  "use_galore": false,
+  "galore_target_modules": null,
+  "galore_rank": 128,
+  "galore_update_proj_gap": 50,
+  "galore_scale": 1.0,
+  "galore_proj_type": "std",
+  "galore_optim_per_parameter": false,
+  "galore_with_embedding": false,
+  "galore_quantization": false,
+  "galore_proj_quant": false,
+  "galore_proj_bits": 4,
+  "galore_proj_group_size": 256,
+  "galore_cos_threshold": 0.4,
+  "galore_gamma_proj": 2,
+  "galore_queue_size": 5,
+  "adalora_target_r": 8,
+  "adalora_init_r": 12,
+  "adalora_tinit": 0,
+  "adalora_tfinal": 0,
+  "adalora_deltaT": 1,
+  "adalora_beta1": 0.85,
+  "adalora_beta2": 0.85,
+  "adalora_orth_reg_weight": 0.5,
+  "llamapro_num_new_blocks": 4,
+  "llamapro_num_groups": null,
+  "lisa_activated_layers": 0,
+  "lisa_step_interval": 20,
+  "reft_layer_key": null,
+  "reft_layers": null,
+  "reft_rank": 4,
+  "reft_intervention_type": "LoreftIntervention",
+  "reft_args": null,
+  "swanlab_token": null,
+  "swanlab_project": null,
+  "swanlab_workspace": null,
+  "swanlab_exp_name": null,
+  "swanlab_mode": "cloud",
+  "add_version": true,
+  "resume_only_model": false,
+  "create_checkpoint_symlink": false,
+  "packing": false,
+  "lazy_tokenize": true,
+  "loss_type": "grpo",
+  "optimizer": null,
+  "metric": null,
+  "zero_hpz_partition_size": null,
+  "reward_model": null,
+  "reward_adapters": [],
+  "reward_model_type": null,
+  "reward_model_revision": null,
+  "num_ppo_epochs": 4,
+  "whiten_rewards": false,
+  "kl_coef": 0.05,
+  "cliprange": 0.2,
+  "vf_coef": 0.1,
+  "cliprange_value": 0.2,
+  "gamma": 1.0,
+  "lam": 0.95,
+  "num_mini_batches": 1,
+  "local_rollout_forward_batch_size": 64,
+  "num_sample_generations": 10,
+  "response_length": 512,
+  "missing_eos_penalty": null,
+  "epsilon": 0.2,
+  "epsilon_high": 0.28,
+  "num_infer_workers": 1,
+  "vllm_device": [
+    "auto"
+  ],
+  "vllm_gpu_memory_utilization": 0.8,
+  "vllm_max_model_len": 8192,
+  "vllm_max_num_seqs": 8,
+  "vllm_enforce_eager": false,
+  "vllm_limit_mm_per_prompt": null,
+  "vllm_enable_prefix_caching": true,
+  "cosine_min_len_value_wrong": -0.5,
+  "cosine_max_len_value_wrong": 0.0,
+  "cosine_min_len_value_correct": 1.0,
+  "cosine_max_len_value_correct": 0.5,
+  "cosine_max_len": null,
+  "repetition_n_grams": 3,
+  "repetition_max_penalty": -1.0,
+  "use_lmdeploy": false,
+  "lmdeploy_device": "auto",
+  "lmdeploy_session_len": null,
+  "lmdeploy_cache_max_entry_count": 0.8,
+  "async_generate": false,
+  "tensor_parallel_size": 1,
+  "sleep_level": 0,
+  "move_model_batches": null,
+  "offload_optimizer": false,
+  "offload_model": false,
+  "gc_collect_after_offload": false,
+  "multi_turn_func": null,
+  "dynamic_sample": false,
+  "max_resample_times": 3,
+  "overlong_filter": false,
+  "soft_max_length": null,
+  "soft_cache_length": 20,
+  "scale_rewards": true,
+  "wandb_log_unique_prompts": null,
+  "vllm_server_host": null,
+  "vllm_server_port": 8000,
+  "vllm_server_timeout": 240.0,
+  "num_generations": 8,
+  "max_completion_length": 50,
+  "ds3_gather_for_generation": true,
+  "reward_funcs": [
+    "Response_no_think_reward"
+  ],
+  "reward_weights": null,
+  "log_completions": true,
+  "use_vllm": true,
+  "num_iterations": 2,
+  "rlhf_type": "grpo",
+  "ref_model": null,
+  "ref_model_type": null,
+  "ref_model_revision": null,
+  "beta": 0.0001,
+  "label_smoothing": 0,
+  "rpo_alpha": 1.0,
+  "cpo_alpha": 1.0,
+  "simpo_gamma": 1,
+  "desirable_weight": 1.0,
+  "undesirable_weight": 1.0,
+  "center_rewards_coefficient": null,
+  "rank": 0,
+  "global_world_size": 1,
+  "local_world_size": 1,
+  "model_suffix": "Llama-3.2-11B-Vision-Instruct",
+  "model_info": "ModelInfo(model_type='llama3_2_vision', model_dir='/data1/chenjiazun/LLM/Llama-3.2-11B-Vision-Instruct', torch_dtype=torch.bfloat16, max_model_len=131072, quant_method=None, quant_bits=None, rope_scaling={'factor': 8.0, 'high_freq_factor': 4.0, 'low_freq_factor': 1.0, 'original_max_position_embeddings': 8192, 'rope_type': 'llama3'}, config=None, task_type='causal_lm', num_labels=None)",
+  "model_meta": "ModelMeta(model_type='llama3_2_vision', model_groups=[ModelGroup(models=[Model(ms_model_id='LLM-Research/Llama-3.2-11B-Vision-Instruct', hf_model_id='meta-llama/Llama-3.2-11B-Vision-Instruct', model_path=None, ms_revision=None, hf_revision=None), Model(ms_model_id='LLM-Research/Llama-3.2-90B-Vision-Instruct', hf_model_id='meta-llama/Llama-3.2-90B-Vision-Instruct', model_path=None, ms_revision=None, hf_revision=None), Model(ms_model_id='LLM-Research/Llama-3.2-11B-Vision', hf_model_id='meta-llama/Llama-3.2-11B-Vision', model_path=None, ms_revision=None, hf_revision=None), Model(ms_model_id='LLM-Research/Llama-3.2-90B-Vision', hf_model_id='meta-llama/Llama-3.2-90B-Vision', model_path=None, ms_revision=None, hf_revision=None)], ignore_patterns=None, requires=None, tags=[])], template='llama3_2_vision', get_function=<function get_model_tokenizer_llama3_2_vision at 0x7e5165f700d0>, model_arch='llama3_2_vision', architectures=['MllamaForConditionalGeneration'], additional_saved_files=[], torch_dtype=None, is_multimodal=True, is_reward=False, task_type=None, ignore_patterns=None, requires=['transformers>=4.45'], tags=[])",
+  "model_dir": "/data1/chenjiazun/LLM/Llama-3.2-11B-Vision-Instruct",
+  "hub": "<class 'swift.hub.hub.MSHub'>",
+  "evaluation_strategy": "steps",
+  "training_args": "GRPOConfig(output_dir='/data2/kdd/crag/cjz/agents/output/response没kl全程reward2/v10-20250518-175827', overwrite_output_dir=False, do_train=False, do_eval=True, do_predict=False, eval_strategy=<IntervalStrategy.STEPS: 'steps'>, prediction_loss_only=False, per_device_train_batch_size=8, per_device_eval_batch_size=8, per_gpu_train_batch_size=None, per_gpu_eval_batch_size=None, gradient_accumulation_steps=16, eval_accumulation_steps=None, eval_delay=0, torch_empty_cache_steps=None, learning_rate=0.0002, weight_decay=0.1, adam_beta1=0.9, adam_beta2=0.95, adam_epsilon=1e-08, max_grad_norm=0.5, num_train_epochs=4.0, max_steps=-1, lr_scheduler_type=<SchedulerType.COSINE: 'cosine'>, lr_scheduler_kwargs=None, warmup_ratio=0.05, warmup_steps=0, log_level='passive', log_level_replica='warning', log_on_each_node=True, logging_dir='/data2/kdd/crag/cjz/agents/output/response没kl全程reward2/v10-20250518-175827/runs', logging_strategy=<IntervalStrategy.STEPS: 'steps'>, logging_first_step=True, logging_steps=1, logging_nan_inf_filter=True, save_strategy=<SaveStrategy.STEPS: 'steps'>, save_steps=5, save_total_limit=None, save_safetensors=True, save_on_each_node=False, save_only_model=False, restore_callback_states_from_checkpoint=False, no_cuda=False, use_cpu=False, use_mps_device=False, seed=42, data_seed=42, jit_mode_eval=False, use_ipex=False, bf16=True, fp16=False, fp16_opt_level='O1', half_precision_backend='auto', bf16_full_eval=False, fp16_full_eval=False, tf32=None, local_rank=0, ddp_backend=None, tpu_num_cores=None, tpu_metrics_debug=False, debug=[], dataloader_drop_last=True, eval_steps=500, dataloader_num_workers=4, dataloader_prefetch_factor=10, past_index=-1, run_name='/data2/kdd/crag/cjz/agents/output/response没kl全程reward2/v10-20250518-175827', disable_tqdm=False, remove_unused_columns=False, label_names=None, load_best_model_at_end=False, metric_for_best_model='reward', greater_is_better=True, ignore_data_skip=False, fsdp=[], fsdp_min_num_params=0, fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}, tp_size=0, fsdp_transformer_layer_cls_to_wrap=None, accelerator_config=AcceleratorConfig(split_batches=False, dispatch_batches=False, even_batches=True, use_seedable_sampler=True, non_blocking=False, gradient_accumulation_kwargs=None, use_configured_state=False), deepspeed={'fp16': {'enabled': 'auto', 'loss_scale': 0, 'loss_scale_window': 1000, 'initial_scale_power': 16, 'hysteresis': 2, 'min_loss_scale': 1}, 'bf16': {'enabled': 'auto'}, 'zero_optimization': {'stage': 3, 'offload_optimizer': {'device': 'none', 'pin_memory': True}, 'offload_param': {'device': 'none', 'pin_memory': True}, 'overlap_comm': False, 'contiguous_gradients': True, 'sub_group_size': 1000000000.0, 'reduce_bucket_size': 'auto', 'zero_quantized_weights': False, 'zero_quantized_gradients': False, 'stage3_prefetch_bucket_size': 0, 'stage3_param_persistence_threshold': 'auto', 'stage3_max_live_parameters': 1000000000.0, 'stage3_max_reuse_distance': 1000000000.0, 'stage3_gather_16bit_weights_on_model_save': True}, 'gradient_accumulation_steps': 'auto', 'gradient_clipping': 'auto', 'steps_per_print': 2000, 'train_batch_size': 'auto', 'train_micro_batch_size_per_gpu': 'auto', 'wall_clock_breakdown': False}, label_smoothing_factor=0.0, optim=<OptimizerNames.ADAMW_TORCH: 'adamw_torch'>, optim_args=None, adafactor=False, group_by_length=False, length_column_name='length', report_to=['tensorboard'], ddp_find_unused_parameters=None, ddp_bucket_cap_mb=None, ddp_broadcast_buffers=None, dataloader_pin_memory=True, dataloader_persistent_workers=False, skip_memory_metrics=True, use_legacy_prediction_loop=False, push_to_hub=False, resume_from_checkpoint='/data2/kdd/crag/cjz/agents/output/response没kl全程reward2/v9-20250518-172344/checkpoint-15', hub_model_id=None, hub_strategy=<HubStrategy.EVERY_SAVE: 'every_save'>, hub_token=None, hub_private_repo=None, hub_always_push=False, gradient_checkpointing=True, gradient_checkpointing_kwargs=None, include_inputs_for_metrics=False, include_for_metrics=[], eval_do_concat_batches=True, fp16_backend='auto', push_to_hub_model_id=None, push_to_hub_organization=None, push_to_hub_token=None, mp_parameters='', auto_find_batch_size=False, full_determinism=False, torchdynamo=None, ray_scope='last', ddp_timeout=1800, torch_compile=False, torch_compile_backend=None, torch_compile_mode=None, include_tokens_per_second=None, include_num_input_tokens_seen=None, neftune_noise_alpha=None, optim_target_modules=None, batch_eval_metrics=False, eval_on_start=False, use_liger_kernel=False, eval_use_gather_object=False, average_tokens_across_devices=None, model_init_kwargs=None, disable_dropout=False, max_prompt_length=512, num_generations=8, max_completion_length=50, ds3_gather_for_generation=True, shuffle_dataset=True, temperature=1.0, top_p=0.9, top_k=50, min_p=None, repetition_penalty=1.1, cache_implementation=None, use_vllm=True, vllm_server_host=None, vllm_server_port=8000, vllm_server_timeout=240.0, vllm_guided_decoding_regex=None, beta=0.0001, num_iterations=2, epsilon=0.2, epsilon_high=0.28, reward_weights=None, scale_rewards=True, loss_type='grpo', mask_truncated_completions=False, sync_ref_model=False, ref_model_mixup_alpha=0.6, ref_model_sync_steps=512, use_liger_loss=False, log_completions=True, num_completions_to_print=None, wandb_log_unique_prompts=None, vllm_device=['auto'], vllm_gpu_memory_utilization=0.8, vllm_dtype=None, vllm_max_model_len=8192, vllm_enable_prefix_caching=True, check_model=True, acc_strategy='token', train_dataloader_shuffle=True, metric_warmup_step=0, fsdp_num=1, acc_steps=1, eval_use_evalscope=False, eval_datasets=[], eval_limit=None, eval_datasets_args=None, eval_generation_config=None, train_type='lora', optimizer=None, local_repo_path=None, galore_config=None, num_infer_workers=1, vllm_max_num_seqs=8, vllm_enforce_eager=False, vllm_limit_mm_per_prompt={}, cosine_min_len_value_wrong=-0.5, cosine_max_len_value_wrong=0.0, cosine_min_len_value_correct=1.0, cosine_max_len_value_correct=0.5, cosine_max_len=50, repetition_n_grams=3, repetition_max_penalty=-1.0, use_lmdeploy=False, lmdeploy_device='auto', lmdeploy_session_len=None, lmdeploy_cache_max_entry_count=0.8, async_generate=False, tensor_parallel_size=1, sleep_level=0, move_model_batches=None, offload_optimizer=False, offload_model=False, gc_collect_after_offload=False, multi_turn_func=None, dynamic_sample=False, max_resample_times=3, overlong_filter=False, soft_max_length=None, soft_cache_length=20, dataset_shuffle=True, stop_words=[])"
+}

latest

@@ -0,0 +1 @@
+global_step60

rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2e7e5b117772cf3e740827ee783ebfdfbe65ff2112ad28200667a239048b726
+size 14356

scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:994d867989c1edc3f014136481ad1ace3980227fef482846b88cb97090b177f3
+size 1056

trainer_state.json

@@ -0,0 +1,1190 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.6648199445983379,
+  "eval_steps": 500,
+  "global_step": 60,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.046875,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 16.6796875,
+      "completions/min_length": 3.0,
+      "epoch": 0.0110803324099723,
+      "grad_norm": 1.3400153486450683,
+      "kl": 0.0,
+      "learning_rate": 1.1111111111111112e-05,
+      "loss": -9.313225746154785e-10,
+      "memory(GiB)": 77.38,
+      "reward": -0.8125,
+      "reward_std": 0.38430261611938477,
+      "rewards/Response_no_think_reward/mean": -0.8125,
+      "rewards/Response_no_think_reward/std": 0.6958821415901184,
+      "step": 1,
+      "train_speed(iter/s)": 0.000753
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "epoch": 0.0221606648199446,
+      "grad_norm": 1.3454285871298992,
+      "kl": 0.0,
+      "learning_rate": 2.2222222222222223e-05,
+      "loss": -9.313225746154785e-10,
+      "memory(GiB)": 78.17,
+      "step": 2,
+      "train_speed(iter/s)": 0.001317
+    },
+    {
+      "clip_ratio/high_max": 0.0009920635493472219,
+      "clip_ratio/high_mean": 0.0009920635493472219,
+      "clip_ratio/low_mean": 0.0012351538171060383,
+      "clip_ratio/low_min": 0.0012351538171060383,
+      "clip_ratio/region_mean": 0.00222721736645326,
+      "completions/clipped_ratio": 0.0390625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 13.03125,
+      "completions/min_length": 3.0,
+      "epoch": 0.0332409972299169,
+      "grad_norm": 1.1739378406354548,
+      "kl": 0.0013750234156759689,
+      "learning_rate": 3.3333333333333335e-05,
+      "loss": 0.0005532659706659615,
+      "memory(GiB)": 78.17,
+      "reward": -0.0078125,
+      "reward_std": 0.7346676588058472,
+      "rewards/Response_no_think_reward/mean": -0.0078125,
+      "rewards/Response_no_think_reward/std": 1.411405086517334,
+      "step": 3,
+      "train_speed(iter/s)": 0.001417
+    },
+    {
+      "clip_ratio/high_max": 0.0006313131307251751,
+      "clip_ratio/high_mean": 0.0006313131307251751,
+      "clip_ratio/low_mean": 0.0009640990756452084,
+      "clip_ratio/low_min": 0.0009640990756452084,
+      "clip_ratio/region_mean": 0.0015954122063703835,
+      "epoch": 0.0443213296398892,
+      "grad_norm": 1.1825343839681348,
+      "kl": 0.0013630353023472708,
+      "learning_rate": 4.4444444444444447e-05,
+      "loss": -0.0013326075859367847,
+      "memory(GiB)": 78.17,
+      "step": 4,
+      "train_speed(iter/s)": 0.001735
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.005316304974257946,
+      "clip_ratio/low_min": 0.005316304974257946,
+      "clip_ratio/region_mean": 0.005316304974257946,
+      "completions/clipped_ratio": 0.0078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 13.3828125,
+      "completions/min_length": 3.0,
+      "epoch": 0.055401662049861494,
+      "grad_norm": 1.3399484117828366,
+      "kl": 0.0029429409933072748,
+      "learning_rate": 5.555555555555556e-05,
+      "loss": 0.0014606926124542952,
+      "memory(GiB)": 78.25,
+      "reward": -0.15625,
+      "reward_std": 0.9831876158714294,
+      "rewards/Response_no_think_reward/mean": -0.15625,
+      "rewards/Response_no_think_reward/std": 1.3541275262832642,
+      "step": 5,
+      "train_speed(iter/s)": 0.001752
+    },
+    {
+      "clip_ratio/high_max": 0.0014204545877873898,
+      "clip_ratio/high_mean": 0.0014204545877873898,
+      "clip_ratio/low_mean": 0.0008680555620230734,
+      "clip_ratio/low_min": 0.0008680555620230734,
+      "clip_ratio/region_mean": 0.002288510149810463,
+      "completions/clipped_ratio": 0.03125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 14.484375,
+      "completions/min_length": 2.0,
+      "epoch": 0.0664819944598338,
+      "grad_norm": 1.600021993041462,
+      "kl": 0.007914411224192008,
+      "learning_rate": 6.666666666666667e-05,
+      "loss": -0.0006625128444284201,
+      "memory(GiB)": 77.89,
+      "reward": -0.1328125,
+      "reward_std": 0.7212049961090088,
+      "rewards/Response_no_think_reward/mean": -0.1328125,
+      "rewards/Response_no_think_reward/std": 1.3653247356414795,
+      "step": 6,
+      "train_speed(iter/s)": 0.009313
+    },
+    {
+      "clip_ratio/high_max": 0.020065165997948498,
+      "clip_ratio/high_mean": 0.020065165997948498,
+      "clip_ratio/low_mean": 0.03063795086927712,
+      "clip_ratio/low_min": 0.03063795086927712,
+      "clip_ratio/region_mean": 0.050703116634394974,
+      "epoch": 0.07756232686980609,
+      "grad_norm": 0.9793694784898841,
+      "kl": 0.03891617391491309,
+      "learning_rate": 7.777777777777778e-05,
+      "loss": -0.00832418818026781,
+      "memory(GiB)": 77.89,
+      "step": 7,
+      "train_speed(iter/s)": 0.0084
+    },
+    {
+      "clip_ratio/high_max": 0.0005165289039723575,
+      "clip_ratio/high_mean": 0.0005165289039723575,
+      "clip_ratio/low_mean": 0.0012338250526227057,
+      "clip_ratio/low_min": 0.0012338250526227057,
+      "clip_ratio/region_mean": 0.0017503539565950632,
+      "completions/clipped_ratio": 0.0234375,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 14.734375,
+      "completions/min_length": 3.0,
+      "epoch": 0.0886426592797784,
+      "grad_norm": 1.3164563077884395,
+      "kl": 0.05195357115007937,
+      "learning_rate": 8.888888888888889e-05,
+      "loss": -0.00010046618990600109,
+      "memory(GiB)": 78.03,
+      "reward": -0.2734375,
+      "reward_std": 0.7209804058074951,
+      "rewards/Response_no_think_reward/mean": -0.2734375,
+      "rewards/Response_no_think_reward/std": 1.27183198928833,
+      "step": 8,
+      "train_speed(iter/s)": 0.004412
+    },
+    {
+      "clip_ratio/high_max": 0.011744739342248067,
+      "clip_ratio/high_mean": 0.011744739342248067,
+      "clip_ratio/low_mean": 0.021301756787579507,
+      "clip_ratio/low_min": 0.021301756787579507,
+      "clip_ratio/region_mean": 0.033046496275346726,
+      "epoch": 0.0997229916897507,
+      "grad_norm": 0.7228491518243964,
+      "kl": 0.14250310882925987,
+      "learning_rate": 0.0001,
+      "loss": -0.009305480867624283,
+      "memory(GiB)": 78.03,
+      "step": 9,
+      "train_speed(iter/s)": 0.004494
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0234375,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 14.140625,
+      "completions/min_length": 3.0,
+      "epoch": 0.11080332409972299,
+      "grad_norm": 0.7948888253152248,
+      "kl": 0.04085417848546058,
+      "learning_rate": 0.00011111111111111112,
+      "loss": 0.00047483143862336874,
+      "memory(GiB)": 78.03,
+      "reward": 0.0546875,
+      "reward_std": 0.6801962852478027,
+      "rewards/Response_no_think_reward/mean": 0.0546875,
+      "rewards/Response_no_think_reward/std": 1.4380027055740356,
+      "step": 10,
+      "train_speed(iter/s)": 0.002899
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 17.5078125,
+      "completions/min_length": 4.0,
+      "epoch": 0.12188365650969529,
+      "grad_norm": 0.0004263039062328966,
+      "kl": 0.04189836711157113,
+      "learning_rate": 0.00012222222222222224,
+      "loss": 5.555100869969465e-06,
+      "memory(GiB)": 77.88,
+      "reward": 0.0,
+      "reward_std": 0.0,
+      "rewards/Response_no_think_reward/mean": 0.0,
+      "rewards/Response_no_think_reward/std": 0.0,
+      "step": 11,
+      "train_speed(iter/s)": 0.017791
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "epoch": 0.1329639889196676,
+      "grad_norm": 0.001194388957130102,
+      "kl": 0.06704438821179792,
+      "learning_rate": 0.00013333333333333334,
+      "loss": 9.91918750514742e-06,
+      "memory(GiB)": 77.88,
+      "step": 12,
+      "train_speed(iter/s)": 0.014858
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 19.8515625,
+      "completions/min_length": 2.0,
+      "epoch": 0.1440443213296399,
+      "grad_norm": 0.00032184673828801084,
+      "kl": 0.07992906332947314,
+      "learning_rate": 0.00014444444444444444,
+      "loss": 7.416723747155629e-06,
+      "memory(GiB)": 77.92,
+      "reward": 0.0,
+      "reward_std": 0.0,
+      "rewards/Response_no_think_reward/mean": 0.0,
+      "rewards/Response_no_think_reward/std": 0.0,
+      "step": 13,
+      "train_speed(iter/s)": 0.010543
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "epoch": 0.15512465373961218,
+      "grad_norm": 0.0009186451648310945,
+      "kl": 0.13713529019150883,
+      "learning_rate": 0.00015555555555555556,
+      "loss": 1.3553013559430838e-05,
+      "memory(GiB)": 77.92,
+      "step": 14,
+      "train_speed(iter/s)": 0.009842
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0546875,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 19.9609375,
+      "completions/min_length": 3.0,
+      "epoch": 0.16620498614958448,
+      "grad_norm": 0.0007871986446541219,
+      "kl": 0.10129890241660178,
+      "learning_rate": 0.0001666666666666667,
+      "loss": 1.3785504052066244e-05,
+      "memory(GiB)": 77.92,
+      "reward": 0.0,
+      "reward_std": 0.0,
+      "rewards/Response_no_think_reward/mean": 0.0,
+      "rewards/Response_no_think_reward/std": 0.0,
+      "step": 15,
+      "train_speed(iter/s)": 0.008057
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0028207504947204143,
+      "clip_ratio/low_min": 0.0028207504947204143,
+      "clip_ratio/region_mean": 0.0028207504947204143,
+      "completions/clipped_ratio": 0.109375,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 21.1796875,
+      "completions/min_length": 3.0,
+      "epoch": 0.1772853185595568,
+      "grad_norm": 1.5221332407610604,
+      "kl": 0.12953427506727166,
+      "learning_rate": 0.00017777777777777779,
+      "loss": 0.0005249045789241791,
+      "memory(GiB)": 77.88,
+      "reward": -0.140625,
+      "reward_std": 0.747445285320282,
+      "rewards/Response_no_think_reward/mean": -0.140625,
+      "rewards/Response_no_think_reward/std": 1.3500328063964844,
+      "step": 16,
+      "train_speed(iter/s)": 0.0232
+    },
+    {
+      "clip_ratio/high_max": 0.014893613493768498,
+      "clip_ratio/high_mean": 0.014893613493768498,
+      "clip_ratio/low_mean": 0.08722472144290805,
+      "clip_ratio/low_min": 0.08722472144290805,
+      "clip_ratio/region_mean": 0.10211833566427231,
+      "epoch": 0.1883656509695291,
+      "grad_norm": 1.2273096796368343,
+      "kl": 9.949020208441652,
+      "learning_rate": 0.00018888888888888888,
+      "loss": -0.009794662706553936,
+      "memory(GiB)": 77.88,
+      "step": 17,
+      "train_speed(iter/s)": 0.019341
+    },
+    {
+      "clip_ratio/high_max": 0.00042517005931586027,
+      "clip_ratio/high_mean": 0.00042517005931586027,
+      "clip_ratio/low_mean": 0.005602272489340976,
+      "clip_ratio/low_min": 0.005602272489340976,
+      "clip_ratio/region_mean": 0.006027442548656836,
+      "completions/clipped_ratio": 0.078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 16.5234375,
+      "completions/min_length": 3.0,
+      "epoch": 0.1994459833795014,
+      "grad_norm": 1.6723327645413764,
+      "kl": 19.73964432394132,
+      "learning_rate": 0.0002,
+      "loss": 0.0016888766549527645,
+      "memory(GiB)": 77.92,
+      "reward": -0.5078125,
+      "reward_std": 0.46106040477752686,
+      "rewards/Response_no_think_reward/mean": -0.5078125,
+      "rewards/Response_no_think_reward/std": 1.0038986206054688,
+      "step": 18,
+      "train_speed(iter/s)": 0.012746
+    },
+    {
+      "clip_ratio/high_max": 0.008699395693838596,
+      "clip_ratio/high_mean": 0.008699395693838596,
+      "clip_ratio/low_mean": 0.09401972405612469,
+      "clip_ratio/low_min": 0.09401972405612469,
+      "clip_ratio/region_mean": 0.10271912068128586,
+      "epoch": 0.21052631578947367,
+      "grad_norm": 0.9659705365958429,
+      "kl": 28.792964363470674,
+      "learning_rate": 0.00019999578095183124,
+      "loss": -0.005614398047327995,
+      "memory(GiB)": 77.92,
+      "step": 19,
+      "train_speed(iter/s)": 0.011863
+    },
+    {
+      "clip_ratio/high_max": 0.0009057971183210611,
+      "clip_ratio/high_mean": 0.0009057971183210611,
+      "clip_ratio/low_mean": 0.0047033663431648165,
+      "clip_ratio/low_min": 0.0047033663431648165,
+      "clip_ratio/region_mean": 0.005609163461485878,
+      "completions/clipped_ratio": 0.0234375,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 17.703125,
+      "completions/min_length": 4.0,
+      "epoch": 0.22160664819944598,
+      "grad_norm": 2.0349350259537244,
+      "kl": 0.8193748028716072,
+      "learning_rate": 0.00019998312416333227,
+      "loss": 0.0007828018860891461,
+      "memory(GiB)": 77.92,
+      "reward": -0.40625,
+      "reward_std": 0.5371267199516296,
+      "rewards/Response_no_think_reward/mean": -0.40625,
+      "rewards/Response_no_think_reward/std": 1.1734429597854614,
+      "step": 20,
+      "train_speed(iter/s)": 0.009276
+    },
+    {
+      "clip_ratio/high_max": 0.017557535058585927,
+      "clip_ratio/high_mean": 0.017557535058585927,
+      "clip_ratio/low_mean": 0.0823797988705337,
+      "clip_ratio/low_min": 0.0823797988705337,
+      "clip_ratio/region_mean": 0.09993733279407024,
+      "epoch": 0.23268698060941828,
+      "grad_norm": 5.547205800030314,
+      "kl": 155.78269671928138,
+      "learning_rate": 0.00019996203070249516,
+      "loss": 0.028484804555773735,
+      "memory(GiB)": 79.88,
+      "step": 21,
+      "train_speed(iter/s)": 0.008942
+    },
+    {
+      "clip_ratio/high_max": 0.00173778529278934,
+      "clip_ratio/high_mean": 0.00173778529278934,
+      "clip_ratio/low_mean": 0.004023336310638115,
+      "clip_ratio/low_min": 0.004023336310638115,
+      "clip_ratio/region_mean": 0.005761121603427455,
+      "completions/clipped_ratio": 0.0390625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 16.359375,
+      "completions/min_length": 4.0,
+      "epoch": 0.24376731301939059,
+      "grad_norm": 1.1706684735088349,
+      "kl": 0.29493264015764,
+      "learning_rate": 0.00019993250234920636,
+      "loss": 0.0013634033966809511,
+      "memory(GiB)": 79.88,
+      "reward": 0.1640625,
+      "reward_std": 0.6870359778404236,
+      "rewards/Response_no_think_reward/mean": 0.1640625,
+      "rewards/Response_no_think_reward/std": 1.4240750074386597,
+      "step": 22,
+      "train_speed(iter/s)": 0.007618
+    },
+    {
+      "clip_ratio/high_max": 0.01892658555880189,
+      "clip_ratio/high_mean": 0.01892658555880189,
+      "clip_ratio/low_mean": 0.06755852687638253,
+      "clip_ratio/low_min": 0.06755852687638253,
+      "clip_ratio/region_mean": 0.0864851123187691,
+      "epoch": 0.2548476454293629,
+      "grad_norm": 0.5069456216354334,
+      "kl": 0.7371144236531109,
+      "learning_rate": 0.0001998945415950969,
+      "loss": -0.015337318181991577,
+      "memory(GiB)": 79.88,
+      "step": 23,
+      "train_speed(iter/s)": 0.007472
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.00311830299324356,
+      "clip_ratio/low_min": 0.00311830299324356,
+      "clip_ratio/region_mean": 0.00311830299324356,
+      "completions/clipped_ratio": 0.0625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 15.4296875,
+      "completions/min_length": 5.0,
+      "epoch": 0.2659279778393352,
+      "grad_norm": 1.2087135470902275,
+      "kl": 0.9872721234764867,
+      "learning_rate": 0.00019984815164333163,
+      "loss": 0.0027771026361733675,
+      "memory(GiB)": 79.88,
+      "reward": -0.109375,
+      "reward_std": 0.3798917531967163,
+      "rewards/Response_no_think_reward/mean": -0.109375,
+      "rewards/Response_no_think_reward/std": 1.1310532093048096,
+      "step": 24,
+      "train_speed(iter/s)": 0.006636
+    },
+    {
+      "clip_ratio/high_max": 0.0029115322104189545,
+      "clip_ratio/high_mean": 0.0029115322104189545,
+      "clip_ratio/low_mean": 0.07530217664316297,
+      "clip_ratio/low_min": 0.07530217664316297,
+      "clip_ratio/region_mean": 0.07821370894089341,
+      "epoch": 0.2770083102493075,
+      "grad_norm": 0.702071136643931,
+      "kl": 26.457206293552645,
+      "learning_rate": 0.00019979333640833947,
+      "loss": -0.003078086068853736,
+      "memory(GiB)": 79.88,
+      "step": 25,
+      "train_speed(iter/s)": 0.006569
+    },
+    {
+      "clip_ratio/high_max": 0.0023347785463556647,
+      "clip_ratio/high_mean": 0.0023347785463556647,
+      "clip_ratio/low_mean": 0.01951372090843506,
+      "clip_ratio/low_min": 0.01951372090843506,
+      "clip_ratio/region_mean": 0.021848499542102218,
+      "completions/clipped_ratio": 0.015625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 14.6953125,
+      "completions/min_length": 4.0,
+      "epoch": 0.2880886426592798,
+      "grad_norm": 1.6527510231667384,
+      "kl": 7.873713016277179,
+      "learning_rate": 0.00019973010051548275,
+      "loss": 0.0009247222333215177,
+      "memory(GiB)": 79.88,
+      "reward": 0.015625,
+      "reward_std": 0.7059091925621033,
+      "rewards/Response_no_think_reward/mean": 0.015625,
+      "rewards/Response_no_think_reward/std": 1.2610729932785034,
+      "step": 26,
+      "train_speed(iter/s)": 0.005996
+    },
+    {
+      "clip_ratio/high_max": 0.01952397992135957,
+      "clip_ratio/high_mean": 0.01952397992135957,
+      "clip_ratio/low_mean": 0.1283843811834231,
+      "clip_ratio/low_min": 0.1283843811834231,
+      "clip_ratio/region_mean": 0.1479083604644984,
+      "epoch": 0.29916897506925205,
+      "grad_norm": 12.590003285629427,
+      "kl": 741.7941563371569,
+      "learning_rate": 0.000199658449300667,
+      "loss": 0.036141443997621536,
+      "memory(GiB)": 79.88,
+      "step": 27,
+      "train_speed(iter/s)": 0.005965
+    },
+    {
+      "clip_ratio/high_max": 0.0036726996186189353,
+      "clip_ratio/high_mean": 0.0036726996186189353,
+      "clip_ratio/low_mean": 0.005622756987577304,
+      "clip_ratio/low_min": 0.005622756987577304,
+      "clip_ratio/region_mean": 0.00929545663530007,
+      "completions/clipped_ratio": 0.0390625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 13.8203125,
+      "completions/min_length": 4.0,
+      "epoch": 0.31024930747922436,
+      "grad_norm": 2.1964117931286085,
+      "kl": 11.635613721329719,
+      "learning_rate": 0.00019957838880989078,
+      "loss": 0.002877343213185668,
+      "memory(GiB)": 79.88,
+      "reward": 0.390625,
+      "reward_std": 0.6595449447631836,
+      "rewards/Response_no_think_reward/mean": 0.390625,
+      "rewards/Response_no_think_reward/std": 1.1379936933517456,
+      "step": 28,
+      "train_speed(iter/s)": 0.005568
+    },
+    {
+      "clip_ratio/high_max": 0.017762314528226852,
+      "clip_ratio/high_mean": 0.017762314528226852,
+      "clip_ratio/low_mean": 0.047843097941949964,
+      "clip_ratio/low_min": 0.047843097941949964,
+      "clip_ratio/region_mean": 0.06560541247017682,
+      "epoch": 0.32132963988919666,
+      "grad_norm": 1.3735654410758578,
+      "kl": 80.49331146203622,
+      "learning_rate": 0.00019948992579873538,
+      "loss": -0.005675393156707287,
+      "memory(GiB)": 79.88,
+      "step": 29,
+      "train_speed(iter/s)": 0.005557
+    },
+    {
+      "clip_ratio/high_max": 0.00618883827701211,
+      "clip_ratio/high_mean": 0.00618883827701211,
+      "clip_ratio/low_mean": 0.005286419298499823,
+      "clip_ratio/low_min": 0.005286419298499823,
+      "clip_ratio/region_mean": 0.011475257691927254,
+      "completions/clipped_ratio": 0.015625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 11.484375,
+      "completions/min_length": 3.0,
+      "epoch": 0.33240997229916897,
+      "grad_norm": 1.1914361173581312,
+      "kl": 3.281426345667569,
+      "learning_rate": 0.00019939306773179497,
+      "loss": 0.0003165462985634804,
+      "memory(GiB)": 79.88,
+      "reward": 0.5625,
+      "reward_std": 0.4207349121570587,
+      "rewards/Response_no_think_reward/mean": 0.5625,
+      "rewards/Response_no_think_reward/std": 1.3265905380249023,
+      "step": 30,
+      "train_speed(iter/s)": 0.005222
+    },
+    {
+      "clip_ratio/high_max": 0.05031776078976691,
+      "clip_ratio/high_mean": 0.05031776078976691,
+      "clip_ratio/low_mean": 0.041757271508686244,
+      "clip_ratio/low_min": 0.041757271508686244,
+      "clip_ratio/region_mean": 0.09207503264769912,
+      "epoch": 0.34349030470914127,
+      "grad_norm": 0.9882009232899025,
+      "kl": 9.828547842378612,
+      "learning_rate": 0.0001992878227820465,
+      "loss": -0.0019915460143238306,
+      "memory(GiB)": 79.88,
+      "step": 31,
+      "train_speed(iter/s)": 0.005221
+    },
+    {
+      "clip_ratio/high_max": 0.0014815493486821651,
+      "clip_ratio/high_mean": 0.0014815493486821651,
+      "clip_ratio/low_mean": 0.00400914296915289,
+      "clip_ratio/low_min": 0.00400914296915289,
+      "clip_ratio/region_mean": 0.005490692317835055,
+      "completions/clipped_ratio": 0.046875,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 19.078125,
+      "completions/min_length": 4.0,
+      "epoch": 0.3545706371191136,
+      "grad_norm": 1.1670914229830547,
+      "kl": 3.5491041047498584,
+      "learning_rate": 0.00019917419983016025,
+      "loss": 0.0021022851578891277,
+      "memory(GiB)": 79.88,
+      "reward": 0.25,
+      "reward_std": 0.7367533445358276,
+      "rewards/Response_no_think_reward/mean": 0.25,
+      "rewards/Response_no_think_reward/std": 1.3631829023361206,
+      "step": 32,
+      "train_speed(iter/s)": 0.004948
+    },
+    {
+      "clip_ratio/high_max": 0.02484438387909904,
+      "clip_ratio/high_mean": 0.02484438387909904,
+      "clip_ratio/low_mean": 0.0674317826051265,
+      "clip_ratio/low_min": 0.0674317826051265,
+      "clip_ratio/region_mean": 0.09227616689167917,
+      "epoch": 0.3656509695290859,
+      "grad_norm": 0.9959477925686483,
+      "kl": 51.25826009634329,
+      "learning_rate": 0.00019905220846375032,
+      "loss": -0.002379298210144043,
+      "memory(GiB)": 79.88,
+      "step": 33,
+      "train_speed(iter/s)": 0.004958
+    },
+    {
+      "clip_ratio/high_max": 0.0034505740622989833,
+      "clip_ratio/high_mean": 0.0034505740622989833,
+      "clip_ratio/low_mean": 0.0069533721543848515,
+      "clip_ratio/low_min": 0.0069533721543848515,
+      "clip_ratio/region_mean": 0.010403946158476174,
+      "completions/clipped_ratio": 0.03125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 12.4375,
+      "completions/min_length": 4.0,
+      "epoch": 0.3767313019390582,
+      "grad_norm": 1.1883165046483204,
+      "kl": 6.391523744328879,
+      "learning_rate": 0.00019892185897656578,
+      "loss": 0.0010706284083425999,
+      "memory(GiB)": 79.88,
+      "reward": 0.078125,
+      "reward_std": 0.3766237795352936,
+      "rewards/Response_no_think_reward/mean": 0.078125,
+      "rewards/Response_no_think_reward/std": 1.1126288175582886,
+      "step": 34,
+      "train_speed(iter/s)": 0.004716
+    },
+    {
+      "clip_ratio/high_max": 0.01320959790609777,
+      "clip_ratio/high_mean": 0.01320959790609777,
+      "clip_ratio/low_mean": 0.059380761347711086,
+      "clip_ratio/low_min": 0.059380761347711086,
+      "clip_ratio/region_mean": 0.07259036041796207,
+      "epoch": 0.3878116343490305,
+      "grad_norm": 0.527136254278729,
+      "kl": 34.938005739822984,
+      "learning_rate": 0.00019878316236762196,
+      "loss": -0.005408844910562038,
+      "memory(GiB)": 79.88,
+      "step": 35,
+      "train_speed(iter/s)": 0.004731
+    },
+    {
+      "clip_ratio/high_max": 0.002702809404581785,
+      "clip_ratio/high_mean": 0.002702809404581785,
+      "clip_ratio/low_mean": 0.0033881079871207476,
+      "clip_ratio/low_min": 0.0033881079871207476,
+      "clip_ratio/region_mean": 0.006090917449910194,
+      "completions/clipped_ratio": 0.0078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 10.703125,
+      "completions/min_length": 4.0,
+      "epoch": 0.3988919667590028,
+      "grad_norm": 1.0383854818250358,
+      "kl": 10.416337528149597,
+      "learning_rate": 0.00019863613034027224,
+      "loss": 0.0015072962269186974,
+      "memory(GiB)": 79.88,
+      "reward": 0.5,
+      "reward_std": 0.4501557946205139,
+      "rewards/Response_no_think_reward/mean": 0.5,
+      "rewards/Response_no_think_reward/std": 1.1294203996658325,
+      "step": 36,
+      "train_speed(iter/s)": 0.004567
+    },
+    {
+      "clip_ratio/high_max": 0.04597766371443868,
+      "clip_ratio/high_mean": 0.04597766371443868,
+      "clip_ratio/low_mean": 0.029037245316430926,
+      "clip_ratio/low_min": 0.029037245316430926,
+      "clip_ratio/region_mean": 0.07501490786671638,
+      "epoch": 0.4099722991689751,
+      "grad_norm": 1.198632656262216,
+      "kl": 1.9380782267544419,
+      "learning_rate": 0.00019848077530122083,
+      "loss": -0.00020142836729064584,
+      "memory(GiB)": 79.88,
+      "step": 37,
+      "train_speed(iter/s)": 0.004584
+    },
+    {
+      "clip_ratio/high_max": 0.0011905284482054412,
+      "clip_ratio/high_mean": 0.0011905284482054412,
+      "clip_ratio/low_mean": 0.013364006066694856,
+      "clip_ratio/low_min": 0.013364006066694856,
+      "clip_ratio/region_mean": 0.014554534514900297,
+      "completions/clipped_ratio": 0.0078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 9.28125,
+      "completions/min_length": 3.0,
+      "epoch": 0.42105263157894735,
+      "grad_norm": 1.9489606220013782,
+      "kl": 19.19968014501501,
+      "learning_rate": 0.0001983171103594755,
+      "loss": 0.0004887112881988287,
+      "memory(GiB)": 79.88,
+      "reward": 0.0625,
+      "reward_std": 0.47397348284721375,
+      "rewards/Response_no_think_reward/mean": 0.0625,
+      "rewards/Response_no_think_reward/std": 0.9618715047836304,
+      "step": 38,
+      "train_speed(iter/s)": 0.004438
+    },
+    {
+      "clip_ratio/high_max": 0.0027730573201552033,
+      "clip_ratio/high_mean": 0.0027730573201552033,
+      "clip_ratio/low_mean": 0.11966108158230782,
+      "clip_ratio/low_min": 0.11966108158230782,
+      "clip_ratio/region_mean": 0.12243413866963238,
+      "epoch": 0.43213296398891965,
+      "grad_norm": 1.0018802051645932,
+      "kl": 49.89589491917286,
+      "learning_rate": 0.0001981451493252418,
+      "loss": -0.009153969585895538,
+      "memory(GiB)": 79.88,
+      "step": 39,
+      "train_speed(iter/s)": 0.004457
+    },
+    {
+      "clip_ratio/high_max": 0.0013888889225199819,
+      "clip_ratio/high_mean": 0.0013888889225199819,
+      "clip_ratio/low_mean": 0.01340760855237022,
+      "clip_ratio/low_min": 0.01340760855237022,
+      "clip_ratio/region_mean": 0.014796497474890202,
+      "completions/clipped_ratio": 0.0078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 8.34375,
+      "completions/min_length": 3.0,
+      "epoch": 0.44321329639889195,
+      "grad_norm": 1.5554099620650559,
+      "kl": 4.425425759982318,
+      "learning_rate": 0.0001979649067087574,
+      "loss": 0.006227841135114431,
+      "memory(GiB)": 79.88,
+      "reward": 0.109375,
+      "reward_std": 0.5266703963279724,
+      "rewards/Response_no_think_reward/mean": 0.109375,
+      "rewards/Response_no_think_reward/std": 1.0057659149169922,
+      "step": 40,
+      "train_speed(iter/s)": 0.004236
+    },
+    {
+      "clip_ratio/high_max": 0.017265826056245714,
+      "clip_ratio/high_mean": 0.017265826056245714,
+      "clip_ratio/low_mean": 0.11041530082002282,
+      "clip_ratio/low_min": 0.11041530082002282,
+      "clip_ratio/region_mean": 0.1276811266434379,
+      "epoch": 0.45429362880886426,
+      "grad_norm": 6.231748769450823,
+      "kl": 522.7659375867806,
+      "learning_rate": 0.00019777639771906795,
+      "loss": -0.0034558051265776157,
+      "memory(GiB)": 79.88,
+      "step": 41,
+      "train_speed(iter/s)": 0.004255
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.006427675718441606,
+      "clip_ratio/low_min": 0.006427675718441606,
+      "clip_ratio/region_mean": 0.006427675718441606,
+      "completions/clipped_ratio": 0.015625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 7.34375,
+      "completions/min_length": 5.0,
+      "epoch": 0.46537396121883656,
+      "grad_norm": 1.0020303907992343,
+      "kl": 6.527791675332992,
+      "learning_rate": 0.00019757963826274357,
+      "loss": -0.0012322411639615893,
+      "memory(GiB)": 79.88,
+      "reward": 0.1484375,
+      "reward_std": 0.21095064282417297,
+      "rewards/Response_no_think_reward/mean": 0.1484375,
+      "rewards/Response_no_think_reward/std": 0.628582239151001,
+      "step": 42,
+      "train_speed(iter/s)": 0.004165
+    },
+    {
+      "clip_ratio/high_max": 0.01585913705639541,
+      "clip_ratio/high_mean": 0.01585913705639541,
+      "clip_ratio/low_mean": 0.026421017944812775,
+      "clip_ratio/low_min": 0.026421017944812775,
+      "clip_ratio/region_mean": 0.04228015476837754,
+      "epoch": 0.47645429362880887,
+      "grad_norm": 0.4174666284613835,
+      "kl": 30.479038277953805,
+      "learning_rate": 0.0001973746449425368,
+      "loss": -0.009352116845548153,
+      "memory(GiB)": 79.88,
+      "step": 43,
+      "train_speed(iter/s)": 0.004185
+    },
+    {
+      "clip_ratio/high_max": 0.0026535717770457268,
+      "clip_ratio/high_mean": 0.0026535717770457268,
+      "clip_ratio/low_mean": 0.008503502060193568,
+      "clip_ratio/low_min": 0.008503502060193568,
+      "clip_ratio/region_mean": 0.011157073837239295,
+      "completions/clipped_ratio": 0.0390625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 9.953125,
+      "completions/min_length": 5.0,
+      "epoch": 0.48753462603878117,
+      "grad_norm": 2.3666909999114085,
+      "kl": 21.929430570075056,
+      "learning_rate": 0.0001971614350559814,
+      "loss": 0.0009473273530602455,
+      "memory(GiB)": 79.88,
+      "reward": -0.03125,
+      "reward_std": 0.3844763934612274,
+      "rewards/Response_no_think_reward/mean": -0.03125,
+      "rewards/Response_no_think_reward/std": 0.6750619411468506,
+      "step": 44,
+      "train_speed(iter/s)": 0.0041
+    },
+    {
+      "clip_ratio/high_max": 0.003926915815100074,
+      "clip_ratio/high_mean": 0.003926915815100074,
+      "clip_ratio/low_mean": 0.1417178469710052,
+      "clip_ratio/low_min": 0.1417178469710052,
+      "clip_ratio/region_mean": 0.14564476208761334,
+      "epoch": 0.4986149584487535,
+      "grad_norm": 51.49679854517617,
+      "kl": 4715.846689800266,
+      "learning_rate": 0.00019694002659393305,
+      "loss": 0.22465737164020538,
+      "memory(GiB)": 79.88,
+      "step": 45,
+      "train_speed(iter/s)": 0.004121
+    },
+    {
+      "clip_ratio/high_max": 0.0025718168762978166,
+      "clip_ratio/high_mean": 0.0025718168762978166,
+      "clip_ratio/low_mean": 0.0054531682981178164,
+      "clip_ratio/low_min": 0.0054531682981178164,
+      "clip_ratio/region_mean": 0.008024985319934785,
+      "completions/clipped_ratio": 0.015625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 7.453125,
+      "completions/min_length": 3.0,
+      "epoch": 0.5096952908587258,
+      "grad_norm": 7.932036097415596,
+      "kl": 134.56803331989795,
+      "learning_rate": 0.0001967104382390511,
+      "loss": 0.00671145785599947,
+      "memory(GiB)": 79.88,
+      "reward": 0.109375,
+      "reward_std": 0.21731878817081451,
+      "rewards/Response_no_think_reward/mean": 0.109375,
+      "rewards/Response_no_think_reward/std": 0.6303901672363281,
+      "step": 46,
+      "train_speed(iter/s)": 0.004027
+    },
+    {
+      "clip_ratio/high_max": 0.013239058025646955,
+      "clip_ratio/high_mean": 0.013239058025646955,
+      "clip_ratio/low_mean": 0.03283043531700969,
+      "clip_ratio/low_min": 0.03283043531700969,
+      "clip_ratio/region_mean": 0.04606949305161834,
+      "epoch": 0.5207756232686981,
+      "grad_norm": 2.141404413074471,
+      "kl": 68.84463222045451,
+      "learning_rate": 0.00019647268936422206,
+      "loss": -0.005232019349932671,
+      "memory(GiB)": 79.88,
+      "step": 47,
+      "train_speed(iter/s)": 0.004048
+    },
+    {
+      "clip_ratio/high_max": 0.0005482456181198359,
+      "clip_ratio/high_mean": 0.0005482456181198359,
+      "clip_ratio/low_mean": 0.005675954627804458,
+      "clip_ratio/low_min": 0.005675954627804458,
+      "clip_ratio/region_mean": 0.006224200245924294,
+      "completions/clipped_ratio": 0.0078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 6.5078125,
+      "completions/min_length": 3.0,
+      "epoch": 0.5318559556786704,
+      "grad_norm": 2.4236302292247642,
+      "kl": 80.9618124118424,
+      "learning_rate": 0.00019622680003092503,
+      "loss": 0.005468316841870546,
+      "memory(GiB)": 79.88,
+      "reward": 0.0703125,
+      "reward_std": 0.1054728701710701,
+      "rewards/Response_no_think_reward/mean": 0.0703125,
+      "rewards/Response_no_think_reward/std": 0.50483638048172,
+      "step": 48,
+      "train_speed(iter/s)": 0.003952
+    },
+    {
+      "clip_ratio/high_max": 0.0366715406998992,
+      "clip_ratio/high_mean": 0.0366715406998992,
+      "clip_ratio/low_mean": 0.0379418209195137,
+      "clip_ratio/low_min": 0.0379418209195137,
+      "clip_ratio/region_mean": 0.07461336255073547,
+      "epoch": 0.5429362880886427,
+      "grad_norm": 27.44200416529675,
+      "kl": 31.79825783398701,
+      "learning_rate": 0.0001959727909875389,
+      "loss": 0.08109890669584274,
+      "memory(GiB)": 79.88,
+      "step": 49,
+      "train_speed(iter/s)": 0.003972
+    },
+    {
+      "clip_ratio/high_max": 0.002227982331532985,
+      "clip_ratio/high_mean": 0.002227982331532985,
+      "clip_ratio/low_mean": 0.006371813942678273,
+      "clip_ratio/low_min": 0.006371813942678273,
+      "clip_ratio/region_mean": 0.008599796216003597,
+      "completions/clipped_ratio": 0.0078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 7.3125,
+      "completions/min_length": 5.0,
+      "epoch": 0.554016620498615,
+      "grad_norm": 1.8915627714701582,
+      "kl": 32.8292202873854,
+      "learning_rate": 0.00019571068366759143,
+      "loss": 0.004485747776925564,
+      "memory(GiB)": 79.88,
+      "reward": 0.046875,
+      "reward_std": 0.2615792155265808,
+      "rewards/Response_no_think_reward/mean": 0.046875,
+      "rewards/Response_no_think_reward/std": 0.5731440782546997,
+      "step": 50,
+      "train_speed(iter/s)": 0.003883
+    },
+    {
+      "clip_ratio/high_max": 0.02533616591244936,
+      "clip_ratio/high_mean": 0.02533616591244936,
+      "clip_ratio/low_mean": 0.05068295169621706,
+      "clip_ratio/low_min": 0.05068295169621706,
+      "clip_ratio/region_mean": 0.07601911667734385,
+      "epoch": 0.5650969529085873,
+      "grad_norm": 3.4383916295396406,
+      "kl": 341.1058924796962,
+      "learning_rate": 0.00019544050018795075,
+      "loss": -2.989325366797857e-06,
+      "memory(GiB)": 79.88,
+      "step": 51,
+      "train_speed(iter/s)": 0.003903
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0021551724057644606,
+      "clip_ratio/low_min": 0.0021551724057644606,
+      "clip_ratio/region_mean": 0.0021551724057644606,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 24.0,
+      "completions/mean_length": 6.3046875,
+      "completions/min_length": 6.0,
+      "epoch": 0.5761772853185596,
+      "grad_norm": 0.7839236512036062,
+      "kl": 20.716427901759744,
+      "learning_rate": 0.0001951622633469592,
+      "loss": 0.0010195983340963721,
+      "memory(GiB)": 79.88,
+      "reward": 0.0390625,
+      "reward_std": 0.09916213154792786,
+      "rewards/Response_no_think_reward/mean": 0.0390625,
+      "rewards/Response_no_think_reward/std": 0.3847965598106384,
+      "step": 52,
+      "train_speed(iter/s)": 0.003843
+    },
+    {
+      "clip_ratio/high_max": 0.0052083334885537624,
+      "clip_ratio/high_mean": 0.0052083334885537624,
+      "clip_ratio/low_mean": 0.031070403289049864,
+      "clip_ratio/low_min": 0.031070403289049864,
+      "clip_ratio/region_mean": 0.036278736777603626,
+      "epoch": 0.5872576177285319,
+      "grad_norm": 0.5191411391520966,
+      "kl": 69.86534339073114,
+      "learning_rate": 0.00019487599662250943,
+      "loss": -0.0016538399504497647,
+      "memory(GiB)": 79.88,
+      "step": 53,
+      "train_speed(iter/s)": 0.003863
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0052784152794629335,
+      "clip_ratio/low_min": 0.0052784152794629335,
+      "clip_ratio/region_mean": 0.0052784152794629335,
+      "completions/clipped_ratio": 0.015625,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 7.3046875,
+      "completions/min_length": 3.0,
+      "epoch": 0.5983379501385041,
+      "grad_norm": 2.311740957637672,
+      "kl": 2.841107105177798,
+      "learning_rate": 0.00019458172417006347,
+      "loss": 0.004196230787783861,
+      "memory(GiB)": 79.88,
+      "reward": 0.3984375,
+      "reward_std": 0.26962828636169434,
+      "rewards/Response_no_think_reward/mean": 0.3984375,
+      "rewards/Response_no_think_reward/std": 1.0221164226531982,
+      "step": 54,
+      "train_speed(iter/s)": 0.003692
+    },
+    {
+      "clip_ratio/high_max": 0.017684109043329954,
+      "clip_ratio/high_mean": 0.017684109043329954,
+      "clip_ratio/low_mean": 0.09186162473633885,
+      "clip_ratio/low_min": 0.09186162473633885,
+      "clip_ratio/region_mean": 0.10954573331400752,
+      "epoch": 0.6094182825484764,
+      "grad_norm": 2.0764162059344025,
+      "kl": 263.63854900409933,
+      "learning_rate": 0.00019427947082061432,
+      "loss": 0.006979572586715221,
+      "memory(GiB)": 79.88,
+      "step": 55,
+      "train_speed(iter/s)": 0.003712
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.018976572435349226,
+      "clip_ratio/low_min": 0.018976572435349226,
+      "clip_ratio/region_mean": 0.018976572435349226,
+      "completions/clipped_ratio": 0.0078125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 8.0,
+      "completions/min_length": 3.0,
+      "epoch": 0.6204986149584487,
+      "grad_norm": 2.3049148508787534,
+      "kl": 227.90118128014728,
+      "learning_rate": 0.00019396926207859084,
+      "loss": 0.00933245662599802,
+      "memory(GiB)": 79.88,
+      "reward": -0.1171875,
+      "reward_std": 0.33797404170036316,
+      "rewards/Response_no_think_reward/mean": -0.1171875,
+      "rewards/Response_no_think_reward/std": 0.7594143748283386,
+      "step": 56,
+      "train_speed(iter/s)": 0.003618
+    },
+    {
+      "clip_ratio/high_max": 0.024031779495999217,
+      "clip_ratio/high_mean": 0.024031779495999217,
+      "clip_ratio/low_mean": 0.10989384504500777,
+      "clip_ratio/low_min": 0.10989384504500777,
+      "clip_ratio/region_mean": 0.13392562558874488,
+      "epoch": 0.631578947368421,
+      "grad_norm": 408.32691013228697,
+      "kl": 23.09741603245493,
+      "learning_rate": 0.0001936511241197055,
+      "loss": 1.690829873085022,
+      "memory(GiB)": 79.88,
+      "step": 57,
+      "train_speed(iter/s)": 0.003638
+    },
+    {
+      "clip_ratio/high_max": 0.00028280544211156666,
+      "clip_ratio/high_mean": 0.00028280544211156666,
+      "clip_ratio/low_mean": 0.023381341248750687,
+      "clip_ratio/low_min": 0.023381341248750687,
+      "clip_ratio/region_mean": 0.023664146661758423,
+      "completions/clipped_ratio": 0.046875,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 12.25,
+      "completions/min_length": 3.0,
+      "epoch": 0.6426592797783933,
+      "grad_norm": 2.545625270383155,
+      "kl": 130.25380403241434,
+      "learning_rate": 0.0001933250837887457,
+      "loss": 0.014269158244132996,
+      "memory(GiB)": 79.88,
+      "reward": 0.203125,
+      "reward_std": 0.3562659025192261,
+      "rewards/Response_no_think_reward/mean": 0.203125,
+      "rewards/Response_no_think_reward/std": 1.0600800514221191,
+      "step": 58,
+      "train_speed(iter/s)": 0.003488
+    },
+    {
+      "clip_ratio/high_max": 0.0029069767333567142,
+      "clip_ratio/high_mean": 0.0029069767333567142,
+      "clip_ratio/low_mean": 0.15388594008982182,
+      "clip_ratio/low_min": 0.15388594008982182,
+      "clip_ratio/region_mean": 0.15679291728883982,
+      "epoch": 0.6537396121883656,
+      "grad_norm": 16.25636826382254,
+      "kl": 3078.323552307884,
+      "learning_rate": 0.0001929911685973088,
+      "loss": 0.1047079786658287,
+      "memory(GiB)": 79.88,
+      "step": 59,
+      "train_speed(iter/s)": 0.003508
+    },
+    {
+      "clip_ratio/high_max": 0.0012499999720603228,
+      "clip_ratio/high_mean": 0.0012499999720603228,
+      "clip_ratio/low_mean": 0.032458390574902296,
+      "clip_ratio/low_min": 0.032458390574902296,
+      "clip_ratio/region_mean": 0.03370839054696262,
+      "completions/clipped_ratio": 0.125,
+      "completions/max_length": 51.0,
+      "completions/mean_length": 10.96875,
+      "completions/min_length": 4.0,
+      "epoch": 0.6648199445983379,
+      "grad_norm": 7.944638763107582,
+      "kl": 1613.5422123023309,
+      "learning_rate": 0.00019264940672148018,
+      "loss": 0.06612621992826462,
+      "memory(GiB)": 79.88,
+      "reward": -0.2109375,
+      "reward_std": 0.21648234128952026,
+      "rewards/Response_no_think_reward/mean": -0.2109375,
+      "rewards/Response_no_think_reward/std": 0.4636782705783844,
+      "step": 60,
+      "train_speed(iter/s)": 0.003389
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 360,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 5,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 0.0,
+  "train_batch_size": 8,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d687b8e1a620d39f463f0cdaff5c9b4a3e4b169125aef77a418326896c9811b
+size 10208

zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)