mudassiruddin21 commited on
Commit
2ba17bf
·
verified ·
1 Parent(s): c53c944

Run r1: logs and results

Browse files
runs/r1/engine_logs/20261118-034539-eval-r1.log ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ vllm serve /workspace/emc/models/target --served-model-name target --host 127.0.0.1 --port 43427 --dtype bfloat16 --max-model-len 16384 --gpu-memory-utilization 0.9 --max-num-seqs 4 --seed 0 --no-enable-prefix-caching --generation-config vllm --language-model-only --speculative-config {"method": "mtp", "model": "/workspace/emc/models/baseline-mtp", "num_speculative_tokens": 3}
2
+
3
+ (APIServer pid=13527) INFO 11-18 03:45:58 [api_utils.py:347]
4
+ (APIServer pid=13527) INFO 11-18 03:45:58 [api_utils.py:347] █ █ █▄ ▄█
5
+ (APIServer pid=13527) INFO 11-18 03:45:58 [api_utils.py:347] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.31.0
6
+ (APIServer pid=13527) INFO 11-18 03:45:58 [api_utils.py:347] █▄█▀ █ █ █ █ model /workspace/emc/models/target
7
+ (APIServer pid=13527) INFO 11-18 03:45:58 [api_utils.py:347] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀
8
+ (APIServer pid=13527) INFO 11-18 03:45:58 [api_utils.py:347]
9
+ (APIServer pid=13527) INFO 11-18 03:45:58 [api_utils.py:286] non-default args: {'model_tag': '/workspace/emc/models/target', 'host': '127.0.0.1', 'port': 43427, 'model': '/workspace/emc/models/target', 'dtype': 'bfloat16', 'max_model_len': 16384, 'served_model_name': ['target'], 'generation_config': 'vllm', 'gpu_memory_utilization': 0.9, 'enable_prefix_caching': False, 'language_model_only': True, 'max_num_seqs': 4, 'speculative_config': {'method': 'mtp', 'model': '/workspace/emc/models/baseline-mtp', 'num_speculative_tokens': 3}}
10
+ (APIServer pid=13527) INFO 11-18 03:45:58 [model.py:699] Resolved architecture: Qwen3_5ForConditionalGeneration
11
+ (APIServer pid=13527) INFO 11-18 03:45:58 [model.py:2109] Using max model len 16384
12
+ (APIServer pid=13527) INFO 11-18 03:46:00 [model.py:917] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode.
13
+ (APIServer pid=13527) INFO 11-18 03:46:00 [model.py:974] Disabled mm_prefix attention mode because multimodal inputs are configuration-disabled. Attention backends without mm_prefix support may now be selected.
14
+ (APIServer pid=13527) INFO 11-18 03:46:00 [model.py:699] Resolved architecture: Qwen3_5MTP
15
+ (APIServer pid=13527) INFO 11-18 03:46:00 [model.py:2109] Using max model len 262144
16
+ (APIServer pid=13527) WARNING 11-18 03:46:00 [speculative.py:1375] Enabling num_speculative_tokens > 1 will run multiple times of forward on same MTP layer,which may result in lower acceptance rate
17
+ (APIServer pid=13527) INFO 11-18 03:46:00 [speculative.py:1663] Overriding draft model max model len from 262144 to 16384
18
+ (APIServer pid=13527) INFO 11-18 03:46:00 [scheduler.py:308] Chunked prefill is enabled with max_num_batched_tokens=8192.
19
+ (APIServer pid=13527) INFO 11-18 03:46:00 [kernel.py:408] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'], gelu_and_mul_sparse=['triton', 'native'])
20
+ (EngineCore pid=13626) INFO 11-18 03:46:19 [core.py:129] Initializing a V1 LLM engine (v0.31.0) with config: model='/workspace/emc/models/target', speculative_config=SpeculativeConfig(method='mtp', model='/workspace/emc/models/baseline-mtp', num_spec_tokens=3), tokenizer='/workspace/emc/models/target', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=16384, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=False, aux_output_config=AuxOutputConfig(enable_return_routed_experts=False, max_bytes=None), kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, per_request_spec_decode_metrics='none', kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=target, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'ir_enable_torch_wrap': True, 'splitting_ops': ['vllm::unified_attention_with_output', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::qwen4_exp_ple_short_conv', 'vllm::qwen4_exp_qsa_with_output', 'vllm::linear_attention', 'vllm::qwen_gdn_attention_core', 'vllm::qwen_gdn_attention_core_fused_norm_packed', 'vllm::gdn_attention_core_xpu', 'vllm::olmo_hybrid_gdn_full_forward', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::deepseek_v4_attention', 'vllm::hpc_rope_norm_forward', 'vllm::unified_kv_cache_update', 'vllm::unified_mla_kv_cache_update'], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [8192], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 32, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'], gelu_and_mul_sparse=['triton', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, moe_backend='auto', sparse_indexer_topk_backend='auto', linear_backend='auto', linear_backend_per_quant=None)
21
+ (EngineCore pid=13626) INFO 11-18 03:46:21 [parallel_state.py:1825] world_size=1 rank=0 local_rank=0 distributed_init_method=file:///workspace/emc/tmp/vllm_dist_335ca7a8d4544e29bcb71f721eb56366 backend=nccl
22
+ (EngineCore pid=13626) INFO 11-18 03:46:21 [parallel_state.py:2269] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, ETP rank 0, EP rank N/A, EPLB rank N/A
23
+ (EngineCore pid=13626) INFO 11-18 03:46:21 [gpu_worker.py:479] Using V2 Model Runner
24
+ (EngineCore pid=13626) INFO 11-18 03:46:22 [model_runner.py:375] Loading model from scratch...
25
+ (EngineCore pid=13626) INFO 11-18 03:46:23 [cuda.py:585] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention
26
+ (EngineCore pid=13626) INFO 11-18 03:46:23 [mm_encoder_attention.py:372] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention.
27
+ (EngineCore pid=13626) INFO 11-18 03:46:23 [qwen_gdn_linear_attn.py:177] Using FlashInfer GDN prefill kernel (requested=auto, head_k_dim=128).
28
+ (EngineCore pid=13626) INFO 11-18 03:46:23 [qwen_gdn_linear_attn.py:533] GDN decode kernel: cuda
29
+ (EngineCore pid=13626) INFO 11-18 03:46:26 [cuda.py:526] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
30
+ (EngineCore pid=13626) INFO 11-18 03:46:26 [flash_attn.py:1167] Using FlashAttention version 2
31
+ (EngineCore pid=13626) INFO 11-18 03:46:26 [weight_utils.py:1063] Filesystem type for checkpoints: OVERLAY. Checkpoint size: 51.75 GiB. Available RAM: 156.09 GiB.
32
+ (EngineCore pid=13626) INFO 11-18 03:46:26 [weight_utils.py:1086] Auto-prefetch is disabled because the filesystem (OVERLAY) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
33
+
34
+ Loading safetensors checkpoint shards: 0% Completed | 0/18 [00:00<?, ?it/s]
35
+
36
+ Loading safetensors checkpoint shards: 6% Completed | 1/18 [00:00<00:07, 2.18it/s]
37
+
38
+ Loading safetensors checkpoint shards: 11% Completed | 2/18 [00:00<00:07, 2.25it/s]
39
+
40
+ Loading safetensors checkpoint shards: 17% Completed | 3/18 [00:01<00:05, 2.50it/s]
41
+
42
+ Loading safetensors checkpoint shards: 22% Completed | 4/18 [00:01<00:06, 2.17it/s]
43
+
44
+ Loading safetensors checkpoint shards: 28% Completed | 5/18 [00:02<00:05, 2.52it/s]
45
+
46
+ Loading safetensors checkpoint shards: 33% Completed | 6/18 [00:02<00:05, 2.22it/s]
47
+
48
+ Loading safetensors checkpoint shards: 39% Completed | 7/18 [00:02<00:04, 2.48it/s]
49
+
50
+ Loading safetensors checkpoint shards: 44% Completed | 8/18 [00:03<00:04, 2.21it/s]
51
+
52
+ Loading safetensors checkpoint shards: 50% Completed | 9/18 [00:03<00:03, 2.48it/s]
53
+
54
+ Loading safetensors checkpoint shards: 56% Completed | 10/18 [00:04<00:03, 2.17it/s]
55
+
56
+ Loading safetensors checkpoint shards: 61% Completed | 11/18 [00:04<00:02, 2.44it/s]
57
+
58
+ Loading safetensors checkpoint shards: 67% Completed | 12/18 [00:05<00:02, 2.18it/s]
59
+
60
+ Loading safetensors checkpoint shards: 72% Completed | 13/18 [00:05<00:02, 2.48it/s]
61
+
62
+ Loading safetensors checkpoint shards: 78% Completed | 14/18 [00:06<00:01, 2.21it/s]
63
+
64
+ Loading safetensors checkpoint shards: 83% Completed | 15/18 [00:06<00:01, 2.45it/s]
65
+
66
+ Loading safetensors checkpoint shards: 89% Completed | 16/18 [00:06<00:00, 2.14it/s]
67
+
68
+ Loading safetensors checkpoint shards: 94% Completed | 17/18 [00:07<00:00, 2.39it/s]
69
+
70
+ Loading safetensors checkpoint shards: 100% Completed | 18/18 [00:07<00:00, 2.48it/s]
71
+
72
+ Loading safetensors checkpoint shards: 100% Completed | 18/18 [00:07<00:00, 2.35it/s]
73
+
74
+ (EngineCore pid=13626) INFO 11-18 03:46:34 [default_loader.py:484] Loading weights took 7.69 seconds
75
+ (EngineCore pid=13626) INFO 11-18 03:46:34 [weight_utils.py:1063] Filesystem type for checkpoints: OVERLAY. Checkpoint size: 0.79 GiB. Available RAM: 156.85 GiB.
76
+
77
+ Loading safetensors checkpoint shards: 0% Completed | 0/1 [00:00<?, ?it/s]
78
+
79
+ Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00, 8.43it/s]
80
+
81
+ Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00, 8.42it/s]
82
+
83
+ (EngineCore pid=13626) INFO 11-18 03:46:34 [default_loader.py:484] Loading weights took 0.12 seconds
84
+ (EngineCore pid=13626) INFO 11-18 03:46:35 [model_runner.py:407] Model loading took 51.02 GiB memory and 12.771811 seconds
85
+ (EngineCore pid=13626) INFO 11-18 03:46:35 [topk_topp_sampler.py:96] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0.
86
+ (EngineCore pid=13626) WARNING 11-18 03:46:35 [torch_utils.py:274] OMP_NUM_THREADS=20 is set; leaving Torch threads at 20 for serving. Multi-threaded torch CPU ops during serving can degrade performance through spin-wait contention and cgroup CPU-quota throttling.
87
+ (EngineCore pid=13626) INFO 11-18 03:46:35 [interface.py:989] Setting attention block size to 800 tokens to ensure that attention page size is >= mamba page size.
88
+ (EngineCore pid=13626) INFO 11-18 03:46:35 [interface.py:1013] Padding mamba page size by 0.25% to ensure that mamba page size and attention page size are exactly equal.
89
+ (EngineCore pid=13626) INFO 11-18 03:46:35 [utils.py:343] Using LBNHC KV cache layout.
90
+ (EngineCore pid=13626) INFO 11-18 03:46:43 [backends.py:1090] Using cache directory: /workspace/emc/work/cache/vllm/torch_compile_cache/b82da1837d/rank_0_0/backbone for vLLM's torch.compile
91
+ (EngineCore pid=13626) INFO 11-18 03:46:43 [backends.py:1151] Dynamo bytecode transform time: 7.36 s
runs/r1/logs/eval_baseline.log ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ method: mtp
2
+ parameters: 424.7 M
3
+ weights: 0.849 GB
4
+ KV cache @ full context: 0.067 GB
5
+ memory metric: 0.917 GB
6
+ PASS
7
+ [03:45:39] repeat 1/1: fresh vLLM process
8
+ [03:45:39] starting vLLM (log: /workspace/emc/work/logs/20261118-034539-eval-r1.log)
runs/r1/logs/pipeline.log CHANGED
@@ -84,3 +84,4 @@
84
  [2026-11-18 03:45:29] done eval_main (1652s)
85
  [2026-11-18 03:45:29] start report_main
86
  [2026-11-18 03:45:29] done report_main (0s)
 
 
84
  [2026-11-18 03:45:29] done eval_main (1652s)
85
  [2026-11-18 03:45:29] start report_main
86
  [2026-11-18 03:45:29] done report_main (0s)
87
+ [2026-11-18 03:45:30] start eval_baseline
runs/r1/logs/pipeline.out CHANGED
@@ -84,3 +84,4 @@
84
  [2026-11-18 03:45:29] done eval_main (1652s)
85
  [2026-11-18 03:45:29] start report_main
86
  [2026-11-18 03:45:29] done report_main (0s)
 
 
84
  [2026-11-18 03:45:29] done eval_main (1652s)
85
  [2026-11-18 03:45:29] start report_main
86
  [2026-11-18 03:45:29] done report_main (0s)
87
+ [2026-11-18 03:45:30] start eval_baseline
runs/r1/logs/stats.log CHANGED
@@ -12,3 +12,4 @@
12
  [2026-11-18 03:36:54] ram 4.6/172 GB used (cache 0.3, shmem 0.0) | cgroup 4.6/172 GB | disk 93 GB used | gpu 83.1/95.6 GiB, 100%, 284 W | load 0.0 | largest VLLM::EngineCor 4.3 GB, vllm 2.5 GB, python 1.2 GB
13
  [2026-11-18 03:45:29] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 93 GB used | gpu 0.0/95.6 GiB, 0%, 80 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
14
  [2026-11-18 03:45:30] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 93 GB used | gpu 0.0/95.6 GiB, 0%, 75 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
 
 
12
  [2026-11-18 03:36:54] ram 4.6/172 GB used (cache 0.3, shmem 0.0) | cgroup 4.6/172 GB | disk 93 GB used | gpu 83.1/95.6 GiB, 100%, 284 W | load 0.0 | largest VLLM::EngineCor 4.3 GB, vllm 2.5 GB, python 1.2 GB
13
  [2026-11-18 03:45:29] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 93 GB used | gpu 0.0/95.6 GiB, 0%, 80 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
14
  [2026-11-18 03:45:30] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 93 GB used | gpu 0.0/95.6 GiB, 0%, 75 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
15
+ [2026-11-18 03:46:56] ram 3.6/172 GB used (cache 0.3, shmem 0.0) | cgroup 3.5/172 GB | disk 93 GB used | gpu 53.6/95.6 GiB, 0%, 76 W | load 0.0 | largest VLLM::EngineCor 2.9 GB, vllm 2.4 GB, python 1.2 GB
runs/r1/logs/sync.log CHANGED
@@ -24,3 +24,4 @@ uploaded sub_fp8 to mudassiruddin21/EMC:submissions/r1/main-fp8
24
  uploaded logs/results to mudassiruddin21/EMC:runs/r1
25
  uploaded logs/results to mudassiruddin21/EMC:runs/r1
26
  uploaded logs/results to mudassiruddin21/EMC:runs/r1
 
 
24
  uploaded logs/results to mudassiruddin21/EMC:runs/r1
25
  uploaded logs/results to mudassiruddin21/EMC:runs/r1
26
  uploaded logs/results to mudassiruddin21/EMC:runs/r1
27
+ uploaded logs/results to mudassiruddin21/EMC:runs/r1