mudassiruddin21 commited on
Commit
1819465
·
verified ·
1 Parent(s): b9e90b4

Run r2: logs and results

Browse files
runs/r2/engine_logs/20261118-182934-eval-r1.log ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ vllm serve /workspace/emc/models/target --served-model-name target --host 127.0.0.1 --port 21883 --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=32660) INFO 11-18 18:29:53 [api_utils.py:347]
4
+ (APIServer pid=32660) INFO 11-18 18:29:53 [api_utils.py:347] █ █ █▄ ▄█
5
+ (APIServer pid=32660) INFO 11-18 18:29:53 [api_utils.py:347] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.31.0
6
+ (APIServer pid=32660) INFO 11-18 18:29:53 [api_utils.py:347] █▄█▀ █ █ █ █ model /workspace/emc/models/target
7
+ (APIServer pid=32660) INFO 11-18 18:29:53 [api_utils.py:347] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀
8
+ (APIServer pid=32660) INFO 11-18 18:29:53 [api_utils.py:347]
9
+ (APIServer pid=32660) INFO 11-18 18:29:53 [api_utils.py:286] non-default args: {'model_tag': '/workspace/emc/models/target', 'host': '127.0.0.1', 'port': 21883, '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=32660) INFO 11-18 18:29:53 [model.py:699] Resolved architecture: Qwen3_5ForConditionalGeneration
11
+ (APIServer pid=32660) INFO 11-18 18:29:53 [model.py:2109] Using max model len 16384
12
+ (APIServer pid=32660) INFO 11-18 18:29:55 [model.py:917] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode.
13
+ (APIServer pid=32660) INFO 11-18 18:29:55 [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=32660) INFO 11-18 18:29:55 [model.py:699] Resolved architecture: Qwen3_5MTP
15
+ (APIServer pid=32660) INFO 11-18 18:29:55 [model.py:2109] Using max model len 262144
16
+ (APIServer pid=32660) WARNING 11-18 18:29:55 [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=32660) INFO 11-18 18:29:55 [speculative.py:1663] Overriding draft model max model len from 262144 to 16384
18
+ (APIServer pid=32660) INFO 11-18 18:29:55 [scheduler.py:308] Chunked prefill is enabled with max_num_batched_tokens=8192.
19
+ (APIServer pid=32660) INFO 11-18 18:29:55 [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=32759) INFO 11-18 18:30:13 [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=32759) INFO 11-18 18:30:15 [parallel_state.py:1825] world_size=1 rank=0 local_rank=0 distributed_init_method=file:///workspace/emc/tmp/vllm_dist_05ebc840314b45c2a6c4935f56326083 backend=nccl
22
+ (EngineCore pid=32759) INFO 11-18 18:30:15 [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=32759) INFO 11-18 18:30:15 [gpu_worker.py:479] Using V2 Model Runner
24
+ (EngineCore pid=32759) INFO 11-18 18:30:16 [model_runner.py:375] Loading model from scratch...
25
+ (EngineCore pid=32759) INFO 11-18 18:30:17 [cuda.py:585] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention
26
+ (EngineCore pid=32759) INFO 11-18 18:30:17 [mm_encoder_attention.py:372] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention.
27
+ (EngineCore pid=32759) INFO 11-18 18:30:17 [qwen_gdn_linear_attn.py:177] Using FlashInfer GDN prefill kernel (requested=auto, head_k_dim=128).
28
+ (EngineCore pid=32759) INFO 11-18 18:30:17 [qwen_gdn_linear_attn.py:533] GDN decode kernel: cuda
29
+ (EngineCore pid=32759) INFO 11-18 18:30:20 [cuda.py:526] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
30
+ (EngineCore pid=32759) INFO 11-18 18:30:20 [flash_attn.py:1167] Using FlashAttention version 2
31
+ (EngineCore pid=32759) INFO 11-18 18:30:20 [weight_utils.py:1063] Filesystem type for checkpoints: OVERLAY. Checkpoint size: 51.75 GiB. Available RAM: 156.13 GiB.
32
+ (EngineCore pid=32759) INFO 11-18 18:30:20 [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, 2.36it/s]
37
+
38
+ Loading safetensors checkpoint shards: 11% Completed | 2/18 [00:00<00:07, 2.27it/s]
39
+
40
+ Loading safetensors checkpoint shards: 17% Completed | 3/18 [00:01<00:05, 2.51it/s]
41
+
42
+ Loading safetensors checkpoint shards: 22% Completed | 4/18 [00:01<00:06, 2.15it/s]
43
+
44
+ Loading safetensors checkpoint shards: 28% Completed | 5/18 [00:02<00:05, 2.50it/s]
45
+
46
+ Loading safetensors checkpoint shards: 33% Completed | 6/18 [00:02<00:05, 2.24it/s]
47
+
48
+ Loading safetensors checkpoint shards: 39% Completed | 7/18 [00:02<00:04, 2.56it/s]
49
+
50
+ Loading safetensors checkpoint shards: 44% Completed | 8/18 [00:03<00:04, 2.31it/s]
51
+
52
+ Loading safetensors checkpoint shards: 50% Completed | 9/18 [00:03<00:03, 2.61it/s]
53
+
54
+ Loading safetensors checkpoint shards: 56% Completed | 10/18 [00:04<00:03, 2.32it/s]
55
+
56
+ Loading safetensors checkpoint shards: 61% Completed | 11/18 [00:04<00:02, 2.59it/s]
57
+
58
+ Loading safetensors checkpoint shards: 67% Completed | 12/18 [00:05<00:02, 2.26it/s]
59
+
60
+ Loading safetensors checkpoint shards: 72% Completed | 13/18 [00:05<00:01, 2.53it/s]
61
+
62
+ Loading safetensors checkpoint shards: 78% Completed | 14/18 [00:05<00:01, 2.26it/s]
63
+
64
+ Loading safetensors checkpoint shards: 83% Completed | 15/18 [00:06<00:01, 2.54it/s]
65
+
66
+ Loading safetensors checkpoint shards: 89% Completed | 16/18 [00:06<00:00, 2.25it/s]
67
+
68
+ Loading safetensors checkpoint shards: 94% Completed | 17/18 [00:07<00:00, 2.52it/s]
69
+
70
+ Loading safetensors checkpoint shards: 100% Completed | 18/18 [00:07<00:00, 2.62it/s]
71
+
72
+ Loading safetensors checkpoint shards: 100% Completed | 18/18 [00:07<00:00, 2.43it/s]
73
+
74
+ (EngineCore pid=32759) INFO 11-18 18:30:28 [default_loader.py:484] Loading weights took 7.43 seconds
75
+ (EngineCore pid=32759) INFO 11-18 18:30:28 [weight_utils.py:1063] Filesystem type for checkpoints: OVERLAY. Checkpoint size: 0.79 GiB. Available RAM: 156.89 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, 2.21it/s]
80
+
81
+ Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00, 2.21it/s]
82
+
83
+ (EngineCore pid=32759) INFO 11-18 18:30:28 [default_loader.py:484] Loading weights took 0.46 seconds
84
+ (EngineCore pid=32759) INFO 11-18 18:30:29 [model_runner.py:407] Model loading took 51.02 GiB memory and 12.712775 seconds
85
+ (EngineCore pid=32759) INFO 11-18 18:30:29 [topk_topp_sampler.py:96] FlashInfer top-p/top-k sampling disabled via VLLM_USE_FLASHINFER_SAMPLER=0.
86
+ (EngineCore pid=32759) WARNING 11-18 18:30:29 [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=32759) INFO 11-18 18:30:29 [interface.py:989] Setting attention block size to 800 tokens to ensure that attention page size is >= mamba page size.
88
+ (EngineCore pid=32759) INFO 11-18 18:30:29 [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=32759) INFO 11-18 18:30:29 [utils.py:343] Using LBNHC KV cache layout.
90
+ (EngineCore pid=32759) INFO 11-18 18:30:37 [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=32759) INFO 11-18 18:30:37 [backends.py:1151] Dynamo bytecode transform time: 7.51 s
92
+ (EngineCore pid=32759) INFO 11-18 18:30:52 [backends.py:391] Compiling a graph for compile range (1, 8192) takes 15.25 s
93
+ (EngineCore pid=32759) INFO 11-18 18:30:56 [backends.py:916] collected artifacts: 65 entries, 20 artifacts, 26020849 bytes total
94
+ (EngineCore pid=32759) INFO 11-18 18:30:56 [decorators.py:717] saved AOT compiled function to /workspace/emc/work/cache/vllm/torch_compile_cache/torch_aot_compile/4576de1741488f67af65fd90310b9e3b6d01c1bc0ff22ab9ad27127b1cf561b8/rank_0_0/model
95
+ (EngineCore pid=32759) INFO 11-18 18:30:56 [monitor.py:53] torch.compile took 26.79 s in total
96
+ (EngineCore pid=32759) INFO 11-18 18:31:10 [monitor.py:81] Initial profiling/warmup run took 13.41 s
97
+ (EngineCore pid=32759) INFO 11-18 18:31:10 [backends.py:1090] Using cache directory: /workspace/emc/work/cache/vllm/torch_compile_cache/b82da1837d/rank_0_0/eagle_head for vLLM's torch.compile
98
+ (EngineCore pid=32759) INFO 11-18 18:31:10 [backends.py:1151] Dynamo bytecode transform time: 0.40 s
99
+ (EngineCore pid=32759) INFO 11-18 18:31:12 [backends.py:391] Compiling a graph for compile range (1, 8192) takes 1.95 s
100
+ (EngineCore pid=32759) INFO 11-18 18:31:12 [backends.py:916] collected artifacts: 2 entries, 2 artifacts, 3600315 bytes total
101
+ (EngineCore pid=32759) INFO 11-18 18:31:12 [decorators.py:717] saved AOT compiled function to /workspace/emc/work/cache/vllm/torch_compile_cache/torch_aot_compile/a2d91f09f00eca7918a4a1313b5efe9a834db12eecbb11300ae44c97b5d96aed/rank_0_0/model
102
+ (EngineCore pid=32759) INFO 11-18 18:31:12 [monitor.py:53] torch.compile took 2.55 s in total
103
+ (EngineCore pid=32759) INFO 11-18 18:31:12 [monitor.py:81] Initial profiling/warmup run took 0.09 s
104
+ (EngineCore pid=32759) WARNING 11-18 18:31:14 [kv_cache_utils.py:1574] Add 3 padding layers, may waste at most 6.25% KV cache memory
105
+
106
+ Capturing CUDA graphs (PIECEWISE): 0%| | 0/7 [00:00<?, ?it/s]
107
+ Capturing CUDA graphs (PIECEWISE): 14%|█▍ | 1/7 [00:06<00:38, 6.44s/it]
108
+ Capturing CUDA graphs (PIECEWISE): 29%|██▊ | 2/7 [00:06<00:13, 2.74s/it]
109
+ Capturing CUDA graphs (PIECEWISE): 43%|████▎ | 3/7 [00:06<00:06, 1.55s/it]
110
+ Capturing CUDA graphs (PIECEWISE): 57%|█████▋ | 4/7 [00:06<00:02, 1.00it/s]
111
+ Capturing CUDA graphs (PIECEWISE): 71%|███████▏ | 5/7 [00:07<00:01, 1.43it/s]
112
+ Capturing CUDA graphs (PIECEWISE): 100%|██████████| 7/7 [00:07<00:00, 2.53it/s]
113
+ Capturing CUDA graphs (PIECEWISE): 100%|██████████| 7/7 [00:07<00:00, 1.04s/it]
114
+
115
+ Capturing CUDA graphs (FULL): 0%| | 0/2 [00:00<?, ?it/s]
116
+ Capturing CUDA graphs (FULL): 100%|██████████| 2/2 [00:00<00:00, 13.83it/s]
117
+ Capturing CUDA graphs (FULL): 100%|██████████| 2/2 [00:00<00:00, 13.81it/s]
118
+ (EngineCore pid=32759) INFO 11-18 18:31:22 [speculator.py:144] Capturing model for speculator...
119
+
120
+ Capturing prefill CUDA graphs (PIECEWISE): 0%| | 0/7 [00:00<?, ?it/s]
121
+ Capturing prefill CUDA graphs (PIECEWISE): 100%|██████████| 7/7 [00:00<00:00, 146.00it/s]
122
+
123
+ Capturing prefill CUDA graphs (FULL): 0%| | 0/3 [00:00<?, ?it/s]
124
+ Capturing prefill CUDA graphs (FULL): 100%|██████████| 3/3 [00:00<00:00, 174.17it/s]
125
+
126
+ Capturing decode CUDA graphs (FULL): 0%| | 0/3 [00:00<?, ?it/s]
127
+ Capturing decode CUDA graphs (FULL): 100%|██████████| 3/3 [00:00<00:00, 95.89it/s]
128
+ (EngineCore pid=32759) INFO 11-18 18:31:22 [model_runner.py:1087] Graph capturing finished in 8 secs, took 0.23 GiB
129
+ (EngineCore pid=32759) INFO 11-18 18:31:22 [gpu_worker.py:692] Available KV cache memory: 30.91 GiB
130
+ (EngineCore pid=32759) INFO 11-18 18:31:22 [gpu_worker.py:707] CUDA graph memory profiling is enabled (default since v0.21.0). The current --gpu-memory-utilization=0.9000 is equivalent to --gpu-memory-utilization=0.8975 without CUDA graph memory profiling. To maintain the same effective KV cache size as before, increase --gpu-memory-utilization to 0.9025. To disable, set VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0.
131
+ (EngineCore pid=32759) WARNING 11-18 18:31:22 [kv_cache_utils.py:1574] Add 3 padding layers, may waste at most 6.25% KV cache memory
132
+ (EngineCore pid=32759) INFO 11-18 18:31:22 [kv_cache_utils.py:2464] GPU KV cache size: 295,408 tokens, Maximum concurrency for 16,384 tokens per request: 18.03x
133
+ (EngineCore pid=32759) INFO 11-18 18:31:23 [kernel_warmup.py:172] JIT kernel warmup starting.
134
+ (EngineCore pid=32759) INFO 11-18 18:31:23 [kernel_warmup.py:185] JIT kernel warmup finished in 0.00s.
135
+ (EngineCore pid=32759) INFO 11-18 18:31:23 [qwen_triton_warmup.py:294] Warming up Qwen GDN Triton kernels for model_type=qwen3_5_text.
136
+ (EngineCore pid=32759) INFO 11-18 18:31:23 [qwen_vl_triton_warmup.py:98] Warmed M-RoPE Triton kernels.
137
+ (EngineCore pid=32759) INFO 11-18 18:31:23 [mamba_triton_warmup.py:42] Warmed Mamba batch_memcpy_kernel.
138
+ (EngineCore pid=32759) INFO 11-18 18:31:24 [kernel_warmup.py:437] Using FlashInfer autotune cache file: /workspace/emc/work/cache/vllm/flashinfer_autotune_cache/0.7.0.post1/120f/15d616625102b0759db551924078ebe0ac352718b248a2d422bcc4f02f6daf97/autotune_configs.json
139
+ 2026-11-18 18:31:24,043 - INFO - autotuner.py:1046 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
140
+ (EngineCore pid=32759) INFO 11-18 18:31:24 [kernel_warmup.py:359] Running FlashInfer autotune with 8192 tokens and token buckets (1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 768, 1024, 1280, 1536, 1792, 2048, 2560, 3072, 3584, 4096, 8192).
141
+ 2026-11-18 18:31:24,067 - INFO - autotuner.py:1069 - flashinfer.jit: [Autotuner]: Autotuning process ends
142
+ 2026-11-18 18:31:24,225 - INFO - autotuner.py:3554 - flashinfer.jit: [Autotuner]: Saved 0 configs to /workspace/emc/work/cache/vllm/flashinfer_autotune_cache/0.7.0.post1/120f/15d616625102b0759db551924078ebe0ac352718b248a2d422bcc4f02f6daf97/autotune_configs.json (0 new, 0 from previous config)
143
+
144
+ Capturing CUDA graphs (PIECEWISE): 0%| | 0/7 [00:00<?, ?it/s]
145
+ Capturing CUDA graphs (PIECEWISE): 14%|█▍ | 1/7 [00:00<00:00, 6.70it/s]
146
+ Capturing CUDA graphs (PIECEWISE): 29%|██▊ | 2/7 [00:00<00:00, 6.62it/s]
147
+ Capturing CUDA graphs (PIECEWISE): 43%|████▎ | 3/7 [00:00<00:00, 6.63it/s]
148
+ Capturing CUDA graphs (PIECEWISE): 57%|█████▋ | 4/7 [00:00<00:00, 6.61it/s]
149
+ Capturing CUDA graphs (PIECEWISE): 71%|███████▏ | 5/7 [00:00<00:00, 7.53it/s]
150
+ Capturing CUDA graphs (PIECEWISE): 100%|██████████| 7/7 [00:00<00:00, 8.77it/s]
151
+ Capturing CUDA graphs (PIECEWISE): 100%|██████████| 7/7 [00:00<00:00, 7.81it/s]
152
+
153
+ Capturing CUDA graphs (FULL): 0%| | 0/3 [00:00<?, ?it/s]
154
+ Capturing CUDA graphs (FULL): 67%|██████▋ | 2/3 [00:00<00:00, 14.85it/s]
155
+ Capturing CUDA graphs (FULL): 100%|██████████| 3/3 [00:00<00:00, 14.67it/s]
156
+ (EngineCore pid=32759) INFO 11-18 18:31:27 [speculator.py:144] Capturing model for speculator...
157
+
158
+ Capturing prefill CUDA graphs (PIECEWISE): 0%| | 0/7 [00:00<?, ?it/s]
159
+ Capturing prefill CUDA graphs (PIECEWISE): 100%|██████████| 7/7 [00:00<00:00, 143.33it/s]
160
+
161
+ Capturing prefill CUDA graphs (FULL): 0%| | 0/3 [00:00<?, ?it/s]
162
+ Capturing prefill CUDA graphs (FULL): 100%|██████████| 3/3 [00:00<00:00, 172.06it/s]
163
+
164
+ Capturing decode CUDA graphs (FULL): 0%| | 0/3 [00:00<?, ?it/s]
165
+ Capturing decode CUDA graphs (FULL): 100%|██████████| 3/3 [00:00<00:00, 106.48it/s]
166
+ (EngineCore pid=32759) INFO 11-18 18:31:27 [model_runner.py:1087] Graph capturing finished in 2 secs, took 0.17 GiB
167
+ (EngineCore pid=32759) INFO 11-18 18:31:27 [gpu_worker.py:892] CUDA graph pool memory: 0.17 GiB (actual), 0.24 GiB (estimated), difference: 0.07 GiB (40.7%).
168
+ (EngineCore pid=32759) INFO 11-18 18:31:27 [gpu_worker.py:955] Free memory on device (94.43/94.97 GiB) on startup. Desired GPU memory utilization is (0.9, 85.48 GiB). Actual usage is 51.37 GiB for consumed memory (weights + non-torch), 3.2 GiB for peak activation, and 0.17 GiB for CUDAGraph memory. Replace gpu_memory_utilization config with `--kv-cache-memory=32850334516` (30.59 GiB) to fit into requested memory, or `--kv-cache-memory=42462617600` (39.55 GiB) to fully utilize gpu memory. Current kv cache memory in use is 30.91 GiB.
169
+ (EngineCore pid=32759) INFO 11-18 18:31:28 [jit_monitor.py:84] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
170
+ (EngineCore pid=32759) INFO 11-18 18:31:29 [core.py:373] init engine (profile, create kv cache, warmup model) took 60.26 s (compilation: 29.34 s)
171
+ (EngineCore pid=32759) INFO 11-18 18:31:29 [kv_cache_utils.py:772] kv cache group sizes [16384, 16384, 16384, 800]
172
+ (EngineCore pid=32759) INFO 11-18 18:31:29 [kv_cache_utils.py:773] kv lcm block sizes 409600
173
+ (EngineCore pid=32759) INFO 11-18 18:31:29 [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'])
174
+ (APIServer pid=32660) INFO 11-18 18:31:29 [entry.py:132] Supported tasks: ['generate']
175
+ (APIServer pid=32660) INFO 11-18 18:31:29 [factories.py:76] Scale-out endpoints are disabled. Set --enable-scale-out to enable them.
176
+ (APIServer pid=32660) INFO 11-18 18:31:30 [hf.py:646] Detected the chat template content format to be 'openai'. You can set `--chat-template-content-format` to override this.
177
+ (APIServer pid=32660) WARNING 11-18 18:31:30 [serving.py:221] The chat template requires system-first ordering, so inline system messages in /v1/messages requests (e.g. host Code's per-turn reminders) are merged into the leading system prompt. Each new one changes the prompt prefix, so the rest of the conversation misses the prefix cache. Pass a --chat-template that accepts non-leading system messages to avoid this.
178
+ (APIServer pid=32660) WARNING 11-18 18:31:30 [serving.py:221] The chat template requires system-first ordering, so inline system messages in /v1/messages requests (e.g. host Code's per-turn reminders) are merged into the leading system prompt. Each new one changes the prompt prefix, so the rest of the conversation misses the prefix cache. Pass a --chat-template that accepts non-leading system messages to avoid this.
179
+ (APIServer pid=32660) INFO 11-18 18:31:30 [entry.py:136] Starting vLLM server on http://127.0.0.1:21883
180
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:60] Available routes are:
181
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /openapi.json, Methods: HEAD, GET
182
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /docs, Methods: HEAD, GET
183
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /docs/oauth2-redirect, Methods: HEAD, GET
184
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /redoc, Methods: HEAD, GET
185
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /load, Methods: GET
186
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /version, Methods: GET
187
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /kv_event_sources, Methods: GET
188
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /health, Methods: GET
189
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /metrics, Methods: GET
190
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /tokenize, Methods: POST
191
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /detokenize, Methods: POST
192
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/models, Methods: GET
193
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /ping, Methods: GET
194
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /ping, Methods: POST
195
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /invocations, Methods: POST
196
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/chat/completions, Methods: POST
197
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/chat/completions/batch, Methods: POST
198
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/responses, Methods: POST
199
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/responses/{response_id}, Methods: GET
200
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/responses/{response_id}/cancel, Methods: POST
201
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/completions, Methods: POST
202
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/messages, Methods: POST
203
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /v1/messages/count_tokens, Methods: POST
204
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /generative_scoring, Methods: POST
205
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /scale_elastic_ep, Methods: POST
206
+ (APIServer pid=32660) INFO 11-18 18:31:30 [launcher.py:69] Route: /is_scaling_elastic_ep, Methods: POST
207
+ INFO: Started server process [32660]
208
+ INFO: Waiting for application startup.
209
+ INFO: Application startup complete.
210
+ INFO: 127.0.0.1:52228 - "GET /health HTTP/1.1" 200 OK
211
+ INFO: 127.0.0.1:37132 - "POST /v1/completions HTTP/1.1" 200 OK
212
+ (EngineCore pid=32759) WARNING 11-18 18:31:32 [jit_monitor.py:140] Triton kernel JIT compilation during inference: _compute_local_logits_stats_kernel. This causes a latency spike; consider extending warmup to cover this shape/config.
213
+ (EngineCore pid=32759) WARNING 11-18 18:31:32 [jit_monitor.py:140] Triton kernel JIT compilation during inference: _rejection_kernel. This causes a latency spike; consider extending warmup to cover this shape/config.
214
+ (EngineCore pid=32759) WARNING 11-18 18:31:32 [jit_monitor.py:140] Triton kernel JIT compilation during inference: _resample_kernel. This causes a latency spike; consider extending warmup to cover this shape/config.
215
+ INFO: 127.0.0.1:62990 - "POST /v1/completions HTTP/1.1" 200 OK
216
+ (APIServer pid=32660) INFO 11-18 18:31:40 [loggers.py:320] Engine 000: Avg prompt throughput: 16.5 tokens/s, Avg generation throughput: 41.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
217
+ (APIServer pid=32660) INFO 11-18 18:31:40 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.11, Accepted throughput: 28.23 tokens/s, Drafted throughput: 40.09 tokens/s, Accepted: 319 tokens, Drafted: 453 tokens, Per-position acceptance rate: 0.815, 0.715, 0.583, Avg Draft acceptance rate: 70.4%
218
+ INFO: 127.0.0.1:45201 - "POST /v1/completions HTTP/1.1" 200 OK
219
+ INFO: 127.0.0.1:42567 - "POST /v1/completions HTTP/1.1" 200 OK
220
+ INFO: 127.0.0.1:33888 - "POST /v1/completions HTTP/1.1" 200 OK
221
+ (APIServer pid=32660) INFO 11-18 18:31:50 [loggers.py:320] Engine 000: Avg prompt throughput: 42.7 tokens/s, Avg generation throughput: 62.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
222
+ (APIServer pid=32660) INFO 11-18 18:31:50 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.54, Accepted throughput: 45.19 tokens/s, Drafted throughput: 53.39 tokens/s, Accepted: 452 tokens, Drafted: 534 tokens, Per-position acceptance rate: 0.938, 0.826, 0.775, Avg Draft acceptance rate: 84.6%
223
+ INFO: 127.0.0.1:64093 - "POST /v1/completions HTTP/1.1" 200 OK
224
+ INFO: 127.0.0.1:24131 - "POST /v1/completions HTTP/1.1" 200 OK
225
+ (APIServer pid=32660) INFO 11-18 18:32:00 [loggers.py:320] Engine 000: Avg prompt throughput: 14.0 tokens/s, Avg generation throughput: 52.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
226
+ (APIServer pid=32660) INFO 11-18 18:32:00 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.91, Accepted throughput: 34.50 tokens/s, Drafted throughput: 54.29 tokens/s, Accepted: 345 tokens, Drafted: 543 tokens, Per-position acceptance rate: 0.845, 0.613, 0.448, Avg Draft acceptance rate: 63.5%
227
+ (APIServer pid=32660) INFO 11-18 18:32:10 [loggers.py:320] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 44.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
228
+ (APIServer pid=32660) INFO 11-18 18:32:10 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.40, Accepted throughput: 25.80 tokens/s, Drafted throughput: 55.19 tokens/s, Accepted: 258 tokens, Drafted: 552 tokens, Per-position acceptance rate: 0.712, 0.446, 0.245, Avg Draft acceptance rate: 46.7%
229
+ INFO: 127.0.0.1:53249 - "POST /v1/completions HTTP/1.1" 200 OK
230
+ (APIServer pid=32660) INFO 11-18 18:32:20 [loggers.py:320] Engine 000: Avg prompt throughput: 7.8 tokens/s, Avg generation throughput: 41.7 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
231
+ (APIServer pid=32660) INFO 11-18 18:32:20 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.30, Accepted throughput: 23.60 tokens/s, Drafted throughput: 54.60 tokens/s, Accepted: 236 tokens, Drafted: 546 tokens, Per-position acceptance rate: 0.665, 0.396, 0.236, Avg Draft acceptance rate: 43.2%
232
+ (APIServer pid=32660) INFO 11-18 18:32:30 [loggers.py:320] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 40.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
233
+ (APIServer pid=32660) INFO 11-18 18:32:30 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.17, Accepted throughput: 21.70 tokens/s, Drafted throughput: 55.50 tokens/s, Accepted: 217 tokens, Drafted: 555 tokens, Per-position acceptance rate: 0.649, 0.351, 0.173, Avg Draft acceptance rate: 39.1%
234
+ INFO: 127.0.0.1:20517 - "GET /metrics HTTP/1.1" 200 OK
235
+ INFO: 127.0.0.1:39762 - "POST /v1/completions HTTP/1.1" 200 OK
236
+ (APIServer pid=32660) INFO 11-18 18:32:40 [loggers.py:320] Engine 000: Avg prompt throughput: 7.2 tokens/s, Avg generation throughput: 41.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
237
+ (APIServer pid=32660) INFO 11-18 18:32:40 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.26, Accepted throughput: 22.90 tokens/s, Drafted throughput: 54.59 tokens/s, Accepted: 229 tokens, Drafted: 546 tokens, Per-position acceptance rate: 0.654, 0.379, 0.225, Avg Draft acceptance rate: 41.9%
238
+ INFO: 127.0.0.1:39583 - "POST /v1/completions HTTP/1.1" 200 OK
239
+ (APIServer pid=32660) INFO 11-18 18:32:50 [loggers.py:320] Engine 000: Avg prompt throughput: 3.9 tokens/s, Avg generation throughput: 50.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
240
+ (APIServer pid=32660) INFO 11-18 18:32:50 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.77, Accepted throughput: 32.30 tokens/s, Drafted throughput: 54.59 tokens/s, Accepted: 323 tokens, Drafted: 546 tokens, Per-position acceptance rate: 0.780, 0.599, 0.396, Avg Draft acceptance rate: 59.2%
241
+ INFO: 127.0.0.1:18085 - "POST /v1/completions HTTP/1.1" 200 OK
242
+ (APIServer pid=32660) INFO 11-18 18:33:00 [loggers.py:320] Engine 000: Avg prompt throughput: 6.9 tokens/s, Avg generation throughput: 63.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
243
+ (APIServer pid=32660) INFO 11-18 18:33:00 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.45, Accepted throughput: 44.80 tokens/s, Drafted throughput: 54.90 tokens/s, Accepted: 448 tokens, Drafted: 549 tokens, Per-position acceptance rate: 0.923, 0.803, 0.721, Avg Draft acceptance rate: 81.6%
244
+ INFO: 127.0.0.1:21601 - "POST /v1/completions HTTP/1.1" 200 OK
245
+ (APIServer pid=32660) INFO 11-18 18:33:10 [loggers.py:320] Engine 000: Avg prompt throughput: 5.0 tokens/s, Avg generation throughput: 62.4 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
246
+ (APIServer pid=32660) INFO 11-18 18:33:10 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.40, Accepted throughput: 44.20 tokens/s, Drafted throughput: 55.20 tokens/s, Accepted: 442 tokens, Drafted: 552 tokens, Per-position acceptance rate: 0.897, 0.777, 0.728, Avg Draft acceptance rate: 80.1%
247
+ INFO: 127.0.0.1:33875 - "POST /v1/completions HTTP/1.1" 200 OK
248
+ (APIServer pid=32660) INFO 11-18 18:33:20 [loggers.py:320] Engine 000: Avg prompt throughput: 10.3 tokens/s, Avg generation throughput: 55.0 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
249
+ (APIServer pid=32660) INFO 11-18 18:33:20 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.00, Accepted throughput: 36.79 tokens/s, Drafted throughput: 55.19 tokens/s, Accepted: 368 tokens, Drafted: 552 tokens, Per-position acceptance rate: 0.848, 0.658, 0.495, Avg Draft acceptance rate: 66.7%
250
+ INFO: 127.0.0.1:31925 - "POST /v1/completions HTTP/1.1" 200 OK
251
+ (APIServer pid=32660) INFO 11-18 18:33:30 [loggers.py:320] Engine 000: Avg prompt throughput: 6.4 tokens/s, Avg generation throughput: 56.7 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
252
+ (APIServer pid=32660) INFO 11-18 18:33:30 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.09, Accepted throughput: 38.29 tokens/s, Drafted throughput: 54.89 tokens/s, Accepted: 383 tokens, Drafted: 549 tokens, Per-position acceptance rate: 0.869, 0.699, 0.525, Avg Draft acceptance rate: 69.8%
253
+ INFO: 127.0.0.1:36276 - "POST /v1/completions HTTP/1.1" 200 OK
254
+ (APIServer pid=32660) INFO 11-18 18:33:40 [loggers.py:320] Engine 000: Avg prompt throughput: 2.6 tokens/s, Avg generation throughput: 47.0 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
255
+ (APIServer pid=32660) INFO 11-18 18:33:40 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.57, Accepted throughput: 28.70 tokens/s, Drafted throughput: 54.90 tokens/s, Accepted: 287 tokens, Drafted: 549 tokens, Per-position acceptance rate: 0.781, 0.481, 0.306, Avg Draft acceptance rate: 52.3%
256
+ INFO: 127.0.0.1:33486 - "POST /v1/completions HTTP/1.1" 200 OK
257
+ INFO: 127.0.0.1:21880 - "POST /v1/completions HTTP/1.1" 200 OK
258
+ (APIServer pid=32660) INFO 11-18 18:33:50 [loggers.py:320] Engine 000: Avg prompt throughput: 26.7 tokens/s, Avg generation throughput: 53.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
259
+ (APIServer pid=32660) INFO 11-18 18:33:50 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.97, Accepted throughput: 35.70 tokens/s, Drafted throughput: 54.29 tokens/s, Accepted: 357 tokens, Drafted: 543 tokens, Per-position acceptance rate: 0.823, 0.652, 0.497, Avg Draft acceptance rate: 65.7%
260
+ INFO: 127.0.0.1:49808 - "POST /v1/completions HTTP/1.1" 200 OK
261
+ (APIServer pid=32660) INFO 11-18 18:34:00 [loggers.py:320] Engine 000: Avg prompt throughput: 5.4 tokens/s, Avg generation throughput: 55.0 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
262
+ (APIServer pid=32660) INFO 11-18 18:34:00 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.99, Accepted throughput: 36.70 tokens/s, Drafted throughput: 55.19 tokens/s, Accepted: 367 tokens, Drafted: 552 tokens, Per-position acceptance rate: 0.864, 0.652, 0.478, Avg Draft acceptance rate: 66.5%
263
+ INFO: 127.0.0.1:50186 - "POST /v1/completions HTTP/1.1" 200 OK
264
+ INFO: 127.0.0.1:55890 - "POST /v1/completions HTTP/1.1" 200 OK
265
+ (APIServer pid=32660) INFO 11-18 18:34:10 [loggers.py:320] Engine 000: Avg prompt throughput: 26.5 tokens/s, Avg generation throughput: 62.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
266
+ (APIServer pid=32660) INFO 11-18 18:34:10 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.45, Accepted throughput: 44.40 tokens/s, Drafted throughput: 54.29 tokens/s, Accepted: 444 tokens, Drafted: 543 tokens, Per-position acceptance rate: 0.934, 0.801, 0.718, Avg Draft acceptance rate: 81.8%
267
+ INFO: 127.0.0.1:56818 - "POST /v1/completions HTTP/1.1" 200 OK
268
+ (APIServer pid=32660) INFO 11-18 18:34:20 [loggers.py:320] Engine 000: Avg prompt throughput: 3.4 tokens/s, Avg generation throughput: 55.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
269
+ (APIServer pid=32660) INFO 11-18 18:34:20 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.03, Accepted throughput: 37.20 tokens/s, Drafted throughput: 54.89 tokens/s, Accepted: 372 tokens, Drafted: 549 tokens, Per-position acceptance rate: 0.885, 0.639, 0.508, Avg Draft acceptance rate: 67.8%
270
+ INFO: 127.0.0.1:42391 - "POST /v1/completions HTTP/1.1" 200 OK
271
+ (APIServer pid=32660) INFO 11-18 18:34:30 [loggers.py:320] Engine 000: Avg prompt throughput: 4.4 tokens/s, Avg generation throughput: 43.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
272
+ (APIServer pid=32660) INFO 11-18 18:34:30 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 2.37, Accepted throughput: 25.20 tokens/s, Drafted throughput: 55.19 tokens/s, Accepted: 252 tokens, Drafted: 552 tokens, Per-position acceptance rate: 0.723, 0.424, 0.223, Avg Draft acceptance rate: 45.7%
273
+ INFO: 127.0.0.1:47152 - "POST /v1/completions HTTP/1.1" 200 OK
274
+ INFO: 127.0.0.1:62673 - "POST /v1/completions HTTP/1.1" 200 OK
275
+ (APIServer pid=32660) INFO 11-18 18:34:40 [loggers.py:320] Engine 000: Avg prompt throughput: 13.5 tokens/s, Avg generation throughput: 55.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
276
+ (APIServer pid=32660) INFO 11-18 18:34:40 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.08, Accepted throughput: 37.69 tokens/s, Drafted throughput: 54.29 tokens/s, Accepted: 377 tokens, Drafted: 543 tokens, Per-position acceptance rate: 0.856, 0.685, 0.541, Avg Draft acceptance rate: 69.4%
277
+ INFO: 127.0.0.1:34437 - "POST /v1/completions HTTP/1.1" 200 OK
278
+ (APIServer pid=32660) INFO 11-18 18:34:50 [loggers.py:320] Engine 000: Avg prompt throughput: 8.3 tokens/s, Avg generation throughput: 58.1 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%
279
+ (APIServer pid=32660) INFO 11-18 18:34:50 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.17, Accepted throughput: 39.70 tokens/s, Drafted throughput: 54.89 tokens/s, Accepted: 397 tokens, Drafted: 549 tokens, Per-position acceptance rate: 0.863, 0.732, 0.574, Avg Draft acceptance rate: 72.3%
280
+ INFO: 127.0.0.1:48741 - "POST /v1/completions HTTP/1.1" 200 OK
runs/r2/logs/eval_baseline.log ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
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
+ [18:29:34] repeat 1/1: fresh vLLM process
8
+ [18:29:34] starting vLLM (log: /workspace/emc/work/logs/20261118-182934-eval-r1.log)
9
+ [18:31:32] vLLM ready after 118s
runs/r2/logs/pipeline.log CHANGED
@@ -52,3 +52,4 @@
52
  [2026-11-18 18:28:52] done eval_main (1652s)
53
  [2026-11-18 18:28:54] start report_main
54
  [2026-11-18 18:28:54] done report_main (0s)
 
 
52
  [2026-11-18 18:28:52] done eval_main (1652s)
53
  [2026-11-18 18:28:54] start report_main
54
  [2026-11-18 18:28:54] done report_main (0s)
55
+ [2026-11-18 18:29:25] start eval_baseline
runs/r2/logs/pipeline.out CHANGED
@@ -52,3 +52,4 @@
52
  [2026-11-18 18:28:52] done eval_main (1652s)
53
  [2026-11-18 18:28:54] start report_main
54
  [2026-11-18 18:28:54] done report_main (0s)
 
 
52
  [2026-11-18 18:28:52] done eval_main (1652s)
53
  [2026-11-18 18:28:54] start report_main
54
  [2026-11-18 18:28:54] done report_main (0s)
55
+ [2026-11-18 18:29:25] start eval_baseline
runs/r2/logs/stats.log CHANGED
@@ -36,3 +36,5 @@
36
  [2026-11-18 18:28:52] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 117 GB used | gpu 0.0/95.6 GiB, 0%, 147 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
37
  [2026-11-18 18:28:54] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 117 GB used | gpu 0.0/95.6 GiB, 0%, 99 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
38
  [2026-11-18 18:28:54] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 117 GB used | gpu 0.0/95.6 GiB, 0%, 99 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
 
 
 
36
  [2026-11-18 18:28:52] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 117 GB used | gpu 0.0/95.6 GiB, 0%, 147 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
37
  [2026-11-18 18:28:54] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 117 GB used | gpu 0.0/95.6 GiB, 0%, 99 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
38
  [2026-11-18 18:28:54] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 117 GB used | gpu 0.0/95.6 GiB, 0%, 99 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
39
+ [2026-11-18 18:29:25] ram 0.2/172 GB used (cache 0.0, shmem 0.0) | cgroup 0.2/172 GB | disk 117 GB used | gpu 0.0/95.6 GiB, 0%, 60 W | load 0.0 | largest python 0.1 GB, python 0.1 GB, python 0.0 GB
40
+ [2026-11-18 18:34:57] ram 4.6/172 GB used (cache 0.3, shmem 0.0) | cgroup 4.6/172 GB | disk 117 GB used | gpu 83.1/95.6 GiB, 100%, 292 W | load 0.0 | largest VLLM::EngineCor 4.3 GB, vllm 2.5 GB, python 1.2 GB
runs/r2/logs/sync.log CHANGED
@@ -30,3 +30,4 @@ uploaded logs/results to mudassiruddin21/EMC:runs/r2
30
  uploaded logs/results to mudassiruddin21/EMC:runs/r2
31
  uploaded logs/results to mudassiruddin21/EMC:runs/r2
32
  uploaded logs/results to mudassiruddin21/EMC:runs/r2
 
 
30
  uploaded logs/results to mudassiruddin21/EMC:runs/r2
31
  uploaded logs/results to mudassiruddin21/EMC:runs/r2
32
  uploaded logs/results to mudassiruddin21/EMC:runs/r2
33
+ uploaded logs/results to mudassiruddin21/EMC:runs/r2