Buckets:

Mercity
/

Desicion-Model-Custom

Files

xet

Mercity/Desicion-Model-Custom / run1 /logs /eval_votes.log

Pranav2748

8 days ago

download

raw

18.6 kB

(MainProcess pid=59443) INFO 10-31 09:50:55 [model.py:733] Resolved architecture: Qwen3_5ForCausalLM
(MainProcess pid=59443) INFO 10-31 09:50:55 [model.py:2152] Using max model len 32768
(MainProcess pid=59443) INFO 10-31 09:50:55 [scheduler.py:307] Chunked prefill is enabled with max_num_batched_tokens=2048.
(MainProcess pid=59443) INFO 10-31 09:50:55 [config.py:688] Mamba cache mode is set to 'align' for Qwen3_5ForCausalLM by default when prefix caching is enabled
(MainProcess pid=59443) INFO 10-31 09:50:56 [kernel.py:467] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'], gelu_and_mul_sparse=['triton', 'native'])
(EngineCore pid=59778) INFO 10-31 09:51:02 [core.py:133] Initializing a V1 LLM engine (v0.32.0) with config: model='checkpoints/votes_only_merged', speculative_config=None, tokenizer='checkpoints/votes_only_merged', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=32768, 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, custom_histogram_buckets=None, 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=17, served_model_name=checkpoints/votes_only_merged, enable_prefix_caching=True, 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::qwen4_exp_qsa_prepare_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, 'cudagraph_decoder_replay': True, 'decoder_replay_cudagraph_capture_sizes': [], 'decoder_replay_trim_threshold': 768, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], '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, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256], '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, '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': 256, '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_rocm_segmented_attn_autotune=False, enable_cutedsl_warmup=True, enable_jit_warmup=True, moe_backend='auto', sparse_indexer_topk_backend='auto', linear_backend='auto', linear_backend_per_quant=None)
(EngineCore pid=59778) INFO 10-31 09:51:03 [parallel_state.py:1906] world_size=1 rank=0 local_rank=0 distributed_init_method=file:///tmp/vllm_dist_1c8f0f7a83324ab2b1fef8874eca4bf3 backend=nccl
(EngineCore pid=59778) INFO 10-31 09:51:03 [parallel_state.py:2350] 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
(EngineCore pid=59778) INFO 10-31 09:51:03 [gpu_worker.py:518] Using V2 Model Runner
(EngineCore pid=59778) INFO 10-31 09:51:05 [model_runner.py:392] Loading model from scratch...
(EngineCore pid=59778) INFO 10-31 09:51:05 [qwen_gdn_linear_attn.py:221] Using Triton/FLA GDN prefill kernel (requested=auto, head_k_dim=128).
(EngineCore pid=59778) INFO 10-31 09:51:06 [cuda.py:506] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
(EngineCore pid=59778) INFO 10-31 09:51:06 [flash_attn.py:1117] Using FlashAttention version 2
(EngineCore pid=59778) INFO 10-31 09:51:07 [base_loader.py:82] Initializing model took 1.759581 seconds
(EngineCore pid=59778) INFO 10-31 09:51:07 [weight_utils.py:1108] Filesystem type for checkpoints: XFS. Checkpoint size: 7.83 GiB. Available RAM: 161.98 GiB.
(EngineCore pid=59778) INFO 10-31 09:51:07 [weight_utils.py:1131] Auto-prefetch is disabled because the filesystem (XFS) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.

Loading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]

Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  1.03it/s]

Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  1.03it/s]

(EngineCore pid=59778) INFO 10-31 09:51:08 [default_loader.py:501] Loading weights took 1.00 seconds
(EngineCore pid=59778) INFO 10-31 09:51:09 [model_runner.py:424] Model loading took 7.94 GiB memory and 4.853671 seconds
(EngineCore pid=59778) INFO 10-31 09:51:09 [topk_topp_sampler.py:124] Using FlashInfer for top-p & top-k sampling.
(EngineCore pid=59778) INFO 10-31 09:51:09 [torch_utils.py:295] Reducing Torch threads from 128 to 1 for serving. Set OMP_NUM_THREADS in the external environment to override.
(EngineCore pid=59778) INFO 10-31 09:51:09 [interface.py:1096] Setting attention block size to 528 tokens to ensure that attention page size is >= mamba page size.
(EngineCore pid=59778) INFO 10-31 09:51:09 [interface.py:1120] Padding mamba page size by 0.76% to ensure that mamba page size and attention page size are exactly equal.
(EngineCore pid=59778) INFO 10-31 09:51:09 [utils.py:347] Using LBNHC KV cache layout.
(EngineCore pid=59778) INFO 10-31 09:51:10 [monitor.py:40] [backbone] Starting torch.compile
(EngineCore pid=59778) INFO 10-31 09:51:16 [backends.py:1104] [backbone] Using cache directory: /root/.cache/vllm/torch_compile_cache/3f1d4624f6/rank_0_0/backbone for vLLM's torch.compile
(EngineCore pid=59778) INFO 10-31 09:51:16 [backends.py:1166] [backbone] Dynamo bytecode transform time: 5.43 s
(EngineCore pid=59778) INFO 10-31 09:51:34 [backends.py:398] [backbone] Compiling a graph for compile range (1, 2048) takes 18.10 s
(EngineCore pid=59778) INFO 10-31 09:51:37 [backends.py:927] collected artifacts: 33 entries, 12 artifacts, 20433988 bytes total
(EngineCore pid=59778) INFO 10-31 09:51:37 [decorators.py:743] [backbone] saved AOT compiled function to /root/.cache/vllm/torch_compile_cache/torch_aot_compile/3eb3c78fc507f33c9366ef8e9bbf26e7f9a0836847e49322673b158d219c79f3/rank_0_0/model
(EngineCore pid=59778) INFO 10-31 09:51:37 [monitor.py:62] [backbone] torch.compile took 26.51 s in total
(EngineCore pid=59778) INFO 10-31 09:53:03 [monitor.py:91] [backbone] Initial profiling/warmup run took 86.30 s

Capturing CUDA graphs (PIECEWISE):   0%|          | 0/35 [00:00<?, ?it/s]
Capturing CUDA graphs (PIECEWISE):   3%|▎         | 1/35 [00:01<01:04,  1.91s/it]
Capturing CUDA graphs (PIECEWISE):   6%|▌         | 2/35 [00:02<00:38,  1.18s/it]
Capturing CUDA graphs (PIECEWISE):   9%|▊         | 3/35 [00:02<00:24,  1.33it/s]
Capturing CUDA graphs (PIECEWISE):  11%|█▏        | 4/35 [00:03<00:17,  1.82it/s]
Capturing CUDA graphs (PIECEWISE):  14%|█▍        | 5/35 [00:03<00:12,  2.32it/s]
Capturing CUDA graphs (PIECEWISE):  17%|█▋        | 6/35 [00:03<00:10,  2.78it/s]
Capturing CUDA graphs (PIECEWISE):  20%|██        | 7/35 [00:03<00:08,  3.21it/s]
Capturing CUDA graphs (PIECEWISE):  23%|██▎       | 8/35 [00:03<00:07,  3.55it/s]
Capturing CUDA graphs (PIECEWISE):  26%|██▌       | 9/35 [00:04<00:07,  3.52it/s]
Capturing CUDA graphs (PIECEWISE):  29%|██▊       | 10/35 [00:04<00:06,  3.58it/s]
Capturing CUDA graphs (PIECEWISE):  31%|███▏      | 11/35 [00:04<00:06,  3.66it/s]
Capturing CUDA graphs (PIECEWISE):  34%|███▍      | 12/35 [00:04<00:06,  3.82it/s]
Capturing CUDA graphs (PIECEWISE):  37%|███▋      | 13/35 [00:05<00:05,  3.98it/s]
Capturing CUDA graphs (PIECEWISE):  40%|████      | 14/35 [00:05<00:04,  4.22it/s]
Capturing CUDA graphs (PIECEWISE):  43%|████▎     | 15/35 [00:05<00:04,  4.71it/s]
Capturing CUDA graphs (PIECEWISE):  46%|████▌     | 16/35 [00:05<00:03,  5.36it/s]
Capturing CUDA graphs (PIECEWISE):  49%|████▊     | 17/35 [00:06<00:08,  2.24it/s]
Capturing CUDA graphs (PIECEWISE):  51%|█████▏    | 18/35 [00:06<00:06,  2.61it/s]
Capturing CUDA graphs (PIECEWISE):  60%|██████    | 21/35 [00:07<00:02,  5.11it/s]
Capturing CUDA graphs (PIECEWISE):  69%|██████▊   | 24/35 [00:07<00:01,  7.73it/s]
Capturing CUDA graphs (PIECEWISE):  77%|███████▋  | 27/35 [00:07<00:00, 10.31it/s]
Capturing CUDA graphs (PIECEWISE):  86%|████████▌ | 30/35 [00:07<00:00, 12.71it/s]
Capturing CUDA graphs (PIECEWISE):  94%|█████████▍| 33/35 [00:07<00:00, 14.72it/s]
Capturing CUDA graphs (PIECEWISE): 100%|██████████| 35/35 [00:07<00:00, 12.19it/s]
Capturing CUDA graphs (PIECEWISE): 100%|██████████| 35/35 [00:07<00:00,  4.41it/s]

Capturing CUDA graphs (FULL):   0%|          | 0/2 [00:00<?, ?it/s]
Capturing CUDA graphs (FULL): 100%|██████████| 2/2 [00:00<00:00, 26.09it/s]
(EngineCore pid=59778) INFO 10-31 09:53:15 [model_runner.py:1127] Graph capturing finished in 9 secs, took 0.30 GiB
(EngineCore pid=59778) INFO 10-31 09:53:16 [gpu_worker.py:759] Available KV cache memory: 59.38 GiB
(EngineCore pid=59778) INFO 10-31 09:53:16 [gpu_worker.py:774] CUDA graph memory profiling is enabled (default since v0.21.0). The current --gpu-memory-utilization=0.9000 is equivalent to --gpu-memory-utilization=0.8960 without CUDA graph memory profiling. To maintain the same effective KV cache size as before, increase --gpu-memory-utilization to 0.9040. To disable, set VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0.
(EngineCore pid=59778) INFO 10-31 09:53:16 [kv_cache_utils.py:2327] GPU KV cache size: 1,749,526 tokens, Maximum concurrency for 32,768 tokens per request: 53.39x
(EngineCore pid=59778) INFO 10-31 09:53:16 [kv_cache_utils.py:853] kv cache group sizes (528, 528, 528, 528)
(EngineCore pid=59778) INFO 10-31 09:53:16 [kv_cache_utils.py:854] kv lcm block sizes 528
(EngineCore pid=59778) INFO 10-31 09:53:16 [kernel_warmup.py:176] JIT kernel warmup starting.
(EngineCore pid=59778) INFO 10-31 09:53:16 [kernel_warmup.py:189] JIT kernel warmup finished in 0.00s.
(EngineCore pid=59778) INFO 10-31 09:53:16 [qwen_triton_warmup.py:296] Warming up Qwen GDN Triton kernels for model_type=qwen3_5_text.
(EngineCore pid=59778) INFO 10-31 09:53:17 [mamba_triton_warmup.py:42] Warmed Mamba batch_memcpy_kernel.

Capturing CUDA graphs (PIECEWISE):   0%|          | 0/35 [00:00<?, ?it/s]
Capturing CUDA graphs (PIECEWISE):   3%|▎         | 1/35 [00:00<00:10,  3.19it/s]
Capturing CUDA graphs (PIECEWISE):   6%|▌         | 2/35 [00:00<00:12,  2.69it/s]
Capturing CUDA graphs (PIECEWISE):   9%|▊         | 3/35 [00:01<00:11,  2.68it/s]
Capturing CUDA graphs (PIECEWISE):  11%|█▏        | 4/35 [00:01<00:11,  2.74it/s]
Capturing CUDA graphs (PIECEWISE):  14%|█▍        | 5/35 [00:01<00:11,  2.69it/s]
Capturing CUDA graphs (PIECEWISE):  17%|█▋        | 6/35 [00:04<00:34,  1.18s/it]
Capturing CUDA graphs (PIECEWISE):  20%|██        | 7/35 [00:04<00:24,  1.16it/s]
Capturing CUDA graphs (PIECEWISE):  23%|██▎       | 8/35 [00:05<00:17,  1.52it/s]
Capturing CUDA graphs (PIECEWISE):  26%|██▌       | 9/35 [00:05<00:14,  1.84it/s]
Capturing CUDA graphs (PIECEWISE):  29%|██▊       | 10/35 [00:05<00:11,  2.19it/s]
Capturing CUDA graphs (PIECEWISE):  31%|███▏      | 11/35 [00:05<00:09,  2.56it/s]
Capturing CUDA graphs (PIECEWISE):  34%|███▍      | 12/35 [00:06<00:07,  3.01it/s]
Capturing CUDA graphs (PIECEWISE):  37%|███▋      | 13/35 [00:06<00:06,  3.51it/s]
Capturing CUDA graphs (PIECEWISE):  40%|████      | 14/35 [00:06<00:05,  4.05it/s]
Capturing CUDA graphs (PIECEWISE):  43%|████▎     | 15/35 [00:06<00:04,  4.73it/s]
Capturing CUDA graphs (PIECEWISE):  46%|████▌     | 16/35 [00:06<00:03,  5.44it/s]
Capturing CUDA graphs (PIECEWISE):  54%|█████▍    | 19/35 [00:06<00:01,  9.40it/s]
Capturing CUDA graphs (PIECEWISE):  63%|██████▎   | 22/35 [00:06<00:01, 12.65it/s]
Capturing CUDA graphs (PIECEWISE):  71%|███████▏  | 25/35 [00:06<00:00, 15.26it/s]
Capturing CUDA graphs (PIECEWISE):  80%|████████  | 28/35 [00:07<00:00, 17.23it/s]
Capturing CUDA graphs (PIECEWISE):  89%|████████▊ | 31/35 [00:07<00:00, 18.80it/s]
Capturing CUDA graphs (PIECEWISE):  97%|█████████▋| 34/35 [00:07<00:00, 19.93it/s]
Capturing CUDA graphs (PIECEWISE): 100%|██████████| 35/35 [00:07<00:00,  4.71it/s]

Capturing CUDA graphs (FULL):   0%|          | 0/19 [00:00<?, ?it/s]
Capturing CUDA graphs (FULL):  16%|█▌        | 3/19 [00:00<00:00, 28.47it/s]
Capturing CUDA graphs (FULL):  32%|███▏      | 6/19 [00:00<00:00, 28.95it/s]
Capturing CUDA graphs (FULL):  47%|████▋     | 9/19 [00:00<00:00, 28.95it/s]
Capturing CUDA graphs (FULL):  63%|██████▎   | 12/19 [00:00<00:00, 28.37it/s]
Capturing CUDA graphs (FULL):  79%|███████▉  | 15/19 [00:00<00:00, 28.49it/s]
Capturing CUDA graphs (FULL):  95%|█████████▍| 18/19 [00:00<00:00, 28.21it/s]
Capturing CUDA graphs (FULL): 100%|██████████| 19/19 [00:00<00:00, 28.40it/s]
(EngineCore pid=59778) INFO 10-31 09:53:45 [model_runner.py:1127] Graph capturing finished in 9 secs, took 0.16 GiB
(EngineCore pid=59778) INFO 10-31 09:53:45 [gpu_worker.py:971] CUDA graph pool memory: 0.16 GiB (actual), 0.32 GiB (estimated), difference: 0.16 GiB (103.8%).
(EngineCore pid=59778) INFO 10-31 09:53:45 [gpu_worker.py:1034] Free memory on device (78.84/79.25 GiB) on startup. Desired GPU memory utilization is (0.9, 71.33 GiB). Actual usage is 10.39 GiB for consumed memory (weights + non-torch), 1.55 GiB for peak activation, and 0.16 GiB for CUDAGraph memory. Replace gpu_memory_utilization config with `--kv-cache-memory=63430039962` (59.07 GiB) to fit into requested memory, or `--kv-cache-memory=71496105984` (66.59 GiB) to fully utilize gpu memory. Current kv cache memory in use is 59.38 GiB.
(EngineCore pid=59778) INFO 10-31 09:53:46 [jit_monitor.py:84] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
(EngineCore pid=59778) INFO 10-31 09:53:46 [core.py:395] init engine (profile, create kv cache, warmup model) took 156.74 s (compilation: 26.51 s)
(EngineCore pid=59778) INFO 10-31 09:53:46 [kernel.py:467] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'], gelu_and_mul_sparse=['triton', 'native'])
(MainProcess pid=59443) WARNING 10-31 09:53:47 [input_processor.py:408] Passing raw prompts to InputProcessor is deprecated and will be removed in the future. You should instead pass the outputs of Renderer.render_cmpl() or Renderer.render_chat().
(EngineCore pid=59778) WARNING 10-31 09:53:47 [jit_monitor.py:140] Triton kernel JIT compilation during inference: _topk_log_softmax_kernel. This causes a latency spike; consider extending warmup to cover this shape/config.
readout votes_only: 5957 items in 114s
(MainProcess pid=59443) INFO 10-31 09:55:40 [core_client.py:770] [shutdown] MPClient: start timeout=default
(MainProcess pid=59443) INFO 10-31 09:55:40 [core_client.py:772] [shutdown] MPClient: stopping engine manager
(MainProcess pid=59443) INFO 10-31 09:55:40 [utils.py:628] [shutdown] Process manager: send sigterm to process EngineCore
(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1454] [shutdown] EngineCore: trigger received signal=SIGTERM
(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1635] [shutdown] EngineCore: start mode=abort timeout=0s
(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1666] [shutdown] EngineCore: request processing complete; starting resource teardown
(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1499] [shutdown] EngineCore: exiting busy loop
(MainProcess pid=59443) INFO 10-31 09:55:43 [core_client.py:774] [shutdown] MPClient: engine manager stopped
(MainProcess pid=59443) INFO 10-31 09:55:43 [core_client.py:775] [shutdown] MPClient: cleaning up background resources
(MainProcess pid=59443) INFO 10-31 09:55:43 [core_client.py:777] [shutdown] MPClient: complete

Xet Storage Details

Size:: 18.6 kB
Xet hash:: 015ec9075c0a74d0c341a938d3a1693a6cb5819be1ac392b077f06f63408867b

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.

	(MainProcess pid=59443) INFO 10-31 09:50:55 [model.py:733] Resolved architecture: Qwen3_5ForCausalLM
	(MainProcess pid=59443) INFO 10-31 09:50:55 [model.py:2152] Using max model len 32768
	(MainProcess pid=59443) INFO 10-31 09:50:55 [scheduler.py:307] Chunked prefill is enabled with max_num_batched_tokens=2048.
	(MainProcess pid=59443) INFO 10-31 09:50:55 [config.py:688] Mamba cache mode is set to 'align' for Qwen3_5ForCausalLM by default when prefix caching is enabled
	(MainProcess pid=59443) INFO 10-31 09:50:56 [kernel.py:467] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'], gelu_and_mul_sparse=['triton', 'native'])
	(EngineCore pid=59778) INFO 10-31 09:51:02 [core.py:133] Initializing a V1 LLM engine (v0.32.0) with config: model='checkpoints/votes_only_merged', speculative_config=None, tokenizer='checkpoints/votes_only_merged', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=32768, 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, custom_histogram_buckets=None, 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=17, served_model_name=checkpoints/votes_only_merged, enable_prefix_caching=True, 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::qwen4_exp_qsa_prepare_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, 'cudagraph_decoder_replay': True, 'decoder_replay_cudagraph_capture_sizes': [], 'decoder_replay_trim_threshold': 768, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], '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, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256], '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, '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': 256, '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_rocm_segmented_attn_autotune=False, enable_cutedsl_warmup=True, enable_jit_warmup=True, moe_backend='auto', sparse_indexer_topk_backend='auto', linear_backend='auto', linear_backend_per_quant=None)
	(EngineCore pid=59778) INFO 10-31 09:51:03 [parallel_state.py:1906] world_size=1 rank=0 local_rank=0 distributed_init_method=file:///tmp/vllm_dist_1c8f0f7a83324ab2b1fef8874eca4bf3 backend=nccl
	(EngineCore pid=59778) INFO 10-31 09:51:03 [parallel_state.py:2350] 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
	(EngineCore pid=59778) INFO 10-31 09:51:03 [gpu_worker.py:518] Using V2 Model Runner
	(EngineCore pid=59778) INFO 10-31 09:51:05 [model_runner.py:392] Loading model from scratch...
	(EngineCore pid=59778) INFO 10-31 09:51:05 [qwen_gdn_linear_attn.py:221] Using Triton/FLA GDN prefill kernel (requested=auto, head_k_dim=128).
	(EngineCore pid=59778) INFO 10-31 09:51:06 [cuda.py:506] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
	(EngineCore pid=59778) INFO 10-31 09:51:06 [flash_attn.py:1117] Using FlashAttention version 2
	(EngineCore pid=59778) INFO 10-31 09:51:07 [base_loader.py:82] Initializing model took 1.759581 seconds
	(EngineCore pid=59778) INFO 10-31 09:51:07 [weight_utils.py:1108] Filesystem type for checkpoints: XFS. Checkpoint size: 7.83 GiB. Available RAM: 161.98 GiB.
	(EngineCore pid=59778) INFO 10-31 09:51:07 [weight_utils.py:1131] Auto-prefetch is disabled because the filesystem (XFS) is not a recognized network FS (NFS/Lustre). If you want to force prefetching, start vLLM with --safetensors-load-strategy=prefetch.
	Loading safetensors checkpoint shards: 0% Completed \| 0/1 [00:00<?, ?it/s]
	Loading safetensors checkpoint shards: 100% Completed \| 1/1 [00:00<00:00, 1.03it/s]
	Loading safetensors checkpoint shards: 100% Completed \| 1/1 [00:00<00:00, 1.03it/s]

	(EngineCore pid=59778) INFO 10-31 09:51:08 [default_loader.py:501] Loading weights took 1.00 seconds
	(EngineCore pid=59778) INFO 10-31 09:51:09 [model_runner.py:424] Model loading took 7.94 GiB memory and 4.853671 seconds
	(EngineCore pid=59778) INFO 10-31 09:51:09 [topk_topp_sampler.py:124] Using FlashInfer for top-p & top-k sampling.
	(EngineCore pid=59778) INFO 10-31 09:51:09 [torch_utils.py:295] Reducing Torch threads from 128 to 1 for serving. Set OMP_NUM_THREADS in the external environment to override.
	(EngineCore pid=59778) INFO 10-31 09:51:09 [interface.py:1096] Setting attention block size to 528 tokens to ensure that attention page size is >= mamba page size.
	(EngineCore pid=59778) INFO 10-31 09:51:09 [interface.py:1120] Padding mamba page size by 0.76% to ensure that mamba page size and attention page size are exactly equal.
	(EngineCore pid=59778) INFO 10-31 09:51:09 [utils.py:347] Using LBNHC KV cache layout.
	(EngineCore pid=59778) INFO 10-31 09:51:10 [monitor.py:40] [backbone] Starting torch.compile
	(EngineCore pid=59778) INFO 10-31 09:51:16 [backends.py:1104] [backbone] Using cache directory: /root/.cache/vllm/torch_compile_cache/3f1d4624f6/rank_0_0/backbone for vLLM's torch.compile
	(EngineCore pid=59778) INFO 10-31 09:51:16 [backends.py:1166] [backbone] Dynamo bytecode transform time: 5.43 s
	(EngineCore pid=59778) INFO 10-31 09:51:34 [backends.py:398] [backbone] Compiling a graph for compile range (1, 2048) takes 18.10 s
	(EngineCore pid=59778) INFO 10-31 09:51:37 [backends.py:927] collected artifacts: 33 entries, 12 artifacts, 20433988 bytes total
	(EngineCore pid=59778) INFO 10-31 09:51:37 [decorators.py:743] [backbone] saved AOT compiled function to /root/.cache/vllm/torch_compile_cache/torch_aot_compile/3eb3c78fc507f33c9366ef8e9bbf26e7f9a0836847e49322673b158d219c79f3/rank_0_0/model
	(EngineCore pid=59778) INFO 10-31 09:51:37 [monitor.py:62] [backbone] torch.compile took 26.51 s in total
	(EngineCore pid=59778) INFO 10-31 09:53:03 [monitor.py:91] [backbone] Initial profiling/warmup run took 86.30 s
	Capturing CUDA graphs (PIECEWISE): 0%\| \| 0/35 [00:00<?, ?it/s] Capturing CUDA graphs (PIECEWISE): 3%\|▎ \| 1/35 [00:01<01:04, 1.91s/it] Capturing CUDA graphs (PIECEWISE): 6%\|▌ \| 2/35 [00:02<00:38, 1.18s/it] Capturing CUDA graphs (PIECEWISE): 9%\|▊ \| 3/35 [00:02<00:24, 1.33it/s] Capturing CUDA graphs (PIECEWISE): 11%\|█▏ \| 4/35 [00:03<00:17, 1.82it/s] Capturing CUDA graphs (PIECEWISE): 14%\|█▍ \| 5/35 [00:03<00:12, 2.32it/s] Capturing CUDA graphs (PIECEWISE): 17%\|█▋ \| 6/35 [00:03<00:10, 2.78it/s] Capturing CUDA graphs (PIECEWISE): 20%\|██ \| 7/35 [00:03<00:08, 3.21it/s] Capturing CUDA graphs (PIECEWISE): 23%\|██▎ \| 8/35 [00:03<00:07, 3.55it/s] Capturing CUDA graphs (PIECEWISE): 26%\|██▌ \| 9/35 [00:04<00:07, 3.52it/s] Capturing CUDA graphs (PIECEWISE): 29%\|██▊ \| 10/35 [00:04<00:06, 3.58it/s] Capturing CUDA graphs (PIECEWISE): 31%\|███▏ \| 11/35 [00:04<00:06, 3.66it/s] Capturing CUDA graphs (PIECEWISE): 34%\|███▍ \| 12/35 [00:04<00:06, 3.82it/s] Capturing CUDA graphs (PIECEWISE): 37%\|███▋ \| 13/35 [00:05<00:05, 3.98it/s] Capturing CUDA graphs (PIECEWISE): 40%\|████ \| 14/35 [00:05<00:04, 4.22it/s] Capturing CUDA graphs (PIECEWISE): 43%\|████▎ \| 15/35 [00:05<00:04, 4.71it/s] Capturing CUDA graphs (PIECEWISE): 46%\|████▌ \| 16/35 [00:05<00:03, 5.36it/s] Capturing CUDA graphs (PIECEWISE): 49%\|████▊ \| 17/35 [00:06<00:08, 2.24it/s] Capturing CUDA graphs (PIECEWISE): 51%\|█████▏ \| 18/35 [00:06<00:06, 2.61it/s] Capturing CUDA graphs (PIECEWISE): 60%\|██████ \| 21/35 [00:07<00:02, 5.11it/s] Capturing CUDA graphs (PIECEWISE): 69%\|██████▊ \| 24/35 [00:07<00:01, 7.73it/s] Capturing CUDA graphs (PIECEWISE): 77%\|███████▋ \| 27/35 [00:07<00:00, 10.31it/s] Capturing CUDA graphs (PIECEWISE): 86%\|████████▌ \| 30/35 [00:07<00:00, 12.71it/s] Capturing CUDA graphs (PIECEWISE): 94%\|█████████▍\| 33/35 [00:07<00:00, 14.72it/s] Capturing CUDA graphs (PIECEWISE): 100%\|██████████\| 35/35 [00:07<00:00, 12.19it/s] Capturing CUDA graphs (PIECEWISE): 100%\|██████████\| 35/35 [00:07<00:00, 4.41it/s]
	Capturing CUDA graphs (FULL): 0%\| \| 0/2 [00:00<?, ?it/s] Capturing CUDA graphs (FULL): 100%\|██████████\| 2/2 [00:00<00:00, 26.09it/s]
	(EngineCore pid=59778) INFO 10-31 09:53:15 [model_runner.py:1127] Graph capturing finished in 9 secs, took 0.30 GiB
	(EngineCore pid=59778) INFO 10-31 09:53:16 [gpu_worker.py:759] Available KV cache memory: 59.38 GiB
	(EngineCore pid=59778) INFO 10-31 09:53:16 [gpu_worker.py:774] CUDA graph memory profiling is enabled (default since v0.21.0). The current --gpu-memory-utilization=0.9000 is equivalent to --gpu-memory-utilization=0.8960 without CUDA graph memory profiling. To maintain the same effective KV cache size as before, increase --gpu-memory-utilization to 0.9040. To disable, set VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0.
	(EngineCore pid=59778) INFO 10-31 09:53:16 [kv_cache_utils.py:2327] GPU KV cache size: 1,749,526 tokens, Maximum concurrency for 32,768 tokens per request: 53.39x
	(EngineCore pid=59778) INFO 10-31 09:53:16 [kv_cache_utils.py:853] kv cache group sizes (528, 528, 528, 528)
	(EngineCore pid=59778) INFO 10-31 09:53:16 [kv_cache_utils.py:854] kv lcm block sizes 528
	(EngineCore pid=59778) INFO 10-31 09:53:16 [kernel_warmup.py:176] JIT kernel warmup starting.
	(EngineCore pid=59778) INFO 10-31 09:53:16 [kernel_warmup.py:189] JIT kernel warmup finished in 0.00s.
	(EngineCore pid=59778) INFO 10-31 09:53:16 [qwen_triton_warmup.py:296] Warming up Qwen GDN Triton kernels for model_type=qwen3_5_text.
	(EngineCore pid=59778) INFO 10-31 09:53:17 [mamba_triton_warmup.py:42] Warmed Mamba batch_memcpy_kernel.
	Capturing CUDA graphs (PIECEWISE): 0%\| \| 0/35 [00:00<?, ?it/s] Capturing CUDA graphs (PIECEWISE): 3%\|▎ \| 1/35 [00:00<00:10, 3.19it/s] Capturing CUDA graphs (PIECEWISE): 6%\|▌ \| 2/35 [00:00<00:12, 2.69it/s] Capturing CUDA graphs (PIECEWISE): 9%\|▊ \| 3/35 [00:01<00:11, 2.68it/s] Capturing CUDA graphs (PIECEWISE): 11%\|█▏ \| 4/35 [00:01<00:11, 2.74it/s] Capturing CUDA graphs (PIECEWISE): 14%\|█▍ \| 5/35 [00:01<00:11, 2.69it/s] Capturing CUDA graphs (PIECEWISE): 17%\|█▋ \| 6/35 [00:04<00:34, 1.18s/it] Capturing CUDA graphs (PIECEWISE): 20%\|██ \| 7/35 [00:04<00:24, 1.16it/s] Capturing CUDA graphs (PIECEWISE): 23%\|██▎ \| 8/35 [00:05<00:17, 1.52it/s] Capturing CUDA graphs (PIECEWISE): 26%\|██▌ \| 9/35 [00:05<00:14, 1.84it/s] Capturing CUDA graphs (PIECEWISE): 29%\|██▊ \| 10/35 [00:05<00:11, 2.19it/s] Capturing CUDA graphs (PIECEWISE): 31%\|███▏ \| 11/35 [00:05<00:09, 2.56it/s] Capturing CUDA graphs (PIECEWISE): 34%\|███▍ \| 12/35 [00:06<00:07, 3.01it/s] Capturing CUDA graphs (PIECEWISE): 37%\|███▋ \| 13/35 [00:06<00:06, 3.51it/s] Capturing CUDA graphs (PIECEWISE): 40%\|████ \| 14/35 [00:06<00:05, 4.05it/s] Capturing CUDA graphs (PIECEWISE): 43%\|████▎ \| 15/35 [00:06<00:04, 4.73it/s] Capturing CUDA graphs (PIECEWISE): 46%\|████▌ \| 16/35 [00:06<00:03, 5.44it/s] Capturing CUDA graphs (PIECEWISE): 54%\|█████▍ \| 19/35 [00:06<00:01, 9.40it/s] Capturing CUDA graphs (PIECEWISE): 63%\|██████▎ \| 22/35 [00:06<00:01, 12.65it/s] Capturing CUDA graphs (PIECEWISE): 71%\|███████▏ \| 25/35 [00:06<00:00, 15.26it/s] Capturing CUDA graphs (PIECEWISE): 80%\|████████ \| 28/35 [00:07<00:00, 17.23it/s] Capturing CUDA graphs (PIECEWISE): 89%\|████████▊ \| 31/35 [00:07<00:00, 18.80it/s] Capturing CUDA graphs (PIECEWISE): 97%\|█████████▋\| 34/35 [00:07<00:00, 19.93it/s] Capturing CUDA graphs (PIECEWISE): 100%\|██████████\| 35/35 [00:07<00:00, 4.71it/s]
	Capturing CUDA graphs (FULL): 0%\| \| 0/19 [00:00<?, ?it/s] Capturing CUDA graphs (FULL): 16%\|█▌ \| 3/19 [00:00<00:00, 28.47it/s] Capturing CUDA graphs (FULL): 32%\|███▏ \| 6/19 [00:00<00:00, 28.95it/s] Capturing CUDA graphs (FULL): 47%\|████▋ \| 9/19 [00:00<00:00, 28.95it/s] Capturing CUDA graphs (FULL): 63%\|██████▎ \| 12/19 [00:00<00:00, 28.37it/s] Capturing CUDA graphs (FULL): 79%\|███████▉ \| 15/19 [00:00<00:00, 28.49it/s] Capturing CUDA graphs (FULL): 95%\|█████████▍\| 18/19 [00:00<00:00, 28.21it/s] Capturing CUDA graphs (FULL): 100%\|██████████\| 19/19 [00:00<00:00, 28.40it/s]
	(EngineCore pid=59778) INFO 10-31 09:53:45 [model_runner.py:1127] Graph capturing finished in 9 secs, took 0.16 GiB
	(EngineCore pid=59778) INFO 10-31 09:53:45 [gpu_worker.py:971] CUDA graph pool memory: 0.16 GiB (actual), 0.32 GiB (estimated), difference: 0.16 GiB (103.8%).
	(EngineCore pid=59778) INFO 10-31 09:53:45 [gpu_worker.py:1034] Free memory on device (78.84/79.25 GiB) on startup. Desired GPU memory utilization is (0.9, 71.33 GiB). Actual usage is 10.39 GiB for consumed memory (weights + non-torch), 1.55 GiB for peak activation, and 0.16 GiB for CUDAGraph memory. Replace gpu_memory_utilization config with `--kv-cache-memory=63430039962` (59.07 GiB) to fit into requested memory, or `--kv-cache-memory=71496105984` (66.59 GiB) to fully utilize gpu memory. Current kv cache memory in use is 59.38 GiB.
	(EngineCore pid=59778) INFO 10-31 09:53:46 [jit_monitor.py:84] Kernel JIT monitor activated; monitored JIT compilations during inference will use mode=warn.
	(EngineCore pid=59778) INFO 10-31 09:53:46 [core.py:395] init engine (profile, create kv cache, warmup model) took 156.74 s (compilation: 26.51 s)
	(EngineCore pid=59778) INFO 10-31 09:53:46 [kernel.py:467] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'], gelu_and_mul_sparse=['triton', 'native'])
	(MainProcess pid=59443) WARNING 10-31 09:53:47 [input_processor.py:408] Passing raw prompts to InputProcessor is deprecated and will be removed in the future. You should instead pass the outputs of Renderer.render_cmpl() or Renderer.render_chat().
	(EngineCore pid=59778) WARNING 10-31 09:53:47 [jit_monitor.py:140] Triton kernel JIT compilation during inference: _topk_log_softmax_kernel. This causes a latency spike; consider extending warmup to cover this shape/config.
	readout votes_only: 5957 items in 114s
	(MainProcess pid=59443) INFO 10-31 09:55:40 [core_client.py:770] [shutdown] MPClient: start timeout=default
	(MainProcess pid=59443) INFO 10-31 09:55:40 [core_client.py:772] [shutdown] MPClient: stopping engine manager
	(MainProcess pid=59443) INFO 10-31 09:55:40 [utils.py:628] [shutdown] Process manager: send sigterm to process EngineCore
	(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1454] [shutdown] EngineCore: trigger received signal=SIGTERM
	(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1635] [shutdown] EngineCore: start mode=abort timeout=0s
	(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1666] [shutdown] EngineCore: request processing complete; starting resource teardown
	(EngineCore pid=59778) INFO 10-31 09:55:40 [core.py:1499] [shutdown] EngineCore: exiting busy loop
	(MainProcess pid=59443) INFO 10-31 09:55:43 [core_client.py:774] [shutdown] MPClient: engine manager stopped
	(MainProcess pid=59443) INFO 10-31 09:55:43 [core_client.py:775] [shutdown] MPClient: cleaning up background resources
	(MainProcess pid=59443) INFO 10-31 09:55:43 [core_client.py:777] [shutdown] MPClient: complete