Skip to content

Instantly share code, notes, and snippets.

@mdierolf
Created September 22, 2025 17:56
Show Gist options
  • Select an option

  • Save mdierolf/257d410401d1058657ae5a7728a9ba29 to your computer and use it in GitHub Desktop.

Select an option

Save mdierolf/257d410401d1058657ae5a7728a9ba29 to your computer and use it in GitHub Desktop.
INFO 09-22 10:52:06 [__init__.py:216] Automatically detected platform cuda.
(APIServer pid=1200988) INFO 09-22 10:52:07 [api_server.py:1801] vLLM API server version 0.10.2rc3.dev236+g38db529f6
(APIServer pid=1200988) INFO 09-22 10:52:07 [utils.py:328] non-default args: {'model_tag': 'Qwen/Qwen3-Next-80B-A3B-Instruct-FP8', 'port': 11434, 'enable_auto_tool_choice': True, 'tool_call_parser': 'hermes', 'model': 'Qwen/Qwen3-Next-80B-A3B-Instruct-FP8', 'trust_remote_code': True, 'max_model_len': 262144, 'gpu_memory_utilization': 0.92}
(APIServer pid=1200988) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored.
(APIServer pid=1200988) INFO 09-22 10:52:27 [__init__.py:710] Resolved architecture: Qwen3NextForCausalLM
(APIServer pid=1200988) `torch_dtype` is deprecated! Use `dtype` instead!
(APIServer pid=1200988) INFO 09-22 10:52:27 [__init__.py:1769] Using max model len 262144
(APIServer pid=1200988) INFO 09-22 10:52:31 [scheduler.py:222] Chunked prefill is enabled with max_num_batched_tokens=8192.
(APIServer pid=1200988) INFO 09-22 10:52:31 [config.py:310] Hybrid or mamba-based model detected: disabling prefix caching since it is not yet supported.
(APIServer pid=1200988) INFO 09-22 10:52:31 [config.py:321] Hybrid or mamba-based model detected: setting cudagraph mode to FULL_AND_PIECEWISE in order to optimize performance.
(APIServer pid=1200988) INFO 09-22 10:52:31 [config.py:390] Setting attention block size to 544 tokens to ensure that attention page size is >= mamba page size.
(APIServer pid=1200988) INFO 09-22 10:52:31 [config.py:411] Padding mamba page size by 1.49% to ensure that mamba page size and attention page size are exactly equal.
INFO 09-22 10:52:37 [__init__.py:216] Automatically detected platform cuda.
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:39 [core.py:648] Waiting for init message from front-end.
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:39 [core.py:75] Initializing a V1 LLM engine (v0.10.2rc3.dev236+g38db529f6) with config: model='Qwen/Qwen3-Next-80B-A3B-Instruct-FP8', speculative_config=None, tokenizer='Qwen/Qwen3-Next-80B-A3B-Instruct-FP8', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=262144, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=fp8, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Qwen/Qwen3-Next-80B-A3B-Instruct-FP8, enable_prefix_caching=False, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2","vllm.mamba_mixer","vllm.short_conv","vllm.linear_attention","vllm.plamo2_mamba_mixer","vllm.gdn_attention"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"cudagraph_mode":[2,1],"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"pass_config":{},"max_capture_size":512,"local_cache_dir":null}
(EngineCore_DP0 pid=1201210) W0922 10:52:39.439000 1201210 torch/utils/cpp_extension.py:2425] TORCH_CUDA_ARCH_LIST is not set, all archs for visible cards are included for compilation.
(EngineCore_DP0 pid=1201210) W0922 10:52:39.439000 1201210 torch/utils/cpp_extension.py:2425] If this is not desired, please set os.environ['TORCH_CUDA_ARCH_LIST'] to specific architectures.
[W922 10:52:44.453574974 ProcessGroupNCCL.cpp:981] Warning: TORCH_NCCL_AVOID_RECORD_STREAMS is the default now, this environment variable is thus deprecated. (function operator())
[rank0]:[W922 10:52:44.846799434 ProcessGroupGloo.cpp:514] Warning: Unable to resolve hostname to a (local) address. Using the loopback address as fallback. Manually set the network interface to bind to with GLOO_SOCKET_IFNAME. (function operator())
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:53 [parallel_state.py:1206] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:53 [topk_topp_sampler.py:58] Using FlashInfer for top-p & top-k sampling.
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:53 [gpu_model_runner.py:2434] Starting to load model Qwen/Qwen3-Next-80B-A3B-Instruct-FP8...
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:53 [gpu_model_runner.py:2466] Loading model from scratch...
(EngineCore_DP0 pid=1201210) `torch_dtype` is deprecated! Use `dtype` instead!
(EngineCore_DP0 pid=1201210) WARNING 09-22 10:52:53 [fp8.py:455] Failed to import DeepGemm kernels.
(EngineCore_DP0 pid=1201210) WARNING 09-22 10:52:53 [fp8.py:478] CutlassBlockScaledGroupedGemm not supported on the current platform.
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:53 [cuda.py:368] Using Flash Attention backend on V1 engine.
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:56 [weight_utils.py:348] Using model weights format ['*.safetensors']
(EngineCore_DP0 pid=1201210) INFO 09-22 10:52:58 [weight_utils.py:369] Time spent downloading weights for Qwen/Qwen3-Next-80B-A3B-Instruct-FP8: 2.068584 seconds
Loading safetensors checkpoint shards: 0% Completed | 0/8 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 12% Completed | 1/8 [00:02<00:19, 2.77s/it]
Loading safetensors checkpoint shards: 25% Completed | 2/8 [00:07<00:23, 3.93s/it]
Loading safetensors checkpoint shards: 38% Completed | 3/8 [00:12<00:22, 4.41s/it]
Loading safetensors checkpoint shards: 50% Completed | 4/8 [00:16<00:17, 4.38s/it]
Loading safetensors checkpoint shards: 62% Completed | 5/8 [00:19<00:11, 3.92s/it]
Loading safetensors checkpoint shards: 75% Completed | 6/8 [00:23<00:07, 3.65s/it]
Loading safetensors checkpoint shards: 88% Completed | 7/8 [00:27<00:04, 4.03s/it]
Loading safetensors checkpoint shards: 100% Completed | 8/8 [00:32<00:00, 4.29s/it]
Loading safetensors checkpoint shards: 100% Completed | 8/8 [00:32<00:00, 4.09s/it]
(EngineCore_DP0 pid=1201210)
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:36 [default_loader.py:268] Loading weights took 33.04 seconds
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:37 [gpu_model_runner.py:2488] Model loading took 74.8852 GiB and 43.043391 seconds
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:40 [backends.py:539] Using cache directory: /mnt/media/llm-cache/vllm/torch_compile_cache/5150e1ef30/rank_0_0/backbone for vLLM's torch.compile
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:40 [backends.py:550] Dynamo bytecode transform time: 3.04 s
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:40 [backends.py:194] Cache the graph for dynamic shape for later use
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:42 [backends.py:215] Compiling a graph for dynamic shape takes 1.83 s
(EngineCore_DP0 pid=1201210) WARNING 09-22 10:53:43 [fused_moe.py:728] Using default MoE config. Performance might be sub-optimal! Config file not found at ['/home/mdierolf/gitprojects/vllm/vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Max-Q_Workstation_Edition,dtype=fp8_w8a8,block_shape=[128,128].json']
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:43 [monitor.py:34] torch.compile takes 4.86 s in total
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:44 [gpu_worker.py:299] Available KV cache memory: 10.33 GiB
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:44 [kv_cache_utils.py:1034] GPU KV cache size: 112,608 tokens
(EngineCore_DP0 pid=1201210) INFO 09-22 10:53:44 [kv_cache_utils.py:1038] Maximum concurrency for 262,144 tokens per request: 1.71x
(EngineCore_DP0 pid=1201210) 2025-09-22 10:53:44,836 - INFO - autotuner.py:256 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
(EngineCore_DP0 pid=1201210) 2025-09-22 10:53:44,999 - INFO - autotuner.py:262 - flashinfer.jit: [Autotuner]: Autotuning process ends
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|███████████████████████████████████████████████████████████████████████████| 67/67 [00:03<00:00, 16.80it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 67/67 [00:38<00:00, 1.74it/s]
(EngineCore_DP0 pid=1201210) INFO 09-22 10:54:28 [gpu_model_runner.py:3280] Graph capturing finished in 43 secs, took 0.12 GiB
(EngineCore_DP0 pid=1201210) INFO 09-22 10:54:28 [gpu_worker.py:392] Free memory on device (94.32/94.97 GiB) on startup. Desired GPU memory utilization is (0.92, 87.37 GiB). Actual usage is 74.89 GiB for weight, 2.08 GiB for peak activation, 0.08 GiB for non-torch memory, and 0.12 GiB for CUDAGraph memory. Replace gpu_memory_utilization config with `--kv-cache-memory=10809381580` to fit into requested memory, or `--kv-cache-memory=18271979008` to fully utilize gpu memory. Current kv cache memory in use is 11090399948 bytes.
(EngineCore_DP0 pid=1201210) INFO 09-22 10:54:28 [core.py:214] init engine (profile, create kv cache, warmup model) took 50.83 seconds
(APIServer pid=1200988) INFO 09-22 10:54:29 [loggers.py:142] Engine 000: vllm cache_config_info with initialization after num_gpu_blocks is: 829
(APIServer pid=1200988) INFO 09-22 10:54:29 [api_server.py:1597] Supported_tasks: ['generate']
(APIServer pid=1200988) WARNING 09-22 10:54:29 [__init__.py:1648] Default sampling parameters have been overridden by the model's Hugging Face generation config recommended from the model creator. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
(APIServer pid=1200988) INFO 09-22 10:54:29 [serving_responses.py:135] Using default chat sampling params from model: {'temperature': 0.7, 'top_k': 20, 'top_p': 0.8}
(APIServer pid=1200988) INFO 09-22 10:54:29 [serving_responses.py:164] "auto" tool choice has been enabled please note that while the parallel_tool_calls client option is preset for compatibility reasons, it will be ignored.
(APIServer pid=1200988) INFO 09-22 10:54:29 [serving_chat.py:97] "auto" tool choice has been enabled please note that while the parallel_tool_calls client option is preset for compatibility reasons, it will be ignored.
(APIServer pid=1200988) INFO 09-22 10:54:29 [serving_chat.py:137] Using default chat sampling params from model: {'temperature': 0.7, 'top_k': 20, 'top_p': 0.8}
(APIServer pid=1200988) INFO 09-22 10:54:30 [serving_completion.py:76] Using default completion sampling params from model: {'temperature': 0.7, 'top_k': 20, 'top_p': 0.8}
(APIServer pid=1200988) INFO 09-22 10:54:30 [api_server.py:1876] Starting vLLM API server 0 on http://0.0.0.0:11434
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment