Skip to content

[bug] CUDA error: the provided PTX was compiled with an unsupported toolchain. #16

Description

@comcj

What happened

(Describe the failure in 1–3 sentences.)

Hardware + OS

  • GPU(s), model and driver version (nvidia-smi -q | head -25)
    Wed May 13 21:12:35 2026
    +-----------------------------------------------------------------------------------------+
    | NVIDIA-SMI 591.86 Driver Version: 591.86 CUDA Version: 13.1 |
    +-----------------------------------------+------------------------+----------------------+
    | GPU Name Driver-Model | Bus-Id Disp.A | Volatile Uncorr. ECC |
    | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
    | | | MIG M. |
    |=========================================+========================+======================|
    | 0 NVIDIA GeForce RTX 5090 WDDM | 00000000:02:00.0 On | N/A |
    | 0% 49C P8 27W / 600W | 2502MiB / 32607MiB | 1% Default |
    | | | N/A |
    +-----------------------------------------+------------------------+----------------------+

  • Windows build (winver, paste the build number)
    版本 Windows 11 专业工作站版
    版本号 25H2
    安装日期 ‎2025/‎11/‎1
    操作系统版本 26200.8117
    体验 Windows 功能体验包 1000.26100.297.0

  • CPU + RAM (relevant only for very large prefills)
    处理器 Intel(R) Core(TM) Ultra 7 265K (3.90 GHz)
    机带 RAM 64.0 GB (63.4 GB 可用)

  • Single GPU? Is the display attached to it?

Install path

H:\qwen3.6-windows-serve

  • Portable launcher zip
  • Wheel-only install into my own venv
  • Built from source

vLLM version (python -c "import vllm; print(vllm.__version__)"):
H:\qwen3.6-windows-server\python>python -c "import vllm; print(vllm.version)
0.20.0

Snapshot you tried

H:\qwen3.6-windows-server\snapshots\start_5090.bat
Py
H:\qwen3.6-windows-server\snapshots\start_5090.py
start_speed / start_127k / start_pp2_160k / other (specify):

Reproducer

(The exact command or click sequence that triggered the failure.)

Logs

Paste the boot section of logs\vllm_server.<port>.log (first ~50 lines)
plus the last ~50 lines around the failure. Don't worry about size, we'd
rather have too much than too little.

[warn] vcvars64.bat not found at C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat - FlashInfer JIT may fail.
[info] MSVC env not available; setting VLLM_USE_FLASHINFER_SAMPLER=0 to skip flashinfer sampler JIT (would crash without cl.exe). Native PyTorch sampler is used instead, slightly slower but reliable.
============================================================
vLLM serve: qwen3.6-27b-autoround  (Blackwell snapshot)
  Model   : E:\ModelScope\vllm\Qwen3.6-27B-int4-AutoRound
  Ctx     : 240000  |  TP: 1  |  PP: 1
  KV dtype: fp8_e4m3  |  MTP: True (n=6)
  Listen  : http://0.0.0.0:5001
============================================================
H:\qwen3.6-windows-server\python\python.exe -c from vllm.entrypoints.cli.main import main; main() serve E:\ModelScope\vllm\Qwen3.6-27B-int4-AutoRound --served-model-name=qwen3.6-27b-autoround --quantization=autoround --max-model-len=240000 --max-num-seqs=1 --max-num-batched-tokens=4128 --block-size=32 --enable-prefix-caching --enable-chunked-prefill --enable-auto-tool-choice --tool-call-parser=qwen3_coder --reasoning-parser=qwen3 --chat-template=H:\qwen3.6-windows-server\templates\qwen3.5-enhanced.jinja --default-chat-template-kwargs={"preserve_thinking": false} --kv-cache-dtype=fp8_e4m3 --tensor-parallel-size=1 --pipeline-parallel-size=1 --gpu-memory-utilization=0.93 --trust-remote-code --attention-backend=auto --no-use-tqdm-on-load --host=0.0.0.0 --port=5001 --data-parallel-rpc-port=8991 --limit-mm-per-prompt={"image":0,"video":0} --speculative-config={"method":"mtp","num_speculative_tokens":6}
============================================================
[launcher] tee stdout -> H:\qwen3.6-windows-server\logs\vllm_server.5001.log (also streaming to this terminal)
[launcher] runtime manifest -> H:\qwen3.6-windows-server\logs\runtime\5001.json
(APIServer pid=76280) INFO 05-13 21:34:51 [utils.py:299]

(APIServer pid=76280) INFO 05-13 21:34:51 [utils.py:299]        █     █     █▄   ▄█

(APIServer pid=76280) INFO 05-13 21:34:51 [utils.py:299]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.20.0

(APIServer pid=76280) INFO 05-13 21:34:51 [utils.py:299]   █▄█▀ █     █     █     █  model   E:\ModelScope\vllm\Qwen3.6-27B-int4-AutoRound

(APIServer pid=76280) INFO 05-13 21:34:51 [utils.py:299]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀

(APIServer pid=76280) INFO 05-13 21:34:51 [utils.py:299]
(APIServer pid=76280) INFO 05-13 21:34:51 [utils.py:233] non-default args: {'model_tag': 'E:\\ModelScope\\vllm\\Qwen3.6-27B-int4-AutoRound', 'chat_template': 'H:\\qwen3.6-windows-server\\templates\\qwen3.5-enhanced.jinja', 'default_chat_template_kwargs': {'preserve_thinking': False}, 'enable_auto_tool_choice': True, 'tool_call_parser': 'qwen3_coder', 'host': '0.0.0.0', 'port': 5001, 'model': 'E:\\ModelScope\\vllm\\Qwen3.6-27B-int4-AutoRound', 'trust_remote_code': True, 'max_model_len': 240000, 'quantization': 'autoround', 'served_model_name': ['qwen3.6-27b-autoround'], 'use_tqdm_on_load': False, 'attention_backend': 'auto', 'reasoning_parser': 'qwen3', 'data_parallel_rpc_port': 8991, 'block_size': 32, 'gpu_memory_utilization': 0.93, 'kv_cache_dtype': 'fp8_e4m3', 'enable_prefix_caching': True, 'limit_mm_per_prompt': {'image': 0, 'video': 0}, 'max_num_batched_tokens': 4128, 'max_num_seqs': 1, 'enable_chunked_prefill': True, 'speculative_config': {'method': 'mtp', 'num_speculative_tokens': 6}}
(APIServer pid=76280) INFO 05-13 21:34:51 [system_utils.py:284] Windows detected, skipping ulimit adjustment.
(APIServer pid=76280) WARNING 05-13 21:34:51 [envs.py:1821] Unknown vLLM environment variable detected: VLLM_DISABLE_CUSTOM_OPS
(APIServer pid=76280) WARNING 05-13 21:34:51 [envs.py:1821] Unknown vLLM environment variable detected: VLLM_MODEL_DIR
(APIServer pid=76280) WARNING 05-13 21:34:51 [envs.py:1821] Unknown vLLM environment variable detected: VLLM_NO_MARLIN
(APIServer pid=76280) WARNING 05-13 21:34:51 [envs.py:1821] Unknown vLLM environment variable detected: VLLM_SLEEP_WHEN_IDLE
(APIServer pid=76280) INFO 05-13 21:34:51 [model.py:563] Resolved architecture: Qwen3_5ForConditionalGeneration
(APIServer pid=76280) INFO 05-13 21:34:51 [model.py:1692] Using max model len 240000
(APIServer pid=76280) INFO 05-13 21:34:52 [nixl_utils.py:20] Setting UCX_RCACHE_MAX_UNRELEASED to '1024' to avoid a rare memory leak in UCX when using NIXL.
(APIServer pid=76280) WARNING 05-13 21:34:52 [nixl_utils.py:34] NIXL is not available
(APIServer pid=76280) WARNING 05-13 21:34:52 [nixl_utils.py:44] NIXL agent config is not available
(APIServer pid=76280) INFO 05-13 21:34:53 [cache.py:261] Using fp8_e4m3 data type to store kv cache. It reduces the GPU memory footprint and boosts the performance. Meanwhile, it may cause accuracy drop without a proper scaling factor
(APIServer pid=76280) INFO 05-13 21:34:53 [model.py:563] Resolved architecture: Qwen3_5MTP
(APIServer pid=76280) INFO 05-13 21:34:53 [model.py:1692] Using max model len 262144
(APIServer pid=76280) WARNING 05-13 21:34:53 [speculative.py:654] Enabling num_speculative_tokens > 1 will run multiple times of forward on same MTP layer,which may result in lower acceptance rate
(APIServer pid=76280) INFO 05-13 21:34:53 [scheduler.py:239] Chunked prefill is enabled with max_num_batched_tokens=4128.
(APIServer pid=76280) WARNING 05-13 21:34:53 [config.py:367] Mamba cache mode is set to 'align' for Qwen3_5ForConditionalGeneration by default when prefix caching is enabled
(APIServer pid=76280) INFO 05-13 21:34:53 [config.py:387] Warning: Prefix caching in Mamba cache 'align' mode is currently enabled. Its support for Mamba layers is experimental. Please report any issues you may observe.
(APIServer pid=76280) INFO 05-13 21:34:53 [vllm.py:840] Asynchronous scheduling is enabled.
(APIServer pid=76280) INFO 05-13 21:34:53 [kernel.py:205] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'])
(APIServer pid=76280) WARNING 05-13 21:34:53 [vllm.py:1402] max_num_scheduled_tokens is set to 4128 based on the speculative decoding settings. This may lead to suboptimal performance. Consider increasing max_num_batched_tokens to accommodate the additional draft token slots, or decrease num_speculative_tokens or max_num_seqs.
(APIServer pid=76280) [transformers] `Qwen2VLImageProcessorFast` is deprecated. The `Fast` suffix for image processors has been removed; use `Qwen2VLImageProcessor` instead.
(APIServer pid=76280) INFO 05-13 21:34:54 [registry.py:126] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode.
(EngineCore pid=44048) INFO 05-13 21:35:00 [core.py:109] Initializing a V1 LLM engine (v0.20.0) with config: model='E:\\ModelScope\\vllm\\Qwen3.6-27B-int4-AutoRound', speculative_config=SpeculativeConfig(method='mtp', model='E:\\ModelScope\\vllm\\Qwen3.6-27B-int4-AutoRound', num_spec_tokens=6), tokenizer='E:\\ModelScope\\vllm\\Qwen3.6-27B-int4-AutoRound', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=240000, 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=inc, quantization_config=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=fp8_e4m3, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='qwen3', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=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), seed=0, served_model_name=qwen3.6-27b-autoround, 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::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::gdn_attention_core_xpu', 'vllm::olmo_hybrid_gdn_full_forward', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::deepseek_v4_attention', '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': [4128], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': 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], '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}, 'max_cudagraph_capture_size': 8, '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']), enable_flashinfer_autotune=True, moe_backend='auto')
(EngineCore pid=44048) WARNING 05-13 21:35:00 [nixl_utils.py:34] NIXL is not available
(EngineCore pid=44048) WARNING 05-13 21:35:00 [nixl_utils.py:44] NIXL agent config is not available
(EngineCore pid=44048) [transformers] `Qwen2VLImageProcessorFast` is deprecated. The `Fast` suffix for image processors has been removed; use `Qwen2VLImageProcessor` instead.
(EngineCore pid=44048) INFO 05-13 21:35:02 [registry.py:126] All limits of multimodal modalities supported by the model are set to 0, running in text-only mode.
(EngineCore pid=44048) INFO 05-13 21:35:02 [parallel_state.py:1440] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.50.198:4041 backend=gloo
[W513 21:35:02.000000000 socket.cpp:764] [c10d] The client socket has failed to connect to [prometheusultra]:4041 (system error: 10049 - �����������У�������ĵ�ַ��Ч��).
(EngineCore pid=44048) INFO 05-13 21:35:02 [parallel_state.py:1753] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A
(EngineCore pid=44048) H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\spec_decode\llm_base_proposer.py:145: UserWarning: expandable_segments not supported on this platform (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\pytorch\c10/cuda/CUDAAllocatorConfig.h:39.)
(EngineCore pid=44048)   self.input_ids = torch.zeros(
(EngineCore pid=44048) WARNING 05-13 21:35:02 [__init__.py:206] min_p and logit_bias parameters won't work with speculative decoding.
(EngineCore pid=44048) INFO 05-13 21:35:02 [gpu_model_runner.py:4777] Starting to load model E:\ModelScope\vllm\Qwen3.6-27B-int4-AutoRound...
(EngineCore pid=44048) INFO 05-13 21:35:02 [cuda.py:423] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention
(EngineCore pid=44048) INFO 05-13 21:35:03 [mm_encoder_attention.py:372] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention.
(EngineCore pid=44048) INFO 05-13 21:35:04 [gptq_marlin.py:387] Using MarlinLinearKernel for GPTQMarlinLinearMethod
(EngineCore pid=44048) INFO 05-13 21:35:04 [gdn_linear_attn.py:153] Using Triton/FLA GDN prefill kernel
(EngineCore pid=44048) INFO 05-13 21:35:04 [cuda.py:368] Using FLASHINFER attention backend out of potential backends: ['FLASHINFER', 'TRITON_ATTN'].
(EngineCore pid=44048) INFO 05-13 21:35:04 [weight_utils.py:904] Filesystem type for checkpoints: unknown. Checkpoint size: 17.69 GiB. Available RAM: 17.94 GiB.
(EngineCore pid=44048) INFO 05-13 21:35:04 [weight_utils.py:934] Auto-prefetch is disabled because the filesystem (unknown) is not a recognized network FS (NFS/Lustre) and the checkpoint size (17.69 GiB) exceeds 90% of available RAM (17.94 GiB).
(EngineCore pid=44048) INFO 05-13 21:35:13 [default_loader.py:384] Loading weights took 8.62 seconds
(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136] EngineCore failed to start.

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136] Traceback (most recent call last):

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\engine\core.py", line 1110, in run_engine_core

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     engine_core = EngineCoreProc(*args, engine_index=dp_rank, **kwargs)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\tracing\otel.py", line 178, in sync_wrapper

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     return func(*args, **kwargs)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]            ^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\engine\core.py", line 876, in __init__

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     super().__init__(

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\engine\core.py", line 118, in __init__

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     self.model_executor = executor_class(vllm_config)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\tracing\otel.py", line 178, in sync_wrapper

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     return func(*args, **kwargs)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]            ^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\executor\abstract.py", line 109, in __init__

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     self._init_executor()

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\executor\uniproc_executor.py", line 52, in _init_executor

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     self.driver_worker.load_model()

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\worker\gpu_worker.py", line 324, in load_model

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     self.model_runner.load_model(load_dummy_weights=load_dummy_weights)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\tracing\otel.py", line 178, in sync_wrapper

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     return func(*args, **kwargs)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]            ^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\v1\worker\gpu_model_runner.py", line 4793, in load_model

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     self.model = model_loader.load_model(

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]                  ^^^^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\tracing\otel.py", line 178, in sync_wrapper

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     return func(*args, **kwargs)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]            ^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\model_executor\model_loader\base_loader.py", line 80, in load_model

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     process_weights_after_loading(model, model_config, target_device)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\model_executor\model_loader\utils.py", line 111, in process_weights_after_loading

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     quant_method.process_weights_after_loading(module)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\model_executor\layers\quantization\gptq_marlin.py", line 486, in process_weights_after_loading

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     self.kernel.process_weights_after_loading(layer)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\model_executor\kernels\linear\mixed_precision\marlin.py", line 167, in process_weights_after_loading

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     self._transform_param(layer, self.w_q_name, transform_w_q)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\model_executor\kernels\linear\mixed_precision\MPLinearKernel.py", line 74, in _transform_param

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     new_param = fn(old_param)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]                 ^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\model_executor\kernels\linear\mixed_precision\marlin.py", line 99, in transform_w_q

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     x.data = ops.gptq_marlin_repack(

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]              ^^^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\vllm\_custom_ops.py", line 1279, in gptq_marlin_repack

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     return torch.ops._C.gptq_marlin_repack(

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]   File "H:\qwen3.6-windows-server\python\Lib\site-packages\torch\_ops.py", line 1269, in __call__

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]     return self._op(*args, **kwargs)

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136]            ^^^^^^^^^^^^^^^^^^^^^^^^^

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136] torch.AcceleratorError: CUDA error: the provided PTX was compiled with an unsupported toolchain.

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136] Search for `cudaErrorUnsupportedPtxVersion' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136] CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136] For debugging consider passing CUDA_LAUNCH_BLOCKING=1

(EngineCore pid=44048) ERROR 05-13 21:35:13 [core.py:1136] Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Verifier output

python windows_tools\verify_install.py --venv .\venv
(paste here)

What you've already tried

(e.g. dropped to enforce-eager, disabled MTP, ran the coherence check, etc.)

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions