基础环境如下：
硬件：Atlas300I Pro（310P3*6） 宿主机驱动 HDK：25.5.1，使用的镜像：vllm-ascend:v0.23.0-310p （容器内CANN版本为9.1.0 ）模型：Qwen3-0.6B、Qwen3.5-4B、Qwen3-Embedding-0.6B等均存在同样问题。
下面同样的命令在 Atlas300I Duo 上都正常，也能推理，但是在Atlas300I Pro出现卡死（尝试了好多方法均未解决），看看大家有遇到类似的问题吗，还是说300I Pro不支持（但是看到论坛有人能跑通），希望能给出指导 。
问题日志如下（每次都是在同样的问题卡死，详见日志）：
INFO 09-28 07:53:22 [__init__.py:238] Platform plugin ascend is activated
INFO 09-28 07:53:22 [platform.py:62] Breakable cudagraph is force disabled on Ascend because DeepSeek V4 PIECEWISE cudagraph is not supported yet.
INFO 09-28 07:53:30 [importing.py:81] Triton not installed or not compatible; certain GPU-related functions will not be available.
(EngineCore pid=335) 2026-09-28 07:53:33,580 - 335 - ServiceProfiler - INFO - VLLM_USE_V1 not set, auto-detected via vLLM 0.23.0+empty: default 1
(EngineCore pid=335) INFO 09-28 07:53:33 [__init__.py:115] Registered model loader `<class 'vllm_ascend.model_loader.netloader.netloader.ModelNetLoaderElastic'>` with load format `netloader`
(EngineCore pid=335) INFO 09-28 07:53:33 [__init__.py:115] Registered model loader `<class 'vllm_ascend.model_loader.rfork.rfork_loader.RForkModelLoader'>` with load format `rfork`
(EngineCore pid=335) INFO 09-28 07:53:33 [core.py:113] Initializing a V1 LLM engine (v0.23.0) with config: model='/home/scdata/vllm/Qwen3.5-4B', speculative_config=SpeculativeConfig(method='mtp', model='/home/scdata/vllm/Qwen3.5-4B', num_spec_tokens=1), tokenizer='/home/scdata/vllm/Qwen3.5-4B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=4096, 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=True, quantization=None, quantization_config=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=npu, 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, 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=qwen3.5, 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': 'vllm_ascend.compilation.compiler_interface.AscendCompiler', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], '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': [2048], '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_DECODE_ONLY: (2, 0)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [2, 16], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 16, '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']), enable_flashinfer_autotune=True, moe_backend='auto', linear_backend='auto')
(EngineCore pid=335) `Qwen2VLImageProcessorFast` is deprecated. The `Fast` suffix for image processors has been removed; use `Qwen2VLImageProcessor` instead.
(EngineCore pid=335) INFO 09-28 07:53:37 [ascend_config.py:893] Dynamic EPLB is False
(EngineCore pid=335) INFO 09-28 07:53:37 [ascend_config.py:894] The number of redundant experts is 0
实际部署使用的文档和命令如下：
启动命令参考官方部署文档：Qwen3.5-Dense (2B/4B/9B) - vLLM Ascend (中文)
容器启动命令（两种方式都使用了，启动容器正常，进入容器后npu-smi info 都没问题）：
方式一：
docker run -itd \
--name vllm-ascend \
--net=host \
--device /dev/davinci2 \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
--shm-size=16g \
-v /home/scdata/:/home/scdata \
ascendai/vllm-ascend:v0.23.0rc1-310p \
bash
方式二：
docker run -d \
--runtime=ascend \
--name qwen-4b \
--shm-size=16g \
-e ASCEND_VISIBLE_DEVICES=0 \
-v /home/scdata:/home/scdata \
-p 8000:8000 \
-it "quay.io/ascend/vllm-ascend:v0.23.0-310p" bash
进入容器启动服务命令（都同样卡死，日志详见下面）：
vllm serve /home/scdata/vllm/Qwen3.5-4B \
--host 127.0.0.1 \
--port 8080 \
--tensor-parallel-size 1 \
--served-model-name qwen3.5 \
--max-num-seqs 32 \
--max-model-len 4096 \
--trust-remote-code \
--gpu-memory-utilization 0.80 \
--mamba-ssm-cache-dtype float16 \
--dtype float16 \
--speculative-config '{"method": "qwen3_5_mtp","num_speculative_tokens":1}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes": [2,16]}' \
--additional-config '{"ascend_compilation_config": {"enable_npugraph_ex": false}}' \
--enforce-eager
点赞
收藏
回复
分享
举报
已解决
基础环境如下：
硬件：Atlas300I Pro（310P3*6） 宿主机驱动 HDK：25.5.1，使用的镜像：vllm-ascend:v0.23.0-310p （容器内CANN版本为9.1.0 ）模型：Qwen3-0.6B、Qwen3.5-4B、Qwen3-Embedding-0.6B等均存在同样问题。
下面同样的命令在 Atlas300I Duo 上都正常，也能推理，但是在Atlas300I Pro出现卡死（尝试了好多方法均未解决），看看大家有遇到类似的问题吗，还是说300I Pro不支持（但是看到论坛有人能跑通），希望能给出指导 。
问题日志如下（每次都是在同样的问题卡死，详见日志）：
INFO 09-28 07:53:22 [__init__.py:238] Platform plugin ascend is activated
INFO 09-28 07:53:22 [platform.py:62] Breakable cudagraph is force disabled on Ascend because DeepSeek V4 PIECEWISE cudagraph is not supported yet.
INFO 09-28 07:53:30 [importing.py:81] Triton not installed or not compatible; certain GPU-related functions will not be available.
(EngineCore pid=335) 2026-09-28 07:53:33,580 - 335 - ServiceProfiler - INFO - VLLM_USE_V1 not set, auto-detected via vLLM 0.23.0+empty: default 1
(EngineCore pid=335) INFO 09-28 07:53:33 [__init__.py:115] Registered model loader `<class 'vllm_ascend.model_loader.netloader.netloader.ModelNetLoaderElastic'>` with load format `netloader`
(EngineCore pid=335) INFO 09-28 07:53:33 [__init__.py:115] Registered model loader `<class 'vllm_ascend.model_loader.rfork.rfork_loader.RForkModelLoader'>` with load format `rfork`
(EngineCore pid=335) INFO 09-28 07:53:33 [core.py:113] Initializing a V1 LLM engine (v0.23.0) with config: model='/home/scdata/vllm/Qwen3.5-4B', speculative_config=SpeculativeConfig(method='mtp', model='/home/scdata/vllm/Qwen3.5-4B', num_spec_tokens=1), tokenizer='/home/scdata/vllm/Qwen3.5-4B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=4096, 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=True, quantization=None, quantization_config=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=npu, 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, 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=qwen3.5, 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': 'vllm_ascend.compilation.compiler_interface.AscendCompiler', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], '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': [2048], '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_DECODE_ONLY: (2, 0)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [2, 16], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 16, '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']), enable_flashinfer_autotune=True, moe_backend='auto', linear_backend='auto')
(EngineCore pid=335) `Qwen2VLImageProcessorFast` is deprecated. The `Fast` suffix for image processors has been removed; use `Qwen2VLImageProcessor` instead.
(EngineCore pid=335) INFO 09-28 07:53:37 [ascend_config.py:893] Dynamic EPLB is False
(EngineCore pid=335) INFO 09-28 07:53:37 [ascend_config.py:894] The number of redundant experts is 0
实际部署使用的文档和命令如下：
启动命令参考官方部署文档：Qwen3.5-Dense (2B/4B/9B) - vLLM Ascend (中文)
容器启动命令（两种方式都使用了，启动容器正常，进入容器后npu-smi info 都没问题）：
方式一：
docker run -itd \
--name vllm-ascend \
--net=host \
--device /dev/davinci2 \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
--shm-size=16g \
-v /home/scdata/:/home/scdata \
ascendai/vllm-ascend:v0.23.0rc1-310p \
bash
方式二：
docker run -d \
--runtime=ascend \
--name qwen-4b \
--shm-size=16g \
-e ASCEND_VISIBLE_DEVICES=0 \
-v /home/scdata:/home/scdata \
-p 8000:8000 \
-it "quay.io/ascend/vllm-ascend:v0.23.0-310p" bash
进入容器启动服务命令（都同样卡死，日志详见下面）：
vllm serve /home/scdata/vllm/Qwen3.5-4B \
--host 127.0.0.1 \
--port 8080 \
--tensor-parallel-size 1 \
--served-model-name qwen3.5 \
--max-num-seqs 32 \
--max-model-len 4096 \
--trust-remote-code \
--gpu-memory-utilization 0.80 \
--mamba-ssm-cache-dtype float16 \
--dtype float16 \
--speculative-config '{"method": "qwen3_5_mtp","num_speculative_tokens":1}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes": [2,16]}' \
--additional-config '{"ascend_compilation_config": {"enable_npugraph_ex": false}}' \
--enforce-eager