[Bug]: FlashMLA: invalid configuration argument

[yewentao256]

Your current environment

The output of python collect_env.py

Your output of `python collect_env.py` here

🐛 Describe the bug

On B200

vllm serve deepseek-ai/DeepSeek-V3.2-Exp -dp 8 --enable-expert-parallel --port 9256

vllm bench serve --model deepseek-ai/DeepSeek-V3.2-Exp --dataset-name random --host 127.0.0.1 --port 9256 --random-input-len 256 --random-output-len 256 --request-rate inf --num-prompts 1024

Will meet error

CUDA error (/home/wentao/vllm-source/cmake-build-release/_deps/flashmla-src/csrc/smxx/mla_combine.cu:201): invalid configuration argument
CUDA error (/home/wentao/vllm-source/cmake-build-release/_deps/flashmla-src/csrc/smxx/mla_combine.cu:201): invalid configuration argument
CUDA error (/home/wentao/vllm-source/cmake-build-release/_deps/flashmla-src/csrc/smxx/mla_combine.cu:201): invalid configuration argument
CUDA error (/home/wentao/vllm-source/cmake-build-release/_deps/flashmla-src/csrc/smxx/mla_combine.cu:201): invalid configuration argument
CUDA error (/home/wentao/vllm-source/cmake-build-release/_deps/flashmla-src/csrc/smxx/mla_combine.cu:201): invalid configuration argument
CUDA error (/home/wentao/vllm-source/cmake-build-release/_deps/flashmla-src/csrc/smxx/mla_combine.cu:201): invalid configuration argument
CUDA error (/home/wentao/vllm-source/cmake-build-release/_deps/flashmla-src/csrc/smxx/mla_combine.cu:201): invalid configuration argument
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:18 [core_client.py:597] Engine core proc EngineCore_DP6 died unexpectedly, shutting down client.
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524] AsyncLLM output_handler failed.
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524] Traceback (most recent call last):
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524]   File "/home/wentao/vllm-source/vllm/v1/engine/async_llm.py", line 478, in output_handler
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524]     outputs = await engine_core.get_output_async()
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524]               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524]   File "/home/wentao/vllm-source/vllm/v1/engine/core_client.py", line 882, in get_output_async
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524]     raise self._format_exception(outputs) from None
�[1;36m(APIServer pid=767475)�[0;0m ERROR 10-16 10:21:21 [async_llm.py:524] vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[LucasWilkinson]

Should eventually be resolved by #26152