vLLM-Ascend在银河麒麟桌面版上不开Eager就报错，开了性能暴跌，换服务器版就正常，这是系统差异导致的吗？

General
1

vllm-ascend（310p) 是否支持银河麒麟桌面版操作系统， 如果不开启eager mode, 就会提示下面这个报错， 但是如果开启eager mode, 吞吐 仅仅只有 5~6 tokens/s ， 如果把 桌面版操作系统换成银河麒麟服务器版本， 不用开启eager的情况下，吞吐速度就会有成倍的提升

ant to get the accurate stacktrace, please set the environment variable ASCEND_LAUNCH_BLOCKING=1.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] Note: ASCEND_LAUNCH_BLOCKING=1 will force ops to run in synchronous mode, resulting in performance degradation. Please unset ASCEND_LAUNCH_BLOCKING in time after debugging.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] [ERROR] 2026-09-04-05:30:44 (PID:243, Device:1, RankID:-1) ERR00100 PTA call acl api failed.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] [PID: 243] 2026-09-04-05:28:07.503.485 Invalid_Argument(EH0012): aclrtAllocatorGetByStream failed. Parameter stream is invalid. Reason: The stream is not registered with any allocator.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] TraceBack (most recent call last):
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]         Asynchronous copy task failed. Reason: The pageable memory copy task does not support graph capture.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]         The argument is invalid. Reason: rtMemcpyAsync execution failed.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] 
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] 
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] During handling of the above exception, another exception occurred:
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] 
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] Traceback (most recent call last):
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]   File "/vllm-workspace/vllm/vllm/v1/executor/multiproc_executor.py", line 999, in worker_busy_loop
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]     output = func(*args, **kwargs)
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]              ^^^^^^^^^^^^^^^^^^^^^
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]   File "/vllm-workspace/vllm-ascend/vllm_ascend/worker/worker.py", line 778, in compile_or_warm_up_model
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]     npugraph_memory_bytes = self.model_runner.capture_model()
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]   File "/vllm-workspace/vllm-ascend/vllm_ascend/worker/model_runner_v1.py", line 5010, in capture_model
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]     with _torch_cuda_wrapper(), _replace_gpu_model_runner_function_wrapper(parent_module_name):
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]          ^^^^^^^^^^^^^^^^^^^^^
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]   File "/usr/local/python3.12.13/lib/python3.12/contextlib.py", line 158, in __exit__
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]     self.gen.throw(value)
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]   File "/vllm-workspace/vllm-ascend/vllm_ascend/worker/model_runner_v1.py", line 5117, in _torch_cuda_wrapper
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]     raise RuntimeError(f"NPUModelRunner init failed, error is {e}")
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] RuntimeError: NPUModelRunner init failed, error is NPUModelRunner failed, error is The Inner error is reported as above. The process exits for this inner error, and the current copy params are srclen=8, dstlen=8, kind=1.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] Since the operator is called asynchronously, the stacktrace may be inaccurate. If you want to get the accurate stacktrace, please set the environment variable ASCEND_LAUNCH_BLOCKING=1.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] Note: ASCEND_LAUNCH_BLOCKING=1 will force ops to run in synchronous mode, resulting in performance degradation. Please unset ASCEND_LAUNCH_BLOCKING in time after debugging.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] [ERROR] 2026-09-04-05:30:44 (PID:243, Device:1, RankID:-1) ERR00100 PTA call acl api failed.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] [PID: 243] 2026-09-04-05:28:07.503.485 Invalid_Argument(EH0012): aclrtAllocatorGetByStream failed. Parameter stream is invalid. Reason: The stream is not registered with any allocator.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] TraceBack (most recent call last):
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]         Asynchronous copy task failed. Reason: The pageable memory copy task does not support graph capture.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]         The argument is invalid. Reason: rtMemcpyAsync execution failed.
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007] 
(Worker_TP1 pid=243) ERROR 09-04 05:30:46 [multiproc_executor.py:1007]
2

目前数据源中没有明确说明 vllm-ascend（310p）是否官方支持银河麒麟桌面版操作系统，仅有对硬件、驱动、环境变量等常见问题的排查建议。你遇到的“aclrtAllocatorGetByStream failed…stream is not registered with any allocator”以及“pageable memory copy task does not support graph capture”等报错，似乎与桌面版操作系统的内存管理或驱动兼容性有关，导致 graph capture（非eager模式）失败，只能退回到eager模式，吞吐下降明显。据你的描述，换成银河麒麟服务器版后问题消失，吞吐提升，这进一步说明桌面版OS与NPU驱动/ACL API的兼容性可能存在问题。

我的理解是，vLLM官方文档和相关issue主要建议遇到此类底层报错时，优先排查驱动、硬件、操作系统兼容性，并未针对银河麒麟桌面版做过适配说明。建议优先使用银河麒麟服务器版等官方推荐的企业级Linux发行版，避免桌面版OS带来的兼容性问题。如需进一步定位，可尝试升级驱动、核查ACL API支持列表，或联系NPU/操作系统厂商获取支持。你需要更详细的排查建议或日志分析方法吗？

Sources:

Sources:

Hint: Mention @Herald in the post for followups.