[Bug]: Wrong output tensor shape when running DeepSeek model with V1 engine

[yiz-liu]

Your current environment

The environment is unrelated to this issue; therefore, including environment logs is unnecessary.

🐛 Describe the bug

The following error is encountered when running the DeepSeek model:

ERROR 05-07 16:27:13 [core.py:402] EngineCore encountered a fatal error.
ERROR 05-07 16:27:13 [core.py:402] Traceback (most recent call last):
ERROR 05-07 16:27:13 [core.py:402]   File "workspace/vllm/vllm/v1/engine/core.py", line 393, in run_engine_core
ERROR 05-07 16:27:13 [core.py:402]     engine_core.run_busy_loop()
ERROR 05-07 16:27:13 [core.py:402]   File "workspace/vllm/vllm/v1/engine/core.py", line 417, in run_busy_loop
ERROR 05-07 16:27:13 [core.py:402]     self._process_engine_step()
ERROR 05-07 16:27:13 [core.py:402]   File "workspace/vllm/vllm/v1/engine/core.py", line 442, in _process_engine_step
ERROR 05-07 16:27:13 [core.py:402]     outputs = self.step_fn()
ERROR 05-07 16:27:13 [core.py:402]               ^^^^^^^^^^^^^^
ERROR 05-07 16:27:13 [core.py:402]   File "workspace/vllm/vllm/v1/engine/core.py", line 205, in step
ERROR 05-07 16:27:13 [core.py:402]     output = self.model_executor.execute_model(scheduler_output)
ERROR 05-07 16:27:13 [core.py:402]              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 05-07 16:27:13 [core.py:402]   File "workspace/vllm/vllm/v1/executor/multiproc_executor.py", line 158, in execute_model
ERROR 05-07 16:27:13 [core.py:402]     (output, ) = self.collective_rpc("execute_model",
ERROR 05-07 16:27:13 [core.py:402]                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 05-07 16:27:13 [core.py:402]   File "workspace/vllm/vllm/v1/executor/multiproc_executor.py", line 215, in collective_rpc
ERROR 05-07 16:27:13 [core.py:402]     result = get_response(w, dequeue_timeout)
ERROR 05-07 16:27:13 [core.py:402]              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 05-07 16:27:13 [core.py:402]   File "workspace/vllm/vllm/v1/executor/multiproc_executor.py", line 202, in get_response
ERROR 05-07 16:27:13 [core.py:402]     raise RuntimeError(
ERROR 05-07 16:27:13 [core.py:402] RuntimeError: Worker failed with error 'call aclnnInplaceCopy failed, detail:EZ1001: [PID: 1123276] 2025-05-07-16:27:13.714.340 128 and 2048 cannot broadcast.
ERROR 05-07 16:27:13 [core.py:402]         TraceBack (most recent call last):
ERROR 05-07 16:27:13 [core.py:402]         the size of tensor self [17,2048] must match the size of tensor src [17,16,128].

[yiz-liu]

After debugging, we discovered that the shape of the attn_output tensor changes when chunked prefill is enabled. We resolved this issue by applying view_as. For full details, please refer to PR #773.

[Yikun]

Thanks, I added this to v0.8.5rc1 release note: https://github.com/vllm-project/vllm-ascend/releases/tag/v0.8.5rc1

[wangxiyuan]

please try 0.8.5rc1