EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fix uncompatible between fc1 and non-sp-padding (#7643)

Browse Source 
cherry pick https://github.com/vllm-project/vllm-ascend/pull/7614
### What this PR does / why we need it?
fix uncompatible between fc1 and non-sp-padding
After PR
[non-sp-padding](https://github.com/vllm-project/vllm-ascend/pull/7297),
kimi2.5 open flashcomm1 will raise an error : The expanded size of the
tensor do not match the existing size at non-singleton dimension 0.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.18.0
- vLLM-Ascend main: 9976e685b7

Signed-off-by: Wangbei25 <wangbei41@huawie.com>
Co-authored-by: Wangbei25 <wangbei41@huawie.com>
This commit is contained in:
Wangbei25
2026-03-25 23:23:37 +08:00
committed by GitHub
parent 2ad0ca52a6
commit dd55736ee4
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1976,7 +1976,7 @@ class NPUModelRunner(GPUModelRunner):
_, num_tokens_across_dp, synced_cudagraph_mode = self._sync_batch_across_dp(
num_tokens_padded=num_tokens_padded,
cudagraph_mode=cudagraph_mode.value,
allow_dp_padding=cudagraph_mode != CUDAGraphMode.NONE,
allow_dp_padding=(cudagraph_mode != CUDAGraphMode.NONE) or enable_sp(self.vllm_config),
)
# Extract DP padding if there is any