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

[bugfix] Fix moe bug: allgather error. (#3279)

Browse Source 
It will crash when deepseek model executed in A2.


- vLLM version: v0.11.0rc3
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.0

---------

Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
This commit is contained in:
weijinqian0
2025-09-30 18:45:09 +08:00
committed by GitHub
parent b8c58d68e1
commit 474fa737c8
2 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/ops/moe/token_dispatcher.py
View File

@@ -383,7 +383,7 @@ class TokenDispatcherWithAllGather(MoETokenDispatcher):
assert self.original_shape is not None
final_hidden_states = torch_npu.npu_moe_token_unpermute(
permuted_tokens=hidden_states,
sorted_indices=self.expanded_row_idx,
sorted_indices=torch.abs(self.expanded_row_idx),
probs=self.topk_weights)
if len(self.original_shape) == 3:
final_hidden_states = final_hidden_states.view(self.original_shape)