[Bugfix] Fix precision issues in moe_mlp (vllm-ascend main) (#5025)

### What this PR does / why we need it?
Use group_list[0] to replace group_diff[0] in function
"cumsum_group_list" (moe_mlp.py).
The purpose is to modify it to the correct logic of converting cumsum to
count.

### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: tanqingshan (A)  <50050625@china.huawei.com>
Co-authored-by: tanqingshan (A) <50050625@china.huawei.com>
This commit is contained in:
Clorist33
2025-12-16 08:39:54 +08:00
committed by GitHub
parent b662d914a4
commit d43cabc2b1

View File

@@ -46,7 +46,7 @@ def cumsum_group_list(group_list: torch.Tensor,
return group_list.cumsum(dim=0)
if src_list_type == 0 and dst_list_type == 1:
group_diff = torch.diff(group_list)
new_group = torch.cat([group_diff[0].unsqueeze(0), group_diff], dim=0)
new_group = torch.cat([group_list[0].unsqueeze(0), group_diff], dim=0)
return new_group
if src_list_type == 2 and dst_list_type == 0:
experts = pad(group_list[:, 0], (1, 0))