biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Use 'sum' reduction instead of 'avg' in Async TP reduce-scatter (#33088)

Browse Source 
Signed-off-by: Xingran Wang <wangxingran123456@outlook.com>
Signed-off-by: Hongjian Zhang <hirokenovo@gmail.com>
Co-authored-by: Hongjian Zhang <hirokenovo@gmail.com>
This commit is contained in:
Wang Xingran
2026-02-27 15:06:08 +08:00
committed by GitHub
parent a572baff5e
commit 07bdabef03
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/compilation/passes/fusion/collective_fusion.py
View File

@@ -53,7 +53,7 @@ class GEMMReduceScatterPattern(BasePattern):
gemm_rs = torch.ops.symm_mem.fused_matmul_reduce_scatter(
mul,
mm_weight,
"avg",
"sum",
scatter_dim=0,
group_name=self.tp.device_group.group_name,
)
@@ -150,7 +150,7 @@ class ScaledMMReduceScatterPattern(BasePattern):
mat2,
scale_a,
scale_b,
"avg",
"sum",
scatter_dim, # orig_scatter_dim
scatter_dim, # scatter_dim_after_maybe_reshape
self.tp.device_group.group_name,
@@ -285,7 +285,7 @@ class CutlassScaledMMReduceScatterPattern(BasePattern):
mat2,
scale_a,
scale_b,
"avg",
"sum",
scatter_dim, # orig_scatter_dim
scatter_dim, # scatter_dim_after_maybe_reshape
self.tp.device_group.group_name,