A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 15:15 UTC
[Bugfix] Use 'sum' reduction instead of 'avg' in Async TP reduce-scatter (#33088)
Signed-off-by: Xingran Wang <wangxingran123456@outlook.com> Signed-off-by: Hongjian Zhang <hirokenovo@gmail.com> Co-authored-by: Hongjian Zhang <hirokenovo@gmail.com>
Wang Xingran committed
Feb 27, 2026 at 15:06 UTC
07bdabef03c78c9dc6beb549185930888ec6f2ce
1 file changed
+3
-3
vllm/compilation/passes/fusion/collective_fusion.py
+3
-3
@@ -53,7 +53,7 @@ class GEMMReduceScatterPattern(BasePattern):
53
gemm_rs = torch.ops.symm_mem.fused_matmul_reduce_scatter(
54
mul,
55
mm_weight,
56
- "avg",
56
+ "sum",
57
scatter_dim=0,
58
group_name=self.tp.device_group.group_name,
59
)
@@ -150,7 +150,7 @@ class ScaledMMReduceScatterPattern(BasePattern):
150
mat2,
151
scale_a,
152
scale_b,
153
- "avg",
153
+ "sum",
154
scatter_dim, # orig_scatter_dim
155
scatter_dim, # scatter_dim_after_maybe_reshape
156
self.tp.device_group.group_name,
@@ -285,7 +285,7 @@ class CutlassScaledMMReduceScatterPattern(BasePattern):
285
mat2,
286
scale_a,
287
scale_b,
288
- "avg",
288
+ "sum",
289
scatter_dim, # orig_scatter_dim
290
scatter_dim, # scatter_dim_after_maybe_reshape
291
self.tp.device_group.group_name,