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 4, 03:45 UTC
[ROCm] Fix AITER AR+RMSNorm no-residual fusion (#41972)
Signed-off-by: Aakif Nawaz <aakif.nawaz@amd.com>
akii96 committed
May 7, 2026 at 23:14 UTC
3af561ec0a00aabd407cf782a60f17e86c1a253f
1 file changed
+1
-1
vllm/compilation/passes/fusion/allreduce_rms_fusion.py
+1
-1
@@ -939,7 +939,7 @@ class AiterAllreduceFusedRMSNormPattern(BasePattern, VllmPatternReplacement):
939
def _replacement(
940
input: torch.Tensor, weight: torch.Tensor
941
) -> tuple[torch.Tensor, torch.Tensor]:
942
- residual = torch.empty_like(input)
942
+ residual = torch.zeros_like(input)
943
allreduce = self.FUSED_AR_RMSNORM_OP(
944
input_=input,
945
residual=residual,