[ROCm] Fix AITER AR+RMSNorm no-residual fusion (#41972)

Signed-off-by: Aakif Nawaz <aakif.nawaz@amd.com>
This commit is contained in:
akii96
2026-05-07 13:14:58 -07:00
committed by GitHub
parent c936548ce6
commit 3af561ec0a
@@ -939,7 +939,7 @@ class AiterAllreduceFusedRMSNormPattern(BasePattern, VllmPatternReplacement):
def _replacement(
input: torch.Tensor, weight: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
residual = torch.empty_like(input)
residual = torch.zeros_like(input)
allreduce = self.FUSED_AR_RMSNORM_OP(
input_=input,
residual=residual,