[perf] Add gemma RMS AR fusion (#42646)

Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
This commit is contained in:
Jiahan Chang (Cyrus)
2026-06-04 01:33:59 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Jiangyun Zhu
parent 1bdc60ed53
commit d0975a4b50
3 changed files with 225 additions and 16 deletions
+4 -14
View File
@@ -159,20 +159,10 @@ class GemmaRMSNorm(CustomOp):
residual: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
"""PyTorch-native implementation equivalent to forward()."""
orig_dtype = x.dtype
weight = self.weight.data.float() + 1.0
if residual is not None:
x = (
x.float() + residual.float()
if orig_dtype == torch.float16
else x + residual
)
residual = x
# ir.ops.rms_norm handles fp32 upcast internally
out = ir.ops.rms_norm(x, weight, self.variance_epsilon)
return (
out.to(orig_dtype) if residual is None else (out.to(orig_dtype), residual)
)
weight = self.weight.float() + 1.0
if residual is None:
return ir.ops.rms_norm(x, weight, self.variance_epsilon)
return ir.ops.fused_add_rms_norm(x, residual, weight, self.variance_epsilon)
def forward_cuda(
self,