fix(marlin_nvfp4): only apply routed_scaling_factor in moe_sum_reduce (#31762)

Signed-off-by: Muqi Li <muqi1029@gmail.com>
2026-07-22 08:45:42 +08:00
parent 03342e7732
commit a2c38175a4
1 changed files with 6 additions and 1 deletions
@@ -373,7 +373,12 @@ class FusedMoE(torch.nn.Module):
            self.moe_runner_config.inplace = False

        self.should_fuse_routed_scaling_factor_in_topk = (
-            isinstance(self.quant_method, ModelOptNvFp4FusedMoEMethod)
+            (
+                isinstance(self.quant_method, ModelOptNvFp4FusedMoEMethod)
+                and not getattr(
+                    self.quant_method, "_moe_runner_backend", get_moe_runner_backend()
+                ).is_marlin()
+            )
            or (
                isinstance(self.quant_method, Fp8MoEMethod)
                and (