[AMD][Bugfix] Skip invalid fused MoE reduction for direct top-1 output (#36309)

2026-08-26 18:52:15 -07:00
parent b658d64d5b
commit 4f59a8dcfa
1 changed files with 3 additions and 1 deletions
@@ -908,7 +908,9 @@ def _fused_moe_kernel_sequence(
                    routed_scaling_factor,
                )
    elif _is_hip:
-        if _use_aiter:
+        if topk == 1 and routed_scaling_factor == 1.0 and not _use_intermediate:
+            pass  # we wrote directly into out_hidden_states
+        elif _use_aiter:
            moe_sum(
                intermediate_cache3.view(*intermediate_cache3.shape),
                out_hidden_states,