diff --git a/python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe.py b/python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe.py
index 8cc0f4ff7..546f79b58 100644
--- a/python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe.py
+++ b/python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe.py
@@ -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,