diff --git a/vllm/model_executor/layers/fused_moe/modular_kernel.py b/vllm/model_executor/layers/fused_moe/modular_kernel.py
index f2e6e2560e70..b7b652273728 100644
--- a/vllm/model_executor/layers/fused_moe/modular_kernel.py
+++ b/vllm/model_executor/layers/fused_moe/modular_kernel.py
@@ -10,6 +10,7 @@
 import torch
 
 import vllm.envs as envs
+from vllm._aiter_ops import rocm_aiter_ops
 from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe.activation import (
     MoEActivation,
@@ -1394,6 +1395,15 @@ def apply(
             expert_tokens_meta=expert_tokens_meta,
         )
 
+        # Aiter don't use temp buffer, so can skip copy if output is contiguous
+        if (
+            rocm_aiter_ops.is_fused_moe_enabled()
+            and not self.inplace
+            and fused_out.shape == output.shape
+            and fused_out.is_contiguous()
+        ):
+            output = fused_out
+
         return self._finalize(
             output,
             fused_out,
diff --git a/vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py b/vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py
index d24bda101ffa..bf75b80b339e 100644
--- a/vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py
+++ b/vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py
@@ -437,4 +437,4 @@ def apply(
             num_local_tokens=num_local_tokens,
             output_dtype=output.dtype,
         )
-        output.copy_(result)
+        output.data = result
diff --git a/vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py b/vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py
index 4cebe608a6b4..5c5a923648f0 100644
--- a/vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py
+++ b/vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py
@@ -69,7 +69,8 @@ def apply(
             f"But got output={output.size()}, "
             f"used_expert_output={fused_expert_output.size()}"
         )
-        output.copy_(fused_expert_output, non_blocking=True)
+        if output.data_ptr() != fused_expert_output.data_ptr():
+            output.copy_(fused_expert_output, non_blocking=True)
         return output