diff --git a/python/sglang/srt/layers/quantization/fp8.py b/python/sglang/srt/layers/quantization/fp8.py
index 1e12baff1..c7874830d 100644
--- a/python/sglang/srt/layers/quantization/fp8.py
+++ b/python/sglang/srt/layers/quantization/fp8.py
@@ -998,23 +998,15 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 w2_weight_scale, requires_grad=False
             )
             layer.w2_input_scale = None
-            if _use_aiter:
-                # add this section for MI300
-                # Pre-shuffle weights
-                layer.w13_weight.data = shuffle_weight(
-                    layer.w13_weight.contiguous(), (16, 16)
-                )
-                layer.w2_weight.data = shuffle_weight(
-                    layer.w2_weight.contiguous(), (16, 16)
-                )
-        elif _use_aiter:
+
+        if _use_aiter:
             # Pre-shuffle weights
-            layer.w13_weight.data = shuffle_weight(
-                layer.w13_weight.contiguous(), (16, 16)
-            )
-            layer.w2_weight.data = shuffle_weight(
-                layer.w2_weight.contiguous(), (16, 16)
-            )
+            t = shuffle_weight(layer.w13_weight, (16, 16))
+            layer.w13_weight.copy_(t)
+            del t
+            t = shuffle_weight(layer.w2_weight, (16, 16))
+            layer.w2_weight.copy_(t)
+            del t
         elif _is_cpu:
             assert (
                 _is_cpu_amx_available