[AMD] fix the regression issue for DeepseekV3 on MI300 (#14383)

2025-12-04 15:30:11 +08:00
parent fa78c44a3d
commit d6c490192d
1 changed files with 9 additions and 0 deletions
--- a/python/sglang/srt/layers/quantization/fp8.py
+++ b/python/sglang/srt/layers/quantization/fp8.py
@@ -762,6 +762,15 @@ 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:
                # Pre-shuffle weights
                layer.w13_weight.data = shuffle_weight(