Fix the accuracy issue when running mxfp4 dsv3 model and enable ep (#15304)

Co-authored-by: wunhuang <wunhuang@amd.com>
2025-12-17 17:53:22 +08:00
parent e9abb52576
commit 8c34e18140
2 changed files with 2 additions and 0 deletions
--- a/python/sglang/srt/layers/quantization/mxfp4.py
+++ b/python/sglang/srt/layers/quantization/mxfp4.py
@@ -866,5 +866,6 @@ class Mxfp4DynamicQuantMoEMethod(FusedMoEMethodBase):
                else ActivationType.Gelu
            ),
            doweight_stage1=False,
+            expert_mask=layer.expert_mask_gpu,
        )
        return StandardCombineInput(hidden_states=output)
--- a/python/sglang/srt/layers/quantization/quark/quark_moe.py
+++ b/python/sglang/srt/layers/quantization/quark/quark_moe.py
@@ -245,6 +245,7 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
                else ActivationType.Gelu
            ),
            doweight_stage1=False,
+            expert_mask=layer.expert_mask_gpu,
        )
        return StandardCombineInput(hidden_states=output)