fix moe-ep accuracy issue for fp8 (#2489)

2024-12-16 20:54:02 +08:00
parent a0592c059f
commit b532a5fd16
1 changed files with 4 additions and 0 deletions
--- a/python/sglang/srt/layers/ep_moe/layer.py
+++ b/python/sglang/srt/layers/ep_moe/layer.py
@@ -644,6 +644,10 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
                        "QuantConfig has static quantization, but found "
                        "activation scales are None."
                    )
+                layer.w13_weight_scale = torch.nn.Parameter(
+                    torch.max(layer.w13_weight_scale, dim=1).values,
+                    requires_grad=False,
+                )
            return

    def apply(