From b532a5fd16d0c2b0d7945bffdc3beab1d7018975 Mon Sep 17 00:00:00 2001
From: xiaobochen <35516720+xiaobochen123@users.noreply.github.com>
Date: Mon, 16 Dec 2024 20:54:02 +0800
Subject: [PATCH] fix moe-ep accuracy issue for fp8 (#2489)

---
 python/sglang/srt/layers/ep_moe/layer.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/python/sglang/srt/layers/ep_moe/layer.py b/python/sglang/srt/layers/ep_moe/layer.py
index eca119845a7..3c477fdc2ef 100644
--- a/python/sglang/srt/layers/ep_moe/layer.py
+++ b/python/sglang/srt/layers/ep_moe/layer.py
@@ -644,6 +644,10 @@ def process_weights_after_loading(self, layer: Module) -> None:
                         "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(