From 108182cb8199190a66f14c04ff8f0cc669ea7af2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=D0=90=D1=80=D1=82=D0=B5=D0=BC=20=D0=A1=D0=B0=D0=B2=D0=BA?=
 =?UTF-8?q?=D0=B8=D0=BD?= <58187114+OrangeRedeng@users.noreply.github.com>
Date: Thu, 23 Jul 2026 05:31:32 +0300
Subject: [PATCH] [Bugfix] [NPU] Fix w4a8 MoE performance degradation (#32113)

---
 .../srt/hardware_backend/npu/quantization/moe_methods.py   | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py b/python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py
index aa307b19c..f08d1af6c 100644
--- a/python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py
+++ b/python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py
@@ -428,14 +428,15 @@ class NPUW4A8Int8MoEMethod(_NPUMoEMethodBase):
                     layer,
                     f"{weight_prefix}_scale_bias",
                     torch.nn.Parameter(
-                        bias.data.transpose(1, 2).sum(dim=1).contiguous(),
+                        bias.data.contiguous(),
                         requires_grad=False,
                     ),
                 )
 
         # Process weight
         weight = getattr(layer, f"{weight_prefix}_weight")
-        weight.data = npu_format_cast(weight.data.transpose(1, 2))
+        weight.data = weight.data.transpose(1, 2).contiguous()
+        weight.data = npu_format_cast(weight.data)
         weight.data = self._pack_to_int32(weight.data)
 
         # Set dispatcher output dtype
@@ -491,7 +492,7 @@ class NPUW4A8Int8MoEMethod(_NPUMoEMethodBase):
             f"Last dimension of weight must be divisible by 4 for int8→int32 packing, "
             f"got shape {weight.shape}"
         )
-        return weight.contiguous().view(torch.int32)
+        return weight.view(torch.int32).contiguous()
 
     def apply(
         self,