From a358374ae93785ba8a6df4c5cf427d89003ac28d Mon Sep 17 00:00:00 2001
From: icarus_zh <37328560+zhsurpass@users.noreply.github.com>
Date: Mon, 27 Jul 2026 09:20:28 +0800
Subject: [PATCH] [NPU][Fix Issue]: Send expert weights contiguous tensor
 across cards during EPLB rebalance (#32001)

---
 .../sglang/srt/hardware_backend/npu/quantization/moe_methods.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

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 f08d1af6c..f49d53534 100644
--- a/python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py
+++ b/python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py
@@ -332,7 +332,7 @@ class NPUW8A8Int8MoEMethod(_NPUMoEMethodBase):
 
         # Process weight
         weight: torch.Tensor = getattr(layer, f"{weight_prefix}_weight")
-        weight.data = npu_format_cast(weight.data.transpose(1, 2))
+        weight.data = npu_format_cast(weight.data.transpose(1, 2).contiguous())
 
         # Set dispatcher output dtype
         if weight_prefix == "w13":