From 5a7c1b8ec632a12f8984a3df1ef75c38e2174c13 Mon Sep 17 00:00:00 2001
From: JiaruiChang5268 <71279366+JiaruiChang5268@users.noreply.github.com>
Date: Tue, 10 Mar 2026 21:08:37 +0800
Subject: [PATCH] [NPU] replace swiglu with custom kernel

---
 .../npu/quantization/fused_moe_method_npu.py                | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/hardware_backend/npu/quantization/fused_moe_method_npu.py b/python/sglang/srt/hardware_backend/npu/quantization/fused_moe_method_npu.py
index 924d45d47..f9f3ce4a7 100644
--- a/python/sglang/srt/hardware_backend/npu/quantization/fused_moe_method_npu.py
+++ b/python/sglang/srt/hardware_backend/npu/quantization/fused_moe_method_npu.py
@@ -115,6 +115,8 @@ def npu_fused_experts(
 def npu_fused_moe_without_routing_weights_bf16(
     layer, hidden_states, group_list_type, group_list, output_dtype
 ):
+    from sgl_kernel_npu.activation.swiglu_quant import swiglu_quant
+
     # gmm1: gate_up_proj
     hidden_states = torch.ops.npu.npu_grouped_matmul(
         x=[hidden_states],
@@ -125,7 +127,9 @@ def npu_fused_moe_without_routing_weights_bf16(
         group_list=group_list,
         output_dtype=output_dtype,
     )[0]
-    hidden_states = torch.ops.npu.npu_swiglu(hidden_states)
+    hidden_states, _ = swiglu_quant(
+        hidden_states, group_list, group_list_type, need_quant=False
+    )
     # gmm2: down_proj
     hidden_states = torch.ops.npu.npu_grouped_matmul(
         x=[hidden_states],