From 9ffc0cc67ee54520d9d675af3f8e82cbae8ef0a7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=9C=8B=E6=B5=B7=E7=9A=84=E4=BA=BA?=
 <37328560+zhsurpass@users.noreply.github.com>
Date: Tue, 28 Apr 2026 09:08:19 +0800
Subject: [PATCH] [NPU] Support GLM-4.5V (#22961)

---
 python/sglang/srt/models/glm4_moe.py | 22 +++++++++++++++++-----
 1 file changed, 17 insertions(+), 5 deletions(-)

diff --git a/python/sglang/srt/models/glm4_moe.py b/python/sglang/srt/models/glm4_moe.py
index a83a7b36d..155173731 100644
--- a/python/sglang/srt/models/glm4_moe.py
+++ b/python/sglang/srt/models/glm4_moe.py
@@ -314,6 +314,18 @@ class Glm4MoeAttention(nn.Module):
         else:
             if self.attn.layer_id == forward_batch.token_to_kv_pool.start_layer:
                 self.rotary_emb.get_cos_sin_with_position(positions)
+            if self.use_qk_norm:
+                eps = self.q_norm.variance_epsilon
+                q_weight = self.q_norm.weight
+                k_weight = self.k_norm.weight
+                q_bias = getattr(self.q_norm, "bias", None)
+                k_bias = getattr(self.k_norm, "bias", None)
+            else:
+                eps = None
+                q_weight = None
+                k_weight = None
+                q_bias = None
+                k_bias = None
             q, k, v = split_qkv_rmsnorm_rope(
                 qkv,
                 self.rotary_emb.position_sin,
@@ -321,11 +333,11 @@ class Glm4MoeAttention(nn.Module):
                 self.q_size,
                 self.kv_size,
                 self.head_dim,
-                eps=self.q_norm.variance_epsilon,
-                q_weight=self.q_norm.weight,
-                k_weight=self.k_norm.weight,
-                q_bias=getattr(self.q_norm, "bias", None),
-                k_bias=getattr(self.k_norm, "bias", None),
+                eps=eps,
+                q_weight=q_weight,
+                k_weight=k_weight,
+                q_bias=q_bias,
+                k_bias=k_bias,
             )
 
         inner_state = q, k, v, forward_batch