From ff51acd67b228bae35d73ce8d36a8f44c8cf9697 Mon Sep 17 00:00:00 2001
From: silencejade <222910030+silencejade@users.noreply.github.com>
Date: Tue, 30 Jun 2026 15:55:50 +0800
Subject: [PATCH] [NPU] Qwen3-VL-8B use split_qkv_rmsnorm_rope for extend
 (#29627)

---
 python/sglang/srt/models/qwen3.py | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/python/sglang/srt/models/qwen3.py b/python/sglang/srt/models/qwen3.py
index ba6652915..dfa440621 100644
--- a/python/sglang/srt/models/qwen3.py
+++ b/python/sglang/srt/models/qwen3.py
@@ -287,10 +287,7 @@ class Qwen3Attention(nn.Module):
                 positions, hidden_states, forward_batch
             )
             save_kv_cache = False
-        elif (
-            not _is_npu
-            or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed()
-        ):
+        elif not _is_npu:
             q, k, v = self.forward_prepare_native(
                 positions=positions,
                 hidden_states=hidden_states,