From f85cc94d82a5a7b55a2fe3ad1cd18dd0b13b1763 Mon Sep 17 00:00:00 2001
From: silencejade <222910030+silencejade@users.noreply.github.com>
Date: Mon, 29 Jun 2026 14:35:19 +0800
Subject: [PATCH] [NPU] Qwen3-VL-30B use split_qkv_rmsnorm_rope for extend
 (#29505)

---
 python/sglang/srt/models/qwen3_moe.py | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/python/sglang/srt/models/qwen3_moe.py b/python/sglang/srt/models/qwen3_moe.py
index a29df9f46..6b749191e 100644
--- a/python/sglang/srt/models/qwen3_moe.py
+++ b/python/sglang/srt/models/qwen3_moe.py
@@ -661,10 +661,7 @@ class Qwen3MoeAttention(nn.Module):
     ):
         if hidden_states.shape[0] == 0:
             return hidden_states, forward_batch, None
-        if (
-            not _is_npu
-            or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed()
-        ):
+        if not _is_npu:
             return self.forward_prepare_native(
                 positions=positions,
                 hidden_states=hidden_states,