From 8810152e88899379809f5849f83488e421aa28a5 Mon Sep 17 00:00:00 2001
From: Mick <mickjagger19@icloud.com>
Date: Tue, 9 Dec 2025 04:56:20 +0800
Subject: [PATCH] vlm: Use fa3 as the default backend for qwen3 vl (#14634)

---
 python/sglang/srt/utils/common.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index f6b06ca87..cf7264225 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -315,7 +315,6 @@ try:
 except:
     is_intel_amx_backend_available = False
 
-
 try:
     # move torch._C._cpu._is_amx_tile_supported() from cpu_has_amx_support
     # to support torch compile
@@ -2707,14 +2706,16 @@ def is_fa3_default_architecture(hf_config):
     if not isinstance(architectures, list) or not architectures:
         return False
     default_archs = {
-        "Qwen2ForCausalLM",
         "Llama4ForConditionalGeneration",
         "LlamaForCausalLM",
         "Olmo2ForCausalLM",
         "Gemma2ForCausalLM",
         "Gemma3ForConditionalGeneration",
+        "Qwen2ForCausalLM",
         "Qwen3ForCausalLM",
         "Qwen3MoeForCausalLM",
+        "Qwen3VLForConditionalGeneration",
+        "Qwen3VLMoeForConditionalGeneration",
         "Glm4MoeForCausalLM",
         "Glm4vForConditionalGeneration",
         "Glm4vMoeForConditionalGeneration",