From 05de73efd1ab6e7fa8218e51471bc0f674c22d36 Mon Sep 17 00:00:00 2001
From: Lucia Fang <116399278+luccafong@users.noreply.github.com>
Date: Fri, 1 May 2026 15:49:30 -0700
Subject: [PATCH] [core/model] Use explicit model arch for Llama4 attention
 backend auto-selection (#24232)

---
 python/sglang/srt/server_args.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 8e30444ea..6b8ac5277 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -81,6 +81,10 @@ MIMO_V2_MODEL_ARCHS = (
     "MiMoV2ForCausalLM",
     "MiMoV2FlashForCausalLM",
 )
+LLAMA4_MODEL_ARCHS = (
+    "Llama4ForConditionalGeneration",
+    "Llama4ForCausalLM",
+)
 
 SAMPLING_BACKEND_CHOICES = {"flashinfer", "pytorch", "ascend"}
 
@@ -2034,7 +2038,7 @@ class ServerArgs:
                 logger.warning(
                     "Disable hybrid SWA memory for Step3p5ForCausalLM model with hierarchical cache"
                 )
-        elif "Llama4" in model_arch and self.device != "cpu":
+        elif model_arch in LLAMA4_MODEL_ARCHS and self.device != "cpu":
             # Auto-select attention backend for Llama4 if not specified
             if self.attention_backend is None:
                 if is_sm100_supported():