From 3d484be547e016ca0f0dea379fe800292565ab30 Mon Sep 17 00:00:00 2001
From: ratish <114130421+Ratish1@users.noreply.github.com>
Date: Mon, 15 Dec 2025 20:42:26 +0400
Subject: [PATCH] fix(attention): Prevent trtllm_mha auto-selection with eagle3
 speculative decoding (#15127)

---
 python/sglang/srt/server_args.py | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index b777ad3be..4498f4214 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -1493,7 +1493,14 @@ class ServerArgs:
                     and is_fa3_default_architecture(self.model_config.hf_config)
                 ):
                     self.attention_backend = "fa3"
-                elif is_sm100_supported() and is_no_spec_infer_or_topk_one(self):
+                elif (
+                    is_sm100_supported()
+                    and is_no_spec_infer_or_topk_one(self)
+                    and (
+                        self.speculative_algorithm is None
+                        or self.speculative_eagle_topk is not None
+                    )
+                ):
                     self.attention_backend = "trtllm_mha"
                 elif is_hip():
                     self.attention_backend = "aiter"