minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix(attention): Prevent trtllm_mha auto-selection with eagle3 speculative decoding (#15127)

Browse Source
This commit is contained in:
ratish
2025-12-15 16:42:26 +00:00
committed by GitHub
parent c0d94440b7
commit 3d484be547
1 changed files with 8 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+8 -1
View File
@@ -1493,7 +1493,14 @@ class ServerArgs:
and is_fa3_default_architecture(self.model_config.hf_config) and is_fa3_default_architecture(self.model_config.hf_config)
): ):
self.attention_backend = "fa3" 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" self.attention_backend = "trtllm_mha"
elif is_hip(): elif is_hip():
self.attention_backend = "aiter" self.attention_backend = "aiter"