Skip to content

Fix attention backend logic for Qwen3-Next on SM100 #14560

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Fridge003 merged 5 commits into from
Dec 7, 2025
Merged

Fix attention backend logic for Qwen3-Next on SM100 #14560

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
734a417
init
Chen-0210 Dec 7, 2025
e09d7fc
Apply suggestion from @gemini-code-assist[bot]
Chen-0210 Dec 7, 2025
9733d8d
fix
Chen-0210 Dec 7, 2025
a179445
init
Chen-0210 Dec 7, 2025
9214417
Merge branch 'main' into fix-atten-backend
b8zhong Dec 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 15 additions & 0 deletions python/sglang/srt/server_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1282,6 +1282,21 @@ def _handle_model_specific_adjustments(self):
logger.info(
"Use flashinfer_trtllm as MoE runner backend on sm100 for Qwen3NextForCausalLM"
)
if self.attention_backend is None:
self.attention_backend = "triton"
logger.info(
"Use triton as attention backend on sm100 for Qwen3NextForCausalLM"
)
if (
not self.disable_radix_cache
and self.attention_backend == "trtllm_mha"
):
logger.warning(
"Disabling radix cache since trtllm_mha does not support page_size = 1, which is required by MambaRadixCache. "
"Try to use --attention-backend triton if radix cache is necessary."
)
self.disable_radix_cache = True
self.disable_overlap_schedule = False
elif model_arch in [
"NemotronHForCausalLM",
"FalconH1ForCausalLM",
Expand Down
Loading