Skip to content

fix: default Inkling to fa4 on Hopper #35292

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
ispobock merged 4 commits into from
Oct 11, 2026
Merged

fix: default Inkling to fa4 on Hopper #35292

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
d531d95
fix: boot Inkling on Hopper without an explicit attention backend
alphabetc1 Sep 2, 2026
7932f3d
fix: default Inkling to fa4 on Hopper and key the full prefill graph …
alphabetc1 Oct 11, 2026
9400a26
refactor: inline the Inkling Hopper fa4 default and drop the unit test
alphabetc1 Oct 11, 2026
d94e9b5
refactor: keep the Inkling prefill-graph hook unchanged
alphabetc1 Oct 11, 2026
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
7 changes: 5 additions & 2 deletions python/sglang/srt/arg_groups/model_overrides/inkling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,15 +58,18 @@ def _inkling_overrides(server_args: Any, hf_config: Any) -> dict:
# handle_mamba_radix_cache rejects extra_buffer on a model without it.
if not cfg.disable_radix_cache:
overrides["uses_mamba_radix_cache"] = True
# Inkling attention runs only on the fa4 (Blackwell) or triton backends --
# Inkling attention runs only on the fa4 (Hopper/Blackwell) or triton backends --
# models/inkling_common/attn.py asserts attention_backend in {fa4, triton}.
# The generic resolver would otherwise pick trtllm_mha (SM100) / fa3
# (Hopper), so a bare launch fails on the first attention forward. Pin a
# supported default when the user left every attention-backend flag unset
# (mirrors the MiniMax-M3 SM100 fa4-default above); an explicit
# --attention-backend / --prefill/decode-attention-backend still wins.
if is_attention_backend_not_set(cfg):
inkling_attn_backend = "fa4" if get_platform().is_sm100 else "triton"
if get_platform().is_sm90 or get_platform().is_sm100:
inkling_attn_backend = "fa4"
else:
inkling_attn_backend = "triton"
overrides["attention_backend"] = inkling_attn_backend
logger.info(
f"Use {inkling_attn_backend} as the attention backend for Inkling "
Expand Down
Loading