[TRTLLM-9805][feat] Skip Softmax Attention. - #9821

Merged

bobboli merged 32 commits into

NVIDIA:mainfrom

bobboli:feat/skip_softmax_attention

Dec 21, 2025

Commits on Dec 17, 2025

Initial commit.

bobboli
authored and
Tom-Zheng
committed
Address comments.

bobboli
authored and
Tom-Zheng
committed
Delete sm70 and sm72 in fmha_v2 makefile.

bobboli
authored and
Tom-Zheng
committed
Delete sm70 and sm72 in fmha_v2 makefile.

bobboli
authored and
Tom-Zheng
committed
No need for manually adding stop tokens.

bobboli
authored and
Tom-Zheng
committed
Delete longbench_V2 in examples.

bobboli
authored and
Tom-Zheng
committed
fix supports_backend

bobboli
authored and
Tom-Zheng
committed
Adjust attention OP args

bobboli
authored and
Tom-Zheng
committed
Run pre-commit -a.

bobboli
authored and
Tom-Zheng
committed
Small fix.

bobboli
authored and
Tom-Zheng
committed
dynamic threshold of skip-softmax

bobboli
authored and
Tom-Zheng
committed
Wrap skip &= with may_skip.

bobboli
authored and
Tom-Zheng
committed
Support ctx-len aware skip-softmax attention.

bobboli
authored and
Tom-Zheng
committed
Support ctx-len aware skip-softmax attention.

bobboli
authored and
Tom-Zheng
committed
[feat] skip_softmax attention for trtllm-gen based attention on Blackwell
Tom-Zheng
committed
Skip softmax as a compile-time feature of fmha_v2.

bobboli
authored and
Tom-Zheng
committed
pre-commit -a

bobboli
authored and
Tom-Zheng
committed
Fix compile error.

bobboli
authored and
Tom-Zheng
committed
Separate prefill and decode threshold.

bobboli
authored and
Tom-Zheng
committed
Adjust LongBenchV2 samplingparams.

bobboli
authored and
Tom-Zheng
committed
Add temperature and top_p as cmdline args for LongBenchV2.

bobboli
authored and
Tom-Zheng
committed
Fix compile error.

bobboli
authored and
Tom-Zheng
committed
Disable skip_softmax for test_fmha.

bobboli
authored and
Tom-Zheng
committed
Fix compile error.

bobboli
authored and
Tom-Zheng
committed
Update cubins.

bobboli
authored and
Tom-Zheng
committed
Should not modify setup.py.

bobboli
authored and
Tom-Zheng
committed
update cubins
Tom-Zheng
committed
Add timeout for test_fmha.
bobboli
committed
Waive test_qwen3_eagle3.
bobboli
committed
Merge branch 'main' into feat/skip_softmax_attention
bobboli
authored

Commits on Dec 19, 2025

Merge branch 'main' into feat/skip_softmax_attention
EmmaQiaoCh
authored

Commits on Dec 20, 2025

Merge branch 'main' into feat/skip_softmax_attention
bobboli
authored