mirror of https://github.com/tile-ai/tilelang.git synced 2026-10-22 06:50:25 +08:00

Files

T

Tong WUandcoderabbitai[bot] 3ab93cd76b [Enhancement] Keep max score attention across blocks in FlashAttention for better numerical stablity (#1269 )

* Implement max score retention across blocks in FlashAttention for improved stability

* fix manual pipeline parameters

* Update examples/flash_attention/example_gqa_fwd_varlen.py

Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

* fix typo

* more

* fix a previous typo

---------

Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

2025-11-17 21:53:19 +08:00

ops

[Example] Add vertical slash sparse attention pattern (#762 )

2025-08-28 14:34:17 +08:00

example_vertical_slash_sparse_attn.py

[Enhancement] Keep max score attention across blocks in FlashAttention for better numerical stablity (#1269 )

2025-11-17 21:53:19 +08:00

README.md

[Example] Add vertical slash sparse attention pattern (#762 )

2025-08-28 14:34:17 +08:00

test_vs_sparse_attn.py

[CI][Refactor] Merge test CI workflow files into one (#973 )

2025-10-15 22:12:41 +08:00

README.md

Performance Benchmark

Hardware & Environment

Hardware: NVIDIA H100 PCIe
CUDA version: 12.8.1
PyTorch Version: 2.7.1+cu128
Triton Version: 3.3.1

Performance Results

BATCH_SIZE=1, HEAD=1, DIM=64

SEQ_LEN	VS_LIST	Triton Time	TileLang Time	Speedup
8192	[1000, 200]	0.168 ms	0.105 ms	1.60x
8192	[1000, 600]	0.207 ms	0.119 ms	1.74x
8192	[800, 600]	0.207 ms	0.122 ms	1.70x

16384	[1000, 200]	0.261 ms	0.167 ms	1.56x
16384	[1000, 600]	0.419 ms	0.258 ms	1.62x
16384	[800, 600]	0.422 ms	0.255 ms	1.65x

32768	[1000, 200]	0.374 ms	0.248 ms	1.51x
32768	[1000, 600]	0.823 ms	0.554 ms	1.49x
32768	[800, 600]	0.826 ms	0.558 ms	1.48x

65536	[1000, 200]	0.637 ms	0.524 ms	1.22x
65536	[1000, 600]	1.758 ms	1.501 ms	1.17x
65536	[800, 600]	1.783 ms	1.489 ms	1.20x