Version 0.4.2 is released on 19th Nov 2024

What's Changed

• ulysses in benchmark by @feifeibear in #104
• flash_attn3 not directly import by @feifeibear in #105
• version to 0.4.2 by @feifeibear in #106

Full Changelog: 0.4.1...0.4.2

0.4.1 is released on Nov 15th 2024

What's Changed

• feat: add use_sync switch to ulysses by @Eigensystem in #103

Full Changelog: 0.4.0...0.4.1

0.4.0 version is released on Nov 15th 2024

The 0.4.0 version starts to support FlashAttention V3 on Hopper GPUs. This version also has been tested on low-memory (24GB) GPUs.

What's Changed

• update all_to_all by @Lay2000 in #87
• upgrade version to 0.3.6 by @feifeibear in #89
• prevent dispatch issues in SequenceParallel to improve performance by @uclalch in #92
• sync after all2all for device memory saving by @feifeibear in #93
• version 0.3.7 by @feifeibear in #94
• flash attention 3 by @feifeibear in #95
• FA3: update to the latest FA3 API. by @feifeibear in #96
• add dit benchmark script by @feifeibear in #97
• 1114v2 by @feifeibear in #98
• update readme for FA3 by @feifeibear in #99
• version to 0.4.0 by @feifeibear in #100
• dump version to 0.4.1 by @feifeibear in #101
• dump to 0.4.0 by @feifeibear in #102

New Contributors

• @Lay2000 made their first contribution in #87
• @uclalch made their first contribution in #92

Full Changelog: 0.3.5...0.4.0

0.3.5

What's Changed

• revert version 0.3.2 by @feifeibear in #83
• version 0.3.5 by @feifeibear in #84

Full Changelog: 0.3.3...0.3.5

0.3.2 released

What's Changed

• remove amd installation to an individual doc by @feifeibear in #76
• auto publish python package when release on github by @feifeibear in #77
• version 0.3.2 by @feifeibear in #78
• remove useless workflow by @feifeibear in #79
• version to 0.3.2 by @feifeibear in #80
• polish publish workflow by @feifeibear in #81

Full Changelog: v0.3.1...0.3.2

v0.3.1 released at 2024.09.14

stripe_extract_local, basic_extract_local, zigzag_extract_local works for tensors dimension >=2.

v0.3 released on 27th August 2024!

upgrade flash_attn >= 2.6.0

v0.2 released on 24th June 2024!

1. Ulysses supports T4 and V100.
2. Updates some directory structures.

v0.1

Sequence parallel attention adopting a hybrid ulysses and ring attention approach.
Support GQA
Support QKV packed.