v2.6.2

What's Changed

• Fix a wrong reference to seqlen_k variable in the varlen kernel by @cakeng in #18
• Fix ima for split-kv kernel by @bfontain in #20

New Contributors

• @cakeng made their first contribution in #18
• @bfontain made their first contribution in #20

Full Changelog: v2.6.1...v2.6.2

v2.6.1

What's Changed

• Adds Python 3.12 to publish.yml by @mgoin in #10
• Sync with FA v2.6.0 to support soft capping by @WoosukKwon in #13
• Support non-default CUDA version by @WoosukKwon in #14
• Bump up to v2.6.1 by @WoosukKwon in #15

New Contributors

• @mgoin made their first contribution in #10

Full Changelog: v2.6.0...v2.6.1

v2.6.0

What's Changed

• Upgrade to torch 2.3.1 by @WoosukKwon in #5
• Upgrade to v2.5.9.post1 by @WoosukKwon in #6
• use global function rather than lambda by @youkaichao in #7
• Update torch to 2.4 by @SageMoore in #8
• Add CUDA 11.8 by @WoosukKwon in #9

New Contributors

• @youkaichao made their first contribution in #7
• @SageMoore made their first contribution in #8

Full Changelog: v2.5.9...v2.6.0

v2.5.9.post1

What's Changed

• Upgrade to torch 2.3.1 by @WoosukKwon in #5
• Upgrade to v2.5.9.post1 by @WoosukKwon in #6

Full Changelog: v2.5.9...v2.5.9.post1

v2.5.9

What's Changed

• Fix out kwarg shape check with ngroups swapped by @Yard1 in #4

Full Changelog: v2.5.8.post3...v2.5.9

v2.5.8.post3

What's Changed

• Expose out in python API by @Yard1 in #2
• Don't use kwargs in autograd functions by @Yard1 in #3

New Contributors

• @Yard1 made their first contribution in #2

Full Changelog: v2.5.8.post2...v2.5.8.post3

v2.5.8.post2

Full Changelog: v2.5.8.post1...v2.5.8.post2

v2.5.8.post1

What's Changed

• Sync up by @WoosukKwon in #1

New Contributors

• @WoosukKwon made their first contribution in #1

Full Changelog: https://github.com/vllm-project/flash-attention/commits/v2.5.8.post1