Liger-Kernel
Liger-Kernel copied to clipboard
2024 Q4 Roadmap
trafficstars
As the community grows, keeping track of issues and PRs becomes more and more challenging. This pinned issue will serve as the central place to manage the progress in 2024 Q4 (~2024/12). Here we only list the important and top level issues/PRs.
Progress Tracker
Format: [title] [link], [contributor] / [reviewer (maintainer)]
New Features: Liger to Post Training
https://github.com/linkedin/Liger-Kernel/issues/371
Models
- llama 3.2 vision (https://github.com/linkedin/Liger-Kernel/pull/282), @tyler-romero / @shivam15s
- pixtral 12B (https://github.com/linkedin/Liger-Kernel/pull/253), @AndreSlavescu / @ByronHsu
- Flux model (https://github.com/linkedin/Liger-Kernel/issues/73), call for help / @qingquansong
- DeepseekV2 (https://github.com/linkedin/Liger-Kernel/issues/129), call for help / @qingquansong
- Gemma2 (https://github.com/linkedin/Liger-Kernel/issues/127), call for help / @yundai424
- Qwen2-VL (https://github.com/linkedin/Liger-Kernel/pull/283), @tyler-romero / @shivam15s
- Jamba (https://github.com/linkedin/Liger-Kernel/pull/214), @yubofredwang / @ByronHsu
Kernels
- TVD loss, https://github.com/linkedin/Liger-Kernel/pull/324 / @qingquansong @lancerts
- JSD loss, https://github.com/linkedin/Liger-Kernel/pull/300 / @qingquansong @lancerts
- GroupNorm (https://github.com/linkedin/Liger-Kernel/pull/225), @denti / @shivam15s
- Z Loss in cross entropy (https://github.com/linkedin/Liger-Kernel/pull/239), @Tcc0403 / @shivam15s
- Flash Attention in Triton (https://github.com/linkedin/Liger-Kernel/pull/275), @remi-or / @shivam15s
- Conv2d (https://github.com/linkedin/Liger-Kernel/pull/228), @AndreSlavescu / @lancerts
- Triton mm int8 x int2(https://github.com/linkedin/Liger-Kernel/pull/195), @mekkcyber / @ByronHsu
Testing
assert_verbose_allclose(https://github.com/linkedin/Liger-Kernel/pull/261), @Tcc0403 / @ByronHsu
Patching
- Weights are not copied for instance patching (https://github.com/linkedin/Liger-Kernel/pull/280), @shimizust / @ByronHsu
Community Sync
TBD