Skip to content

Releases: openpsi-project/ReaLHF

v0.3.0

05 Sep 06:44
f54a9f1
Compare
Choose a tag to compare

What's Changed

  • [MAJOR] Refactor configuration and dataflow graph to make it more user-friendly. by @garrett4wade in #25
  • [MAJOR] Add MoE model (Mixtral) support. by @nuzant in #39
  • [MAJOR] Support minibatched execution in interface implementations to save GPU memory by @garrett4wade in #37
  • [MAJOR] Fully support CUDAGraph generation in PPO by @nuzant in #49
  • [New Feature] Support arbitrary symmetric allocations and fix MFC time log in workers by @garrett4wade in #60
  • [New Feature] Add utilities for profiling individual interfaces by @garrett4wade in #56
  • [Docs] v0.3.0 docs about implementation details and the code architecture by @garrett4wade in #57
  • Add a Reference EMA example for the ICML 2024 paper by @garrett4wade in #38
  • Fix mistral model conversion by @garrett4wade in #51

Minor and Bug Fixes by @garrett4wade

  • Fix several system-wide performance issues.
  • Add more configuration validation in experiments.
  • Fix package dependency issues in requirements.txt and Dockerfile.
  • Update the docker image to 24.03-py3, required by CUDAGraph.

Full Changelog: v0.2.0...v0.3.0

v0.2.0

04 Jul 10:14
46d7dc3
Compare
Choose a tag to compare

What's Changed

  • Support more dense models: GPT-2, Gemma, Qwen2, Mistral.

  • Support fast generation with CUDAgraph.

  • Support distributed experiments with Ray.

  • Bug fixes with the Megatron training backend and the C++ extension.

Please check the updated documentation for details.

Full Changelog: https://github.com/openpsi-project/ReaLHF/commits/v0.2.0