Skip to content

Milestones

List view

  • Support more customized models; Support and optimize more automatic parallelism for HF model usage.

    Due by December 31, 2025
    1/2 issues closed
  • Verify training convergence at various tasks and settings; Match or enhance the convergence compared to other frameworks; Start from on-policy mode to async mode; New techniques to enhance training convergence for async mode.

    Due by December 31, 2025
    2/3 issues closed
  • Analyze memory consumption saving under various optimization techniques; Compare memory consumption with other frameworks like Verl; More memory optimization techniques; Enable efficient long token training.

    Due by December 31, 2025
    1/3 issues closed