All pull requests

Pull requests

• Add experimental Jeva trainer

  #8040·kashif opened Oct 30, 2026Collaborator·4 of 8 tasks·

  ·
  1

• Bump deepspeed to >=0.19.8

  🔧 maintenance
  #8039·kashif opened Oct 30, 2026Collaborator·2 of 8 tasks·

  ·
  3

• Stop the async trainers from hanging when the rollout worker dies

  #8038·sergiopaniego opened Oct 30, 2026Member·3 of 8 tasks·

  ·
  4

  

• Narrow the Mistral Large 4 FP8 xfail to triton 3.9

  🔧 maintenance
  #8037·albertvillanova opened Oct 30, 2026Member·

  ·
  1

• Run the OpenEnv harness tests in the experimental CI workflow

  🔧 maintenance
  #8036·albertvillanova opened Oct 30, 2026Member·1 task·

  ·
  1

• Document the openenv>=0.7.0 requirement for training on harnesses

  📚 documentation
  #8035·albertvillanova opened Oct 30, 2026Member·

  ·
  1

• Pin the AsyncGRPO Harbor example to the OpenEnv 0.7.0 release

  #8034·albertvillanova opened Oct 30, 2026Member·

  ·
  1

• Parse GLM-4-MoE completions ending with a stop token with the legacy response schema

  #8032·albertvillanova opened Oct 30, 2026Member·2/2·

  ·
  1

  1

• Drop GLM-4-MoE's <|observation|> stop token from parsed responses

  #8030·albertvillanova opened Oct 30, 2026Member·1/2·

  ·
  1

  1

• Allow dropped group ids in the rollout buffer in the OpenEnv session failure test

  🔧 maintenance
  #8029·albertvillanova opened Oct 30, 2026Member·

  ·
  1

  1

• Recognize TRL training chat templates for response parsing

  #8027·albertvillanova opened Oct 30, 2026Member·

  ·
  1

  3

• Refuse tree packing for recurrent models without layer_types

  #8026·AmineDiro opened Oct 29, 2026Member·

  ·

• Support the standard dataset format in the built-in reward functions

  #8025·RerankerGuo opened Oct 29, 2026Contributor·5 of 8 tasks·

  ·
  1

• Drop idle connections before vLLM does in the async rollout workers

  #8016·sergiopaniego opened Oct 29, 2026Member·2 of 8 tasks·

  ·
  1

• Fail fast when an AsyncDistillation teacher has no room to score a sequence

  #8015·sergiopaniego opened Oct 29, 2026Member·4 of 8 tasks·

  ·
  5

• Use the released Open-MOPD experts in the MOPD example

  #8014·sergiopaniego opened Oct 29, 2026Member·3 of 8 tasks·

  ·
  4

• Bump doc-builder to get the docs machine-translated

  🔧 maintenance
  #8013·mishig25 opened Oct 29, 2026Contributor·

  ·
  1

• Fail CI test jobs on self-hosted runners when pytest does not complete

  🔧 maintenance
  #8006·albertvillanova opened Oct 29, 2026Member·

  ·
  1

  1

• Fix IW-OPD skipping the vLLM weight sync after resuming from a checkpoint

  #7994·sahil-mangla opened Oct 28, 2026·4 of 8 tasks·

  ·
  1

  1

• feat / fix: support get_kernels kwargs to support architectures such as Qwen3.8

  #7987·younesbelkada opened Oct 28, 2026Collaborator·

  ·
  1

• Support assistant-only loss for the DeepSeek-V3 and DeepSeek-V3-0324 chat templates

  #7978·albertvillanova opened Oct 28, 2026Member·2/2·

  ·
  1

• Correct the docs of the GLM-4-MoE and Qwen3 training templates

  #7975·albertvillanova opened Oct 28, 2026Member·

  ·
  1

• Recognize the original Qwen3 chat template

  #7974·albertvillanova opened Oct 28, 2026Member·

  ·
  1

• Parse Qwen2.5 base tool calls

  #7972·albertvillanova opened Oct 28, 2026Member·2/2·

  ·
  1

• Support assistant-only loss for the Qwen2.5 base chat template

  #7971·albertvillanova opened Oct 28, 2026Member·1/2·

  ·
  1