All pull requests

Pull requests

• Seed before creating the model in experimental trainers

  #7401·albertvillanova opened Sep 25, 2026Member·

  ·
  1

• Align tiny LFM2 config with LiquidAI/LFM2-1.2B

  🔧 maintenance
  #7400·albertvillanova opened Sep 25, 2026Member·

  ·
  2

• Align tiny DeepSeek-V3-0528 config with deepseek-ai/DeepSeek-R1-0528

  🔧 maintenance
  #7399·albertvillanova opened Sep 25, 2026Member·

  ·
  3

• Align tiny DeepSeek-V3 config with deepseek-ai/DeepSeek-R1

  🔧 maintenance
  #7398·albertvillanova opened Sep 25, 2026Member·

  ·
  3

• Remove the nightly cache cleanup workflow

  🔧 maintenance
  #7395·albertvillanova opened Sep 25, 2026Member·

  ·
  1

• Preserve explicit model_init_kwargs in CLI scripts

  #7394·DaoyuanLi2816 opened Sep 25, 2026Contributor·

  ·

• Honor resume_from_checkpoint in CLI scripts

  #7391·DaoyuanLi2816 opened Sep 25, 2026Contributor·

  ·

• Score DPO and KTO tokens with the fused LM head

  #7390·qgallouedec opened Sep 24, 2026Member·4/4·

  ·
  7

• Score GRPO and RLOO tokens with the fused LM head

  #7389·qgallouedec opened Sep 24, 2026Member·3/4·

  ·
  5

• Add an opt-in fused LM head to the patched forward

  #7387·qgallouedec opened Sep 24, 2026Member·2/4·

  ·
  17

• Call trainer.end() at the end of the scripts and examples

  #7384·qgallouedec opened Sep 24, 2026Member·

  ·
  2

• FIX: test_offloading_with_peft_models on L40. Register model params in activation offloading tests

  🔧 maintenance
  #7380·AmineDiro opened Sep 24, 2026Member·

  ·
  1

• Test the vLLM client against the documented vllm serve command

  🔧 maintenance
  #7378·albertvillanova opened Sep 24, 2026Member·

  ·
  1

• Support assistant-only loss for vision datasets in SFT

  ✨ enhancement
  #7377·albertvillanova opened Sep 24, 2026Member·

  ·
  1

  1

• Keep the next turn's first token out of the Llava-Next assistant mask

  🐛 bug
  #7375·albertvillanova opened Sep 24, 2026Member·

  ·
  1

  1

• [OpenReward] Return None for unrewarded rollouts (fixes #7364)

  #7370·tanvir-ux opened Sep 24, 2026·

  ·
  1

• Avoid device_map="auto" default on MPS in create_model_from_path

  #7368·SammyTourani opened Sep 24, 2026·

  ·
  1

• [WIP] Direction for optimized trainers

  #7367·qgallouedec opened Sep 24, 2026Member·

  ·
  10

• Discover environment tools on the class, not the instance

  #7360·Rome-1 opened Sep 23, 2026·

  ·

• Stream GRPO log-probs by default

  #7358·kashif opened Sep 23, 2026Collaborator·

  ·
  1

• Correct the rollout_func docstring about repeated prompts

  #7357·kuotunyu opened Sep 23, 2026·

  ·
  1

• Add KLPO trainer (KL-Regularized Policy Optimization)

  #7356·mmjerge opened Sep 23, 2026·

  ·
  3

• Preserve dataset_text_field for vision SFT datasets

  #7355·HelloHaoWu opened Sep 23, 2026·

  ·
  1

  3

• Keep the DoRA magnitude vector in float32 under QLoRA

  #7342·qgallouedec opened Sep 23, 2026Member·

  ·
  1

  4

• Preserve explicit image placeholders in mixed message content

  #7330·DaoyuanLi2816 opened Sep 22, 2026Contributor·

  ·