All pull requests

Pull requests

• AsyncGRPO: keep filling a micro-batch when another row has free room

  #7793·clementtisseau opened Oct 19, 2026·5 of 8 tasks·

  ·
  1

• Add a decision model example with an SFTTrainer subclass

  #7784·rycerzes opened Oct 18, 2026Contributor·4 of 8 tasks·

  ·
  1

  2

• feat: log raw importance ratios and fraction of truncation/masking in vLLM importance sampling correction

  #7778·muupan opened Oct 17, 2026Contributor·4 of 8 tasks·

  ·
  1

  3

• Load the tokenizer for models that take no images

  #7774·qgallouedec opened Oct 16, 2026Member·

  ·
  5

• Warn when LoRA skips the dense and shared expert layers of MoE models on transformers v5 with peft 0.19 to 0.21.2

  #7771·albertvillanova opened Oct 16, 2026Member·

  ·
  1

  3

• Raise when LoRA would skip the routed experts of MoE models on transformers v5 with peft<0.19

  #7769·albertvillanova opened Oct 16, 2026Member·

  ·
  1

  5

• Test that LoRA on an MoE model also trains its dense and shared expert layers

  🔧 maintenance
  #7768·albertvillanova opened Oct 16, 2026Member·

  ·
  2

• Test Qwen3.5 loaded as a text-only causal LM

  🔧 maintenance
  #7763·qgallouedec opened Oct 16, 2026Member·

  ·
  7

• Allow FlashAttention 2 in Async GRPO tree attention

  #7762·DimensionSTP opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·
  4

• Asyncgrpo score centering

  #7760·AmineDiro opened Oct 15, 2026Member·

  ·
  10

• Fix empty-target training in SDPO and SDFT

  #7758·DimensionSTP opened Oct 15, 2026Contributor·3 of 8 tasks·

  ·
  1

• Add an LFM2.5-8B-A1B research agent example (SFT + GRPO)

  #7754·qgallouedec opened Oct 15, 2026Member·

  ·
  5

• Require torch>=2.7.0

  #7752·albertvillanova opened Oct 15, 2026Member·

  ·
  2

• Fix the default IS mode in the GRPOConfig docstring

  #7751·sergiopaniego opened Oct 15, 2026Member·4 of 8 tasks·

  ·
  3

• Example / Experimental: Single GPU LoRA + single harness

  #7749·younesbelkada opened Oct 15, 2026Contributor·

  ·
  32

• Prevent VLM policies in Online DPO from generating the image placeholder token

  #7708·albertvillanova opened Oct 13, 2026Member·

  ·
  1

• Shuffle the KTO training data once so batches mix labels

  #7707·hpdkhoa opened Oct 13, 2026·6 of 8 tasks·

  ·
  1

  1

• Prevent VLM policies in GOLD from generating the image placeholder token

  #7699·albertvillanova opened Oct 13, 2026Member·

  ·
  1

• Group iterable datasets per process when Accelerate gives each process its own shards

  #7697·Strongich opened Oct 13, 2026Contributor·

  ·
  1

• Tree attention support for Linear attention models through FLA

  #7686·AmineDiro opened Oct 13, 2026Member·

  ·
  1

• SDFT: add DOPD advantage-gap token routing

  #7656·ucalyptus opened Oct 12, 2026Contributor·

  ·

• SDFT/SDPO: distillation plumbing fixes

  #7655·ucalyptus opened Oct 12, 2026Contributor·

  ·
  9

• Fix cross-rank metric averaging in the experimental self-distillation trainers

  #7654·ucalyptus opened Oct 12, 2026Contributor·

  ·
  5

• Fix: DPOTrainer crashes on the first step when use_weighting=True is combined with a loss_type containing "sft"

  #7651·Sergasgr opened Oct 12, 2026·5 of 8 tasks·

  ·
  1

• Filter zero-advantage groups in AsyncGRPO

  #7609·AmineDiro opened Oct 9, 2026Member·

  ·
  9