All pull requests

Pull requests

• fix: add missing imports in dpo.py

  CLA Signed
  #2998·mayuriphad opened Sep 17, 2026·

  ·
  1

• fix(config): raise informative InstantiationError on unexpected kwargs (#2152)

  CLA Signed
  #2996·somuai opened Sep 15, 2026·

  ·

• Fix eleuther_eval issues: GPT-2 fetch, VLM truncation, hardcoded max_gen_toks (#2990)

  CLA Signed
  #2995·mayuriphad opened Sep 14, 2026·

  ·
  1

  1

• docs: add Simplified Chinese translation (README.zh.md)

  CLA Signed
  #2994·JasonYeYuhe opened Sep 13, 2026·

  ·
  1

• Add add_end_token to Gemma Tokenizer

  CLA Signed
  #2993·mayuriphad opened Sep 9, 2026·

  ·
  1

  1

• fix(eval): support configurable and harness-requested generation budget in eleuther_eval

  CLA Signed
  #2992·somuai opened Sep 7, 2026·

  ·
  1

  3

• Fix multimodal evaluation output and generation length

  #2991·feng1201 opened Sep 6, 2026·

  ·
  1

• fix: clarify DistributedCheckpointer recipe configuration

  #2988·cananoo opened Aug 31, 2026·

  ·
  1

  1

• Add end token option to Mistral tokenizer

  CLA Signed
  #2987·Marchematics opened Aug 30, 2026·

  ·
  4

• [Draft] Add reference-free SimPO loss

  CLA Signed
  #2986·marcusacosta opened Aug 26, 2026·

  ·

• Validate unsupported LoRA attention module names

  CLA Signed
  #2985·marcusacosta opened Aug 25, 2026·

  ·
  2

• docs: fix broken 7B_full config link in recipe deep-dive

  CLA Signed
  #2983·richboyneedcash opened Aug 5, 2026·

  ·
  3

• fix(dev): support current torchao and torchrl releases

  CLA Signed
  #2982·n-dlms opened Aug 3, 2026·

  ·

• fix(grpo): replay buffer contract corruptions advantages (#2943)

  CLA Signed
  #2981·n-dlms opened Aug 2, 2026·

  ·
  1

• perf(recipes): skip scale_grads when gradient_accumulation_steps == 1

  CLA Signed
  #2980·n-dlms opened Jul 30, 2026·

  ·
  1

• feat(config): warn on CLI overrides introducing unknown top-level keys

  CLA Signed
  #2979·n-dlms opened Jul 30, 2026·

  ·
  1

• docs(readme): switch optimization flags table to Linear Cross Entropy

  CLA Signed
  #2978·n-dlms opened Jul 30, 2026·

  ·
  1

• Fix Mistral and Gemma end-token control

  CLA Signed
  #2977·patrickswedish opened Jul 26, 2026·

  ·
  1

  5

• feat: add optional torchembed fused Triton kernel for RotaryPositionalEmbeddings (8–23× faster)

  CLA Signed
  #2970·py-ai-dev opened Jul 12, 2026·

  ·

• fix: respect Gemma tokenizer inference mode.

  CLA Signed
  #2969·ofeliacode opened Jul 9, 2026·

  ·
  1

  6

• Fix NameError in torchtune rlhf/loss/dpo.py from missing imports

  CLA Signedmeta-exported
  #2966·wjh70301-meta opened Jun 17, 2026·

  ·
  1

• Fix missing imports in rlhf/loss/dpo.py

  CLA Signedfb-exportedmeta-exported
  #2965·tmacd-meta opened May 27, 2026·

  ·
  1

• fix(logging): use RANK env var in log_rank_zero before dist init

  CLA Signed
  #2964·kratos0718 opened May 22, 2026·

  ·
  1

  4

• Add TrackioLogger metric backend

  #2963·abidlabs opened Apr 27, 2026·

  ·
  1

• fix: restrict config logging to rank zero to prevent spam

  CLA Signed
  #2950·fatih-uzlmz opened Feb 14, 2026·

  ·