All pull requests

Pull requests

• FIX: make gemma4 prefix-tuning warning describe why layers are skipped

  triaged
  #3873·peft-jambot opened Oct 2, 2026Contributor·

  ·

• CI Fix several nightly GPU run errors

  triaged
  #3870·BenjaminBossan opened Oct 2, 2026Member·

  ·
  1

• FIX LoHa wrong gradients with use_effective_conv2d=True

  triaged
  #3869·Sergasgr opened Oct 2, 2026Contributor·

  ·
  1

• FIX Restore inference mode when saving adapter config fails

  triaged
  #3866·DaoyuanLi2816 opened Oct 2, 2026Contributor·

  ·
  1

  7

• FIX Validate the requested adapter configuration during hotswap

  triaged
  #3865·DaoyuanLi2816 opened Oct 2, 2026Contributor·

  ·
  1

• FIX Preserve trainability when loading mixed adapters

  triaged
  #3862·aswanth-07 opened Oct 1, 2026Contributor·

  ·
  1

• FIX Avoid making ranknum in AdaLoRA trainable by accident

  triaged
  #3859·BenjaminBossan opened Oct 1, 2026Member·

  ·
  1

• CI: Bump the third-party-actions group with 6 updates

  dependenciesgithub_actionstriaged
  #3853·dependabot[bot] opened Oct 1, 2026Bot·

  ·

• Fix Trainable Tokens adapters with disjoint target modules

  triaged
  #3848·DaoyuanLi2816 opened Sep 30, 2026Contributor·

  ·
  1

  7

• aLoRA: speed up forward pass

  triaged
  #3847·githubnemo opened Sep 29, 2026Collaborator·

  ·
  3

• FIX add_weighted_adapter for lora_bias, DoRA and KaSA

  triaged
  #3844·LK-maker-007 opened Sep 29, 2026Contributor·

  ·
  1

• FIX add_weighted_adapter with rank_pattern for element-wise combinations

  triaged
  #3841·amogh-nagri-11 opened Sep 29, 2026Contributor·

  ·
  1

  4

• FIX Adapter shape for FSDP2 layers on meshes not named "fsdp"

  triaged
  #3838·hivaze opened Sep 28, 2026Contributor·

  ·
  1

  23

• FIX Several fixes to Conv2d layers, extend test coverage

  triaged
  #3837·BenjaminBossan opened Sep 28, 2026Member·

  ·
  2

  1

• Add state dict regression cases for overlapping rank/alpha pattern keys

  triaged
  #3834·jayzuccarelli opened Sep 28, 2026·

  ·
  1

  

• FIX strict_adapter_check for load_adapter and from_pretrained

  triaged
  #3832·rupeshpoojary9 opened Sep 28, 2026Contributor·

  ·
  1

  14

• Keep PeftModel state consistent after deleting the last adapter

  triaged
  #3826·simpleqt opened Sep 27, 2026Contributor·

  ·
  6

  

• FIX modules_to_save must match whole module names

  triaged
  #3824·AHSharan opened Sep 26, 2026Contributor·

  ·
  1

  7

• FIX Clear error when PeftModel.set_adapter gets multiple adapter names

  triaged
  #3822·stanleys12 opened Sep 26, 2026·

  ·
  1

• FIX all-linear skipping dense and shared-expert MoE layers

  triaged
  #3819·rupeshpoojary9 opened Sep 26, 2026Contributor·

  ·
  1

  2

  

• save_and_load: fix false-positive warning with LoRA bias

  triaged
  #3817·githubnemo opened Sep 25, 2026Collaborator·

  ·
  7

• FIX Respect adapter module training state in explicit inference mode

  triaged
  #3809·eSVeeF opened Sep 23, 2026Contributor·

  ·
  1

  30

• FEAT Add a 'healthcheck' method to PEFT models

  triaged
  #3798·BenjaminBossan opened Sep 23, 2026Member·

  ·
  25

• FIX safe_merge leaving the base layer modified, and merging several adapters at once on torchao

  triaged
  #3779·yupengtang opened Sep 21, 2026·

  ·
  2

  32

  

• FIX MultiheadAttention LoRA merge() applies out_proj delta twice

  triaged
  #3774·Vedant-Agarwal opened Sep 20, 2026Contributor·

  ·
  29