Issues

• CPT loss uses the default adapter config instead of the active adapter config

  huggingface/peft#3879 · Filter by author DaoyuanLi2816 opened on Oct 4, 2026

• 4-bit unmerge_adapter does not restore the base weight

  huggingface/peft#3878 · Filter by author chrikrah opened on Oct 3, 2026

• Add support for BoRA

  huggingface/peft#3877 · Filter by author uis246 opened on Oct 3, 2026

• get_peft_model_state_dict returns an empty dict for FSDP1-wrapped models since 0.21.0 (adapter checkpoints / weight syncs silently empty)

  huggingface/peft#3876 · Filter by author murphy977 opened on Oct 3, 2026
  1

• MoE recipe in the LoRA docs gives adapters that vLLM serves with the wrong expert scaling

  huggingface/peft#3874 · Filter by author surajsharan opened on Oct 2, 2026

• _reshard_fsdp_modules skips the whole FSDP unit if any adapter inside it is merged, not just the merged one

  huggingface/peft#3872 · Filter by author rupeshpoojary9 opened on Oct 2, 2026
  2

• Astra preprocessing scales activations by max value, not max abs value, can overflow fp16 covariance

  huggingface/peft#3871 · Filter by author rupeshpoojary9 opened on Oct 2, 2026
  5

• New method proposal: PaCA (Partial Connection Adaptation, ICLR 2025)

  huggingface/peft#3868 · Filter by author WooSunghyeon opened on Oct 2, 2026

• DoRA hotswap raises a bare KeyError for compatible adapters

  huggingface/peft#3867 · Filter by author DaoyuanLi2816 opened on Oct 2, 2026

• LoHa with use_effective_conv2d=True computes wrong gradients for hada_w1_a and hada_w2_a

  huggingface/peft#3863 · Filter by author Sergasgr opened on Oct 1, 2026
  1

• DeFT / TinyLoRA: adding a second adapter to a Conv1D layer breaks the layer (TinyLoRA forward crashes, DeFT merge crashes)

  huggingface/peft#3861 · Filter by author Zhuoxi2000 opened on Oct 1, 2026
  1

• PeftModelForQuestionAnswering silently drops token_type_ids and position_ids

  huggingface/peft#3857 · Filter by author Sergasgr opened on Oct 1, 2026
  1

• Failed config save leaves live adapter in inference mode

  huggingface/peft#3854 · Filter by author DaoyuanLi2816 opened on Oct 1, 2026
  1

• Trainable Tokens adapters with different target modules raise KeyError on forward and merge

  huggingface/peft#3842 · Filter by author DaoyuanLi2816 opened on Sep 29, 2026
  1

  

• delete_adapter skips adapter layers whose module name contains the tuner prefix (e.g. lora_foobar)

  huggingface/peft#3835 · Filter by author JoeyTan21 opened on Sep 28, 2026
  1

• OFT module_dropout has had no effect since #2575: the MultiplicativeDropoutLayer is created but never called

  huggingface/peft#3830 · Filter by author Nicholas022400701 opened on Sep 27, 2026
  1

  

• OFT Conv2d forward is wrong for kernels larger than 1x1: identity init changes the output and does not match the merged model

  huggingface/peft#3829 · Filter by author Nicholas022400701 opened on Sep 27, 2026
  1

  

• delete_adapter on the last adapter leaves active_adapter pointing at a deleted name

  huggingface/peft#3828 · Filter by author simpleqt opened on Sep 27, 2026
  2

  

• OSFLayer.merge() skips check_adapters_to_merge(), double-applies the delta on re-merge

  huggingface/peft#3825 · Filter by author AmirF194 opened on Sep 27, 2026
  3

• [Bug] PeftMixedModel loading unfreezes inference-only adapters

  huggingface/peft#3823 · Filter by author aswanth-07 opened on Sep 26, 2026
  2

• hotswap_adapter validates the active adapter instead of the requested inactive target

  huggingface/peft#3821 · Filter by author DaoyuanLi2816 opened on Sep 26, 2026
  1

• Overlapping rank_pattern / alpha_pattern keys resolve by dict order, and save_pretrained(sort_keys=True) changes that order on reload

  huggingface/peft#3818 · Filter by author jayzuccarelli opened on Sep 25, 2026
  6

  

• BOFT Conv2d forward ignores dilation, 2D stride/padding and padding_mode

  huggingface/peft#3813 · Filter by author huiq777 opened on Sep 24, 2026
  4

  

• Adapter loading never warns about unexpected keys, and strict=True is silently ignored

  huggingface/peft#3804 · Filter by author hivaze opened on Sep 23, 2026
  10

• LoRA on an FSDP2-sharded model gets the local shard shape when the mesh dimension is not named fsdp

  huggingface/peft#3803 · Filter by author hivaze opened on Sep 23, 2026
  4