Issues

`unload()` returns the trained `modules_to_save` copy instead of the original module

huggingface/peft#3955 · opened

on Oct 13, 2026

1

Prompt-learning methods crash on transformers 4.43/4.44: the pre-get_text_config fallback uses `in` on a PretrainedConfig

huggingface/peft#3951 · opened

on Oct 13, 2026

MoE conversion no longer targets routed experts when shared experts or dense layers share the target name

huggingface/peft#3945 · opened

on Oct 12, 2026

1

LoRA on nn.MultiheadAttention: lora_bias=True never trains the bias, and forward fails on a bias-free MHA

huggingface/peft#3935 · opened

on Oct 11, 2026

1

get_peft_model_state_dict gathers the whole sharded base model on DTensor-sharded models

huggingface/peft#3930 · opened

on Oct 11, 2026

5

`ensure_weight_tying=True` silently stops tying the LoRA weights when the model is in bf16/fp16

huggingface/peft#3929 · opened

on Oct 11, 2026

CPT forward mutates the caller-provided input_type_mask in place

huggingface/peft#3928 · opened

on Oct 11, 2026

1

Saving / loading fails with multiple LoRA adapters when only one uses `trainable_token_indices`

huggingface/peft#3919 · opened

on Oct 9, 2026

2

`merge_and_unload()`/`unload()` raise AttributeError on RoBERTa sequence classifiers for all methods except LoRA, AdaLoRA and HiRA

huggingface/peft#3918 · opened

on Oct 9, 2026

1

LoRA on MoE experts (target_parameters) fails under tensor/expert parallelism

huggingface/peft#3914 · opened

on Oct 9, 2026

2

LoRA-FA optimizer crashes when an adapted layer has no gradient

huggingface/peft#3900 · opened

on Oct 6, 2026

1

Default CPTConfig masks all input labels and returns NaN training loss

huggingface/peft#3899 · opened

on Oct 6, 2026

3

Inference-only prompt adapters remain trainable after loading

huggingface/peft#3898 · opened

on Oct 6, 2026

1

Adapters are saved under transformers v5's renamed module paths, while `save_pretrained` writes base models in the original checkpoint layout

huggingface/peft#3896 · opened

on Oct 5, 2026

9

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

huggingface/peft#3879 · opened

on Oct 4, 2026

4

Add support for BoRA

huggingface/peft#3877 · opened

on Oct 3, 2026

1

`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 · opened

on Oct 3, 2026

4

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

huggingface/peft#3874 · opened

on Oct 2, 2026

1

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

huggingface/peft#3871 · opened

on Oct 2, 2026

5

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

huggingface/peft#3868 · opened

on Oct 2, 2026

1

DoRA hotswap raises a bare KeyError for compatible adapters

huggingface/peft#3867 · opened

on Oct 2, 2026

3

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

huggingface/peft#3861 · opened

on Oct 1, 2026

1

PeftModelForQuestionAnswering silently drops `token_type_ids` and `position_ids`

huggingface/peft#3857 · opened

on Oct 1, 2026

2

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

huggingface/peft#3835 · 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 · opened

on Sep 27, 2026

3

Navigation Menu

Repository navigation

Issues

`unload()` returns the trained `modules_to_save` copy instead of the original module

Prompt-learning methods crash on transformers 4.43/4.44: the pre-get_text_config fallback uses `in` on a PretrainedConfig

MoE conversion no longer targets routed experts when shared experts or dense layers share the target name

LoRA on nn.MultiheadAttention: lora_bias=True never trains the bias, and forward fails on a bias-free MHA

get_peft_model_state_dict gathers the whole sharded base model on DTensor-sharded models

`ensure_weight_tying=True` silently stops tying the LoRA weights when the model is in bf16/fp16

CPT forward mutates the caller-provided input_type_mask in place

Saving / loading fails with multiple LoRA adapters when only one uses `trainable_token_indices`

`merge_and_unload()`/`unload()` raise AttributeError on RoBERTa sequence classifiers for all methods except LoRA, AdaLoRA and HiRA

LoRA on MoE experts (target_parameters) fails under tensor/expert parallelism

LoRA-FA optimizer crashes when an adapted layer has no gradient

Default CPTConfig masks all input labels and returns NaN training loss

Inference-only prompt adapters remain trainable after loading

Adapters are saved under transformers v5's renamed module paths, while `save_pretrained` writes base models in the original checkpoint layout

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

Add support for BoRA

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

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

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

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

DoRA hotswap raises a bare KeyError for compatible adapters

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

PeftModelForQuestionAnswering silently drops `token_type_ids` and `position_ids`

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

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