peft · Issues· 81 open

Merging a subset of the active adapters silently disables the remaining active adapters in forward

`modules_to_save` raises TypeError when a name is a suffix of a LoRA sub-module (e.g. "out", "A")

Failed adapter injection leaves orphaned adapter layers behind

KappaTuneSelector does not dequantize bitsandbytes 8-bit weights, computing condition numbers on raw int8 codes

Silently slower delta-rule fallback when training a Qwen3.8-27B LoRA with examples/sft under FSDP on 8x B200

Transformers v5 MoE conversion incorrectly rewrites fully qualified shared-expert LoRA targets

bnb LoRA merge with lora_bias=True drops scaling factor on bias

PiSSA / MiCA / CorDA init bypasses quantized-layer guard on 4-bit float-storage layers

add_weighted_adapter with rank_pattern produces adapters that cannot be reloaded (linear, ties, dare, magnitude_prune)

target_parameters docs: broken example and a wrong regex claim

`get_peft_model` and `add_adapter` do not preserve eval mode for newly injected adapter modules

Bug: Hotswapping error with rank pattern

Pooled threat: safe merging not working properly

Proposal: Benchmarking on different PEFT techniques

Expand model test matrices to cover LoHa, LoKr, and LN Tuning