$ the-wire · showcase
PEFT grows a strict adapter check, diffusers resets its cache hooks
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
PEFT made adapter/checkpoint mismatches visible for the first time, and a second PEFT fix stops FSDP2 shards from being written into .bin files, while diffusers closed a state leak in ModularPipeline.
ENH Add strict_adapter_check for model loading (#3832) huggingface/peft
load_adapter never warned about unexpected keys, and a strict argument passed to either load_adapter or from_pretrained silently landed in **kwargs and was dropped. Both now take strict_adapter_check ("ignore", "warn", or "raise") and filter out keys expected to differ for structural reasons, including VB-LoRA's shared vector bank, the prompt encoder path, and TinyLoRA/UniLoRA references, befor...
FIX Gather FSDP2 adapter DTensors when saving huggingface/peft
FSDP2/HSDP adapters hold DTensors without setting _tp_size, so saving could fail against safetensors or silently serialize distributed shards into .bin. The fix gathers only the selected adapter entries into detached, contiguous CPU tensors via a vendored Transformers helper, leaving ordinary tensor dicts on their existing device, layout and autograd behavior; every participating rank must now ...
Reset stateful cache hooks after every ModularPipeline call (#15212) huggingface/diffusers
Standard pipelines clear cache hook state in maybe_free_model_hooks; ModularPipeline had no equivalent, so Cosmos3 carried a __call__ override that newer pipelines started copying. The reset moved into the base __call__ and the override is gone, which matters if you vendored that pattern into your own pipeline.
FIX AdaLoRA Conv2d with groups and padding_mode (#4084) huggingface/peft
AdaLoRA's Conv2d path now passes padding mode through correctly, and groups != 1 raises an error instead of proceeding with unsupported behavior.
Pin doc-builder to its v0.6.0 release (#8154) huggingface/trl
Elsewhere: trl pinned doc-builder to v0.6.0, and diffusers addressed a protobuf problem in its training-examples CI.
Action items
- → Drop any pip-installed strict= keyword from load_adapter and from_pretrained calls and migrate to strict_adapter_chec... huggingface/peft [plan]
- → Ensure every participating rank calls save_pretrained when saving FSDP2/HSDP adapters huggingface/peft [plan]
