PEFT 0.18.0 release introduces RoAd, ALoRA, Arrow, WaveFT, DeLoRA, OSF; Transformers v5 compatibility and Python 3.9 deprecation

AI Impact Summary

PEFT 0.18.0 adds RoAd, ALoRA, Arrow with GenKnowSub, WaveFT, DeLoRA, and OSF, expanding and optimizing the PEFT method suite for fine-tuning large models. It also enforces a hard upgrade path: Transformers v5 will be incompatible with PEFT < 0.18.0 and Python 3.9 is no longer supported, requiring downstream services to upgrade dependencies. RoAd introduces mixed-adapter batches, enabling per-sample adapter routing within a single forward pass, which can materially improve serving efficiency for multi-adapter workloads. Expect some behavioral changes in OFT and weight-tying handling, and plan to retrain or validate adapters after upgrading to avoid numeric differences and regressions.