🤗 PEFT adds new LoRA adapter merging methods (Concatenation, Linear, SVD, TIES)

AI Impact Summary

🤗 PEFT expands LoRA adapter experimentation with new merging methods (Concatenation, Linear, SVD, TIES), enabling on-the-fly composition and support for adapters with differing ranks. The integration leverages mergekit and Diffusers set_adapters to blend or switch adapters without full checkpoint merges, reducing memory pressure and accelerating iteration cycles. This provides engineering teams a practical path to optimize task-specific adapters directly on hardware-constrained deployments, potentially boosting performance without costly model retraining.