MiCA is now part of Hugging Face PEFT

The MiCA (Minor Component Adaptation) method has been merged into the main branch of the Hugging Face PEFT library, allowing users to install it directly from source. It is exposed through the existing LoRA interface by setting `init_lora_weights="mica"`.

• MiCA adapts weights using the minor singular subspace rather than dominant directions, initializing as a no-op and keeping matrix B frozen while training only A.
• Experiments show approximately 90% higher knowledge uptake, 20% less catastrophic forgetting, and 80% fewer trainable parameters compared to standard LoRA.
• Recommended hyperparameters for MiCA are roughly half the rank (`r_mica ≈ r_lora / 2`) and double the learning rate of a working LoRA setup.
• The method is primarily intended for continued pretraining or domain-adaptive pretraining, with a workflow suggesting merging the adapter before instruction tuning.

This integration provides a more parameter-efficient alternative to LoRA for injecting new knowledge into pretrained models while preserving existing behavior during initialization.