Trainable Adapters
Trainable adapters are the small LoRA matrices A and B (and optionally scaling) injected into specific linear layers that are the only parameters updated during fine-tuning.
What is Trainable Adapters?
Trainable adapters are the small LoRA matrices A and B (and optionally scaling) injected into specific linear layers that are the only parameters updated during fine-tuning.
Trainable adapters are the small LoRA matrices A and B (and optionally scaling) injected into specific linear layers that are the only parameters updated during fine-tuning.
Where is it used?
Adapters are attached to `q_proj`, `k_proj`, `v_proj`, and `o_proj` in Llama/Mistral via `peft`; `unsloth` fuses adapters for 2x speedup; adapters can be swapped at inference with `load_adapter`.
How to build it
Inspect `model.peft_config` and `model.adapter_model()` to list injected layers; save with `model.save_pretrained("adapter")` producing a small ~10MB `adapter_model.safetensors` file.