loaders.adapter

loaders.adapter

Adapter loading functionality, including LoRA / QLoRA and associated utils

Functions

Name	Description
read_saved_adapter_config	The adapter config saved under lora_model_dir, or None when it is not readable.
reinit_lora_from_seed	Re-draw every fresh lora_A from a generator seeded by the module name.
setup_quantized_meta_for_peft	Replaces quant_state.to with a dummy function to prevent PEFT from moving quant_state to meta device
setup_quantized_peft_meta_for_training	Replaces dummy quant_state.to method with the original function to allow training to continue

read_saved_adapter_config

loaders.adapter.read_saved_adapter_config(lora_model_dir, from_hub=False)

The adapter config saved under lora_model_dir, or None when it is not readable.

A hub id has no local file; with from_hub the config is fetched the way PEFT will.

reinit_lora_from_seed

loaders.adapter.reinit_lora_from_seed(model, seed)

Re-draw every fresh lora_A from a generator seeded by the module name.

PEFT draws the adapters from the process RNG in module order, so a value depends on everything drawn before it: the rank, the parallel layout (expert parallelism sizes the expert adapter locally, shifting the stream for every later module) and the model’s module order. Seeding per module makes the init a function of the seed alone. Expert (target_parameters) adapters are drawn over the GLOBAL expert count and sliced to this rank’s experts. Returns the number of adapters re-initialised.

setup_quantized_meta_for_peft

loaders.adapter.setup_quantized_meta_for_peft(model)

Replaces quant_state.to with a dummy function to prevent PEFT from moving quant_state to meta device

setup_quantized_peft_meta_for_training

loaders.adapter.setup_quantized_peft_meta_for_training(model)

Replaces dummy quant_state.to method with the original function to allow training to continue