text-generation-webui/modules/LoRA.py

28 lines
909 B
Python
Raw Normal View History

2023-03-17 01:35:53 +01:00
from pathlib import Path
import modules.shared as shared
from modules.models import load_model
def add_lora_to_model(lora_name):
2023-03-18 14:55:24 +01:00
from peft import PeftModel
2023-03-17 01:35:53 +01:00
# Is there a more efficient way of returning to the base model?
if lora_name == "None":
2023-03-17 15:43:11 +01:00
print("Reloading the model to remove the LoRA...")
2023-03-17 01:35:53 +01:00
shared.model, shared.tokenizer = load_model(shared.model_name)
else:
print(f"Adding the LoRA {lora_name} to the model...")
2023-03-23 04:55:33 +01:00
2023-03-17 21:45:28 +01:00
params = {}
2023-03-23 04:55:33 +01:00
if shared.args.load_in_8bit:
params['device_map'] = {'': 0}
2023-03-23 05:05:13 +01:00
elif not shared.args.cpu:
2023-03-23 04:55:33 +01:00
params['device_map'] = 'auto'
params['dtype'] = shared.model.dtype
2023-03-17 21:45:28 +01:00
shared.model = PeftModel.from_pretrained(shared.model, Path(f"loras/{lora_name}"), **params)
2023-03-23 05:05:13 +01:00
if not shared.args.load_in_8bit and not shared.args.cpu:
2023-03-23 04:55:33 +01:00
shared.model.half()