text-generation-webui/modules/LoRA.py

44 lines
1.5 KiB
Python
Raw Normal View History

2023-03-17 01:35:53 +01:00
from pathlib import Path
2023-03-25 05:18:32 +01:00
import torch
2023-03-17 01:35:53 +01:00
import modules.shared as shared
from modules.models import load_model
2023-03-24 01:56:26 +01:00
from modules.text_generation import clear_torch_cache
2023-03-17 01:35:53 +01:00
2023-03-24 01:56:26 +01:00
def reload_model():
shared.model = shared.tokenizer = None
clear_torch_cache()
shared.model, shared.tokenizer = load_model(shared.model_name)
2023-03-17 01:35:53 +01:00
def add_lora_to_model(lora_name):
2023-03-18 14:55:24 +01:00
from peft import PeftModel
2023-03-24 02:02:09 +01:00
# If a LoRA had been previously loaded, or if we want
# to unload a LoRA, reload the model
if shared.lora_name not in ['None', ''] or lora_name in ['None', '']:
2023-03-24 02:02:09 +01:00
reload_model()
shared.lora_name = lora_name
2023-03-24 01:56:26 +01:00
if lora_name not in ['None', '']:
print(f"Adding the LoRA {lora_name} to the model...")
2023-03-17 21:45:28 +01:00
params = {}
2023-03-23 20:49:41 +01:00
if not shared.args.cpu:
2023-03-23 04:55:33 +01:00
params['dtype'] = shared.model.dtype
2023-03-23 20:49:41 +01:00
if hasattr(shared.model, "hf_device_map"):
params['device_map'] = {"base_model.model."+k: v for k, v in shared.model.hf_device_map.items()}
elif shared.args.load_in_8bit:
params['device_map'] = {'': 0}
2023-03-23 04:55:33 +01:00
shared.model = PeftModel.from_pretrained(shared.model, Path(f"{shared.args.lora_dir}/{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()
2023-03-23 20:49:41 +01:00
if not hasattr(shared.model, "hf_device_map"):
2023-03-25 05:18:32 +01:00
if torch.has_mps:
device = torch.device('mps')
shared.model = shared.model.to(device)
else:
shared.model = shared.model.cuda()