2023-03-16 20:35:53 -04:00
|
|
|
from pathlib import Path
|
|
|
|
|
2023-03-25 00:18:32 -04:00
|
|
|
import torch
|
2023-03-29 21:50:58 -04:00
|
|
|
from peft import PeftModel
|
2023-03-25 00:18:32 -04:00
|
|
|
|
2023-03-16 20:35:53 -04:00
|
|
|
import modules.shared as shared
|
2023-04-07 20:36:04 -04:00
|
|
|
from modules.models import reload_model
|
2023-03-23 20:56:26 -04:00
|
|
|
|
2023-04-06 23:15:45 -04:00
|
|
|
|
2023-03-16 20:35:53 -04:00
|
|
|
def add_lora_to_model(lora_name):
|
|
|
|
|
2023-03-23 21:02:09 -04:00
|
|
|
# If a LoRA had been previously loaded, or if we want
|
|
|
|
# to unload a LoRA, reload the model
|
2023-03-26 23:04:43 -04:00
|
|
|
if shared.lora_name not in ['None', ''] or lora_name in ['None', '']:
|
2023-03-23 21:02:09 -04:00
|
|
|
reload_model()
|
|
|
|
shared.lora_name = lora_name
|
2023-03-23 20:56:26 -04:00
|
|
|
|
2023-03-26 23:04:43 -04:00
|
|
|
if lora_name not in ['None', '']:
|
2023-03-17 10:39:48 -04:00
|
|
|
print(f"Adding the LoRA {lora_name} to the model...")
|
2023-03-17 16:45:28 -04:00
|
|
|
params = {}
|
2023-03-23 15:49:41 -04:00
|
|
|
if not shared.args.cpu:
|
2023-03-22 23:55:33 -04:00
|
|
|
params['dtype'] = shared.model.dtype
|
2023-03-23 15:49:41 -04:00
|
|
|
if hasattr(shared.model, "hf_device_map"):
|
2023-04-06 23:15:45 -04:00
|
|
|
params['device_map'] = {"base_model.model." + k: v for k, v in shared.model.hf_device_map.items()}
|
2023-03-23 15:49:41 -04:00
|
|
|
elif shared.args.load_in_8bit:
|
|
|
|
params['device_map'] = {'': 0}
|
2023-04-06 23:15:45 -04:00
|
|
|
|
2023-03-24 20:30:18 -04:00
|
|
|
shared.model = PeftModel.from_pretrained(shared.model, Path(f"{shared.args.lora_dir}/{lora_name}"), **params)
|
2023-03-23 00:05:13 -04:00
|
|
|
if not shared.args.load_in_8bit and not shared.args.cpu:
|
2023-03-22 23:55:33 -04:00
|
|
|
shared.model.half()
|
2023-03-23 15:49:41 -04:00
|
|
|
if not hasattr(shared.model, "hf_device_map"):
|
2023-03-25 00:18:32 -04:00
|
|
|
if torch.has_mps:
|
|
|
|
device = torch.device('mps')
|
|
|
|
shared.model = shared.model.to(device)
|
|
|
|
else:
|
|
|
|
shared.model = shared.model.cuda()
|