2023-03-17 01:35:53 +01:00
|
|
|
from pathlib import Path
|
|
|
|
|
2023-03-25 05:18:32 +01:00
|
|
|
import torch
|
2023-03-30 03:50:58 +02:00
|
|
|
from peft import PeftModel
|
2023-03-25 05:18:32 +01:00
|
|
|
|
2023-03-17 01:35:53 +01:00
|
|
|
import modules.shared as shared
|
2023-03-24 01:56:26 +01:00
|
|
|
|
2023-04-07 05:15:45 +02:00
|
|
|
|
2023-04-14 19:52:06 +02:00
|
|
|
def add_lora_to_model(lora_names):
|
2023-04-26 02:20:26 +02:00
|
|
|
shared.lora_names = list(lora_names)
|
2023-04-14 19:52:06 +02:00
|
|
|
prior_set = set(shared.lora_names)
|
|
|
|
added_set = set(lora_names) - prior_set
|
|
|
|
removed_set = prior_set - set(lora_names)
|
2023-03-17 01:35:53 +01:00
|
|
|
|
2023-04-26 02:20:26 +02:00
|
|
|
# If no LoRA needs to be added or removed, exit
|
2023-04-14 19:52:06 +02:00
|
|
|
if len(added_set) == 0 and len(removed_set) == 0:
|
|
|
|
return
|
2023-03-24 01:56:26 +01:00
|
|
|
|
2023-04-26 02:20:26 +02:00
|
|
|
# Add a LoRA when another LoRA is already present
|
2023-04-14 19:52:06 +02:00
|
|
|
if len(removed_set) == 0 and len(prior_set) > 0:
|
|
|
|
print(f"Adding the LoRA(s) named {added_set} to the model...")
|
|
|
|
for lora in added_set:
|
|
|
|
shared.model.load_adapter(Path(f"{shared.args.lora_dir}/{lora}"), lora)
|
2023-04-26 02:20:26 +02:00
|
|
|
|
2023-04-14 19:52:06 +02:00
|
|
|
return
|
|
|
|
|
2023-04-26 02:20:26 +02:00
|
|
|
# If any LoRA needs to be removed, start over
|
2023-04-14 19:52:06 +02:00
|
|
|
if len(removed_set) > 0:
|
|
|
|
shared.model.disable_adapter()
|
|
|
|
|
|
|
|
if len(lora_names) > 0:
|
|
|
|
print("Applying the following LoRAs to {}: {}".format(shared.model_name, ', '.join(lora_names)))
|
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"):
|
2023-04-07 05:15:45 +02:00
|
|
|
params['device_map'] = {"base_model.model." + k: v for k, v in shared.model.hf_device_map.items()}
|
2023-03-23 20:49:41 +01:00
|
|
|
elif shared.args.load_in_8bit:
|
|
|
|
params['device_map'] = {'': 0}
|
2023-04-07 05:15:45 +02:00
|
|
|
|
2023-04-14 19:52:06 +02:00
|
|
|
shared.model = PeftModel.from_pretrained(shared.model, Path(f"{shared.args.lora_dir}/{lora_names[0]}"), **params)
|
|
|
|
|
|
|
|
for lora in lora_names[1:]:
|
|
|
|
shared.model.load_adapter(Path(f"{shared.args.lora_dir}/{lora}"), lora)
|
|
|
|
|
2023-03-23 05:05:13 +01:00
|
|
|
if not shared.args.load_in_8bit and not shared.args.cpu:
|
2023-04-26 02:20:26 +02:00
|
|
|
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()
|