text-generation-webui/modules/LoRA.py

53 lines
1.9 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
2023-03-30 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
def add_lora_to_model(lora_names):
prior_set = set(shared.lora_names)
added_set = set(lora_names) - prior_set
removed_set = prior_set - set(lora_names)
shared.lora_names = list(lora_names)
2023-03-17 01:35:53 +01:00
# Nothing to do = skip.
if len(added_set) == 0 and len(removed_set) == 0:
return
2023-03-24 01:56:26 +01:00
# Only adding, and already peft? Do it the easy way.
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)
return
# If removing anything, disable all and re-add.
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"):
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}
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:
2023-03-23 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()