2023-05-17 16:12:12 +02:00
|
|
|
from pathlib import Path
|
|
|
|
|
2023-06-02 06:33:55 +02:00
|
|
|
from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
|
2023-05-17 16:12:12 +02:00
|
|
|
|
|
|
|
import modules.shared as shared
|
2023-05-22 03:42:34 +02:00
|
|
|
from modules.logging_colors import logger
|
2023-05-17 16:12:12 +02:00
|
|
|
from modules.models import get_max_memory_dict
|
|
|
|
|
|
|
|
|
|
|
|
def load_quantized(model_name):
|
|
|
|
path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
|
|
|
|
pt_path = None
|
|
|
|
|
|
|
|
# Find the model checkpoint
|
2023-06-02 06:33:55 +02:00
|
|
|
if shared.args.checkpoint:
|
|
|
|
pt_path = Path(shared.args.checkpoint)
|
|
|
|
else:
|
|
|
|
for ext in ['.safetensors', '.pt', '.bin']:
|
|
|
|
found = list(path_to_model.glob(f"*{ext}"))
|
|
|
|
if len(found) > 0:
|
|
|
|
if len(found) > 1:
|
|
|
|
logger.warning(f'More than one {ext} model has been found. The last one will be selected. It could be wrong.')
|
|
|
|
|
|
|
|
pt_path = found[-1]
|
|
|
|
break
|
2023-05-17 20:52:23 +02:00
|
|
|
|
|
|
|
if pt_path is None:
|
2023-05-22 03:42:34 +02:00
|
|
|
logger.error("The model could not be loaded because its checkpoint file in .bin/.pt/.safetensors format could not be located.")
|
2023-05-17 20:52:23 +02:00
|
|
|
return
|
2023-05-17 16:12:12 +02:00
|
|
|
|
2023-06-02 06:33:55 +02:00
|
|
|
use_safetensors = pt_path.suffix == '.safetensors'
|
|
|
|
if not (path_to_model / "quantize_config.json").exists():
|
|
|
|
quantize_config = BaseQuantizeConfig(
|
|
|
|
bits=bits if (bits := shared.args.wbits) > 0 else 4,
|
|
|
|
group_size=gs if (gs := shared.args.groupsize) > 0 else -1,
|
|
|
|
desc_act=shared.args.desc_act
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
quantize_config = None
|
|
|
|
|
2023-05-17 16:12:12 +02:00
|
|
|
# Define the params for AutoGPTQForCausalLM.from_quantized
|
|
|
|
params = {
|
|
|
|
'model_basename': pt_path.stem,
|
|
|
|
'device': "cuda:0" if not shared.args.cpu else "cpu",
|
|
|
|
'use_triton': shared.args.triton,
|
|
|
|
'use_safetensors': use_safetensors,
|
2023-05-29 15:20:18 +02:00
|
|
|
'trust_remote_code': shared.args.trust_remote_code,
|
2023-06-02 06:33:55 +02:00
|
|
|
'max_memory': get_max_memory_dict(),
|
|
|
|
'quantize_config': quantize_config
|
2023-05-17 16:12:12 +02:00
|
|
|
}
|
|
|
|
|
2023-06-02 06:33:55 +02:00
|
|
|
logger.info(f"The AutoGPTQ params are: {params}")
|
2023-05-17 16:12:12 +02:00
|
|
|
model = AutoGPTQForCausalLM.from_quantized(path_to_model, **params)
|
2023-06-11 22:52:23 +02:00
|
|
|
|
|
|
|
# These lines fix the multimodal extension when used with AutoGPTQ
|
2023-06-14 23:44:43 +02:00
|
|
|
if hasattr(model, 'model'):
|
|
|
|
if not hasattr(model, 'dtype'):
|
|
|
|
if hasattr(model.model, 'dtype'):
|
|
|
|
model.dtype = model.model.dtype
|
2023-06-11 22:52:23 +02:00
|
|
|
|
2023-06-14 23:44:43 +02:00
|
|
|
if hasattr(model.model, 'model') and hasattr(model.model.model, 'embed_tokens'):
|
|
|
|
if not hasattr(model, 'embed_tokens'):
|
|
|
|
model.embed_tokens = model.model.model.embed_tokens
|
2023-06-11 22:52:23 +02:00
|
|
|
|
2023-06-14 23:44:43 +02:00
|
|
|
if not hasattr(model.model, 'embed_tokens'):
|
|
|
|
model.model.embed_tokens = model.model.model.embed_tokens
|
2023-06-11 22:52:23 +02:00
|
|
|
|
2023-05-17 16:12:12 +02:00
|
|
|
return model
|