text-generation-webui/modules/GPTQ_loader.py

156 lines
5.9 KiB
Python
Raw Normal View History

import inspect
import re
2023-03-12 15:12:34 +01:00
import sys
from pathlib import Path
import accelerate
import torch
2023-03-28 19:38:55 +02:00
import transformers
2023-03-28 22:34:15 +02:00
from transformers import AutoConfig, AutoModelForCausalLM
2023-03-12 15:12:34 +01:00
import modules.shared as shared
sys.path.insert(0, str(Path("repositories/GPTQ-for-LLaMa")))
import llama_inference_offload
2023-03-28 19:38:55 +02:00
from modelutils import find_layers
2023-03-28 22:34:15 +02:00
from quant import make_quant
2023-03-28 19:38:55 +02:00
2023-03-28 21:45:38 +02:00
def _load_quant(model, checkpoint, wbits, groupsize=-1, faster_kernel=False, exclude_layers=['lm_head'], kernel_switch_threshold=128):
2023-03-28 19:38:55 +02:00
def noop(*args, **kwargs):
pass
config = AutoConfig.from_pretrained(model)
torch.nn.init.kaiming_uniform_ = noop
torch.nn.init.uniform_ = noop
torch.nn.init.normal_ = noop
2023-03-28 19:38:55 +02:00
torch.set_default_dtype(torch.half)
transformers.modeling_utils._init_weights = False
torch.set_default_dtype(torch.half)
model = AutoModelForCausalLM.from_config(config)
torch.set_default_dtype(torch.float)
model = model.eval()
layers = find_layers(model)
for name in exclude_layers:
if name in layers:
del layers[name]
gptq_args = inspect.getfullargspec(make_quant).args
2023-03-28 19:38:55 +02:00
make_quant_kwargs = {
'module': model,
'names': layers,
'bits': wbits,
}
if 'groupsize' in gptq_args:
make_quant_kwargs['groupsize'] = groupsize
if 'faster' in gptq_args:
make_quant_kwargs['faster'] = faster_kernel
if 'kernel_switch_threshold' in gptq_args:
make_quant_kwargs['kernel_switch_threshold'] = kernel_switch_threshold
make_quant(**make_quant_kwargs)
del layers
2023-03-28 19:38:55 +02:00
print('Loading model ...')
if checkpoint.endswith('.safetensors'):
from safetensors.torch import load_file as safe_load
model.load_state_dict(safe_load(checkpoint), strict=False)
2023-03-28 19:38:55 +02:00
else:
model.load_state_dict(torch.load(checkpoint), strict=False)
2023-03-28 19:38:55 +02:00
model.seqlen = 2048
print('Done.')
return model
2023-03-12 15:12:34 +01:00
2023-03-13 20:11:32 +01:00
def load_quantized(model_name):
if not shared.args.model_type:
2023-03-13 20:11:32 +01:00
# Try to determine model type from model name
2023-03-30 02:47:36 +02:00
name = model_name.lower()
if any((k in name for k in ['llama', 'alpaca', 'vicuna'])):
model_type = 'llama'
2023-03-30 02:47:36 +02:00
elif any((k in name for k in ['opt-', 'galactica'])):
model_type = 'opt'
2023-03-30 02:47:36 +02:00
elif any((k in name for k in ['gpt-j', 'pygmalion-6b'])):
2023-03-28 19:38:55 +02:00
model_type = 'gptj'
else:
print("Can't determine model type from model name. Please specify it manually using --model_type "
2023-03-13 20:11:32 +01:00
"argument")
exit()
else:
model_type = shared.args.model_type.lower()
2023-03-13 20:11:32 +01:00
2023-04-05 06:21:40 +02:00
if shared.args.pre_layer and model_type == 'llama':
load_quant = llama_inference_offload.load_quant
2023-03-28 19:38:55 +02:00
elif model_type in ('llama', 'opt', 'gptj'):
2023-04-05 06:21:40 +02:00
if shared.args.pre_layer:
print("Warning: ignoring --pre_layer because it only works for llama model type.")
2023-03-28 19:38:55 +02:00
load_quant = _load_quant
2023-03-12 15:12:34 +01:00
else:
2023-03-28 19:38:55 +02:00
print("Unknown pre-quantized model type specified. Only 'llama', 'opt' and 'gptj' are supported")
exit()
2023-03-12 15:12:34 +01:00
# Now we are going to try to locate the quantized model file.
path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
found_pts = list(path_to_model.glob("*.pt"))
found_safetensors = list(path_to_model.glob("*.safetensors"))
pt_path = None
if len(found_pts) == 1:
pt_path = found_pts[0]
elif len(found_safetensors) == 1:
pt_path = found_safetensors[0]
2023-03-13 20:11:32 +01:00
else:
if path_to_model.name.lower().startswith('llama-7b'):
pt_model = f'llama-7b-{shared.args.wbits}bit'
elif path_to_model.name.lower().startswith('llama-13b'):
pt_model = f'llama-13b-{shared.args.wbits}bit'
elif path_to_model.name.lower().startswith('llama-30b'):
pt_model = f'llama-30b-{shared.args.wbits}bit'
elif path_to_model.name.lower().startswith('llama-65b'):
pt_model = f'llama-65b-{shared.args.wbits}bit'
else:
pt_model = f'{model_name}-{shared.args.wbits}bit'
2023-03-12 15:12:34 +01:00
# Try to find the .safetensors or .pt both in the model dir and in the subfolder
for path in [Path(p + ext) for ext in ['.safetensors', '.pt'] for p in [f"{shared.args.model_dir}/{pt_model}", f"{path_to_model}/{pt_model}"]]:
if path.exists():
print(f"Found {path}")
pt_path = path
break
2023-03-12 15:12:34 +01:00
if not pt_path:
print("Could not find the quantized model in .pt or .safetensors format, exiting...")
2023-03-12 15:12:34 +01:00
exit()
2023-03-20 20:40:08 +01:00
# qwopqwop200's offload
if model_type == 'llama' and shared.args.pre_layer:
model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, shared.args.pre_layer)
else:
2023-03-28 21:45:38 +02:00
threshold = False if model_type == 'gptj' else 128
model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, kernel_switch_threshold=threshold)
2023-03-12 15:12:34 +01:00
2023-03-20 20:40:08 +01:00
# accelerate offload (doesn't work properly)
if shared.args.gpu_memory:
memory_map = list(map(lambda x: x.strip(), shared.args.gpu_memory))
max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
max_memory = {}
for i in range(len(memory_map)):
max_memory[i] = f'{memory_map[i]}GiB' if not re.match('.*ib$', memory_map[i].lower()) else memory_map[i]
max_memory['cpu'] = max_cpu_memory
2023-03-12 15:12:34 +01:00
device_map = accelerate.infer_auto_device_map(model, max_memory=max_memory, no_split_module_classes=["LlamaDecoderLayer"])
print("Using the following device map for the 4-bit model:", device_map)
# https://huggingface.co/docs/accelerate/package_reference/big_modeling#accelerate.dispatch_model
model = accelerate.dispatch_model(model, device_map=device_map, offload_buffers=True)
2023-03-20 20:40:08 +01:00
# No offload
elif not shared.args.cpu:
model = model.to(torch.device('cuda:0'))
2023-03-12 15:12:34 +01:00
return model