text-generation-webui/modules/GPTQ_loader.py

210 lines
7.7 KiB
Python
Raw Normal View History

import inspect
2023-05-04 02:43:17 +02:00
import logging
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")))
try:
import llama_inference_offload
except ImportError:
logging.error('Failed to load GPTQ-for-LLaMa')
logging.error('See https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md')
sys.exit(-1)
try:
from modelutils import find_layers
except ImportError:
from utils import find_layers
try:
from quant import make_quant
is_triton = False
except ImportError:
import quant
is_triton = True
2023-03-28 22:34:15 +02:00
2023-03-28 19:38:55 +02:00
2023-04-17 04:26:52 +02:00
# This function is a replacement for the load_quant function in the
# GPTQ-for_LLaMa repository. It supports more models and branches.
def _load_quant(model, checkpoint, wbits, groupsize=-1, faster_kernel=False, exclude_layers=None, kernel_switch_threshold=128, eval=True):
exclude_layers = exclude_layers or ['lm_head']
2023-03-28 19:38:55 +02:00
def noop(*args, **kwargs):
pass
config = AutoConfig.from_pretrained(model, trust_remote_code=shared.args.trust_remote_code)
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, trust_remote_code=shared.args.trust_remote_code)
2023-03-28 19:38:55 +02:00
torch.set_default_dtype(torch.float)
if eval:
model = model.eval()
2023-03-28 19:38:55 +02:00
layers = find_layers(model)
for name in exclude_layers:
if name in layers:
del layers[name]
if not is_triton:
gptq_args = inspect.getfullargspec(make_quant).args
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)
else:
quant.make_quant_linear(model, layers, wbits, groupsize)
del layers
2023-03-28 19:38:55 +02:00
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)
2023-04-12 17:26:06 +02:00
if is_triton:
2023-04-22 17:27:30 +02:00
if shared.args.quant_attn:
quant.make_quant_attn(model)
2023-04-22 17:27:30 +02:00
if eval and shared.args.fused_mlp:
quant.make_fused_mlp(model)
2023-04-17 04:26:52 +02:00
2023-04-22 17:27:30 +02:00
if shared.args.warmup_autotune:
quant.autotune_warmup_linear(model, transpose=not eval)
2023-04-22 17:27:30 +02:00
if eval and shared.args.fused_mlp:
quant.autotune_warmup_fused(model)
2023-04-12 17:26:06 +02:00
2023-03-28 19:38:55 +02:00
model.seqlen = 2048
return model
2023-03-12 15:12:34 +01:00
2023-04-17 04:26:52 +02:00
# Used to locate the .pt/.safetensors quantized file
def find_quantized_model_file(model_name):
2023-05-04 20:17:20 +02:00
if shared.args.checkpoint:
return Path(shared.args.checkpoint)
2023-04-17 04:26:52 +02:00
path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
pt_path = None
priority_name_list = [
Path(f'{shared.args.model_dir}/{model_name}{hyphen}{shared.args.wbits}bit{group}{ext}')
for group in ([f'-{shared.args.groupsize}g', ''] if shared.args.groupsize > 0 else [''])
for ext in ['.safetensors', '.pt']
for hyphen in ['-', f'/{model_name}-', '/']
]
for path in priority_name_list:
if path.exists():
pt_path = path
break
# If the model hasn't been found with a well-behaved name, pick the last .pt
# or the last .safetensors found in its folder as a last resort
if not pt_path:
found_pts = list(path_to_model.glob("*.pt"))
found_safetensors = list(path_to_model.glob("*.safetensors"))
pt_path = None
if len(found_pts) > 0:
if len(found_pts) > 1:
2023-05-04 02:43:17 +02:00
logging.warning('More than one .pt model has been found. The last one will be selected. It could be wrong.')
2023-04-17 04:26:52 +02:00
pt_path = found_pts[-1]
elif len(found_safetensors) > 0:
if len(found_pts) > 1:
2023-05-04 02:43:17 +02:00
logging.warning('More than one .safetensors model has been found. The last one will be selected. It could be wrong.')
2023-04-17 04:26:52 +02:00
pt_path = found_safetensors[-1]
return pt_path
# The function that loads the model in modules/models.py
2023-03-13 20:11:32 +01:00
def load_quantized(model_name):
# Find the model type
if not shared.args.model_type:
2023-03-30 02:47:36 +02:00
name = model_name.lower()
2023-04-24 01:32:22 +02:00
if any((k in name for k in ['llama', 'alpaca', 'vicuna', 'llava'])):
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:
2023-05-04 02:43:17 +02:00
logging.error("Can't determine model type from model name. Please specify it manually using --model_type argument")
2023-03-13 20:11:32 +01:00
exit()
else:
model_type = shared.args.model_type.lower()
2023-03-13 20:11:32 +01:00
# Select the appropriate load_quant function
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:
2023-05-04 02:43:17 +02:00
logging.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:
2023-05-04 02:43:17 +02:00
logging.error("Unknown pre-quantized model type specified. Only 'llama', 'opt' and 'gptj' are supported")
exit()
2023-03-12 15:12:34 +01:00
2023-04-17 04:26:52 +02:00
# Find the quantized model weights file (.pt/.safetensors)
path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
2023-04-17 04:26:52 +02:00
pt_path = find_quantized_model_file(model_name)
2023-03-12 15:12:34 +01:00
if not pt_path:
2023-05-04 02:43:17 +02:00
logging.error("Could not find the quantized model in .pt or .safetensors format, exiting...")
2023-03-12 15:12:34 +01:00
exit()
2023-04-10 04:19:28 +02:00
else:
2023-05-04 02:43:17 +02:00
logging.info(f"Found the following quantized model: {pt_path}")
2023-03-12 15:12:34 +01:00
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)
2023-04-12 19:48:17 +02:00
if shared.args.gpu_memory or torch.cuda.device_count() > 1:
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
else:
max_memory = accelerate.utils.get_balanced_memory(model)
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"])
2023-05-04 02:43:17 +02:00
logging.info("Using the following device map for the quantized 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