text-generation-webui/modules/models.py

189 lines
8.2 KiB
Python
Raw Normal View History

2023-02-23 17:28:30 +01:00
import json
import os
import time
import zipfile
from pathlib import Path
import numpy as np
import torch
import transformers
2023-03-16 17:34:23 +01:00
from accelerate import infer_auto_device_map, init_empty_weights
from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
BitsAndBytesConfig)
2023-02-23 18:41:42 +01:00
import modules.shared as shared
2023-02-23 17:28:30 +01:00
transformers.logging.set_verbosity_error()
2023-02-23 17:28:30 +01:00
local_rank = None
if shared.args.flexgen:
2023-03-16 14:18:34 +01:00
from flexgen.flex_opt import CompressionConfig, ExecutionEnv, OptLM, Policy
2023-02-23 17:28:30 +01:00
if shared.args.deepspeed:
import deepspeed
2023-02-23 18:41:42 +01:00
from transformers.deepspeed import (HfDeepSpeedConfig,
is_deepspeed_zero3_enabled)
2023-02-23 17:28:30 +01:00
from modules.deepspeed_parameters import generate_ds_config
# Distributed setup
local_rank = shared.args.local_rank if shared.args.local_rank is not None else int(os.getenv("LOCAL_RANK", "0"))
world_size = int(os.getenv("WORLD_SIZE", "1"))
torch.cuda.set_device(local_rank)
deepspeed.init_distributed()
ds_config = generate_ds_config(shared.args.bf16, 1 * world_size, shared.args.nvme_offload_dir)
dschf = HfDeepSpeedConfig(ds_config) # Keep this object alive for the Transformers integration
2023-03-13 18:00:38 +01:00
2023-02-23 17:28:30 +01:00
def load_model(model_name):
print(f"Loading {model_name}...")
t0 = time.time()
2023-02-28 03:03:35 +01:00
shared.is_RWKV = model_name.lower().startswith('rwkv-')
2023-02-23 17:28:30 +01:00
# Default settings
2023-03-13 18:00:38 +01:00
if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.gptq_bits, shared.args.auto_devices, shared.args.disk, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None, shared.args.deepspeed, shared.args.flexgen, shared.is_RWKV]):
2023-02-23 17:28:30 +01:00
if any(size in shared.model_name.lower() for size in ('13b', '20b', '30b')):
model = AutoModelForCausalLM.from_pretrained(Path(f"models/{shared.model_name}"), device_map='auto', load_in_8bit=True)
else:
model = AutoModelForCausalLM.from_pretrained(Path(f"models/{shared.model_name}"), low_cpu_mem_usage=True, torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16).cuda()
# FlexGen
elif shared.args.flexgen:
2023-02-26 20:53:41 +01:00
# Initialize environment
env = ExecutionEnv.create(shared.args.disk_cache_dir)
2023-02-23 17:28:30 +01:00
# Offloading policy
policy = Policy(1, 1,
shared.args.percent[0], shared.args.percent[1],
shared.args.percent[2], shared.args.percent[3],
shared.args.percent[4], shared.args.percent[5],
2023-03-04 05:04:02 +01:00
overlap=True, sep_layer=True, pin_weight=shared.args.pin_weight,
2023-02-23 17:28:30 +01:00
cpu_cache_compute=False, attn_sparsity=1.0,
compress_weight=shared.args.compress_weight,
comp_weight_config=CompressionConfig(
num_bits=4, group_size=64,
group_dim=0, symmetric=False),
compress_cache=False,
comp_cache_config=CompressionConfig(
num_bits=4, group_size=64,
group_dim=2, symmetric=False))
2023-02-26 20:53:41 +01:00
model = OptLM(f"facebook/{shared.model_name}", env, "models", policy)
2023-02-23 17:28:30 +01:00
# DeepSpeed ZeRO-3
elif shared.args.deepspeed:
model = AutoModelForCausalLM.from_pretrained(Path(f"models/{shared.model_name}"), torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16)
model = deepspeed.initialize(model=model, config_params=ds_config, model_parameters=None, optimizer=None, lr_scheduler=None)[0]
model.module.eval() # Inference
print(f"DeepSpeed ZeRO-3 is enabled: {is_deepspeed_zero3_enabled()}")
2023-02-28 03:03:35 +01:00
# RMKV model (not on HuggingFace)
elif shared.is_RWKV:
2023-03-06 12:45:49 +01:00
from modules.RWKV import RWKVModel, RWKVTokenizer
2023-02-28 03:03:35 +01:00
2023-03-01 16:08:55 +01:00
model = RWKVModel.from_pretrained(Path(f'models/{model_name}'), dtype="fp32" if shared.args.cpu else "bf16" if shared.args.bf16 else "fp16", device="cpu" if shared.args.cpu else "cuda")
2023-03-06 12:45:49 +01:00
tokenizer = RWKVTokenizer.from_pretrained(Path('models'))
2023-03-01 16:08:55 +01:00
2023-03-06 12:45:49 +01:00
return model, tokenizer
2023-02-28 03:03:35 +01:00
2023-03-13 18:00:38 +01:00
# Quantized model
elif shared.args.gptq_bits > 0:
from modules.GPTQ_loader import load_quantized
2023-03-13 20:11:40 +01:00
model = load_quantized(model_name)
2023-03-09 19:50:26 +01:00
2023-02-23 17:28:30 +01:00
# Custom
else:
params = {"low_cpu_mem_usage": True}
2023-02-23 17:28:30 +01:00
if not shared.args.cpu and not torch.cuda.is_available():
print("Warning: torch.cuda.is_available() returned False.\nThis means that no GPU has been detected.\nFalling back to CPU mode.\n")
2023-02-23 17:28:30 +01:00
shared.args.cpu = True
if shared.args.cpu:
params["torch_dtype"] = torch.float32
2023-02-23 17:28:30 +01:00
else:
params["device_map"] = 'auto'
if shared.args.load_in_8bit:
params['quantization_config'] = BitsAndBytesConfig(load_in_8bit=True, llm_int8_enable_fp32_cpu_offload=True)
elif shared.args.bf16:
params["torch_dtype"] = torch.bfloat16
else:
params["torch_dtype"] = torch.float16
2023-02-23 17:28:30 +01:00
if shared.args.gpu_memory:
memory_map = shared.args.gpu_memory
2023-03-16 17:34:23 +01:00
max_memory = {}
for i in range(len(memory_map)):
max_memory[i] = f'{memory_map[i]}GiB'
max_memory['cpu'] = f'{shared.args.cpu_memory or 99}GiB'
params['max_memory'] = max_memory
else:
2023-03-16 17:34:23 +01:00
total_mem = (torch.cuda.get_device_properties(0).total_memory / (1024*1024))
suggestion = round((total_mem-1000) / 1000) * 1000
if total_mem - suggestion < 800:
2023-02-23 17:28:30 +01:00
suggestion -= 1000
suggestion = int(round(suggestion/1000))
print(f"\033[1;32;1mAuto-assiging --gpu-memory {suggestion} for your GPU to try to prevent out-of-memory errors.\nYou can manually set other values.\033[0;37;0m")
2023-03-16 17:34:23 +01:00
max_memory = {0: f'{suggestion}GiB', 'cpu': f'{shared.args.cpu_memory or 99}GiB'}
params['max_memory'] = max_memory
2023-02-23 17:28:30 +01:00
if shared.args.disk:
params["offload_folder"] = shared.args.disk_cache_dir
checkpoint = Path(f'models/{shared.model_name}')
if shared.args.load_in_8bit and params.get('max_memory', None) is not None and params['device_map'] == 'auto':
config = AutoConfig.from_pretrained(checkpoint)
with init_empty_weights():
model = AutoModelForCausalLM.from_config(config)
model.tie_weights()
params['device_map'] = infer_auto_device_map(
model,
dtype=torch.int8,
max_memory=params['max_memory'],
no_split_module_classes = model._no_split_modules
)
model = AutoModelForCausalLM.from_pretrained(checkpoint, **params)
2023-02-23 17:28:30 +01:00
# Loading the tokenizer
if shared.model_name.lower().startswith(('gpt4chan', 'gpt-4chan', '4chan')) and Path("models/gpt-j-6B/").exists():
2023-02-23 17:28:30 +01:00
tokenizer = AutoTokenizer.from_pretrained(Path("models/gpt-j-6B/"))
else:
tokenizer = AutoTokenizer.from_pretrained(Path(f"models/{shared.model_name}/"))
tokenizer.truncation_side = 'left'
print(f"Loaded the model in {(time.time()-t0):.2f} seconds.")
return model, tokenizer
def load_soft_prompt(name):
if name == 'None':
shared.soft_prompt = False
shared.soft_prompt_tensor = None
else:
with zipfile.ZipFile(Path(f'softprompts/{name}.zip')) as zf:
zf.extract('tensor.npy')
zf.extract('meta.json')
j = json.loads(open('meta.json', 'r').read())
print(f"\nLoading the softprompt \"{name}\".")
for field in j:
if field != 'name':
if type(j[field]) is list:
print(f"{field}: {', '.join(j[field])}")
else:
print(f"{field}: {j[field]}")
print()
tensor = np.load('tensor.npy')
Path('tensor.npy').unlink()
Path('meta.json').unlink()
tensor = torch.Tensor(tensor).to(device=shared.model.device, dtype=shared.model.dtype)
tensor = torch.reshape(tensor, (1, tensor.shape[0], tensor.shape[1]))
shared.soft_prompt = True
shared.soft_prompt_tensor = tensor
return name