Open jamesjiangxing opened 8 months ago
from plugins.settings import settings
import os from typing import Dict, Tuple, Union, Optional
from torch.nn import Module from transformers import AutoModel
def auto_configure_device_map(num_gpus: int) -> Dict[str, int]:
# transformer.final_layernorm 和 lm_head 占用1层
# transformer.layers 占用 28 层
# 总共30层分配到num_gpus张卡上
num_trans_layers = 28
per_gpu_layers = 30 / num_gpus
# bugfix: 在linux中调用torch.embedding传入的weight,input不在同一device上,导致RuntimeError
# windows下 model.device 会被设置成 transformer.word_embeddings.device
# linux下 model.device 会被设置成 lm_head.device
# 在调用chat或者stream_chat时,input_ids会被放到model.device上
# 如果transformer.word_embeddings.device和model.device不同,则会导致RuntimeError
# 因此这里将transformer.word_embeddings,transformer.final_layernorm,lm_head都放到第一张卡上
device_map = {'transformer.word_embeddings': 0,
'transformer.final_layernorm': 0, 'lm_head': 0}
used = 2
gpu_target = 0
for i in range(num_trans_layers):
if used >= per_gpu_layers:
gpu_target += 1
used = 0
assert gpu_target < num_gpus
device_map[f'transformer.layers.{i}'] = gpu_target
used += 1
return device_map
def load_model_on_gpus(checkpoint_path: Union[str, os.PathLike], num_gpus: int = 2, device_map: Optional[Dict[str, int]] = None, kwargs) -> Module: if num_gpus < 2 and device_map is None: model = AutoModel.from_pretrained(checkpoint_path, trust_remote_code=True, kwargs).half().cuda() else: from accelerate import dispatch_model
model = AutoModel.from_pretrained(checkpoint_path, trust_remote_code=True, **kwargs).half()
if device_map is None:
device_map = auto_configure_device_map(num_gpus)
model = dispatch_model(model, device_map=device_map)
return
def chat_init(history): history_formatted = None if history is not None: history_formatted = [] tmp = [] for i, old_chat in enumerate(history): if len(tmp) == 0 and old_chat['role'] == "user": tmp.append(old_chat['content']) elif old_chat['role'] == "AI" or old_chat['role'] == 'assistant': tmp.append(old_chat['content']) history_formatted.append(tuple(tmp)) tmp = [] else: continue return history_formatted
def chat_one(prompt, history_formatted, max_length, top_p, temperature, zhishiku=False): for response, history in model.stream_chat(tokenizer, prompt, history_formatted, max_length=max_length, top_p=top_p, temperature=temperature): yield response
def load_model(): global model, tokenizer from transformers import AutoModel, AutoTokenizer tokenizer = AutoTokenizer.from_pretrained( settings.Path, local_files_only=True, trust_remote_code=True)
# settings.Path, local_files_only=True, trust_remote_code=True)
model = load_model_on_gpus(settings.Path, num_gpus=2)
if not (settings.Lora == '' or settings.Lora == None):
print('Lora模型地址', settings.Lora)
from peft import PeftModel
model = PeftModel.from_pretrained(model, settings.Lora)
device, precision = settings.Strategy.split()
# 根据设备执行不同的操作
if device == 'cpu':
# 如果是cpu,不做任何操作
pass
elif device == 'cuda':
# 如果是gpu,把模型移动到显卡
import torch
if not (precision.startswith('fp16i') and torch.cuda.get_device_properties(0).total_memory < 1.4e+10):
model = model.cuda()
else:
# 如果是其他设备,报错并退出程序
print('Error: 不受支持的设备')
exit()
# 根据精度执行不同的操作
if precision == 'fp16':
# 如果是fp16,把模型转化为半精度
model = model.half()
elif precision == 'fp32':
# 如果是fp32,把模型转化为全精度
model = model.float()
elif precision.startswith('fp16i'):
# 如果是fp16i开头,把模型转化为指定的精度
# 从字符串中提取精度的数字部分
bits = int(precision[5:])
# 调用quantize方法,传入精度参数
model = model.quantize(bits)
if device == 'cuda':
model = model.cuda()
model = model.half()
elif precision.startswith('fp32i'):
# 如果是fp32i开头,把模型转化为指定的精度
# 从字符串中提取精度的数字部分
bits = int(precision[5:])
# 调用quantize方法,传入精度参数
model = model.quantize(bits)
if device == 'cuda':
model = model.cuda()
model = model.float()
else:
# 如果是其他精度,报错并退出程序
print('Error: 不受支持的精度')
exit()
#model = model.eval()
Describe the bug 在两张GPU卡上加载Chatglm3-6b 出错.
To Reproduce Steps to reproduce the behavior:
Expected behavior A clear and concise description of what you expected to happen.
Screenshots