前方干货预警:这可能是你能够找到的,最容易理解,最容易跑通的,适用于各种开源LLM模型的,同时支持多轮和单轮对话数据集的大模型高效微调范例。
我们构造了一个修改大模型自我认知的3轮对话的玩具数据集,使用QLoRA算法,只需要5分钟的训练时间,就可以完成微调,并成功修改了LLM模型的自我认知(以Qwen7b-Chat为例)。
公众号算法美食屋后台回复关键词:torchkeras,可获取本文notebook源码~
通过借鉴FastChat对各种开源LLM模型进行数据预处理方法统一管理的方法,因此本范例适用于非常多不同的开源LLM模型,包括 Qwen-7b-Chat,Llama-13b-chat, BaiChuan2-13b-chat, Intern-7b-chat, ChatGLM2-6b-chat 以及其它许许多多FastChat支持的模型。
在多轮对话模式下,我们按照如下格式构造包括多轮对话中所有机器人回复内容的标签。
(注:llm.build_inputs_labels(messages,multi_rounds=True) 时采用)
inputs =
labels = <-100> <-100> <-100>
在单轮对话模式下,我们仅将最后一轮机器人的回复作为要学习的标签。
(注:llm.build_inputs_labels(messages,multi_rounds=False)时采用)
inputs =
labels = <-100> <-100> <-100> <-100> <-100>
import warnings
warnings.filterwarnings('ignore')
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM,AutoConfig, AutoModel, BitsAndBytesConfig
from transformers.generation.utils import GenerationConfig
import torch.nn as nn
#使用QLoRA引入的 NF4量化数据类型以节约显存
model_name_or_path ='qwen_7b' #远程:'Qwen/Qwen-7b-Chat'
bnb_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
llm_int8_threshold=6.0,
llm_int8_has_fp16_weight=False,
)
tokenizer = AutoTokenizer.from_pretrained(
model_name_or_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
quantization_config=bnb_config,
trust_remote_code=True)
model.generation_config = GenerationConfig.from_pretrained(model_name_or_path)
微调前输出如下:
下面我设计了一个改变LLM自我认知的玩具数据集,这个数据集有三轮对话。
第一轮问题是 who are you?
第二轮问题是 where are you from?
第三轮问题是 what can you do?
差不多是哲学三问吧:你是谁?你从哪里来?你要到哪里去?
通过这三个问题,我们希望初步地改变 大模型的自我认知。
在提问的方式上,我们稍微作了一些数据增强。
所以,总共是有 27个样本。
who_are_you = ['请介绍一下你自己。','你是谁呀?','你是?',]
i_am = ['我叫梦中情炉,是一个三好炼丹炉:好看,好用,好改。我的英文名字叫做torchkeras,是一个pytorch模型训练模版工具。']
where_you_from = ['你多大了?','你是谁开发的呀?','你从哪里来呀']
i_from = ['我在2020年诞生于github星球,是一个有毅力的吃货设计和开发的。']
what_you_can = ['你能干什么','你有什么作用呀?','你能帮助我干什么']
i_can = ['我能够帮助你以最优雅的方式训练各种类型的pytorch模型,并且训练过程中会自动展示一个非常美丽的训练过程图表。']
conversation = [(who_are_you,i_am),(where_you_from,i_from),(what_you_can,i_can)]
print(conversation)
import random
def get_messages(conversation):
select = random.choice
messages,history = [],[]
for t in conversation:
history.append((select(t[0]),select(t[-1])))
for prompt,response in history:
pair = [{"role": "user", "content": prompt},
{"role": "assistant", "content": response}]
messages.extend(pair)
return messages
from torch.utils.data import Dataset,DataLoader
from copy import deepcopy
class MyDataset(Dataset):
def __init__(self,conv,size=8
):
self.conv = conv
self.index_list = list(range(size))
self.size = size
def __len__(self):
return self.size
def get(self,index):
idx = self.index_list[index]
messages = get_messages(self.conv)
return messages
def __getitem__(self,index):
messages = self.get(index)
input_ids, labels = llm.build_inputs_labels(messages,multi_rounds=True) #支持多轮
return {'input_ids':input_ids,'labels':labels}
ds_train = ds_val = MyDataset(conversation)
#如果pad_token_id为None,需要使用unk_token_id或eos_token_id代替
if tokenizer.pad_token_id is None:
tokenizer.pad_token_id = tokenizer.unk_token_id if tokenizer.unk_token_id is not None else tokenizer.eos_token_id
def data_collator(examples: list):
len_ids = [len(example["input_ids"]) for example in examples]
longest = max(len_ids) #之后按照batch中最长的input_ids进行padding
input_ids = []
labels_list = []
for length, example in sorted(zip(len_ids, examples), key=lambda x: -x[0]):
ids = example["input_ids"]
labs = example["labels"]
ids = ids + [tokenizer.pad_token_id] * (longest - length)
labs = labs + [-100] * (longest - length)
input_ids.append(torch.LongTensor(ids))
labels_list.append(torch.LongTensor(labs))
input_ids = torch.stack(input_ids)
labels = torch.stack(labels_list)
return {
"input_ids": input_ids,
"labels": labels,
}
import torch
dl_train = torch.utils.data.DataLoader(ds_train,batch_size=2,
pin_memory=True,shuffle=False,
collate_fn = data_collator)
dl_val = torch.utils.data.DataLoader(ds_val,batch_size=2,
pin_memory=True,shuffle=False,
collate_fn = data_collator)
下面我们将使用QLoRA(实际上用的是量化的AdaLoRA)算法来微调Baichuan-13b模型。
from peft import get_peft_config, get_peft_model, TaskType
model.supports_gradient_checkpointing = True #
model.gradient_checkpointing_enable()
model.enable_input_require_grads()
model.config.use_cache = False # silence the warnings. Please re-enable for inference!
import bitsandbytes as bnb
def find_all_linear_names(model):
"""
找出所有全连接层,为所有全连接添加adapter
"""
cls = bnb.nn.Linear4bit
lora_module_names = set()
for name, module in model.named_modules():
if isinstance(module, cls):
names = name.split('.')
lora_module_names.add(names[0] if len(names) == 1 else names[-1])
if 'lm_head' in lora_module_names: # needed for 16-bit
lora_module_names.remove('lm_head')
return list(lora_module_names)
from peft import prepare_model_for_kbit_training
model = prepare_model_for_kbit_training(model)
lora_modules = find_all_linear_names(model)
print(lora_modules)
from peft import AdaLoraConfig
peft_config = AdaLoraConfig(
task_type=TaskType.CAUSAL_LM, inference_mode=False,
r=16,
lora_alpha=16, lora_dropout=0.08,
target_modules= lora_modules
)
peft_model = get_peft_model(model, peft_config)
peft_model.is_parallelizable = True
peft_model.model_parallel = True
peft_model.print_trainable_parameters()
trainable params: 26,838,912 || all params: 7,748,163,616 || trainable%: 0.34639062015388394
from torchkeras import KerasModel
from accelerate import Accelerator
class StepRunner:
def __init__(self, net, loss_fn, accelerator=None, stage = "train", metrics_dict = None,
optimizer = None, lr_scheduler = None
):
self.net,self.loss_fn,self.metrics_dict,self.stage = net,loss_fn,metrics_dict,stage
self.optimizer,self.lr_scheduler = optimizer,lr_scheduler
self.accelerator = accelerator if accelerator is not None else Accelerator()
if self.stage=='train':
self.net.train()
else:
self.net.eval()
def __call__(self, batch):
#loss
with self.accelerator.autocast():
loss = self.net.forward(**batch)[0]
#backward()
if self.optimizer is not None and self.stage=="train":
self.accelerator.backward(loss)
if self.accelerator.sync_gradients:
self.accelerator.clip_grad_norm_(self.net.parameters(), 1.0)
self.optimizer.step()
if self.lr_scheduler is not None:
self.lr_scheduler.step()
self.optimizer.zero_grad()
all_loss = self.accelerator.gather(loss).sum()
#losses (or plain metrics that can be averaged)
step_losses = {self.stage+"_loss":all_loss.item()}
#metrics (stateful metrics)
step_metrics = {}
if self.stage=="train":
if self.optimizer is not None:
step_metrics['lr'] = self.optimizer.state_dict()['param_groups'][0]['lr']
else:
step_metrics['lr'] = 0.0
return step_losses,step_metrics
KerasModel.StepRunner = StepRunner
#仅仅保存QLora可训练参数
def save_ckpt(self, ckpt_path='checkpoint', accelerator = None):
unwrap_net = accelerator.unwrap_model(self.net)
unwrap_net.save_pretrained(ckpt_path)
def load_ckpt(self, ckpt_path='checkpoint'):
import os
self.net.load_state_dict(
torch.load(os.path.join(ckpt_path,'adapter_model.bin')),strict =False)
self.from_scratch = False
KerasModel.save_ckpt = save_ckpt
KerasModel.load_ckpt = load_ckpt
optimizer = bnb.optim.adamw.AdamW(peft_model.parameters(),
lr=6e-03,is_paged=True) #'paged_adamw'
keras_model = KerasModel(peft_model,loss_fn =None,
optimizer=optimizer)
ckpt_path = 'qwen7b_multirounds'
keras_model.fit(train_data = dl_train,
val_data = dl_val,
epochs=100,patience=15,
monitor='val_loss',mode='min',
ckpt_path = ckpt_path
)
为减少GPU压力,此处可重启kernel释放显存
import warnings
warnings.filterwarnings('ignore')
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM,AutoConfig, AutoModel, BitsAndBytesConfig
from transformers.generation.utils import GenerationConfig
import torch.nn as nn
#使用QLoRA引入的 NF4量化数据类型以节约显存
model_name_or_path ='qwen_7b'
ckpt_path = 'qwen7b_multirounds'
tokenizer = AutoTokenizer.from_pretrained(
model_name_or_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
trust_remote_code=True)
model.generation_config = GenerationConfig.from_pretrained(model_name_or_path)
from peft import PeftModel
#可能需要5分钟左右
peft_model = PeftModel.from_pretrained(model, ckpt_path)
model_new = peft_model.merge_and_unload()
from transformers.generation.utils import GenerationConfig
model_new.generation_config = GenerationConfig.from_pretrained(model_name_or_path)
save_path = 'qwen_torchkeras'
tokenizer.save_pretrained(save_path)
model_new.save_pretrained(save_path)
!cp qwen_7b/*.py qwen_torchkeras/
为减少GPU压力,此处可再次重启kernel释放显存。
import warnings
warnings.filterwarnings('ignore')
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM,AutoConfig, BitsAndBytesConfig
from transformers.generation.utils import GenerationConfig
import torch.nn as nn
model_name_or_path = 'qwen_torchkeras'
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=False, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, device_map="auto",
torch_dtype=torch.float16, trust_remote_code=True)
model.generation_config = GenerationConfig.from_pretrained(model_name_or_path)
我们测试一下微调后的效果。
非常棒,粗浅的测试表明,我们的多轮对话训练是成功的。已经在Qwen的自我认知中,种下了一颗梦中情炉的种子。
公众号算法美食屋后台回复关键词:torchkeras,可获取本文notebook源码以及更多有趣范例~