transformers==4.26.1

torch==1.13.1

deepspeed==0.8.2

peft==0.2.

import os
import torch
import random
import datasets
import numpy as np

from tqdm import tqdm
from typing import Dict
from torch.utils.data import DataLoader
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    DataCollatorForSeq2Seq,
    TrainingArguments,
    Trainer
)
from peft import (
    LoraConfig,
    TaskType,
    get_peft_model,
    get_peft_model_state_dict,
    set_peft_model_state_dict
)

def set_random_seed(seed):
    if seed is not None and seed > 0:
        random.seed(seed)
        np.random.seed(seed)
        torch.manual_seed(seed)
        torch.random.manual_seed(seed)
        torch.cuda.manual_seed(seed)
        torch.cuda.manual_seed_all(seed)
        torch.backends.cudnn.deterministic = True

set_random_seed(1234)

dataset = datasets.load_dataset("json", data_files=DATA_PATH)
def tokenize(text: str, add_eos_token=True):
    result = tokenizer(
        text,
        truncation=True,
        max_length=MAX_LENGTH,
        padding=False,
        return_tensors=None)
    # 判断是否要添加eos_token
    if (result["input_ids"][-1] != tokenizer.eos_token_id
        and len(result["input_ids"]) < MAX_LENGTH
        and add_eos_token):
        result["input_ids"].append(tokenizer.eos_token_id)
        result["attention_mask"].append(1)
    result["labels"] = result["input_ids"].copy()
    return result

def preprocess(example: Dict, train_on_inputs: bool = False):
    prompt = example["input"]
    response = example["target"]
    text = PATTERN.format(prompt, response)
    tokenized_inp = tokenize(text)
    # 若train_on_inputs为False，则将label中与input相关的token替换为-100
    if not train_on_inputs:
        tokenized_prompt = tokenize(prompt,add_eos_token=False)
        prompt_tokens_len = len(tokenized_prompt["input_ids"])
        tokenized_inp["labels"] = [-100]*prompt_tokens_len + tokenized_inp["labels"][prompt_tokens_len:]
    return tokenized_inp

train_data = dataset["train"].shuffle().map(preprocess, remove_columns=["id", "input", "target"])
print(train_data[0])

collate_fn = DataCollatorForSeq2Seq(tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True)

# 转换模型
model = get_peft_model(model, lora_config)
model.config.use_cache = False
old_state_dict = model.state_dict
model.state_dict = (
    lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
).__get__(model, type(model))
# 打印模型中的可训练参数
model.print_trainable_parameters()

args = TrainingArguments(
    output_dir=OUTPUT_DIR, # checkpoint的存储目录
    per_device_train_batch_size=BATCH_SIZE, # 单设备上的batch size
    gradient_accumulation_steps=GRADIENT_ACCUMULATION_STEPS, # 梯度累加的step数
    warmup_steps=100,
    num_train_epochs=EPOCHS,
    learning_rate=LEARNING_RATE,
    fp16=True, # 使用混合精度训练
    logging_steps=50,
    evaluation_strategy="no", # 不进行评估
    save_strategy="steps",
    save_steps=2000, # 保存checkpoint的step数
    save_total_limit=5, # 最多保存5个checkpoint
    deepspeed=DS_CONFIG
)

trainer = Trainer(
    model=model,
    train_dataset=train_data,
    eval_dataset=None,
    args=args,
    data_collator=collate_fn
)
trainer.train()
model.save_pretrained("best_model")