chat llm llama2 chatgpt

更新：

完整合并后文件下载：https://www.codewithgpu.com/m/file/llama2-13b-Chinese-chat

项目在中文sharegpt数据集上训练得到的llama2 Chinese chat 13b，为减轻文件大小负担这里只放出了adapter的权重
请拉取https://huggingface.co/TheBloke/Llama-2-13B-fp16 作为基础权重，使用如下脚步执行合并得到可工作的总权重：

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name_or_path = '/data/TheBloke/Llama-2-13B-fp16'
adapter_name_or_path = '/data/llama2-13b-Chinese-chat'
save_path = '/data/llama2-13b-Chinese-chat_v1'

tokenizer = AutoTokenizer.from_pretrained(
    model_name_or_path,
    trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    model_name_or_path,
    trust_remote_code=True,
    low_cpu_mem_usage=True,
    torch_dtype=torch.float16,
    device_map='auto'
)
print("load model success")
model = PeftModel.from_pretrained(model, adapter_name_or_path)
print("load adapter success")
model = model.merge_and_unload()
print("merge success")

tokenizer.save_pretrained(save_path)
model.save_pretrained(save_path)
print("save done.")

合并后，体验对话：

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch


def main():
    model_name = '/data/llama2-13b-Chinese-chat_v1'

    device = 'cuda'
    max_new_tokens = 500    # 每轮对话最多生成多少个token
    history_max_len = 2000  # 模型记忆的最大token长度
    top_p = 0.9
    temperature = 0.35 # 越大模型越浪
    repetition_penalty = 1.2 # 如果模型出现重复说话可以调节该系数

    # 加载模型
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        trust_remote_code=True,
        low_cpu_mem_usage=True,
        torch_dtype=torch.float16,
        device_map='auto'
    ).to(device).eval()
    tokenizer = AutoTokenizer.from_pretrained(
        model_name,
        trust_remote_code=True,
        # llama不支持fast
        use_fast=False if model.config.model_type == 'llama' else True
    )
    # 记录所有历史记录
    history_token_ids = tokenizer('<s>', return_tensors="pt").input_ids

    # 开始对话
    user_input = input('User：')
    while True:
        user_input = '{}</s>'.format(user_input)
        user_input_ids = tokenizer(user_input, return_tensors="pt", add_special_tokens=False).input_ids
        history_token_ids = torch.concat((history_token_ids, user_input_ids), dim=1)
        model_input_ids = history_token_ids[:, -history_max_len:].to(device)
        with torch.no_grad():
            outputs = model.generate(
                input_ids=model_input_ids, max_new_tokens=max_new_tokens, do_sample=True, top_p=top_p,
                temperature=temperature, repetition_penalty=repetition_penalty, eos_token_id=tokenizer.eos_token_id
            )
        model_input_ids_len = model_input_ids.size(1)
        response_ids = outputs[:, model_input_ids_len:]
        history_token_ids = torch.concat((history_token_ids, response_ids.cpu()), dim=1)
        response = tokenizer.batch_decode(response_ids)
        print("Bot：" + response[0].strip().replace('</s>', ""))
        user_input = input('User：')


if __name__ == '__main__':
    main()

推荐继续二次训练以针对性调优对话效果~

Training procedure

The following bitsandbytes quantization config was used during training:

Framework versions

感谢：