Firefly Qwen 7b

🚀 Firefly微調通義千問Qwen - 7B模型項目

本項目使用Firefly對通義千問Qwen - 7B模型進行微調。訓練數據約有一百多萬輪對話數據，包含項目分享的moss數據以及2萬條school math數據。

更多詳情請見項目Firefly

🚀 快速開始

本項目提供了單輪對話和多輪對話的示例代碼，你可以根據需求進行使用。

💻 使用示例

基礎用法 - 單輪對話

單輪對話不具有對話歷史的記憶功能。

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
"""
單輪對話，不具有對話歷史的記憶功能
"""


def main():
    model_name = 'YeungNLP/firefly-qwen-7b'

    max_new_tokens = 500
    top_p = 0.9
    temperature = 0.35
    repetition_penalty = 1.0
    device = 'cuda'
    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
    )
    # QWenTokenizer比較特殊，pad_token_id、bos_token_id、eos_token_id均為None。eod_id對應的token為<|endoftext|>
    if tokenizer.__class__.__name__ == 'QWenTokenizer':
        tokenizer.pad_token_id = tokenizer.eod_id
        tokenizer.bos_token_id = tokenizer.eod_id
        tokenizer.eos_token_id = tokenizer.eod_id

    text = input('User：')
    while True:
        text = text.strip()
        # chatglm使用官方的數據組織格式
        if model.config.model_type == 'chatglm':
            text = '[Round 1]\n\n問：{}\n\n答：'.format(text)
            input_ids = tokenizer(text, return_tensors="pt", add_special_tokens=False).input_ids.to(device)
        # 為了兼容qwen-7b，因為其對eos_token進行tokenize，無法得到對應的eos_token_id
        else:
            input_ids = tokenizer(text, return_tensors="pt", add_special_tokens=False).input_ids.to(device)
            bos_token_id = torch.tensor([[tokenizer.bos_token_id]], dtype=torch.long).to(device)
            eos_token_id = torch.tensor([[tokenizer.eos_token_id]], dtype=torch.long).to(device)
            input_ids = torch.concat([bos_token_id, input_ids, eos_token_id], dim=1)
        with torch.no_grad():
            outputs = model.generate(
                input_ids=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
            )
        outputs = outputs.tolist()[0][len(input_ids[0]):]
        response = tokenizer.decode(outputs)
        response = response.strip().replace(tokenizer.eos_token, "").strip()
        print("Firefly：{}".format(response))
        text = input('User：')


if __name__ == '__main__':
    main()

高級用法 - 多輪對話

多輪對話具備對話歷史的記憶功能。

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch


def main():
    model_name = 'YeungNLP/firefly-qwen-7b'

    device = 'cuda'
    max_new_tokens = 500    # 每輪對話最多生成多少個token
    history_max_len = 1000  # 模型記憶的最大token長度
    top_p = 0.9
    temperature = 0.35
    repetition_penalty = 1.0

    # 加載模型
    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
    )
    # QWenTokenizer比較特殊，pad_token_id、bos_token_id、eos_token_id均為None。eod_id對應的token為<|endoftext|>
    if tokenizer.__class__.__name__ == 'QWenTokenizer':
        tokenizer.pad_token_id = tokenizer.eod_id
        tokenizer.bos_token_id = tokenizer.eod_id
        tokenizer.eos_token_id = tokenizer.eod_id

    # 記錄所有歷史記錄
    if model.config.model_type != 'chatglm':
        history_token_ids = torch.tensor([[tokenizer.bos_token_id]], dtype=torch.long)
    else:
        history_token_ids = torch.tensor([[]], dtype=torch.long)

    # 開始對話
    utterance_id = 0    # 記錄當前是第幾輪對話，為了契合chatglm的數據組織格式
    user_input = input('User：')
    while True:
        utterance_id += 1
        # chatglm使用官方的數據組織格式
        if model.config.model_type == 'chatglm':
            user_input = '[Round {}]\n\n問：{}\n\n答：'.format(utterance_id, user_input)
            user_input_ids = tokenizer(user_input, return_tensors="pt", add_special_tokens=False).input_ids
        # firefly的數據組織格式
        # 為了兼容qwen-7b，因為其對eos_token進行tokenize，無法得到對應的eos_token_id
        else:
            input_ids = tokenizer(user_input, return_tensors="pt", add_special_tokens=False).input_ids
            eos_token_id = torch.tensor([[tokenizer.eos_token_id]], dtype=torch.long)
            user_input_ids = torch.concat([input_ids, eos_token_id], dim=1)
        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("Firefly：" + response[0].strip().replace(tokenizer.eos_token, ""))
        user_input = input('User：')


if __name__ == '__main__':
    main()