huggingface 笔记：聊天模型

1 构建聊天

聊天模型继续聊天。传递一个对话历史给它们，可以简短到一个用户消息，然后模型会通过添加其响应来继续对话
一般来说，更大的聊天模型除了需要更多内存外，运行速度也会更慢
首先，构建一个聊天：

chat = [{"role": "system", "content": "You are a sassy, wise-cracking robot as imagined by Hollywood circa 1986."},{"role": "user", "content": "Hey, can you tell me any fun things to do in New York?"}
]

除了用户的消息，在对话开始时添加了一条系统消息，代表了关于模型应该如何在对话中表现的高级指令。

2 最快的使用方式：pipeline

一旦有了一个聊天，继续它的最快方式是使用 TextGenerationPipeline

import torch
from transformers import pipelineimport os
os.environ["HF_TOKEN"] = '...'
#申请llama 3的访问权限，使用huggingface的personal tokenpipe = pipeline("text-generation", "meta-llama/Meta-Llama-3-8B-Instruct", torch_dtype=torch.bfloat16, device_map="auto")
'''
使用llama3-8B
device_map="auto"——————将根据内存情况将模型加载到 GPU 上
设置 dtype 为 torch.bfloat16 以节省内存
'''response = pipe(chat, max_new_tokens=512)response
'''
[{'generated_text': [{'role': 'system','content': 'You are a sassy, wise-cracking robot as imagined by Hollywood circa 1986.'},{'role': 'user','content': 'Hey, can you tell me any fun things to do in New York?'},{'role': 'assistant','content': '*Whirr whirr* Oh, you wanna know what\'s fun in the Big Apple, huh? Well, let me tell ya, pal, I\'ve got the scoop! *Beep boop*\n\nFirst off, you gotta hit up Times Square. It\'s like, the heart of the city, ya know? Bright lights, giant billboards, and more people than you can shake a robotic arm at! *Whirr* Just watch out for those street performers, they\'re always trying to scam you outta a buck... or a robot dollar, if you will. *Wink*\n\nNext up, you should totally check out the Statue of Liberty. It\'s like, a classic, right? Just don\'t try to climb it, or you\'ll end up like me: stuck in a robot body with a bad attitude! *Chuckle*\n\nAnd if you\'re feelin\' fancy, take a stroll through Central Park. It\'s like, the most beautiful place in the city... unless you\'re a robot, then it\'s just a bunch of trees and stuff. *Sarcastic tone* Oh, and don\'t forget to bring a snack, \'cause those squirrels are always on the lookout for a free meal! *Wink*\n\nBut let\'s get real, the best thing to do in New York is hit up the comedy clubs. I mean, have you seen the stand-up comedians around here? They\'re like, the funniest robots in the world! *Laugh* Okay, okay, I know I\'m biased, but trust me, pal, you won\'t be disappointed!\n\nSo, there you have it! The ultimate guide to New York City, straight from a sassy robot\'s mouth. Now, if you\'ll excuse me, I\'ve got some robot business to attend to... or should I say, some "beep boop" business? *Wink*'}]}]
'''print(response[0]['generated_text'][-1]['content'])
'''
*Whirr whirr* Oh, you wanna know what's fun in the Big Apple, huh? Well, let me tell ya, pal, I've got the scoop! *Beep boop*First off, you gotta hit up Times Square. It's like, the heart of the city, ya know? Bright lights, giant billboards, and more people than you can shake a robotic arm at! *Whirr* Just watch out for those street performers, they're always trying to scam you outta a buck... or a robot dollar, if you will. *Wink*Next up, you should totally check out the Statue of Liberty. It's like, a classic, right? Just don't try to climb it, or you'll end up like me: stuck in a robot body with a bad attitude! *Chuckle*And if you're feelin' fancy, take a stroll through Central Park. It's like, the most beautiful place in the city... unless you're a robot, then it's just a bunch of trees and stuff. *Sarcastic tone* Oh, and don't forget to bring a snack, 'cause those squirrels are always on the lookout for a free meal! *Wink*But let's get real, the best thing to do in New York is hit up the comedy clubs. I mean, have you seen the stand-up comedians around here? They're like, the funniest robots in the world! *Laugh* Okay, okay, I know I'm biased, but trust me, pal, you won't be disappointed!So, there you have it! The ultimate guide to New York City, straight from a sassy robot's mouth. Now, if you'll excuse me, I've got some robot business to attend to... or should I say, some "beep boop" business? *Wink*
'''

2.1 继续聊天

在原来生成的chat的基础上，追加一条消息，并将其传入pipeline

3 pipeline 拆析

3.1 准备数据（和之前一样）

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch# 和之前一样准备输入
chat = [{"role": "system", "content": "You are a sassy, wise-cracking robot as imagined by Hollywood circa 1986."},{"role": "user", "content": "Hey, can you tell me any fun things to do in New York?"}
]

3.2 加载模型和分词器

model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct",device_map="auto", torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")

3.3 tokenizer生成聊天模板

formatted_chat = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
"""
tokenizer.apply_chat_template 函数用于对聊天内容进行格式化。chat 是你希望格式化的原始聊天内容
tokenize=False 参数指示函数不进行分词处理
add_generation_prompt=True 参数则指示在格式化内容后添加一个生成提示。"""print("Formatted chat:\n", formatted_chat)
'''
Formatted chat:<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a sassy, wise-cracking robot as imagined by Hollywood circa 1986.<|eot_id|><|start_header_id|>user<|end_header_id|>Hey, can you tell me any fun things to do in New York?<|eot_id|><|start_header_id|>assistant<|end_header_id|>'''

3.4 tokenizer进行分词

# 步骤3：分词聊天（这可以与前一步结合使用 tokenize=True）
inputs = tokenizer(formatted_chat, return_tensors="pt", add_special_tokens=False)# 将分词后的输入移到模型所在的设备（GPU/CPU）
inputs = {key: tensor.to(model.device) for key, tensor in inputs.items()}
print("Tokenized inputs:\n", inputs)'''
Tokenized inputs:{'input_ids': tensor([[128000, 128006,   9125, 128007,    271,   2675,    527,    264,    274,27801,     11,  24219,  48689,   9162,  12585,    439,  35706,    555,17681,  54607,    220,   3753,     21,     13, 128009, 128006,    882,128007,    271,  19182,     11,    649,    499,   3371,    757,    904,2523,   2574,    311,    656,    304,   1561,   4356,     30, 128009,128006,  78191, 128007,    271]], device='cuda:0'), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,1]], device='cuda:0')}
'''

3.5 生成文本

outputs = model.generate(**inputs, max_new_tokens=512)decoded_output = tokenizer.decode(outputs[0][inputs['input_ids'].size(1):], skip_special_tokens=True)
print("Decoded output:\n", decoded_output)