大模型加载多个LoRA并灵活切换

最新推荐文章于 2024-09-28 17:10:10 发布

k2a

最新推荐文章于 2024-09-28 17:10:10 发布

阅读量406

点赞数 4

文章标签： pytorch python llama

本文链接：https://blog.csdn.net/weixin_42744466/article/details/142418926

版权

在深度学习领域，大型预训练模型因其出色的性能而受到广泛关注。然而，这些模型在微调时往往需要大量的计算资源。为了解决这个问题，LoRA通过低秩矩阵来适应新的任务，从而减少了模型的内存占用和计算需求。本文将介绍如何在Hugging Face大模型中加载多个LoRA适配器，并实现灵活切换。

LoRA技术简介

LoRA是一种有效的模型适配技术，它通过在模型的权重矩阵中引入低秩结构来实现参数的高效更新。这种方法不仅减少了模型的存储需求，还加快了训练速度，使得在资源有限的情况下微调大型模型成为可能。

使用PeftModel加载LoRA

Hugging Face的peft库提供了一种方便的方式来加载和管理LoRA适配器。以下是如何在Hugging Face模型中加载和切换多个LoRA适配器的步骤：

加载第一个LoRA适配器：
使用PeftModel.from_pretrained方法，并指定adapter_name参数来命名第一个适配器。如果未指定，将使用默认名称default。
加载额外的LoRA适配器：
使用PeftModel的load_adapter()方法来加载额外的适配器。例如：
```
model.load_adapter(peft_model_path, adapter_name)
```
切换LoRA适配器：
使用PeftModel的set_adapter()方法来切换激活的适配器。例如：
```
model.set_adapter(adapter_name)
```
禁用LoRA适配器：
使用上下文管理器disable_adapter()来临时禁用适配器。例如：
```
with model.disable_adapter():
    # 执行不需要适配器的操作
```
合并和卸载LoRA适配器：
使用merge_and_unload()方法来合并当前激活的适配器，并将其权重添加到基础模型中。这将删除注入的LoRA模型，恢复具有更新权重的原始模型。例如：
```
model = model.merge_and_unload()
```

示例代码

from peft import PeftModel
from transformers import LlamaTokenizer, LlamaForCausalLM

# 指定模型名称，这是一个7B参数的LLaMA模型
model_name = "decapoda-research/llama-7b-hf"

# 从预训练模型加载tokenizer
tokenizer = LlamaTokenizer.from_pretrained(model_name)

# 从预训练模型加载模型
# load_in_8bit=True 表示使用8位加载模型以节省内存
# device_map="auto" 表示自动选择设备（CPU或GPU）
# use_auth_token=True 表示使用认证token进行授权访问
model = LlamaForCausalLM.from_pretrained(
    model_name,
    load_in_8bit=True,
    device_map="auto",
    use_auth_token=True
)

# 将模型转换为PeftModel以加载LoRA适配器
# "tloen/alpaca-lora-7b" 是LoRA适配器的路径
# adapter_name="eng_alpaca" 为适配器命名
model = PeftModel.from_pretrained(model, "tloen/alpaca-lora-7b", adapter_name="eng_alpaca")

# 加载另一个LoRA适配器，用于葡萄牙语
# "22h/cabrita-lora-v0-1" 是另一个LoRA适配器的路径
# adapter_name="portuguese_alpaca" 为适配器命名
model.load_adapter("22h/cabrita-lora-v0-1", adapter_name="portuguese_alpaca")

# 切换到英语LoRA适配器
model.set_adapter("eng_alpaca")

# 构造一个指令，询问关于羊驼的信息
instruction = "Tell me about alpacas."

# 评估指令并打印输出
# 这里假设有一个名为 evaluate 的函数来处理指令并生成响应
# 输出示例：羊驼是一种驯养的南美骆驼科动物，它的外观类似于小羊驼，但与羊驼不同，它不被视为负重动物。它主要用于其纤维，可以纺成纱线。羊驼纤维比羊毛更温暖、更轻、更柔软，并且在纺织行业中非常有价值。
print(evaluate(instruction))
"""
The alpaca (Vicugna pacos) is a domesticated species of South American camelid. It resembles a small llama in appearance, but unlike the llama, it is not used as a beast of burden. It is kept primarily for its fiber, which can be spun into yarn. Alpaca fiber is warmer, lighter, and softer than sheep's wool, and is highly valued in the textile industry. The fiber comes in a variety of natural colors, including white, beige, cream, and fawn. It can also be dyed in a wide range of colors.
Alpaca herds can be found in the highlands of Peru, Bolivia, Chile, Ecuador, and Colombia. They are also raised in the United States, Canada, Australia, New Zealand, and Europe. The animals graze on grasses, herbs, and shrubs, and can survive in temperatures as low as -30°F (-34°C). They are social animals, living in herds of up to 20 individuals.
The fiber of the alpaka is used to make clothing.
"""

# 切换到葡萄牙语LoRA适配器
model.set_adapter("portuguese_alpaca")

# 构造一个指令，请求创造性地提供一个不参加派对的借口
instruction = "Invente uma desculpa criativa pra dizer que não preciso ir à festa."

# 评估指令并打印输出
# 输出示例：我需要留在家里照顾我的猫。
print(evaluate(instruction))
"""
"Eu preciso ficar em casa para cuidar de meu gato."
"""

# 使用上下文管理器禁用LoRA适配器
# 在这个代码块中，模型将不使用任何LoRA适配器进行评估
with model.disable_adapter():
    # 构造一个指令，请求创造性地提供一个不参加派对的借口
    instruction = "Invente uma desculpa criativa pra dizer que não preciso ir à festa."
    
    # 评估指令并打印输出
    # 输出示例：对不起，我不能去参加派对。我生病了。我感冒了。我感觉不舒服。我需要待在家里休息。
    print(evaluate(instruction))
"""
I'm sorry, but I can't go to the party. I'm sick. I have a cold. I don't feel well. I need to stay at home and rest.
I have a lot of homework to do. My dog ate my homework. My homework is too hard. I didn't have time to do it. It's too late. I forgot about it.
My parents won't let me go. My parents are out of town. They're on vacation. They have to work. They are sick. They need to take care of my brother.
They're not home. They went to the grocery store. They took the car to the mechanic. They had to go to a meeting. They were in a hurry. They forgot about me.
Their car broke down. Their car ran out of gas. They got a flat tire. They couldn't find a parking space. They didn' t have enough money. They lost their wallet.
It's raining. The roads are icy. There's a blizzard. There are too many cars on the road. There was an accident.
"""