DeepSeek 微调——LoRA 与全参数

DeepSeek 作为强大的大模型，提供了优质的基础能力，但在某些特定任务上，直接使用预训练模型可能无法满足需求。本篇文章将介绍 LoRA（Low-Rank Adaptation）、全参数微调 等微调策略，并提供详细的代码示例，帮助开发者高效定制 DeepSeek 以适应特定任务。

LoRA 微调 DeepSeek

LoRA（Low-Rank Adaptation）是一种高效的参数高效微调方法。其核心思想是在预训练权重的基础上添加可训练的低秩适配层，从而减少计算开销。

安装依赖

pip install torch transformers peft accelerate

加载 DeepSeek 模型

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "deepseek-ai/deepseek-mistral-7b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

LoRA 配置

from peft import LoraConfig, get_peft_model

# 配置 LoRA 训练参数
lora_config = LoraConfig(
    r=8,  # 低秩矩阵的秩
    lora_alpha=32,  # LoRA 缩放因子
    lora_dropout=0.1,  # dropout 率
    bias="none",
    target_modules=["q_proj", "v_proj"],  # 仅对部分层进行微调
)

# 应用 LoRA
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()

训练 DeepSeek LoRA

from transformers import Trainer, TrainingArguments

training_args = TrainingArguments(
    output_dir="./lora_model",
    per_device_train_batch_size=4,
    num_train_epochs=3,
    save_steps=100,
    logging_dir="./logs",
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=my_train_dataset,  # 替换为你的数据集
)
trainer.train()

加载数据集

from datasets import Dataset

# 加载
datasets = Dataset.load_from_disk("./wiki_cn_filtered/")
# 划分

# 数据处理
def process_func(example):
    MAX_LENGTH = 256
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer("\n".join(["Human: " + example["instruction"], example["input"]]).strip() + "\n\nAssistant: ")
    response = tokenizer(example["output"] + tokenizer.eos_token)
    input_ids = instruction["input_ids"] + response["input_ids"]
    attention_mask = instruction["attention_mask"] + response["attention_mask"]
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"]
    if len(input_ids) > MAX_LENGTH:
        input_ids = input_ids[:MAX_LENGTH]
        attention_mask = attention_mask[:MAX_LENGTH]
        labels = labels[:MAX_LENGTH]
    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels
    }

my_train_dataset = datasets.map(process_func, batched=True, remove_columns=datasets.column_names)

全参数微调 DeepSeek

全参数微调适用于 数据量大、任务复杂 的场景，需要对模型所有参数进行更新，计算资源消耗较高。

环境准备

pip install deepspeed transformers torch

加载 DeepSeek 模型

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "deepseek-ai/deepseek-mistral-7b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

配置训练参数

from transformers import TrainingArguments

training_args = TrainingArguments(
    output_dir="./full_finetune",
    per_device_train_batch_size=2,
    num_train_epochs=3,
    save_strategy="epoch",
    report_to="tensorboard",
    logging_dir="./logs",
    deepspeed="./ds_config.json"  # DeepSpeed 加速
)

训练模型

from transformers import Trainer

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=my_train_dataset,  # 替换为你的数据集
)
trainer.train()