Lora微LLAMA模型实战

引言

本文介绍如何复现Alpaca-lora，即基于alpaca数据集用lora方法微调Llama模型。

环境准备

实验环境用的是lanyun，新用户点击注册可以送算力。

下载huggingface上的模型是一个令人头疼的问题，但在lanyun上可以通过在终端运行source /etc/network_turbo 配置加速下载：

如上图，速度还是很快的。

如果不是lanyun上可以尝试：export HF_ENDPOINT=https://hf-mirror.com ，但可能不太稳定。

Cuda版本、pytorch版本如下：

安装依赖

将下面的内容复制到requirements.txt中：

accelerate
appdirs
loralib
bitsandbytes
black
black[jupyter]
datasets
fire
peft
transformers>=4.28.0
sentencepiece
gradio

比如我这里复制到 /root/lanyun-tmp/目录下。然后依次执行：

source /etc/network_turbo 
pip install -r requirements.txt

其输出可能为：

Collecting appdirs (from -r requirements.txt (line 2))
...
Successfully installed aiofiles-23.2.1 aiohappyeyeballs-2.6.1 aiohttp-3.11.13 aiosignal-1.3.2 annotated-types-0.7.0 appdirs-1.4.4 async-timeout-5.0.1 bitsandbytes-0.45.3 black-25.1.0 click-8.1.8 datasets-3.4.0 dill-0.3.8 fastapi-0.115.11 ffmpy-0.5.0 fire-0.7.0 frozenlist-1.5.0 gradio-5.21.0 gradio-client-1.7.2 groovy-0.1.2 loralib-0.1.2 markdown-it-py-3.0.0 mdurl-0.1.2 multidict-6.1.0 multiprocess-0.70.16 mypy-extensions-1.0.0 orjson-3.10.15 pandas-2.2.3 pathspec-0.12.1 peft-0.14.0 propcache-0.3.0 pyarrow-19.0.1 pydantic-2.10.6 pydantic-core-2.27.2 pydub-0.25.1 python-multipart-0.0.20 pytz-2025.1 requests-2.32.3 rich-13.9.4 ruff-0.11.0 safehttpx-0.1.6 semantic-version-2.10.0 sentencepiece-0.2.0 shellingham-1.5.4 starlette-0.46.1 termcolor-2.5.0 tokenize-rt-6.1.0 tomlkit-0.13.2 tqdm-4.67.1 typer-0.15.2 tzdata-2025.1 uvicorn-0.34.0 websockets-15.0.1 xxhash-3.5.0 yarl-1.18.3

等待依赖下载完毕。

模型格式转换

首先需要将LLaMA原始权重文件转换为Transformers库对应的模型文件格式，但我们也可以选择别人转换好的，比如 https://huggingface.co/dfurman/LLaMA-7B：

LLaMA-7B is a base model for text generation with 6.7B parameters and a 1T token training corpus. It was built and released by the FAIR team at Meta AI alongside the paper "LLaMA: Open and Efficient Foundation Language Models".This model repo was converted to work with the transformers package. It is under a bespoke non-commercial license, please see the LICENSE file for more details.

下面编写代码下载模型：

download_model.py:

import transformers
import torchmodel_name = "dfurman/llama-7b"tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
streamer = transformers.TextStreamer(tokenizer)model = transformers.LlamaForCausalLM.from_pretrained(model_name,torch_dtype=torch.bfloat16,device_map="auto"
)

等待执行完毕我们就下载好了想要的模型格式。

训练

单卡训练

克隆alpaca-lora项目的源码：

git clone https://github.com/tloen/alpaca-lora.git
cd alpaca-lora

修改alpaca-lora目录下的 finetune.py文件，将prepare_model_for_int8_training替换为prepare_model_for_kbit_training，不然新版(0.14.0)的peft会报错。

然后在该目录下执行：

python finetune.py \--base_model 'dfurman/llama-7b' \--data_path 'yahma/alpaca-cleaned' \--output_dir './lora-alpaca'

这里的dfurman/llama-7b是我们刚才下载好的模型；yahma/alpaca-cleaned，参考4项目任务原始的alpaca数据集质量不高，因此他们对该数据集进行了一个清理，得到了更高质量的alpaca-cleaned。

Training Alpaca-LoRA model with params:
base_model: dfurman/llama-7b
data_path: yahma/alpaca-cleaned
output_dir: ./lora-alpaca
batch_size: 128
micro_batch_size: 4
num_epochs: 3
learning_rate: 0.0003
cutoff_len: 256
val_set_size: 2000
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules: ['q_proj', 'v_proj']
train_on_inputs: True
add_eos_token: False
group_by_length: False
wandb_project: 
wandb_run_name: 
wandb_watch: 
wandb_log_model: 
resume_from_checkpoint: False
prompt template: alpaca
...                                                                                   | 5/1164 [01:49<6:57:53, 21.63s/it]

从上面可以看到一些默认的参数配置，但是需要训练7个小时左右，太慢了。finetune.py的代码是支持单机多卡的，因此我们重新创建一个4卡的实例。

多卡训练

下面来一步一步在Lanyun上操作一下：

这里我们选择了4卡，并且选择好了Cuda等版本。

等待创建完毕后：

点击JupyterLab进入代码环境。

进入后我们可以看到这样的解码，这里直接点击Terminal进入终端环境。

第一步执行：

source /etc/network_turbo

第二步克隆项目：

git clone https://github.com/tloen/alpaca-lora.git
cd alpaca-lora

第三步安装依赖：

pip install -r requirements.txt

第四步修改alpaca-lora目录下的 finetune.py文件，将prepare_model_for_int8_training替换为prepare_model_for_kbit_training，主要修改有两处。

第五步利用数据并行，在4卡上进行训练：

nohup torchrun --nproc_per_node=4 --master_port=29005 finetune.py \--base_model 'dfurman/llama-7b' \--data_path 'yahma/alpaca-cleaned' \--num_epochs=10 \--cutoff_len=512 \--group_by_length \--output_dir='./lora-alpaca' \--lora_target_modules='[q_proj,k_proj,v_proj,o_proj]' \--lora_r=16 \--micro_batch_size=8 > output.log 2>&1 &

同时这里参考 https://huggingface.co/tloen/alpaca-lora-7b 上的例子调整下参数。

[2025-03-16 16:18:21,847] torch.distributed.run: [WARNING] 
[2025-03-16 16:18:21,847] torch.distributed.run: [WARNING] *****************************************
[2025-03-16 16:18:21,847] torch.distributed.run: [WARNING] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
[2025-03-16 16:18:21,847] torch.distributed.run: [WARNING] *****************************************
Training Alpaca-LoRA model with params:
base_model: dfurman/llama-7b
data_path: yahma/alpaca-cleaned
output_dir: ./lora-alpaca
batch_size: 128
micro_batch_size: 8
num_epochs: 10
learning_rate: 0.0003
cutoff_len: 512
val_set_size: 2000
lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj']
train_on_inputs: True
add_eos_token: False
group_by_length: True
wandb_project: 
wandb_run_name: 
wandb_watch: 
wandb_log_model: 
resume_from_checkpoint: False
prompt template: alpaca
...

这次会自动从huggingface上下载模型dfurman/llama-7b并开始单机多卡训练。

 trainable params: 16,777,216 || all params: 6,755,192,832 || trainable%: 0.24840%|▌                                                                                  4/3880 [00:30<6:16:57,  7.08s/it]

显存使用如上，每个卡都用了16.8G。

4卡训练了5个小时左右，终于训练好了。

推理

在仓库根目录下执行：

python generate.py     --load_8bit     --base_model 'dfurman/llama-7b'     --lora_weights 'lora-alpaca'

AttributeError: module 'gradio' has no attribute 'inputs'

遇到了上面的错误，这是因为仓库的代码有点老，一种比较简单的方法是降低版本：

pip install gradio==3.43.1

Running on local URL:  http://0.0.0.0:7860To create a public link, set `share=True` in `launch()`.
IMPORTANT: You are using gradio version 3.43.1, however version 4.44.1 is available, please upgrade.
--------

顶着各种警告，终于跑起来了。

但是我们在Lanyun上无法访问这个端口，如果是个人电脑可以直接打开了。要在Lanyun上访问，需要通过端口映射开放端口：

找到generate.py中195行这句代码，添加指定的server_port：

    ).queue().launch(server_name="0.0.0.0", share=share_gradio, server_port=17860)

启动成功后点击端口映射中的访问即可：

finetune.py文件分析

该项目下的finetune.py脚本值得我们学习一下：

import os
import sys
from typing import Listimport fire
import torch
import transformers
from datasets import load_dataset"""
Unused imports:
import torch.nn as nn
import bitsandbytes as bnb
"""from peft import (LoraConfig,get_peft_model,get_peft_model_state_dict,prepare_model_for_int8_training,set_peft_model_state_dict,
)
from transformers import LlamaForCausalLM, LlamaTokenizer
# 自定义的提示词工具
from utils.prompter import Prompterdef train(# model/data paramsbase_model: str = "",  # the only required argumentdata_path: str = "yahma/alpaca-cleaned", # 会从huggingface上去下载output_dir: str = "./lora-alpaca", # 训练超参batch_size: int = 128, # 梯度累积后的批大小micro_batch_size: int = 4, # 实际的批大小num_epochs: int = 3, # 训练轮次learning_rate: float = 3e-4, cutoff_len: int = 256, # 最长长度val_set_size: int = 2000, # 验证集大小# lora 超参lora_r: int = 8, # 低秩矩阵的维度lora_alpha: int = 16, # 低秩矩阵的比例因子lora_dropout: float = 0.05, # LoRA层的dropout概率# 应用lora到 query 和 value的投影层(Linear层)lora_target_modules: List[str] = ["q_proj","v_proj",],# llm 超参train_on_inputs: bool = True,  # if False, masks out inputs in lossadd_eos_token: bool = False,group_by_length: bool = False,  # faster, but produces an odd training loss curve# wandb log 相关参数wandb_project: str = "",wandb_run_name: str = "",wandb_watch: str = "",  # options: false | gradients | allwandb_log_model: str = "",  # options: false | trueresume_from_checkpoint: str = None,  # either training checkpoint or final adapterprompt_template_name: str = "alpaca",  # The prompt template to use, will default to alpaca.
):if int(os.environ.get("LOCAL_RANK", 0)) == 0:print(f"Training Alpaca-LoRA model with params:\n"f"base_model: {base_model}\n"f"data_path: {data_path}\n"f"output_dir: {output_dir}\n"f"batch_size: {batch_size}\n"f"micro_batch_size: {micro_batch_size}\n"f"num_epochs: {num_epochs}\n"f"learning_rate: {learning_rate}\n"f"cutoff_len: {cutoff_len}\n"f"val_set_size: {val_set_size}\n"f"lora_r: {lora_r}\n"f"lora_alpha: {lora_alpha}\n"f"lora_dropout: {lora_dropout}\n"f"lora_target_modules: {lora_target_modules}\n"f"train_on_inputs: {train_on_inputs}\n"f"add_eos_token: {add_eos_token}\n"f"group_by_length: {group_by_length}\n"f"wandb_project: {wandb_project}\n"f"wandb_run_name: {wandb_run_name}\n"f"wandb_watch: {wandb_watch}\n"f"wandb_log_model: {wandb_log_model}\n"f"resume_from_checkpoint: {resume_from_checkpoint or False}\n"f"prompt template: {prompt_template_name}\n")assert (base_model), "Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'"#gradient_accumulation_steps = batch_size // micro_batch_size# 自定义了提示词工具类prompter = Prompter(prompt_template_name)device_map = "auto"# 分布式训练时指定的设备数量world_size = int(os.environ.get("WORLD_SIZE", 1))# 判断是否为分布式训练ddp = world_size != 1if ddp:device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}gradient_accumulation_steps = gradient_accumulation_steps // world_sizeuse_wandb = len(wandb_project) > 0 or ("WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0)if len(wandb_project) > 0:os.environ["WANDB_PROJECT"] = wandb_projectif len(wandb_watch) > 0:os.environ["WANDB_WATCH"] = wandb_watchif len(wandb_log_model) > 0:os.environ["WANDB_LOG_MODEL"] = wandb_log_model# 使用transformers 加载Llama模型model = LlamaForCausalLM.from_pretrained(base_model,load_in_8bit=True,torch_dtype=torch.float16,device_map=device_map,)# 加载分词器tokenizer = LlamaTokenizer.from_pretrained(base_model)tokenizer.pad_token_id = (0  # unk. we want this to be different from the eos token)tokenizer.padding_side = "left"  # Allow batched inferencedef tokenize(prompt, add_eos_token=True):# there's probably a way to do this with the tokenizer settings# but again, gotta move fastresult = tokenizer(prompt,truncation=True,max_length=cutoff_len,padding=False,return_tensors=None,)if (result["input_ids"][-1] != tokenizer.eos_token_idand len(result["input_ids"]) < cutoff_lenand add_eos_token):result["input_ids"].append(tokenizer.eos_token_id)result["attention_mask"].append(1)result["labels"] = result["input_ids"].copy()return resultdef generate_and_tokenize_prompt(data_point):# 得到输入提示词full_prompt = prompter.generate_prompt(data_point["instruction"],data_point["input"],data_point["output"],)tokenized_full_prompt = tokenize(full_prompt)if not train_on_inputs:user_prompt = prompter.generate_prompt(data_point["instruction"], data_point["input"])tokenized_user_prompt = tokenize(user_prompt, add_eos_token=add_eos_token)user_prompt_len = len(tokenized_user_prompt["input_ids"])if add_eos_token:user_prompt_len -= 1tokenized_full_prompt["labels"] = [-100] * user_prompt_len + tokenized_full_prompt["labels"][user_prompt_len:]  # could be sped up, probablyreturn tokenized_full_prompt#  适配 INT8 训练，减少显存占用model = prepare_model_for_int8_training(model)# Lora配置config = LoraConfig(r=lora_r,lora_alpha=lora_alpha,target_modules=lora_target_modules,lora_dropout=lora_dropout,bias="none", task_type="CAUSAL_LM", # 任务类型为因果语言模型)# model = get_peft_model(model, config)if data_path.endswith(".json") or data_path.endswith(".jsonl"):data = load_dataset("json", data_files=data_path)else:data = load_dataset(data_path)# 从断点恢复if resume_from_checkpoint:# Check the available weights and load themcheckpoint_name = os.path.join(resume_from_checkpoint, "pytorch_model.bin")  # Full checkpointif not os.path.exists(checkpoint_name):checkpoint_name = os.path.join(resume_from_checkpoint, "adapter_model.bin")  # only LoRA model - LoRA config above has to fitresume_from_checkpoint = (False  # So the trainer won't try loading its state)# The two files above have a different name depending on how they were saved, but are actually the same.if os.path.exists(checkpoint_name):print(f"Restarting from {checkpoint_name}")adapters_weights = torch.load(checkpoint_name)set_peft_model_state_dict(model, adapters_weights)else:print(f"Checkpoint {checkpoint_name} not found")model.print_trainable_parameters()  # Be more transparent about the % of trainable params.if val_set_size > 0:train_val = data["train"].train_test_split(test_size=val_set_size, shuffle=True, seed=42)train_data = (train_val["train"].shuffle().map(generate_and_tokenize_prompt))val_data = (train_val["test"].shuffle().map(generate_and_tokenize_prompt))else:train_data = data["train"].shuffle().map(generate_and_tokenize_prompt)val_data = Noneif not ddp and torch.cuda.device_count() > 1:# keeps Trainer from trying its own DataParallelism when more than 1 gpu is availablemodel.is_parallelizable = Truemodel.model_parallel = Truetrainer = transformers.Trainer(model=model,train_dataset=train_data,eval_dataset=val_data,args=transformers.TrainingArguments(per_device_train_batch_size=micro_batch_size,gradient_accumulation_steps=gradient_accumulation_steps,warmup_steps=100,num_train_epochs=num_epochs,learning_rate=learning_rate,fp16=True,logging_steps=10,optim="adamw_torch",evaluation_strategy="steps" if val_set_size > 0 else "no",save_strategy="steps",eval_steps=200 if val_set_size > 0 else None,save_steps=200,output_dir=output_dir,save_total_limit=3,load_best_model_at_end=True if val_set_size > 0 else False,ddp_find_unused_parameters=False if ddp else None,group_by_length=group_by_length,report_to="wandb" if use_wandb else None,run_name=wandb_run_name if use_wandb else None,),data_collator=transformers.DataCollatorForSeq2Seq(tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True), # pad_to_multiple_of = 8 对齐到 8 的倍数)model.config.use_cache = Falseold_state_dict = model.state_dictmodel.state_dict = (lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())).__get__(model, type(model))if torch.__version__ >= "2" and sys.platform != "win32":# 加速模型推理和训练model = torch.compile(model)trainer.train(resume_from_checkpoint=resume_from_checkpoint)model.save_pretrained(output_dir)print("\n If there's a warning about missing keys above, please disregard :)")if __name__ == "__main__":fire.Fire(train)