Skip to content

实现了Baichuan-Chat微调,Lora、QLora等各种微调方式,一键运行。

License

Notifications You must be signed in to change notification settings

ArtificialZeng/Baichuan-Chat-Tuning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

百川微调指南

软件依赖

  • Python 3.8+ 和 PyTorch 1.13.1+
  • 🤗Transformers, Datasets, Accelerate, PEFT 和 TRL
  • sentencepiece 和 tiktoken
  • jieba, rouge-chinese 和 nltk (用于评估)
  • gradio 和 matplotlib (用于网页端交互)
  • uvicorn, fastapi 和 sse-starlette (用于 API)

以及 强而有力的 GPU

如何使用

数据准备(可跳过)

关于数据集文件的格式,请参考 data/example_dataset 文件夹的内容。构建自定义数据集时,既可以使用单个 .json 文件,也可以使用一个数据加载脚本和多个文件。

注意:使用自定义数据集时,请更新 data/dataset_info.json 文件,该文件的格式请参考 data/README.md

环境搭建(可跳过)

git clone https://github.com/ArtificialZeng/Baichuan-Chat-Tuning
conda create -n baichuan_etuning python=3.10
conda activate baichuan_etuning
cd Baichuan-Chat-Tuning
pip install -r requirements.txt

如果要在 Windows 平台上开启量化 LoRA(QLoRA),需要安装预编译的 bitsandbytes 库, 支持 CUDA 11.1 到 12.1.

pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl

浏览器一键微调/测试

CUDA_VISIBLE_DEVICES=0 python src/train_web.py

目前网页 UI 仅支持单卡训练。如果要多卡训练,请用以下的命令行形式:

Baichuan指令监督微调(SFT - 一般这个用的最多,预训练脚本在下面)

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset alpaca_gpt4_zh \
    --template baichuan \
    --finetuning_type lora \
    --output_dir path_to_sft_checkpoint \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --lora_target W_pack \
    --fp16

预训练

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage pt \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset wiki_demo \
    --template default \
    --finetuning_type lora \
    --output_dir path_to_pt_checkpoint \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16

Baichuan Efficient Tuning

GitHub Repo stars GitHub Code License GitHub last commit PyPI GitHub pull request

[ English | 中文 ]

更新日志

[23/08/12] 现在我们支持了 RoPE 插值来扩展 LLaMA 模型的上下文长度。请尝试使用 --rope_scaling linear 参数训练模型或使用 --rope_scaling dynamic 参数评估模型。

[23/08/11] 现在我们支持了指令模型的 DPO 训练。详情请参阅此示例(实验性功能)。

[23/08/03] 现在我们支持了 Qwen-7B 模型的训练。请尝试使用 --model_name_or_path Qwen/Qwen-7B-Chat--lora_target c_attn 参数。使用 Qwen-7B-Chat 模型时请添加 --template chatml 参数。

[23/07/31] 现在我们支持了数据流式加载。请尝试使用 --streaming--max_steps 10000 参数来流式加载数据集。

[23/07/29] 我们在 Hugging Face 发布了两个 13B 指令微调模型。详细内容请查阅我们的 Hugging Face 项目(LLaMA-2 / Baichuan)。

[23/07/19] 现在我们支持了 LLaMA-2 模型的训练。请尝试使用 --model_name_or_path meta-baichuan/Llama-2-7b-hf 参数。使用 LLaMA-2-chat 模型时请添加 --template baichuan2 参数。

[23/07/18] 我们开发了支持训练和测试的一体化浏览器界面。请尝试使用 train_web.py 在您的浏览器中微调模型。感谢 @KanadeSiina@codemayq 在该功能开发中付出的努力。

[23/07/11] 现在我们支持了 Baichuan-13B 模型的训练。请尝试使用 --model_name_or_path baichuan-inc/Baichuan-13B-Base--lora_target W_pack 参数。使用 Baichuan-13B-Chat 模型时请添加 --template baichuan 参数。

[23/07/09] 我们开源了 FastEdit ⚡🩹,一个简单易用的、能迅速编辑大模型事实记忆的工具包。如果您感兴趣请关注我们的 FastEdit 项目。

[23/07/07] 现在我们支持了 InternLM-7B 模型的训练。请尝试使用 --model_name_or_path internlm/internlm-7b 参数。使用 InternLM-chat 模型时请添加 --template intern 参数。

[23/07/05] 现在我们支持了 Falcon-7B/40B 模型的训练。请尝试使用 --model_name_or_path tiiuae/falcon-7b--lora_target query_key_value 参数。

[23/06/29] 我们提供了一个可复现的指令模型微调示例,详细内容请查阅 Hugging Face 项目

[23/06/22] 我们对齐了示例 APIOpenAI API 的格式,您可以将微调模型接入任意基于 ChatGPT 的应用中。

[23/06/15] 现在我们支持了 Baichuan-7B 模型的训练。请尝试使用 --model_name_or_path baichuan-inc/Baichuan-7B--lora_target W_pack 参数。

[23/06/03] 现在我们实现了 4 比特的 LoRA 训练(也称 QLoRA)。请尝试使用 --quantization_bit 4 参数进行 4 比特量化微调。

[23/05/31] 现在我们支持了 BLOOM & BLOOMZ 模型的训练。请尝试使用 --model_name_or_path bigscience/bloomz-7b1-mt--lora_target query_key_value 参数。

模型

模型名 模型大小 默认模块 Template
LLaMA 7B/13B/33B/65B q_proj,v_proj -
LLaMA-2 7B/13B/70B q_proj,v_proj baichuan2
BLOOM 560M/1.1B/1.7B/3B/7.1B/176B query_key_value -
BLOOMZ 560M/1.1B/1.7B/3B/7.1B/176B query_key_value -
Falcon 7B/40B query_key_value -
Baichuan 7B/13B W_pack baichuan
InternLM 7B q_proj,v_proj intern
Qwen 7B c_attn chatml
XVERSE 13B q_proj,v_proj -
ChatGLM2 6B query_key_value chatglm2
  • 默认模块--lora_target 参数的部分可选项。请使用 python src/train_bash.py -h 查看全部可选项。
  • 对于所有“基座”(Base)模型,--template 参数可以是 default, alpaca, vicuna 等任意值。但“对话”(Chat)模型请务必使用对应的模板。

训练方法

方法 全参数训练 部分参数训练 LoRA QLoRA
预训练
指令监督微调
奖励模型训练
PPO 训练
DPO 训练
  • 使用 --quantization_bit 4/8 参数来启用 QLoRA 训练。

数据集

使用方法请参考 data/README.md 文件。

部分数据集的使用需要确认,我们推荐使用下述命令登录您的 Hugging Face 账户。

pip install --upgrade huggingface_hub
huggingface-cli login

奖励模型训练

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage rm \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset comparison_gpt4_zh \
    --template default \
    --finetuning_type lora \
    --resume_lora_training False \
    --checkpoint_dir path_to_sft_checkpoint \
    --output_dir path_to_rm_checkpoint \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-5 \
    --num_train_epochs 1.0 \
    --plot_loss \
    --fp16

PPO 训练

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage ppo \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset alpaca_gpt4_zh \
    --template default \
    --finetuning_type lora \
    --resume_lora_training False \
    --checkpoint_dir path_to_sft_checkpoint \
    --reward_model path_to_rm_checkpoint \
    --output_dir path_to_ppo_checkpoint \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-5 \
    --num_train_epochs 1.0 \
    --plot_loss

DPO 训练

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage dpo \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset comparison_gpt4_zh \
    --template default \
    --finetuning_type lora \
    --resume_lora_training False \
    --checkpoint_dir path_to_sft_checkpoint \
    --output_dir path_to_dpo_checkpoint \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-5 \
    --num_train_epochs 1.0 \
    --plot_loss \
    --fp16

多 GPU 分布式训练

使用 Huggingface Accelerate

accelerate config # 首先配置分布式环境
accelerate launch src/train_bash.py # 参数同上
使用 DeepSpeed ZeRO-2 进行全参数微调的 Accelerate 配置示例
compute_environment: LOCAL_MACHINE
deepspeed_config:
  gradient_accumulation_steps: 4
  gradient_clipping: 0.5
  offload_optimizer_device: none
  offload_param_device: none
  zero3_init_flag: false
  zero_stage: 2
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: fp16
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

使用 DeepSpeed

deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
    --deepspeed ds_config.json \
    ... # 参数同上
使用 DeepSpeed ZeRO-2 进行全参数微调的 DeepSpeed 配置示例
{
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "zero_allow_untested_optimizer": true,
  "fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "initial_scale_power": 16,
    "loss_scale_window": 1000,
    "hysteresis": 2,
    "min_loss_scale": 1
  },  
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 5e8,
    "reduce_scatter": true,
    "reduce_bucket_size": 5e8,
    "overlap_comm": false,
    "contiguous_gradients": true
  }
}

指标评估(BLEU 分数和汉语 ROUGE 分数)

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --model_name_or_path path_to_your_model \
    --do_eval \
    --dataset alpaca_gpt4_zh \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_eval_result \
    --per_device_eval_batch_size 8 \
    --max_samples 100 \
    --predict_with_generate

我们建议在量化模型的评估中使用 --per_device_eval_batch_size=1--max_target_length 128

模型预测

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --model_name_or_path path_to_your_model \
    --do_predict \
    --dataset alpaca_gpt4_zh \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_predict_result \
    --per_device_eval_batch_size 8 \
    --max_samples 100 \
    --predict_with_generate

API 服务

python src/api_demo.py \
    --model_name_or_path path_to_your_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint

关于 API 文档请见 http://localhost:8000/docs

命令行测试

python src/cli_demo.py \
    --model_name_or_path path_to_your_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint

浏览器测试

python src/web_demo.py \
    --model_name_or_path path_to_your_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint

导出微调模型

python src/export_model.py \
    --model_name_or_path path_to_your_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_export

TODO

  • 实现 flash attention (torch / xformers / flashattn)。
  • 在推理阶段使用 Multi-query attention 进行加速。
  • 支持 RLHF 的全参数微调。

协议

本仓库的代码依照 Apache-2.0 协议开源。

使用模型权重时,请遵循对应的模型协议:

引用

如果您觉得此项目有帮助,请考虑以下列格式引用

@Misc{baichuan-efficient-tuning,
  title = {LLaMA Efficient Tuning},
  author = {hiyouga},
  howpublished = {\url{https://github.com/hiyouga/Baichuan-Chat-Tuning}},
  year = {2023}
}

致谢

本项目是 ChatGLM-Efficient-Tuning 的同类项目。采用了类似的代码结构和训练方法。

Star History

Star History Chart

About

实现了Baichuan-Chat微调,Lora、QLora等各种微调方式,一键运行。

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages