2023-10-12 21:42:29 +08:00
|
|
|
|
# LLaMA Factory: 轻松的大模型训练与评估
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-10-12 21:42:29 +08:00
|
|
|
|
[![GitHub Repo stars](https://img.shields.io/github/stars/hiyouga/LLaMA-Factory?style=social)](https://github.com/hiyouga/LLaMA-Factory/stargazers)
|
|
|
|
|
[![GitHub Code License](https://img.shields.io/github/license/hiyouga/LLaMA-Factory)](LICENSE)
|
|
|
|
|
[![GitHub last commit](https://img.shields.io/github/last-commit/hiyouga/LLaMA-Factory)](https://github.com/hiyouga/LLaMA-Factory/commits/main)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
[![PyPI](https://img.shields.io/pypi/v/llmtuner)](https://pypi.org/project/llmtuner/)
|
2023-09-16 17:33:01 +08:00
|
|
|
|
[![Downloads](https://static.pepy.tech/badge/llmtuner)](https://pypi.org/project/llmtuner/)
|
2023-10-12 21:42:29 +08:00
|
|
|
|
[![GitHub pull request](https://img.shields.io/badge/PRs-welcome-blue)](https://github.com/hiyouga/LLaMA-Factory/pulls)
|
2023-10-12 21:44:28 +08:00
|
|
|
|
[![Discord](https://dcbadge.vercel.app/api/server/e73gccsSd?compact=true&style=flat)](https://discord.gg/e73gccsSd)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-07-22 14:31:16 +08:00
|
|
|
|
👋 加入我们的[微信群](assets/wechat.jpg)。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
\[ [English](README.md) | 中文 \]
|
|
|
|
|
|
2023-10-16 00:28:27 +08:00
|
|
|
|
## LLaMA Board: 通过一站式网页界面快速上手 LLaMA Factory
|
2023-10-15 20:28:14 +08:00
|
|
|
|
|
2023-10-16 00:28:27 +08:00
|
|
|
|
使用 `CUDA_VISIBLE_DEVICES=0 python src/train_web.py` 启动 **LLaMA Board**。(该界面目前仅支持单卡训练)
|
|
|
|
|
|
|
|
|
|
下面是使用单张 GPU 在 10 分钟内更改对话式大型语言模型自我认知的示例。
|
2023-10-15 20:28:14 +08:00
|
|
|
|
|
|
|
|
|
https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846-2d88920d5ba1
|
|
|
|
|
|
2023-07-21 16:57:58 +08:00
|
|
|
|
## 更新日志
|
|
|
|
|
|
2023-10-21 14:24:10 +08:00
|
|
|
|
[23/10/21] 我们支持了 [NEFTune](https://arxiv.org/abs/2310.05914) 优化。试试`--neftune_noise_alpha` 参数来激活 NEFTune,例如,`--neftune_noise_alpha 5`。
|
|
|
|
|
|
2023-09-28 14:39:16 +08:00
|
|
|
|
[23/09/27] 我们针对 LLaMA 模型支持了 [LongLoRA](https://github.com/dvlab-research/LongLoRA) 提出的 **$S^2$-Attn**。请使用 `--shift_attn` 参数以启用该功能。
|
2023-09-27 21:55:50 +08:00
|
|
|
|
|
2023-09-23 21:10:17 +08:00
|
|
|
|
[23/09/23] 我们在项目中集成了 MMLU、C-Eval 和 CMMLU 评估集。使用方法请参阅[此示例](#模型评估)。
|
2023-09-10 20:43:56 +08:00
|
|
|
|
|
2023-10-13 13:53:43 +08:00
|
|
|
|
[23/09/10] 我们针对 LLaMA 模型支持了 **[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)**。如果您使用的是 RTX4090、A100 或 H100 GPU,请使用 `--flash_attn` 参数以启用 FlashAttention-2。
|
2023-08-12 21:00:11 +08:00
|
|
|
|
|
2023-09-23 00:34:17 +08:00
|
|
|
|
[23/08/12] 我们支持了 **RoPE 插值**来扩展 LLaMA 模型的上下文长度。请使用 `--rope_scaling linear` 参数训练模型或使用 `--rope_scaling dynamic` 参数评估模型。
|
2023-08-11 03:02:53 +08:00
|
|
|
|
|
2023-09-23 00:34:17 +08:00
|
|
|
|
[23/08/11] 我们支持了指令模型的 **[DPO 训练](https://arxiv.org/abs/2305.18290)**。使用方法请参阅[此示例](#dpo-训练)。
|
|
|
|
|
|
|
|
|
|
[23/07/31] 我们支持了**数据流式加载**。请尝试使用 `--streaming` 和 `--max_steps 10000` 参数来流式加载数据集。
|
2023-07-31 23:42:32 +08:00
|
|
|
|
|
2023-09-09 13:50:29 +08:00
|
|
|
|
[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](https://huggingface.co/hiyouga/Llama-2-Chinese-13b-chat) / [Baichuan](https://huggingface.co/hiyouga/Baichuan-13B-sft))。
|
2023-08-01 10:08:47 +08:00
|
|
|
|
|
2023-08-18 01:41:17 +08:00
|
|
|
|
[23/07/18] 我们开发了支持训练和测试的**浏览器一体化界面**。请尝试使用 `train_web.py` 在您的浏览器中微调模型。感谢 [@KanadeSiina](https://github.com/KanadeSiina) 和 [@codemayq](https://github.com/codemayq) 在该功能开发中付出的努力。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-08-12 21:29:06 +08:00
|
|
|
|
[23/07/09] 我们开源了 **[FastEdit](https://github.com/hiyouga/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](https://github.com/hiyouga/FastEdit) 项目。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-09-09 13:50:29 +08:00
|
|
|
|
[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 [Baichuan-7B-sft](https://huggingface.co/hiyouga/Baichuan-7B-sft)。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-08-12 21:23:05 +08:00
|
|
|
|
[23/06/22] 我们对齐了[示例 API](src/api_demo.py) 与 [OpenAI API](https://platform.openai.com/docs/api-reference/chat) 的格式,您可以将微调模型接入**任意基于 ChatGPT 的应用**中。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-09-23 00:34:17 +08:00
|
|
|
|
[23/06/03] 我们实现了 4 比特的 LoRA 训练(也称 **[QLoRA](https://github.com/artidoro/qlora)**)。请尝试使用 `--quantization_bit 4` 参数进行 4 比特量化微调。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
## 模型
|
2023-08-07 15:02:02 +08:00
|
|
|
|
|
2023-09-06 21:43:06 +08:00
|
|
|
|
| 模型名 | 模型大小 | 默认模块 | Template |
|
|
|
|
|
| -------------------------------------------------------- | --------------------------- | ----------------- | --------- |
|
|
|
|
|
| [LLaMA](https://github.com/facebookresearch/llama) | 7B/13B/33B/65B | q_proj,v_proj | - |
|
|
|
|
|
| [LLaMA-2](https://huggingface.co/meta-llama) | 7B/13B/70B | q_proj,v_proj | llama2 |
|
|
|
|
|
| [BLOOM](https://huggingface.co/bigscience/bloom) | 560M/1.1B/1.7B/3B/7.1B/176B | query_key_value | - |
|
|
|
|
|
| [BLOOMZ](https://huggingface.co/bigscience/bloomz) | 560M/1.1B/1.7B/3B/7.1B/176B | query_key_value | - |
|
|
|
|
|
| [Falcon](https://huggingface.co/tiiuae/falcon-7b) | 7B/40B | query_key_value | - |
|
2023-09-07 18:54:14 +08:00
|
|
|
|
| [Baichuan](https://github.com/baichuan-inc/Baichuan-13B) | 7B/13B | W_pack | baichuan |
|
2023-09-06 21:43:06 +08:00
|
|
|
|
| [Baichuan2](https://github.com/baichuan-inc/Baichuan2) | 7B/13B | W_pack | baichuan2 |
|
2023-09-21 15:25:29 +08:00
|
|
|
|
| [InternLM](https://github.com/InternLM/InternLM) | 7B/20B | q_proj,v_proj | intern |
|
2023-09-27 21:55:50 +08:00
|
|
|
|
| [Qwen](https://github.com/QwenLM/Qwen-7B) | 7B/14B | c_attn | chatml |
|
2023-09-06 21:43:06 +08:00
|
|
|
|
| [ChatGLM2](https://github.com/THUDM/ChatGLM2-6B) | 6B | query_key_value | chatglm2 |
|
2023-09-22 14:34:13 +08:00
|
|
|
|
| [Phi-1.5](https://huggingface.co/microsoft/phi-1_5) | 1.3B | Wqkv | - |
|
2023-08-07 15:02:02 +08:00
|
|
|
|
|
2023-09-10 21:01:20 +08:00
|
|
|
|
> [!NOTE]
|
2023-09-10 20:43:56 +08:00
|
|
|
|
> **默认模块**应作为 `--lora_target` 参数的默认值,可使用 `--lora_target all` 参数指定全部模块。
|
|
|
|
|
>
|
2023-10-13 13:53:43 +08:00
|
|
|
|
> 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**。
|
2023-10-20 23:28:52 +08:00
|
|
|
|
>
|
|
|
|
|
> 项目所支持模型的完整列表请参阅 [template.py](src/llmtuner/extras/template.py)。
|
2023-08-11 03:02:53 +08:00
|
|
|
|
|
|
|
|
|
## 训练方法
|
|
|
|
|
|
2023-08-17 11:00:22 +08:00
|
|
|
|
| 方法 | 全参数训练 | 部分参数训练 | LoRA | QLoRA |
|
|
|
|
|
| ---------------------- | ------------------ | ------------------ | ------------------ | ------------------ |
|
|
|
|
|
| 预训练 | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
|
|
|
|
|
| 指令监督微调 | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
|
|
|
|
|
| 奖励模型训练 | | | :white_check_mark: | :white_check_mark: |
|
|
|
|
|
| PPO 训练 | | | :white_check_mark: | :white_check_mark: |
|
|
|
|
|
| DPO 训练 | :white_check_mark: | | :white_check_mark: | :white_check_mark: |
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-09-10 21:01:20 +08:00
|
|
|
|
> [!NOTE]
|
2023-09-10 20:43:56 +08:00
|
|
|
|
> 请使用 `--quantization_bit 4/8` 参数来启用 QLoRA 训练。
|
2023-08-12 21:23:05 +08:00
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
## 数据集
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-08-11 03:02:53 +08:00
|
|
|
|
- 用于预训练:
|
2023-07-21 16:57:58 +08:00
|
|
|
|
- [Wiki Demo (en)](data/wiki_demo.txt)
|
2023-07-23 20:01:43 +08:00
|
|
|
|
- [RefinedWeb (en)](https://huggingface.co/datasets/tiiuae/falcon-refinedweb)
|
|
|
|
|
- [StarCoder (en)](https://huggingface.co/datasets/bigcode/starcoderdata)
|
|
|
|
|
- [Wikipedia (en)](https://huggingface.co/datasets/olm/olm-wikipedia-20221220)
|
|
|
|
|
- [Wikipedia (zh)](https://huggingface.co/datasets/pleisto/wikipedia-cn-20230720-filtered)
|
2023-08-11 03:02:53 +08:00
|
|
|
|
- 用于指令监督微调:
|
2023-07-21 16:57:58 +08:00
|
|
|
|
- [Stanford Alpaca (en)](https://github.com/tatsu-lab/stanford_alpaca)
|
|
|
|
|
- [Stanford Alpaca (zh)](https://github.com/ymcui/Chinese-LLaMA-Alpaca)
|
|
|
|
|
- [GPT-4 Generated Data (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
|
|
|
|
|
- [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
|
|
|
|
|
- [Self-cognition (zh)](data/self_cognition.json)
|
|
|
|
|
- [ShareGPT (zh)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection)
|
|
|
|
|
- [Guanaco Dataset (multilingual)](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
|
|
|
|
|
- [BELLE 2M (zh)](https://huggingface.co/datasets/BelleGroup/train_2M_CN)
|
|
|
|
|
- [BELLE 1M (zh)](https://huggingface.co/datasets/BelleGroup/train_1M_CN)
|
|
|
|
|
- [BELLE 0.5M (zh)](https://huggingface.co/datasets/BelleGroup/train_0.5M_CN)
|
|
|
|
|
- [BELLE Dialogue 0.4M (zh)](https://huggingface.co/datasets/BelleGroup/generated_chat_0.4M)
|
|
|
|
|
- [BELLE School Math 0.25M (zh)](https://huggingface.co/datasets/BelleGroup/school_math_0.25M)
|
|
|
|
|
- [BELLE Multiturn Chat 0.8M (zh)](https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M)
|
2023-07-26 17:05:12 +08:00
|
|
|
|
- [LIMA (en)](https://huggingface.co/datasets/GAIR/lima)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
- [CodeAlpaca 20k (en)](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k)
|
|
|
|
|
- [Alpaca CoT (multilingual)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT)
|
2023-09-13 22:30:14 +08:00
|
|
|
|
- [MathInstruct (en)](https://huggingface.co/datasets/TIGER-Lab/MathInstruct)
|
|
|
|
|
- [Firefly 1.1M (zh)](https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
- [Web QA (zh)](https://huggingface.co/datasets/suolyer/webqa)
|
|
|
|
|
- [UltraChat (en)](https://github.com/thunlp/UltraChat)
|
|
|
|
|
- [WebNovel (zh)](https://huggingface.co/datasets/zxbsmk/webnovel_cn)
|
2023-09-01 19:00:45 +08:00
|
|
|
|
- [Ad Gen (zh)](https://huggingface.co/datasets/HasturOfficial/adgen)
|
|
|
|
|
- 用于训练奖励模型或 DPO 训练:
|
2023-07-21 16:57:58 +08:00
|
|
|
|
- [HH-RLHF (en)](https://huggingface.co/datasets/Anthropic/hh-rlhf)
|
|
|
|
|
- [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
|
|
|
|
|
- [GPT-4 Generated Data (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
|
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
使用方法请参考 [data/README.md](data/README_zh.md) 文件。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
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。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
pip install --upgrade huggingface_hub
|
|
|
|
|
huggingface-cli login
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
## 软件依赖
|
|
|
|
|
|
|
|
|
|
- Python 3.8+ 和 PyTorch 1.13.1+
|
|
|
|
|
- 🤗Transformers, Datasets, Accelerate, PEFT 和 TRL
|
2023-09-11 17:31:34 +08:00
|
|
|
|
- sentencepiece, protobuf 和 tiktoken
|
2023-10-09 20:02:50 +08:00
|
|
|
|
- fire, jieba, rouge-chinese 和 nltk (用于评估及预测)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
- gradio 和 matplotlib (用于网页端交互)
|
|
|
|
|
- uvicorn, fastapi 和 sse-starlette (用于 API)
|
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
以及 **强而有力的 GPU**!
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
## 如何使用
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
### 数据准备(可跳过)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
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](https://huggingface.co/docs/datasets/dataset_script)和多个文件。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-09-10 21:01:20 +08:00
|
|
|
|
> [!NOTE]
|
2023-09-10 20:43:56 +08:00
|
|
|
|
> 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`。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
### 环境搭建(可跳过)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
2023-10-12 21:42:29 +08:00
|
|
|
|
git clone https://github.com/hiyouga/LLaMA-Factory.git
|
|
|
|
|
conda create -n llama_factory python=3.10
|
|
|
|
|
conda activate llama_factory
|
|
|
|
|
cd LLaMA-Factory
|
2023-07-21 16:57:58 +08:00
|
|
|
|
pip install -r requirements.txt
|
|
|
|
|
```
|
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
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.
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
|
|
|
|
|
```
|
|
|
|
|
|
2023-08-18 01:51:55 +08:00
|
|
|
|
### 单 GPU 训练
|
|
|
|
|
|
2023-09-10 21:01:20 +08:00
|
|
|
|
> [!IMPORTANT]
|
2023-09-10 20:43:56 +08:00
|
|
|
|
> 如果您使用多张 GPU 训练模型,请移步[多 GPU 分布式训练](#多-gpu-分布式训练)部分。
|
|
|
|
|
|
2023-08-18 01:51:55 +08:00
|
|
|
|
#### 预训练
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
|
|
|
|
|
--stage pt \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--do_train \
|
|
|
|
|
--dataset wiki_demo \
|
|
|
|
|
--finetuning_type lora \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--lora_target q_proj,v_proj \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--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
|
|
|
|
|
```
|
|
|
|
|
|
2023-08-18 01:51:55 +08:00
|
|
|
|
#### 指令监督微调
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
|
|
|
|
|
--stage sft \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--do_train \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--dataset alpaca_gpt4_zh \
|
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--lora_target q_proj,v_proj \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--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 \
|
|
|
|
|
--fp16
|
|
|
|
|
```
|
|
|
|
|
|
2023-08-18 01:51:55 +08:00
|
|
|
|
#### 奖励模型训练
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
|
|
|
|
|
--stage rm \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--do_train \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--dataset comparison_gpt4_zh \
|
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--lora_target q_proj,v_proj \
|
2023-07-28 17:36:00 +08:00
|
|
|
|
--resume_lora_training False \
|
|
|
|
|
--checkpoint_dir path_to_sft_checkpoint \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--output_dir path_to_rm_checkpoint \
|
2023-08-11 03:02:53 +08:00
|
|
|
|
--per_device_train_batch_size 2 \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--gradient_accumulation_steps 4 \
|
|
|
|
|
--lr_scheduler_type cosine \
|
|
|
|
|
--logging_steps 10 \
|
|
|
|
|
--save_steps 1000 \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--learning_rate 1e-6 \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--num_train_epochs 1.0 \
|
|
|
|
|
--plot_loss \
|
|
|
|
|
--fp16
|
|
|
|
|
```
|
|
|
|
|
|
2023-08-18 01:51:55 +08:00
|
|
|
|
#### PPO 训练
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
|
|
|
|
|
--stage ppo \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--do_train \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--dataset alpaca_gpt4_zh \
|
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--lora_target q_proj,v_proj \
|
2023-07-28 17:36:00 +08:00
|
|
|
|
--resume_lora_training False \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--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
|
|
|
|
|
```
|
|
|
|
|
|
2023-08-18 01:51:55 +08:00
|
|
|
|
#### DPO 训练
|
2023-08-11 03:02:53 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
|
|
|
|
|
--stage dpo \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-08-11 03:02:53 +08:00
|
|
|
|
--do_train \
|
|
|
|
|
--dataset comparison_gpt4_zh \
|
|
|
|
|
--template default \
|
|
|
|
|
--finetuning_type lora \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--lora_target q_proj,v_proj \
|
2023-08-11 03:02:53 +08:00
|
|
|
|
--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
|
|
|
|
|
```
|
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
### 多 GPU 分布式训练
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-08-12 21:23:05 +08:00
|
|
|
|
#### 使用 Huggingface Accelerate
|
|
|
|
|
|
2023-07-21 16:57:58 +08:00
|
|
|
|
```bash
|
2023-07-22 14:29:22 +08:00
|
|
|
|
accelerate config # 首先配置分布式环境
|
|
|
|
|
accelerate launch src/train_bash.py # 参数同上
|
2023-07-21 16:57:58 +08:00
|
|
|
|
```
|
|
|
|
|
|
2023-09-10 20:43:56 +08:00
|
|
|
|
<details><summary>LoRA 训练的 Accelerate 配置示例</summary>
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```yaml
|
|
|
|
|
compute_environment: LOCAL_MACHINE
|
2023-09-10 20:43:56 +08:00
|
|
|
|
distributed_type: MULTI_GPU
|
2023-07-21 16:57:58 +08:00
|
|
|
|
downcast_bf16: 'no'
|
2023-09-10 20:43:56 +08:00
|
|
|
|
gpu_ids: all
|
2023-07-21 16:57:58 +08:00
|
|
|
|
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
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
</details>
|
|
|
|
|
|
2023-08-12 21:23:05 +08:00
|
|
|
|
#### 使用 DeepSpeed
|
|
|
|
|
|
|
|
|
|
```bash
|
2023-08-12 21:25:19 +08:00
|
|
|
|
deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
|
|
|
|
|
--deepspeed ds_config.json \
|
|
|
|
|
... # 参数同上
|
2023-08-12 21:23:05 +08:00
|
|
|
|
```
|
|
|
|
|
|
2023-09-10 20:43:56 +08:00
|
|
|
|
<details><summary>使用 DeepSpeed ZeRO-2 进行全参数训练的 DeepSpeed 配置示例</summary>
|
2023-08-12 21:23:05 +08:00
|
|
|
|
|
|
|
|
|
```json
|
|
|
|
|
{
|
2023-09-10 21:01:20 +08:00
|
|
|
|
"train_batch_size": "auto",
|
2023-08-12 21:23:05 +08:00
|
|
|
|
"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
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
</details>
|
|
|
|
|
|
2023-09-22 14:34:13 +08:00
|
|
|
|
### 导出微调后的完整模型
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
2023-08-18 01:51:55 +08:00
|
|
|
|
python src/export_model.py \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
|
|
|
|
--checkpoint_dir path_to_checkpoint \
|
2023-10-19 15:52:24 +08:00
|
|
|
|
--export_dir path_to_export
|
2023-07-21 16:57:58 +08:00
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
### API 服务
|
2023-07-22 14:29:22 +08:00
|
|
|
|
|
2023-07-21 16:57:58 +08:00
|
|
|
|
```bash
|
|
|
|
|
python src/api_demo.py \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
|
|
|
|
--checkpoint_dir path_to_checkpoint
|
|
|
|
|
```
|
|
|
|
|
|
2023-09-10 21:01:20 +08:00
|
|
|
|
> [!NOTE]
|
2023-09-10 20:43:56 +08:00
|
|
|
|
> 关于 API 文档请见 `http://localhost:8000/docs`。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
### 命令行测试
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
python src/cli_demo.py \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
|
|
|
|
--checkpoint_dir path_to_checkpoint
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
### 浏览器测试
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
python src/web_demo.py \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
|
|
|
|
--checkpoint_dir path_to_checkpoint
|
|
|
|
|
```
|
|
|
|
|
|
2023-09-23 21:10:17 +08:00
|
|
|
|
### 模型评估
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
CUDA_VISIBLE_DEVICES=0 python src/evaluate.py \
|
|
|
|
|
--model_name_or_path path_to_llama_model \
|
|
|
|
|
--finetuning_type lora \
|
|
|
|
|
--checkpoint_dir path_to_checkpoint \
|
|
|
|
|
--template vanilla \
|
|
|
|
|
--task ceval \
|
|
|
|
|
--split validation \
|
|
|
|
|
--lang zh \
|
|
|
|
|
--n_shot 5 \
|
|
|
|
|
--batch_size 4
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
### 模型预测
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bash
|
2023-08-18 01:51:55 +08:00
|
|
|
|
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
|
|
|
|
|
--stage sft \
|
2023-08-18 11:43:10 +08:00
|
|
|
|
--model_name_or_path path_to_llama_model \
|
2023-09-23 00:34:17 +08:00
|
|
|
|
--do_predict \
|
2023-08-18 01:51:55 +08:00
|
|
|
|
--dataset alpaca_gpt4_zh \
|
2023-07-31 23:33:00 +08:00
|
|
|
|
--template default \
|
2023-07-21 16:57:58 +08:00
|
|
|
|
--finetuning_type lora \
|
|
|
|
|
--checkpoint_dir path_to_checkpoint \
|
2023-09-23 00:34:17 +08:00
|
|
|
|
--output_dir path_to_predict_result \
|
2023-08-18 01:51:55 +08:00
|
|
|
|
--per_device_eval_batch_size 8 \
|
|
|
|
|
--max_samples 100 \
|
|
|
|
|
--predict_with_generate
|
|
|
|
|
```
|
|
|
|
|
|
2023-09-10 21:01:20 +08:00
|
|
|
|
> [!NOTE]
|
2023-09-23 21:10:17 +08:00
|
|
|
|
> 我们建议在量化模型的预测中使用 `--per_device_eval_batch_size=1` 和 `--max_target_length 128`。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
## 协议
|
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
2023-10-20 23:28:52 +08:00
|
|
|
|
使用模型权重时,请遵循对应的模型协议:[LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [Falcon](LICENSE) / [Baichuan](https://huggingface.co/baichuan-inc/baichuan-7B/resolve/main/baichuan-7B%20%E6%A8%A1%E5%9E%8B%E8%AE%B8%E5%8F%AF%E5%8D%8F%E8%AE%AE.pdf) / [Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/resolve/main/Baichuan%202%E6%A8%A1%E5%9E%8B%E7%A4%BE%E5%8C%BA%E8%AE%B8%E5%8F%AF%E5%8D%8F%E8%AE%AE.pdf) / [InternLM](https://github.com/InternLM/InternLM#open-source-license) / [Qwen](https://huggingface.co/Qwen/Qwen-7B-Chat/blob/main/LICENSE) / [ChatGLM2](https://github.com/THUDM/ChatGLM2-6B/blob/main/MODEL_LICENSE) / [Phi-1.5](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx)
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
## 引用
|
|
|
|
|
|
2023-07-22 14:29:22 +08:00
|
|
|
|
如果您觉得此项目有帮助,请考虑以下列格式引用
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
```bibtex
|
2023-10-12 21:42:29 +08:00
|
|
|
|
@Misc{llama-factory,
|
|
|
|
|
title = {LLaMA Factory},
|
2023-07-21 16:57:58 +08:00
|
|
|
|
author = {hiyouga},
|
2023-10-12 21:42:29 +08:00
|
|
|
|
howpublished = {\url{https://github.com/hiyouga/LLaMA-Factory}},
|
2023-07-21 16:57:58 +08:00
|
|
|
|
year = {2023}
|
|
|
|
|
}
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
## 致谢
|
|
|
|
|
|
2023-10-09 20:02:50 +08:00
|
|
|
|
本项目受益于 [PEFT](https://github.com/huggingface/peft)、[QLoRA](https://github.com/artidoro/qlora) 和 [FastChat](https://github.com/lm-sys/FastChat),感谢以上诸位作者的付出。
|
2023-07-21 16:57:58 +08:00
|
|
|
|
|
|
|
|
|
## Star History
|
|
|
|
|
|
2023-10-12 21:42:29 +08:00
|
|
|
|
![Star History Chart](https://api.star-history.com/svg?repos=hiyouga/LLaMA-Factory&type=Date)
|