release v0.6.0 (real)

Former-commit-id: 34e06bf408ccd21e674f896703f1c7b62e97e1ca

.dockerignore

@@ -0,0 +1,11 @@
+.vscode
+.git
+.github
+.venv
+cache
+data
+examples
+.dockerignore
+.gitattributes
+.gitignore
+Dockerfile

.github/CODE_OF_CONDUCT.md

@@ -0,0 +1,128 @@
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socio-economic status,
+nationality, personal appearance, race, religion, or sexual identity
+and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our
+community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the
+  overall community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or
+  advances of any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email
+  address, without their explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples of representing our community include using an official e-mail address,
+posting via an official social media account, or acting as an appointed
+representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the community leaders responsible for enforcement at
+`hoshihiyouga AT gmail DOT com`.
+All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series
+of actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. This
+includes avoiding interactions in community spaces as well as external channels
+like social media. Violating these terms may lead to a temporary or
+permanent ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. No public or
+private interaction with the people involved, including unsolicited interaction
+with those enforcing the Code of Conduct, is allowed during this period.
+Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior,  harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within
+the community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage],
+version 2.0, available at
+https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
+
+Community Impact Guidelines were inspired by [Mozilla's code of conduct
+enforcement ladder](https://github.com/mozilla/diversity).
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see the FAQ at
+https://www.contributor-covenant.org/faq. Translations are available at
+https://www.contributor-covenant.org/translations.

.github/CONTRIBUTING.md

@@ -0,0 +1,21 @@
+# Contributing to LLaMA Factory
+
+Everyone is welcome to contribute, and we value everybody's contribution. Code contributions are not the only way to help the community. Answering questions, helping others, and improving the documentation are also immensely valuable.
+
+It also helps us if you spread the word! Reference the library in blog posts about the awesome projects it made possible, shout out on Twitter every time it has helped you, or simply ⭐️ the repository to say thank you.
+
+However you choose to contribute, please be mindful and respect our [code of conduct](CODE_OF_CONDUCT.md).
+
+**This guide was heavily inspired by [transformers guide to contributing](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md).**
+
+## Ways to contribute
+
+There are several ways you can contribute to LLaMA Factory:
+
+* Fix outstanding issues with the existing code.
+* Submit issues related to bugs or desired new features.
+* Contribute to the examples or to the documentation.
+
+### Style guide
+
+LLaMA Factory follows the [Google Python Style Guide](https://google.github.io/styleguide/pyguide.html), check it for details.

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,58 @@
+name: "\U0001F41B Bug / Help"
+description: Create a report to help us improve the LLaMA Factory
+body:
+  - type: checkboxes
+    id: reminder
+    attributes:
+      label: Reminder
+      description: |
+        Please ensure you have read the README carefully and searched the existing issues.
+        请确保您已经认真阅读了 README 并且搜索过现有的 Issue。
+
+      options:
+        - label: I have read the README and searched the existing issues.
+          required: true
+
+  - type: textarea
+    id: reproduction
+    validations:
+      required: true
+    attributes:
+      label: Reproduction
+      description: |
+        Please provide code snippets, error messages and stack traces that reproduces the problem.
+        请提供运行参数，错误信息以及异常堆栈以便于我们复现该问题。
+        Remember to use Markdown tags to correctly format your code.
+        请合理使用 Markdown 标签来格式化您的文本。
+
+      placeholder: |
+        python src/train_bash.py ...
+
+  - type: textarea
+    id: expected-behavior
+    validations:
+      required: false
+    attributes:
+      label: Expected behavior
+      description: |
+        Please provide a clear and concise description of what you would expect to happen.
+        请提供您原本的目的，即这段代码的期望行为。
+
+  - type: textarea
+    id: system-info
+    validations:
+      required: false
+    attributes:
+      label: System Info
+      description: |
+        Please share your system info with us. You can run the command **transformers-cli env** and copy-paste its output below.
+        请提供您的系统信息。您可以在命令行运行 **transformers-cli env** 并将其输出复制到该文本框中。
+
+      placeholder: transformers version, platform, python version, ...
+
+  - type: textarea
+    id: others
+    validations:
+      required: false
+    attributes:
+      label: Others

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,7 @@
+# What does this PR do?
+
+Fixes # (issue)
+
+## Before submitting
+
+- [ ] Did you read the [contributor guideline](https://github.com/hiyouga/LLaMA-Factory/blob/main/.github/CONTRIBUTING.md)?

.github/SECURITY.md

@@ -0,0 +1,7 @@
+# Reporting Security Issues
+
+To report a security issue, please use the GitHub Security Advisory ["Report a Vulnerability"](https://github.com/electron/electron/security/advisories/new) tab.
+
+We will send a response indicating the next steps in handling your report. After the initial reply to your report, the security team will keep you informed of the progress towards a fix and full announcement, and may ask for additional information or guidance.
+
+Report security bugs in third-party modules to the person or team maintaining the module.

.github/workflows/tests.yml

@@ -0,0 +1,29 @@
+name: tests
+
+on:
+  push:
+    branches: [ "main" ]
+  pull_request:
+    branches: [ "main" ]
+
+jobs:
+  check_code_quality:
+
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.8"
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install ruff
+
+      - name: Check quality
+        run: |
+            make style && make quality

.gitignore

@@ -157,4 +157,9 @@ cython_debug/
 #  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
-#.idea/
+.idea/
+
+# custom .gitignore
+user.config
+saves/
+cache/

CITATION.cff

@@ -0,0 +1,37 @@
+cff-version: 1.2.0
+date-released: 2024-03
+message: "If you use this software, please cite it as below."
+authors:
+- family-names: "Zheng"
+  given-names: "Yaowei"
+- family-names: "Zhang"
+  given-names: "Richong"
+- family-names: "Zhang"
+  given-names: "Junhao"
+- family-names: "Ye"
+  given-names: "Yanhan"
+- family-names: "Luo"
+  given-names: "Zheyan"
+- family-names: "Ma"
+  given-names: "Yongqiang"
+title: "LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models"
+url: "https://arxiv.org/abs/2403.13372"
+preferred-citation:
+  type: article
+  authors:
+    - family-names: "Zheng"
+      given-names: "Yaowei"
+    - family-names: "Zhang"
+      given-names: "Richong"
+    - family-names: "Zhang"
+      given-names: "Junhao"
+    - family-names: "Ye"
+      given-names: "Yanhan"
+    - family-names: "Luo"
+      given-names: "Zheyan"
+    - family-names: "Ma"
+      given-names: "Yongqiang"
+  journal: "arXiv preprint arXiv:2403.13372"
+  title: "LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models"
+  url: "https://arxiv.org/abs/2403.13372"
+  year: 2024

Dockerfile

@@ -0,0 +1,14 @@
+FROM nvcr.io/nvidia/pytorch:24.01-py3
+
+WORKDIR /app
+
+COPY requirements.txt /app/
+RUN pip install -r requirements.txt
+
+COPY . /app/
+RUN pip install -e .[deepspeed,metrics,bitsandbytes,qwen]
+
+VOLUME [ "/root/.cache/huggingface/", "/app/data", "/app/output" ]
+EXPOSE 7860
+
+CMD [ "python", "src/train_web.py" ]

Makefile

@@ -0,0 +1,11 @@
+.PHONY: quality style
+
+check_dirs := scripts src tests
+
+quality:
+	ruff check $(check_dirs)
+	ruff format --check $(check_dirs)
+
+style:
+	ruff check $(check_dirs) --fix
+	ruff format $(check_dirs)

README.md

@@ -1,34 +1,106 @@
-# LLaMA Factory: Training and Evaluating Large Language Models with Minimal Effort
+![# LLaMA Factory](assets/logo.png)
 
 [![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)
 [![PyPI](https://img.shields.io/pypi/v/llmtuner)](https://pypi.org/project/llmtuner/)
 [![Downloads](https://static.pepy.tech/badge/llmtuner)](https://pypi.org/project/llmtuner/)
+[![Citation](https://img.shields.io/badge/citation-26-green)](#projects-using-llama-factory)
 [![GitHub pull request](https://img.shields.io/badge/PRs-welcome-blue)](https://github.com/hiyouga/LLaMA-Factory/pulls)
-[![Discord](https://dcbadge.vercel.app/api/server/e73gccsSd?compact=true&style=flat)](https://discord.gg/e73gccsSd)
+[![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
+[![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
+[![Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
+[![Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing)
 
 👋 Join our [WeChat](assets/wechat.jpg).
 
 \[ English | [中文](README_zh.md) \]
 
-## LLaMA Board: A One-stop Web UI for Getting Started with LLaMA Factory
+**Fine-tuning a large language model can be easy as...**
 
-Launch **LLaMA Board** via `CUDA_VISIBLE_DEVICES=0 python src/train_web.py`. (multiple GPUs are not supported yet)
+https://github.com/hiyouga/LLaMA-Factory/assets/16256802/9840a653-7e9c-41c8-ae89-7ace5698baf6
 
-Here is an example of altering the self-cognition of an instruction-tuned language model within 10 minutes on a single GPU.
+Choose your path:
 
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846-2d88920d5ba1
+- **Colab**: https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing
+- **Local machine**: Please refer to [usage](#getting-started)
+
+## Table of Contents
+
+- [Features](#features)
+- [Benchmark](#benchmark)
+- [Changelog](#changelog)
+- [Supported Models](#supported-models)
+- [Supported Training Approaches](#supported-training-approaches)
+- [Provided Datasets](#provided-datasets)
+- [Requirement](#requirement)
+- [Getting Started](#getting-started)
+- [Projects using LLaMA Factory](#projects-using-llama-factory)
+- [License](#license)
+- [Citation](#citation)
+- [Acknowledgement](#acknowledgement)
+
+## Features
+
+- **Various models**: LLaMA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
+- **Integrated methods**: (Continuous) pre-training, supervised fine-tuning, reward modeling, PPO and DPO.
+- **Scalable resources**: 32-bit full-tuning, 16-bit freeze-tuning, 16-bit LoRA and 2/4/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8.
+- **Advanced algorithms**: GaLore, DoRA, LongLoRA, LLaMA Pro, LoRA+, LoftQ and Agent tuning.
+- **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA.
+- **Experiment monitors**: LlamaBoard, TensorBoard, Wandb, MLflow, etc.
+- **Faster inference**: OpenAI-style API, Gradio UI and CLI with vLLM worker.
+
+## Benchmark
+
+Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning), LLaMA-Factory's LoRA tuning offers up to **3.7 times faster** training speed with a better Rouge score on the advertising text generation task. By leveraging 4-bit quantization technique, LLaMA-Factory's QLoRA further improves the efficiency regarding the GPU memory.
+
+![benchmark](assets/benchmark.svg)
+
+<details><summary>Definitions</summary>
+
+- **Training Speed**: the number of training samples processed per second during the training. (bs=4, cutoff_len=1024)
+- **Rouge Score**: Rouge-2 score on the development set of the [advertising text generation](https://aclanthology.org/D19-1321.pdf) task. (bs=4, cutoff_len=1024)
+- **GPU Memory**: Peak GPU memory usage in 4-bit quantized training. (bs=1, cutoff_len=1024)
+- We adopt `pre_seq_len=128` for ChatGLM's P-Tuning and `lora_rank=32` for LLaMA-Factory's LoRA tuning.
+
+</details>
 
 ## Changelog
 
-[23/10/21] We supported **[NEFTune](https://arxiv.org/abs/2310.05914)** trick for fine-tuning. Try `--neft_alpha` argument to activate NEFTune, e.g., `--neft_alpha 5`.
+[24/03/21] Our paper "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" is available at arXiv!
+
+[24/03/20] We supported **FSDP+QLoRA** that fine-tunes a 70B model on 2x24GB GPUs. See `examples/fsdp_qlora` for usage.
+
+[24/03/13] We supported **[LoRA+](https://arxiv.org/abs/2402.12354)**. Try `loraplus_lr_ratio=16.0` to enable LoRA+ algorithm.
+
+[24/03/07] We supported gradient low-rank projection (**[GaLore](https://arxiv.org/abs/2403.03507)**) algorithm. Try `--use_galore` to use the memory-efficient optimizer.
+
+[24/03/07] We integrated **[vLLM](https://github.com/vllm-project/vllm)** for faster and concurrent inference. Try `--infer_backend vllm` to enjoy **270%** inference speed. (LoRA is not yet supported, merge it first.)
+
+<details><summary>Full Changelog</summary>
+
+[24/02/28] We supported weight-decomposed LoRA (**[DoRA](https://arxiv.org/abs/2402.09353)**). Try `--use_dora` to activate DoRA training.
+
+[24/02/15] We supported **block expansion** proposed by [LLaMA Pro](https://github.com/TencentARC/LLaMA-Pro). See `examples/extras/llama_pro` for usage.
+
+[24/02/05] Qwen1.5 (Qwen2 beta version) series models are supported in LLaMA-Factory. Check this [blog post](https://qwenlm.github.io/blog/qwen1.5/) for details.
+
+[24/01/18] We supported **agent tuning** for most models, equipping model with tool using abilities by fine-tuning with `--dataset glaive_toolcall`.
+
+[23/12/23] We supported **[unsloth](https://github.com/unslothai/unsloth)**'s implementation to boost LoRA tuning for the LLaMA, Mistral and Yi models. Try `--use_unsloth` argument to activate unsloth patch. It achieves **170%** speed in our benchmark, check [this page](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison) for details.
+
+[23/12/12] We supported fine-tuning the latest MoE model **[Mixtral 8x7B](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1)** in our framework. See hardware requirement [here](#hardware-requirement).
+
+[23/12/01] We supported downloading pre-trained models and datasets from the **[ModelScope Hub](https://modelscope.cn/models)** for Chinese mainland users. See [this tutorial](#use-modelscope-hub-optional) for usage.
+
+[23/10/21] We supported **[NEFTune](https://arxiv.org/abs/2310.05914)** trick for fine-tuning. Try `--neftune_noise_alpha` argument to activate NEFTune, e.g., `--neftune_noise_alpha 5`.
 
 [23/09/27] We supported **$S^2$-Attn** proposed by [LongLoRA](https://github.com/dvlab-research/LongLoRA) for the LLaMA models. Try `--shift_attn` argument to enable shift short attention.
 
 [23/09/23] We integrated MMLU, C-Eval and CMMLU benchmarks in this repo. See [this example](#evaluation) to evaluate your models.
 
-[23/09/10] We supported using **[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)** for the LLaMA models. Try `--flash_attn` argument to enable FlashAttention-2 if you are using RTX4090, A100 or H100 GPUs.
+[23/09/10] We supported **[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)**. Try `--flash_attn` argument to enable FlashAttention-2 if you are using RTX4090, A100 or H100 GPUs.
 
 [23/08/12] We supported **RoPE scaling** to extend the context length of the LLaMA models. Try `--rope_scaling linear` argument in training and `--rope_scaling dynamic` argument at inference to extrapolate the position embeddings.
 
@@ -48,43 +120,54 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 [23/06/03] We supported quantized training and inference (aka **[QLoRA](https://github.com/artidoro/qlora)**). Try `--quantization_bit 4/8` argument to work with quantized models.
 
+</details>
+
 ## Supported Models
 
 | Model                                                    | Model size                  | Default module    | Template  |
 | -------------------------------------------------------- | --------------------------- | ----------------- | --------- |
-| [Baichuan](https://github.com/baichuan-inc/Baichuan-13B) | 7B/13B                      | W_pack            | baichuan  |
-| [Baichuan2](https://github.com/baichuan-inc/Baichuan2)   | 7B/13B                      | W_pack            | baichuan2 |
+| [Baichuan2](https://huggingface.co/baichuan-inc)         | 7B/13B                      | W_pack            | baichuan2 |
 | [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   | -         |
-| [ChatGLM3](https://github.com/THUDM/ChatGLM3)            | 6B                          | query_key_value   | chatglm3  |
-| [Falcon](https://huggingface.co/tiiuae/falcon-7b)        | 7B/40B/180B                 | query_key_value   | -         |
-| [InternLM](https://github.com/InternLM/InternLM)         | 7B/20B                      | q_proj,v_proj     | intern    |
+| [ChatGLM3](https://huggingface.co/THUDM/chatglm3-6b)     | 6B                          | query_key_value   | chatglm3  |
+| [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B                  | q_proj,v_proj     | deepseek  |
+| [Falcon](https://huggingface.co/tiiuae)                  | 7B/40B/180B                 | query_key_value   | falcon    |
+| [Gemma](https://huggingface.co/google)                   | 2B/7B                       | q_proj,v_proj     | gemma     |
+| [InternLM2](https://huggingface.co/internlm)             | 7B/20B                      | wqkv              | intern2   |
 | [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    |
 | [Mistral](https://huggingface.co/mistralai)              | 7B                          | q_proj,v_proj     | mistral   |
-| [Phi-1.5](https://huggingface.co/microsoft/phi-1_5)      | 1.3B                        | Wqkv              | -         |
-| [Qwen](https://github.com/QwenLM/Qwen)                   | 7B/14B                      | c_attn            | qwen      |
-| [XVERSE](https://github.com/xverse-ai)                   | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
+| [Mixtral](https://huggingface.co/mistralai)              | 8x7B                        | q_proj,v_proj     | mistral   |
+| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                       | att_proj          | olmo      |
+| [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                   | q_proj,v_proj     | -         |
+| [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B             | c_attn            | qwen      |
+| [Qwen1.5](https://huggingface.co/Qwen)                   | 0.5B/1.8B/4B/7B/14B/72B     | q_proj,v_proj     | qwen      |
+| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                   | q_proj,v_proj     | -         |
+| [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
+| [Yi](https://huggingface.co/01-ai)                       | 6B/9B/34B                   | q_proj,v_proj     | yi        |
+| [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                 | q_proj,v_proj     | yuan      |
 
 > [!NOTE]
 > **Default module** is used for the `--lora_target` argument, you can use `--lora_target all` to specify all the available modules.
 >
 > For the "base" models, the `--template` argument can be chosen from `default`, `alpaca`, `vicuna` etc. But make sure to use the **corresponding template** for the "chat" models.
 
-Please refer to [template.py](src/llmtuner/extras/template.py) for a full list of models we supported.
+Please refer to [constants.py](src/llmtuner/extras/constants.py) for a full list of models we supported.
+
+You also can add a custom chat template to [template.py](src/llmtuner/data/template.py).
 
 ## Supported Training Approaches
 
-| Approach               |   Full-parameter   | Partial-parameter  |       LoRA         |       QLoRA        |
+| Approach               |     Full-tuning    |    Freeze-tuning   |       LoRA         |       QLoRA        |
 | ---------------------- | ------------------ | ------------------ | ------------------ | ------------------ |
 | Pre-Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | Supervised Fine-Tuning | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
-| Reward Modeling        |                    |                    | :white_check_mark: | :white_check_mark: |
-| PPO Training           |                    |                    | :white_check_mark: | :white_check_mark: |
-| DPO Training           | :white_check_mark: |                    | :white_check_mark: | :white_check_mark: |
+| Reward Modeling        | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
+| PPO Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
+| DPO Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
 > [!NOTE]
-> Use `--quantization_bit 4/8` argument to enable QLoRA.
+> Use `--quantization_bit 4` argument to enable QLoRA.
 
 ## Provided Datasets
 
@@ -106,8 +189,8 @@ Please refer to [template.py](src/llmtuner/extras/template.py) for a full list o
 
 - [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)
-- [Self-cognition (zh)](data/self_cognition.json)
+- [Alpaca GPT4 (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
+- [Self Cognition (zh)](data/self_cognition.json)
 - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
 - [ShareGPT (zh)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection)
 - [Guanaco Dataset (multilingual)](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
@@ -122,10 +205,15 @@ Please refer to [template.py](src/llmtuner/extras/template.py) for a full list o
 - [OpenPlatypus (en)](https://huggingface.co/datasets/garage-bAInd/Open-Platypus)
 - [CodeAlpaca 20k (en)](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k)
 - [Alpaca CoT (multilingual)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT)
+- [OpenOrca (en)](https://huggingface.co/datasets/Open-Orca/OpenOrca)
+- [SlimOrca (en)](https://huggingface.co/datasets/Open-Orca/SlimOrca)
 - [MathInstruct (en)](https://huggingface.co/datasets/TIGER-Lab/MathInstruct)
 - [Firefly 1.1M (zh)](https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M)
+- [Wiki QA (en)](https://huggingface.co/datasets/wiki_qa)
 - [Web QA (zh)](https://huggingface.co/datasets/suolyer/webqa)
 - [WebNovel (zh)](https://huggingface.co/datasets/zxbsmk/webnovel_cn)
+- [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
+- [deepctrl (en&zh)](https://www.modelscope.cn/datasets/deepctrl/deepctrl-sft-data)
 - [Ad Gen (zh)](https://huggingface.co/datasets/HasturOfficial/adgen)
 - [ShareGPT Hyperfiltered (en)](https://huggingface.co/datasets/totally-not-an-llm/sharegpt-hyperfiltered-3k)
 - [ShareGPT4 (en&zh)](https://huggingface.co/datasets/shibing624/sharegpt_gpt4)
@@ -133,6 +221,17 @@ Please refer to [template.py](src/llmtuner/extras/template.py) for a full list o
 - [AgentInstruct (en)](https://huggingface.co/datasets/THUDM/AgentInstruct)
 - [LMSYS Chat 1M (en)](https://huggingface.co/datasets/lmsys/lmsys-chat-1m)
 - [Evol Instruct V2 (en)](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k)
+- [Glaive Function Calling V2 (en)](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2)
+- [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
+- [Open Assistant (de)](https://huggingface.co/datasets/mayflowergmbh/oasst_de)
+- [Dolly 15k (de)](https://huggingface.co/datasets/mayflowergmbh/dolly-15k_de)
+- [Alpaca GPT4 (de)](https://huggingface.co/datasets/mayflowergmbh/alpaca-gpt4_de)
+- [OpenSchnabeltier (de)](https://huggingface.co/datasets/mayflowergmbh/openschnabeltier_de)
+- [Evol Instruct (de)](https://huggingface.co/datasets/mayflowergmbh/evol-instruct_de)
+- [Dolphin (de)](https://huggingface.co/datasets/mayflowergmbh/dolphin_de)
+- [Booksum (de)](https://huggingface.co/datasets/mayflowergmbh/booksum_de)
+- [Airoboros (de)](https://huggingface.co/datasets/mayflowergmbh/airoboros-3.0_de)
+- [Ultrachat (de)](https://huggingface.co/datasets/mayflowergmbh/ultra-chat_de)
 
 </details>
 
@@ -141,6 +240,9 @@ Please refer to [template.py](src/llmtuner/extras/template.py) for a full list o
 - [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)
+- [Orca DPO (en)](https://huggingface.co/datasets/Intel/orca_dpo_pairs)
+- [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
+- [Orca DPO (de)](https://huggingface.co/datasets/mayflowergmbh/intel_orca_dpo_pairs_de)
 
 </details>
 
@@ -155,14 +257,37 @@ huggingface-cli login
 
 ## Requirement
 
-- Python 3.8+ and PyTorch 1.13.1+
-- 🤗Transformers, Datasets, Accelerate, PEFT and TRL
-- sentencepiece, protobuf and tiktoken
-- fire, jieba, rouge-chinese and nltk (used at evaluation and predict)
-- gradio and matplotlib (used in web UI)
-- uvicorn, fastapi and sse-starlette (used in API)
+| Mandatory    | Minimum | Recommend |
+| ------------ | ------- | --------- |
+| python       | 3.8     | 3.10      |
+| torch        | 1.13.1  | 2.2.0     |
+| transformers | 4.37.2  | 4.39.1    |
+| datasets     | 2.14.3  | 2.17.1    |
+| accelerate   | 0.27.2  | 0.28.0    |
+| peft         | 0.9.0   | 0.10.0    |
+| trl          | 0.8.1   | 0.8.1     |
 
-And **powerful GPUs**!
+| Optional     | Minimum | Recommend |
+| ------------ | ------- | --------- |
+| CUDA         | 11.6    | 12.2      |
+| deepspeed    | 0.10.0  | 0.14.0    |
+| bitsandbytes | 0.39.0  | 0.43.0    |
+| flash-attn   | 2.3.0   | 2.5.6     |
+
+### Hardware Requirement
+
+\* *estimated*
+
+| Method | Bits |   7B  |  13B  |  30B  |   70B  |   8x7B |
+| ------ | ---- | ----- | ----- | ----- | ------ | ------ |
+| Full   | AMP  | 120GB | 240GB | 600GB | 1200GB |  900GB |
+| Full   |  16  |  60GB | 120GB | 300GB |  600GB |  400GB |
+| GaLore |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| Freeze |  16  |  20GB |  40GB |  80GB |  200GB |  160GB |
+| LoRA   |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| QLoRA  |   8  |  10GB |  20GB |  40GB |   80GB |   60GB |
+| QLoRA  |   4  |   6GB |  12GB |  24GB |   48GB |   30GB |
+| QLoRA  |   2  |   4GB |   8GB |  16GB |   24GB |   18GB |
 
 ## Getting Started
 
@@ -183,10 +308,34 @@ cd LLaMA-Factory
 pip install -r requirements.txt
 ```
 
-If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you will be required to install a pre-built version of `bitsandbytes` library, which supports CUDA 11.1 to 12.1.
+If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you will be required to install a pre-built version of `bitsandbytes` library, which supports CUDA 11.1 to 12.2, please select the appropriate [release version](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels) based on your CUDA version.
 
 ```bash
-pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
+pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.2.post2-py3-none-win_amd64.whl
+```
+
+To enable FlashAttention-2 on the Windows platform, you need to install the precompiled `flash-attn` library, which supports CUDA 12.1 to 12.2. Please download the corresponding version from [flash-attention](https://github.com/bdashore3/flash-attention/releases) based on your requirements.
+
+### Use ModelScope Hub (optional)
+
+If you have trouble with downloading models and datasets from Hugging Face, you can use LLaMA-Factory together with ModelScope in the following manner.
+
+```bash
+export USE_MODELSCOPE_HUB=1 # `set USE_MODELSCOPE_HUB=1` for Windows
+```
+
+Then you can train the corresponding model by specifying a model ID of the ModelScope Hub. (find a full list of model IDs at [ModelScope Hub](https://modelscope.cn/models))
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
+    --model_name_or_path modelscope/Llama-2-7b-ms \
+    ... # arguments (same as below)
+```
+
+LLaMA Board also supports using the models and datasets on the ModelScope Hub.
+
+```bash
+CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py
 ```
 
 ### Train on a single GPU
@@ -194,13 +343,20 @@ pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/downl
 > [!IMPORTANT]
 > If you want to train models on multiple GPUs, please refer to [Distributed Training](#distributed-training).
 
+
+#### LLaMA Board GUI
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+```
+
 #### Pre-Training
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage pt \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
     --dataset wiki_demo \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
@@ -222,8 +378,8 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage sft \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
     --dataset alpaca_gpt4_en \
     --template default \
     --finetuning_type lora \
@@ -246,65 +402,73 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage rm \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_sft_checkpoint \
+    --create_new_adapter \
     --dataset comparison_gpt4_en \
     --template default \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
-    --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-6 \
-    --num_train_epochs 1.0 \
-    --plot_loss \
-    --fp16
-```
-
-#### PPO Training
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage ppo \
-    --model_name_or_path path_to_llama_model \
-    --do_train \
-    --dataset alpaca_gpt4_en \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --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 \
     --fp16
 ```
 
+#### PPO Training
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
+    --stage ppo \
+    --do_train \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_sft_checkpoint \
+    --create_new_adapter \
+    --dataset alpaca_gpt4_en \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --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 \
+    --top_k 0 \
+    --top_p 0.9 \
+    --logging_steps 10 \
+    --save_steps 1000 \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --plot_loss \
+    --fp16
+```
+
+> [!TIP]
+> Use `--adapter_name_or_path path_to_sft_checkpoint,path_to_ppo_checkpoint` to infer the fine-tuned model.
+
+> [!WARNING]
+> Use `--per_device_train_batch_size=1` for LLaMA-2 models in fp16 PPO training.
+
 #### DPO Training
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage dpo \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_sft_checkpoint \
+    --create_new_adapter \
     --dataset comparison_gpt4_en \
     --template default \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
-    --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 \
@@ -317,19 +481,24 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
+> [!TIP]
+> Use `--adapter_name_or_path path_to_sft_checkpoint,path_to_dpo_checkpoint` to infer the fine-tuned model.
+
 ### Distributed Training
 
 #### Use Huggingface Accelerate
 
 ```bash
-accelerate config # configure the environment
-accelerate launch src/train_bash.py # arguments (same as above)
+accelerate launch --config_file config.yaml src/train_bash.py \
+    --ddp_timeout 180000000 \
+    ... # arguments (same as above)
 ```
 
-<details><summary>Example config for LoRA training</summary>
+<details><summary>Example config.yaml for LoRA training</summary>
 
 ```yaml
 compute_environment: LOCAL_MACHINE
+debug: false
 distributed_type: MULTI_GPU
 downcast_bf16: 'no'
 gpu_ids: all
@@ -348,15 +517,19 @@ use_cpu: false
 
 </details>
 
+> [!TIP]
+> We commend using Accelerate for LoRA tuning.
+
 #### Use DeepSpeed
 
 ```bash
-deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
+deepspeed --num_gpus 8 src/train_bash.py \
     --deepspeed ds_config.json \
+    --ddp_timeout 180000000 \
     ... # arguments (same as above)
 ```
 
-<details><summary>Example config for full-parameter training with DeepSpeed ZeRO-2</summary>
+<details><summary>Example ds_config.json for full-parameter training with DeepSpeed ZeRO-2</summary>
 
 ```json
 {
@@ -368,67 +541,84 @@ deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
   "fp16": {
     "enabled": "auto",
     "loss_scale": 0,
-    "initial_scale_power": 16,
     "loss_scale_window": 1000,
+    "initial_scale_power": 16,
     "hysteresis": 2,
     "min_loss_scale": 1
-  },  
+  },
+  "bf16": {
+    "enabled": "auto"
+  },
   "zero_optimization": {
     "stage": 2,
     "allgather_partitions": true,
     "allgather_bucket_size": 5e8,
+    "overlap_comm": true,
     "reduce_scatter": true,
     "reduce_bucket_size": 5e8,
-    "overlap_comm": false,
-    "contiguous_gradients": true
+    "contiguous_gradients": true,
+    "round_robin_gradients": true
   }
 }
 ```
 
 </details>
 
-### Export model
+> [!TIP]
+> Refer to [examples](examples) for more training scripts.
+
+### Merge LoRA weights and export model
 
 ```bash
-python src/export_model.py \
+CUDA_VISIBLE_DEVICES=0 python src/export_model.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
     --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint \
-    --export_dir path_to_export
+    --export_dir path_to_export \
+    --export_size 2 \
+    --export_legacy_format False
 ```
 
-### API Demo
+> [!WARNING]
+> Merging LoRA weights into a quantized model is not supported.
+
+> [!TIP]
+> Use `--model_name_or_path path_to_export` solely to use the exported model.
+> 
+> Use `--export_quantization_bit 4` and `--export_quantization_dataset data/c4_demo.json` to quantize the model with AutoGPTQ after merging the LoRA weights.
+
+### Inference with OpenAI-style API
 
 ```bash
-python src/api_demo.py \
+CUDA_VISIBLE_DEVICES=0 API_PORT=8000 python src/api_demo.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
-    --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint
+    --finetuning_type lora
 ```
 
-> [!NOTE]
+> [!TIP]
 > Visit `http://localhost:8000/docs` for API documentation.
 
-### CLI Demo
+### Inference with command line
 
 ```bash
-python src/cli_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/cli_demo.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
-    --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint
+    --finetuning_type lora
 ```
 
-### Web Demo
+### Inference with web browser
 
 ```bash
-python src/web_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/web_demo.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
-    --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint
+    --finetuning_type lora
 ```
 
 ### Evaluation
@@ -436,9 +626,9 @@ python src/web_demo.py \
 ```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 \
+    --adapter_name_or_path path_to_checkpoint \
     --template vanilla \
+    --finetuning_type lora \
     --task mmlu \
     --split test \
     --lang en \
@@ -451,44 +641,100 @@ CUDA_VISIBLE_DEVICES=0 python src/evaluate.py \
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage sft \
-    --model_name_or_path path_to_llama_model \
     --do_predict \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --dataset alpaca_gpt4_en \
     --template default \
     --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint \
     --output_dir path_to_predict_result \
-    --per_device_eval_batch_size 8 \
+    --per_device_eval_batch_size 1 \
     --max_samples 100 \
-    --predict_with_generate
+    --predict_with_generate \
+    --fp16
 ```
 
-> [!NOTE]
+> [!WARNING]
+> Use `--per_device_train_batch_size=1` for LLaMA-2 models in fp16 predict.
+
+> [!TIP]
 > We recommend using `--per_device_eval_batch_size=1` and `--max_target_length 128` at 4/8-bit predict.
 
+### Dockerize Training
+
+#### Get ready
+
+Necessary dockerized environment is needed, such as Docker or Docker Compose.
+
+#### Docker support
+
+```bash
+docker build -f ./Dockerfile -t llama-factory:latest .
+
+docker run --gpus=all -v ./hf_cache:/root/.cache/huggingface/ -v ./data:/app/data -v ./output:/app/output -p 7860:7860 --shm-size 16G --name llama_factory -d llama-factory:latest
+```
+
+#### Docker Compose support
+
+```bash
+docker compose -f ./docker-compose.yml up -d
+```
+
+> [!TIP]
+> Details about volume:
+> * hf_cache: Utilize Huggingface cache on the host machine. Reassignable if a cache already exists in a different directory.
+> * data: Place datasets on this dir of the host machine so that they can be selected on LLaMA Board GUI.
+> * output: Set export dir to this location so that the merged result can be accessed directly on the host machine.
+
 ## Projects using LLaMA Factory
 
-- **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: A large language model for Astronomy, based on ChatGLM2-6B and Qwen-14B.
-- **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: A large language model specialized in Chinese legal domain, based on Baichuan-13B, is capable of retrieving and reasoning on legal knowledge.
-- **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: A large language model specialized in Chinese medical domain, based on Baichuan-7B and ChatGLM-6B.
-- **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: A series of large language models for Chinese medical domain, based on LLaMA2-7B and Baichuan-13B.
+1. Wang et al. ESRL: Efficient Sampling-based Reinforcement Learning for Sequence Generation. 2023. [[arxiv]](https://arxiv.org/abs/2308.02223)
+1. Yu et al. Open, Closed, or Small Language Models for Text Classification? 2023. [[arxiv]](https://arxiv.org/abs/2308.10092)
+1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
+1. Luceri et al. Leveraging Large Language Models to Detect Influence Campaigns in Social Media. 2023. [[arxiv]](https://arxiv.org/abs/2311.07816)
+1. Zhang et al. Alleviating Hallucinations of Large Language Models through Induced Hallucinations. 2023. [[arxiv]](https://arxiv.org/abs/2312.15710)
+1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
+1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
+1. Choi et al. FACT-GPT: Fact-Checking Augmentation via Claim Matching with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2402.05904)
+1. Zhang et al. AutoMathText: Autonomous Data Selection with Language Models for Mathematical Texts. 2024. [[arxiv]](https://arxiv.org/abs/2402.07625)
+1. Lyu et al. KnowTuning: Knowledge-aware Fine-tuning for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11176)
+1. Yang et al. LaCo: Large Language Model Pruning via Layer Collaps. 2024. [[arxiv]](https://arxiv.org/abs/2402.11187)
+1. Bhardwaj et al. Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic. 2024. [[arxiv]](https://arxiv.org/abs/2402.11746)
+1. Yang et al. Enhancing Empathetic Response Generation by Augmenting LLMs with Small-scale Empathetic Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11801)
+1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. 2024. [[arxiv]](https://arxiv.org/abs/2402.11809)
+1. Cao et al. Head-wise Shareable Attention for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11819)
+1. Zhang et al. Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languages. 2024. [[arxiv]](https://arxiv.org/abs/2402.12204)
+1. Kim et al. Efficient and Effective Vocabulary Expansion Towards Multilingual Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.14714)
+1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
+1. Huang et al. Key-Point-Driven Data Synthesis with its Enhancement on Mathematical Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2403.02333)
+1. Duan et al. Negating Negatives: Alignment without Human Positive Samples via Distributional Dispreference Optimization. 2024. [[arxiv]](https://arxiv.org/abs/2403.03419)
+1. Xie and Schwertfeger. Empowering Robotics with Large Language Models: osmAG Map Comprehension with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2403.08228)
+1. **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: A large language model for Astronomy, based on ChatGLM2-6B and Qwen-14B.
+1. **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: A large language model specialized in Chinese legal domain, based on Baichuan-13B, is capable of retrieving and reasoning on legal knowledge.
+1. **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: A large language model specialized in Chinese medical domain, based on Baichuan-7B and ChatGLM-6B.
+1. **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: A series of large language models for Chinese medical domain, based on LLaMA2-7B and Baichuan-13B.
+1. **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**: A series of MBTI Personality large language models, capable of giving any LLM 16 different personality types based on different datasets and training methods.
+
+> [!TIP]
+> If you have a project that should be incorporated, please contact via email or create a pull request.
 
 ## License
 
 This repository is licensed under the [Apache-2.0 License](LICENSE).
 
-Please follow the model licenses to use the corresponding model weights: [Baichuan](https://huggingface.co/baichuan-inc/Baichuan-13B-Base/resolve/main/Community%20License%20for%20Baichuan-13B%20Model.pdf) / [Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/resolve/main/Community%20License%20for%20Baichuan2%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [InternLM](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [Phi-1.5](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/LICENSE) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf)
+Please follow the model licenses to use the corresponding model weights: [Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## Citation
 
 If this work is helpful, please kindly cite as:
 
 ```bibtex
-@Misc{llama-factory,
-  title = {LLaMA Factory},
-  author = {hiyouga},
-  howpublished = {\url{https://github.com/hiyouga/LLaMA-Factory}},
-  year = {2023}
+@article{zheng2024llamafactory,
+  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models}, 
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Yongqiang Ma},
+  journal={arXiv preprint arXiv:2403.13372},
+  year={2024},
+  url={http://arxiv.org/abs/2403.13372}
 }
 ```
 

README_zh.md

@@ -1,40 +1,112 @@
-# LLaMA Factory: 轻松的大模型训练与评估
+![# LLaMA Factory](assets/logo.png)
 
 [![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)
 [![PyPI](https://img.shields.io/pypi/v/llmtuner)](https://pypi.org/project/llmtuner/)
 [![Downloads](https://static.pepy.tech/badge/llmtuner)](https://pypi.org/project/llmtuner/)
+[![Citation](https://img.shields.io/badge/citation-26-green)](#使用了-llama-factory-的项目)
 [![GitHub pull request](https://img.shields.io/badge/PRs-welcome-blue)](https://github.com/hiyouga/LLaMA-Factory/pulls)
-[![Discord](https://dcbadge.vercel.app/api/server/e73gccsSd?compact=true&style=flat)](https://discord.gg/e73gccsSd)
+[![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
+[![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
+[![Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
+[![Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing)
 
 👋 加入我们的[微信群](assets/wechat.jpg)。
 
 \[ [English](README.md) | 中文 \]
 
-## LLaMA Board: 通过一站式网页界面快速上手 LLaMA Factory
+**微调大模型可以像这样轻松…**
 
-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）
+https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd-d76c6d0a6594
 
-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。
+选择你的打开方式：
 
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846-2d88920d5ba1
+- **Colab**：https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing
+- **本地机器**：请见[如何使用](#如何使用)
+
+## 目录
+
+- [项目特色](#项目特色)
+- [性能指标](#性能指标)
+- [更新日志](#更新日志)
+- [模型](#模型)
+- [训练方法](#训练方法)
+- [数据集](#数据集)
+- [软硬件依赖](#软硬件依赖)
+- [如何使用](#如何使用)
+- [使用了 LLaMA Factory 的项目](#使用了-llama-factory-的项目)
+- [协议](#协议)
+- [引用](#引用)
+- [致谢](#致谢)
+
+## 项目特色
+
+- **多种模型**：LLaMA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
+- **集成方法**：（增量）预训练、指令监督微调、奖励模型训练、PPO 训练和 DPO 训练。
+- **多种精度**：32 比特全参数微调、16 比特冻结微调、16 比特 LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8 的 2/4/8 比特 QLoRA 微调。
+- **先进算法**：GaLore、DoRA、LongLoRA、LLaMA Pro、LoRA+、LoftQ 和 Agent 微调。
+- **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。
+- **实验监控**：LlamaBoard、TensorBoard、Wandb、MLflow 等等。
+- **极速推理**：基于 vLLM 的 OpenAI 风格 API、浏览器界面和命令行接口。
+
+## 性能指标
+
+与 ChatGLM 官方的 [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning) 微调相比，LLaMA-Factory 的 LoRA 微调提供了 **3.7 倍**的加速比，同时在广告文案生成任务上取得了更高的 Rouge 分数。结合 4 比特量化技术，LLaMA-Factory 的 QLoRA 微调进一步降低了 GPU 显存消耗。
+
+![benchmark](assets/benchmark.svg)
+
+<details><summary>变量定义</summary>
+
+- **Training Speed**: 训练阶段每秒处理的样本数量。（批处理大小=4，截断长度=1024）
+- **Rouge Score**: [广告文案生成](https://aclanthology.org/D19-1321.pdf)任务验证集上的 Rouge-2 分数。（批处理大小=4，截断长度=1024）
+- **GPU Memory**: 4 比特量化训练的 GPU 显存峰值。（批处理大小=1，截断长度=1024）
+- 我们在 ChatGLM 的 P-Tuning 中采用 `pre_seq_len=128`，在 LLaMA-Factory 的 LoRA 微调中采用 `lora_rank=32`。
+
+</details>
 
 ## 更新日志
 
-[23/10/21] 我们支持了 **[NEFTune](https://arxiv.org/abs/2310.05914)** 训练技巧。请使用 `--neft_alpha` 参数启用 NEFTune，例如 `--neft_alpha 5`。
+[24/03/21] 我们的论文 "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" 可在 arXiv 上查看！
+
+[24/03/20] 我们支持了能在 2x24GB GPU 上微调 70B 模型的 **FSDP+QLoRA**。详细用法请参照 `examples/fsdp_qlora`。
+
+[24/03/13] 我们支持了 **[LoRA+](https://arxiv.org/abs/2402.12354)**。请使用 `loraplus_lr_ratio=16.0` 参数开启 LoRA+ 方法。
+
+[24/03/07] 我们支持了梯度低秩投影（**[GaLore](https://arxiv.org/abs/2403.03507)**）算法。请使用 `--use_galore` 参数切换显存高效的优化器。
+
+[24/03/07] 我们集成了 **[vLLM](https://github.com/vllm-project/vllm)** 以实现极速并发推理。请使用 `--infer_backend vllm` 来获得 **270%** 的推理速度。（尚不支持 LoRA，请先合并权重。）
+
+<details><summary>展开日志</summary>
+
+[24/02/28] 我们支持了 **[DoRA](https://arxiv.org/abs/2402.09353)** 微调。请使用 `--use_dora` 参数进行 DoRA 微调。
+
+[24/02/15] 我们支持了 [LLaMA Pro](https://github.com/TencentARC/LLaMA-Pro) 提出的**块扩展**方法。详细用法请参照 `examples/extras/llama_pro`。
+
+[24/02/05] Qwen1.5（Qwen2 测试版）系列模型已在 LLaMA-Factory 中实现微调支持。详情请查阅该[博客页面](https://qwenlm.github.io/zh/blog/qwen1.5/)。
+
+[24/01/18] 我们针对绝大多数模型实现了 **Agent 微调**，微调时指定 `--dataset glaive_toolcall` 即可使模型获得工具调用能力。
+
+[23/12/23] 我们针对 LLaMA, Mistral 和 Yi 模型支持了 **[unsloth](https://github.com/unslothai/unsloth)** 的 LoRA 训练加速。请使用 `--use_unsloth` 参数启用 unsloth 优化。该方法可提供 **170%** 的训练速度，详情请查阅[此页面](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison)。
+
+[23/12/12] 我们支持了微调最新的混合专家模型 **[Mixtral 8x7B](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1)**。硬件需求请查阅[此处](#硬件依赖)。
+
+[23/12/01] 我们支持了从 **[魔搭社区](https://modelscope.cn/models)** 下载预训练模型和数据集。详细用法请参照 [此教程](#使用魔搭社区可跳过)。
+
+[23/10/21] 我们支持了 **[NEFTune](https://arxiv.org/abs/2310.05914)** 训练技巧。请使用 `--neftune_noise_alpha` 参数启用 NEFTune，例如 `--neftune_noise_alpha 5`。
 
 [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。
 
 [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)。
 
-[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。
+[23/09/10] 我们支持了 **[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。
 
 [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。
 
 [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。
+[23/07/31] 我们支持了**数据流式加载**。请使用 `--streaming` 和 `--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](https://huggingface.co/hiyouga/Llama-2-Chinese-13b-chat) / [Baichuan](https://huggingface.co/hiyouga/Baichuan-13B-sft)）。
 
@@ -48,30 +120,41 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 [23/06/03] 我们实现了 4 比特的 LoRA 训练（也称 **[QLoRA](https://github.com/artidoro/qlora)**）。请使用 `--quantization_bit 4` 参数进行 4 比特量化微调。
 
+</details>
+
 ## 模型
 
 | 模型名                                                   | 模型大小                     | 默认模块           | Template  |
 | -------------------------------------------------------- | --------------------------- | ----------------- | --------- |
-| [Baichuan](https://github.com/baichuan-inc/Baichuan-13B) | 7B/13B                      | W_pack            | baichuan  |
-| [Baichuan2](https://github.com/baichuan-inc/Baichuan2)   | 7B/13B                      | W_pack            | baichuan2 |
+| [Baichuan2](https://huggingface.co/baichuan-inc)         | 7B/13B                      | W_pack            | baichuan2 |
 | [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   | -         |
-| [ChatGLM3](https://github.com/THUDM/ChatGLM3)            | 6B                          | query_key_value   | chatglm3  |
-| [Falcon](https://huggingface.co/tiiuae/falcon-7b)        | 7B/40B/180B                 | query_key_value   | -         |
-| [InternLM](https://github.com/InternLM/InternLM)         | 7B/20B                      | q_proj,v_proj     | intern    |
+| [ChatGLM3](https://huggingface.co/THUDM/chatglm3-6b)     | 6B                          | query_key_value   | chatglm3  |
+| [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B                  | q_proj,v_proj     | deepseek  |
+| [Falcon](https://huggingface.co/tiiuae)                  | 7B/40B/180B                 | query_key_value   | falcon    |
+| [Gemma](https://huggingface.co/google)                   | 2B/7B                       | q_proj,v_proj     | gemma     |
+| [InternLM2](https://huggingface.co/internlm)             | 7B/20B                      | wqkv              | intern2   |
 | [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    |
 | [Mistral](https://huggingface.co/mistralai)              | 7B                          | q_proj,v_proj     | mistral   |
-| [Phi-1.5](https://huggingface.co/microsoft/phi-1_5)      | 1.3B                        | Wqkv              | -         |
-| [Qwen](https://github.com/QwenLM/Qwen)                   | 7B/14B                      | c_attn            | qwen      |
-| [XVERSE](https://github.com/xverse-ai)                   | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
+| [Mixtral](https://huggingface.co/mistralai)              | 8x7B                        | q_proj,v_proj     | mistral   |
+| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                       | att_proj          | olmo      |
+| [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                   | q_proj,v_proj     | -         |
+| [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B             | c_attn            | qwen      |
+| [Qwen1.5](https://huggingface.co/Qwen)                   | 0.5B/1.8B/4B/7B/14B/72B     | q_proj,v_proj     | qwen      |
+| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                   | q_proj,v_proj     | -         |
+| [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
+| [Yi](https://huggingface.co/01-ai)                       | 6B/9B/34B                   | q_proj,v_proj     | yi        |
+| [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                 | q_proj,v_proj     | yuan      |
 
 > [!NOTE]
 > **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。
 >
 > 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**。
 
-项目所支持模型的完整列表请参阅 [template.py](src/llmtuner/extras/template.py)。
+项目所支持模型的完整列表请参阅 [constants.py](src/llmtuner/extras/constants.py)。
+
+您也可以在 [template.py](src/llmtuner/data/template.py) 中添加自己的对话模板。
 
 ## 训练方法
 
@@ -79,12 +162,12 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 | ---------------------- | ------------------ | ------------------ | ------------------ | ------------------ |
 | 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: |
+| 奖励模型训练            | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
+| PPO 训练               | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
+| DPO 训练               | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
 > [!NOTE]
-> 请使用 `--quantization_bit 4/8` 参数来启用 QLoRA 训练。
+> 请使用 `--quantization_bit 4` 参数来启用 QLoRA 训练。
 
 ## 数据集
 
@@ -106,8 +189,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 - [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)
-- [Self-cognition (zh)](data/self_cognition.json)
+- [Alpaca GPT4 (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
+- [Self Cognition (zh)](data/self_cognition.json)
 - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
 - [ShareGPT (zh)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection)
 - [Guanaco Dataset (multilingual)](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
@@ -122,10 +205,15 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [OpenPlatypus (en)](https://huggingface.co/datasets/garage-bAInd/Open-Platypus)
 - [CodeAlpaca 20k (en)](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k)
 - [Alpaca CoT (multilingual)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT)
+- [OpenOrca (en)](https://huggingface.co/datasets/Open-Orca/OpenOrca)
+- [SlimOrca (en)](https://huggingface.co/datasets/Open-Orca/SlimOrca)
 - [MathInstruct (en)](https://huggingface.co/datasets/TIGER-Lab/MathInstruct)
 - [Firefly 1.1M (zh)](https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M)
+- [Wiki QA (en)](https://huggingface.co/datasets/wiki_qa)
 - [Web QA (zh)](https://huggingface.co/datasets/suolyer/webqa)
 - [WebNovel (zh)](https://huggingface.co/datasets/zxbsmk/webnovel_cn)
+- [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
+- [deepctrl (en&zh)](https://www.modelscope.cn/datasets/deepctrl/deepctrl-sft-data)
 - [Ad Gen (zh)](https://huggingface.co/datasets/HasturOfficial/adgen)
 - [ShareGPT Hyperfiltered (en)](https://huggingface.co/datasets/totally-not-an-llm/sharegpt-hyperfiltered-3k)
 - [ShareGPT4 (en&zh)](https://huggingface.co/datasets/shibing624/sharegpt_gpt4)
@@ -133,6 +221,17 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [AgentInstruct (en)](https://huggingface.co/datasets/THUDM/AgentInstruct)
 - [LMSYS Chat 1M (en)](https://huggingface.co/datasets/lmsys/lmsys-chat-1m)
 - [Evol Instruct V2 (en)](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k)
+- [Glaive Function Calling V2 (en)](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2)
+- [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
+- [Open Assistant (de)](https://huggingface.co/datasets/mayflowergmbh/oasst_de)
+- [Dolly 15k (de)](https://huggingface.co/datasets/mayflowergmbh/dolly-15k_de)
+- [Alpaca GPT4 (de)](https://huggingface.co/datasets/mayflowergmbh/alpaca-gpt4_de)
+- [OpenSchnabeltier (de)](https://huggingface.co/datasets/mayflowergmbh/openschnabeltier_de)
+- [Evol Instruct (de)](https://huggingface.co/datasets/mayflowergmbh/evol-instruct_de)
+- [Dolphin (de)](https://huggingface.co/datasets/mayflowergmbh/dolphin_de)
+- [Booksum (de)](https://huggingface.co/datasets/mayflowergmbh/booksum_de)
+- [Airoboros (de)](https://huggingface.co/datasets/mayflowergmbh/airoboros-3.0_de)
+- [Ultrachat (de)](https://huggingface.co/datasets/mayflowergmbh/ultra-chat_de)
 
 </details>
 
@@ -141,6 +240,9 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [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)
+- [Orca DPO (en)](https://huggingface.co/datasets/Intel/orca_dpo_pairs)
+- [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
+- [Orca DPO (de)](https://huggingface.co/datasets/mayflowergmbh/intel_orca_dpo_pairs_de)
 
 </details>
 
@@ -153,16 +255,39 @@ pip install --upgrade huggingface_hub
 huggingface-cli login
 ```
 
-## 软件依赖
+## 软硬件依赖
 
-- Python 3.8+ 和 PyTorch 1.13.1+
-- 🤗Transformers, Datasets, Accelerate, PEFT 和 TRL
-- sentencepiece, protobuf 和 tiktoken
-- fire, jieba, rouge-chinese 和 nltk (用于评估及预测)
-- gradio 和 matplotlib (用于网页端交互)
-- uvicorn, fastapi 和 sse-starlette (用于 API)
+| 必需项       | 至少     | 推荐      |
+| ------------ | ------- | --------- |
+| python       | 3.8     | 3.10      |
+| torch        | 1.13.1  | 2.2.0     |
+| transformers | 4.37.2  | 4.39.1    |
+| datasets     | 2.14.3  | 2.17.1    |
+| accelerate   | 0.27.2  | 0.28.0    |
+| peft         | 0.9.0   | 0.10.0    |
+| trl          | 0.8.1   | 0.8.1     |
 
-以及 **强而有力的 GPU**！
+| 可选项       | 至少     | 推荐      |
+| ------------ | ------- | --------- |
+| CUDA         | 11.6    | 12.2      |
+| deepspeed    | 0.10.0  | 0.14.0    |
+| bitsandbytes | 0.39.0  | 0.43.0    |
+| flash-attn   | 2.3.0   | 2.5.6     |
+
+### 硬件依赖
+
+\* *估算值*
+
+| 训练方法 | 精度 |   7B  |  13B  |  30B  |   70B  |   8x7B |
+| ------- | ---- | ----- | ----- | ----- | ------ | ------ |
+| 全参数   | AMP  | 120GB | 240GB | 600GB | 1200GB |  900GB |
+| 全参数   |  16  |  60GB | 120GB | 300GB |  600GB |  400GB |
+| GaLore  |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| 部分参数 |  16  |  20GB |  40GB |  80GB |  200GB |  160GB |
+| LoRA    |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| QLoRA   |   8  |  10GB |  20GB |  40GB |   80GB |   60GB |
+| QLoRA   |   4  |   6GB |  12GB |  24GB |   48GB |   30GB |
+| QLoRA   |   2  |   4GB |   8GB |  16GB |   24GB |   18GB |
 
 ## 如何使用
 
@@ -183,10 +308,34 @@ cd LLaMA-Factory
 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.
+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 12.2, 请根据您的 CUDA 版本情况选择适合的[发布版本](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels)。
 
 ```bash
-pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
+pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.2.post2-py3-none-win_amd64.whl
+```
+
+如果要在 Windows 平台上开启 FlashAttention-2，需要安装预编译的 `flash-attn` 库，支持 CUDA 12.1 到 12.2，请根据需求到 [flash-attention](https://github.com/bdashore3/flash-attention/releases) 下载对应版本安装。
+
+### 使用魔搭社区（可跳过）
+
+如果您在 Hugging Face 模型和数据集的下载中遇到了问题，可以通过下述方法使用魔搭社区。
+
+```bash
+export USE_MODELSCOPE_HUB=1 # Windows 使用 `set USE_MODELSCOPE_HUB=1`
+```
+
+接着即可通过指定模型名称来训练对应的模型。（在[魔搭社区](https://modelscope.cn/models)查看所有可用的模型）
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
+    --model_name_or_path modelscope/Llama-2-7b-ms \
+    ... # 参数同下
+```
+
+LLaMA Board 同样支持魔搭社区的模型和数据集下载。
+
+```bash
+CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py
 ```
 
 ### 单 GPU 训练
@@ -194,13 +343,19 @@ pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/downl
 > [!IMPORTANT]
 > 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。
 
+#### LLaMA Board GUI
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+```
+
 #### 预训练
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage pt \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
     --dataset wiki_demo \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
@@ -222,8 +377,8 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage sft \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
     --dataset alpaca_gpt4_zh \
     --template default \
     --finetuning_type lora \
@@ -246,21 +401,21 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage rm \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_sft_checkpoint \
+    --create_new_adapter \
     --dataset comparison_gpt4_zh \
     --template default \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
-    --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-6 \
+    --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
     --plot_loss \
     --fp16
@@ -271,39 +426,48 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage ppo \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_sft_checkpoint \
+    --create_new_adapter \
     --dataset alpaca_gpt4_zh \
     --template default \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
-    --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 \
+    --top_k 0 \
+    --top_p 0.9 \
     --logging_steps 10 \
     --save_steps 1000 \
     --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
-    --plot_loss
+    --plot_loss \
+    --fp16
 ```
 
+> [!TIP]
+> 使用 `--adapter_name_or_path path_to_sft_checkpoint,path_to_ppo_checkpoint` 来进行微调模型的推理。
+
+> [!WARNING]
+> 如果使用 fp16 精度进行 LLaMA-2 模型的 PPO 训练，请使用 `--per_device_train_batch_size=1`。
+
 #### DPO 训练
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage dpo \
-    --model_name_or_path path_to_llama_model \
     --do_train \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_sft_checkpoint \
+    --create_new_adapter \
     --dataset comparison_gpt4_zh \
     --template default \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
-    --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 \
@@ -316,19 +480,24 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
+> [!TIP]
+> 使用 `--adapter_name_or_path path_to_sft_checkpoint,path_to_dpo_checkpoint` 来进行微调模型的推理。
+
 ### 多 GPU 分布式训练
 
 #### 使用 Huggingface Accelerate
 
 ```bash
-accelerate config # 首先配置分布式环境
-accelerate launch src/train_bash.py # 参数同上
+accelerate launch --config_file config.yaml src/train_bash.py \
+    --ddp_timeout 180000000 \
+    ... # 参数同上
 ```
 
-<details><summary>LoRA 训练的 Accelerate 配置示例</summary>
+<details><summary>使用 Accelerate 进行 LoRA 训练的 config.yaml 示例</summary>
 
 ```yaml
 compute_environment: LOCAL_MACHINE
+debug: false
 distributed_type: MULTI_GPU
 downcast_bf16: 'no'
 gpu_ids: all
@@ -347,15 +516,19 @@ use_cpu: false
 
 </details>
 
+> [!TIP]
+> 我们推荐使用 Accelerate 进行 LoRA 训练。
+
 #### 使用 DeepSpeed
 
 ```bash
-deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
+deepspeed --num_gpus 8 src/train_bash.py \
     --deepspeed ds_config.json \
+    --ddp_timeout 180000000 \
     ... # 参数同上
 ```
 
-<details><summary>使用 DeepSpeed ZeRO-2 进行全参数训练的 DeepSpeed 配置示例</summary>
+<details><summary>使用 DeepSpeed ZeRO-2 进行全参数训练的 ds_config.json 示例</summary>
 
 ```json
 {
@@ -367,67 +540,84 @@ deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
   "fp16": {
     "enabled": "auto",
     "loss_scale": 0,
-    "initial_scale_power": 16,
     "loss_scale_window": 1000,
+    "initial_scale_power": 16,
     "hysteresis": 2,
     "min_loss_scale": 1
-  },  
+  },
+  "bf16": {
+    "enabled": "auto"
+  },
   "zero_optimization": {
     "stage": 2,
     "allgather_partitions": true,
     "allgather_bucket_size": 5e8,
+    "overlap_comm": true,
     "reduce_scatter": true,
     "reduce_bucket_size": 5e8,
-    "overlap_comm": false,
-    "contiguous_gradients": true
+    "contiguous_gradients": true,
+    "round_robin_gradients": true
   }
 }
 ```
 
 </details>
 
-### 导出微调后的完整模型
+> [!TIP]
+> 更多训练脚本请查看 [examples](examples)。
+
+### 合并 LoRA 权重并导出模型
 
 ```bash
-python src/export_model.py \
+CUDA_VISIBLE_DEVICES=0 python src/export_model.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
     --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint \
-    --export_dir path_to_export
+    --export_dir path_to_export \
+    --export_size 2 \
+    --export_legacy_format False
 ```
 
-### API 服务
+> [!WARNING]
+> 尚不支持量化模型的 LoRA 权重合并及导出。
+
+> [!TIP]
+> 仅使用 `--model_name_or_path path_to_export` 来加载导出后的模型。
+> 
+> 合并 LoRA 权重之后可再次使用 `--export_quantization_bit 4` 和 `--export_quantization_dataset data/c4_demo.json` 基于 AutoGPTQ 量化模型。
+
+### 使用 OpenAI 风格 API 推理
 
 ```bash
-python src/api_demo.py \
+CUDA_VISIBLE_DEVICES=0 API_PORT=8000 python src/api_demo.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
-    --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint
+    --finetuning_type lora
 ```
 
-> [!NOTE]
+> [!TIP]
 > 关于 API 文档请见 `http://localhost:8000/docs`。
 
-### 命令行测试
+### 使用命令行推理
 
 ```bash
-python src/cli_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/cli_demo.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
-    --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint
+    --finetuning_type lora
 ```
 
-### 浏览器测试
+### 使用浏览器推理
 
 ```bash
-python src/web_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/web_demo.py \
     --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --template default \
-    --finetuning_type lora \
-    --checkpoint_dir path_to_checkpoint
+    --finetuning_type lora
 ```
 
 ### 模型评估
@@ -435,9 +625,9 @@ python src/web_demo.py \
 ```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 \
+    --adapter_name_or_path path_to_checkpoint \
     --template vanilla \
+    --finetuning_type lora \
     --task ceval \
     --split validation \
     --lang zh \
@@ -450,44 +640,74 @@ CUDA_VISIBLE_DEVICES=0 python src/evaluate.py \
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --stage sft \
-    --model_name_or_path path_to_llama_model \
     --do_predict \
+    --model_name_or_path path_to_llama_model \
+    --adapter_name_or_path path_to_checkpoint \
     --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 \
+    --per_device_eval_batch_size 1 \
     --max_samples 100 \
-    --predict_with_generate
+    --predict_with_generate \
+    --fp16
 ```
 
-> [!NOTE]
+> [!WARNING]
+> 如果使用 fp16 精度进行 LLaMA-2 模型的预测，请使用 `--per_device_eval_batch_size=1`。
+
+> [!TIP]
 > 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`。
 
 ## 使用了 LLaMA Factory 的项目
 
-- **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: 天文大模型 StarWhisper，基于 ChatGLM2-6B 和 Qwen-14B 在天文数据上微调而得。
-- **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: 中文法律领域大模型 DISC-LawLLM，基于 Baichuan-13B 微调而得，具有法律推理和知识检索能力。
-- **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: 孙思邈中文医疗大模型 Sumsimiao，基于 Baichuan-7B 和 ChatGLM-6B 在中文医疗数据上微调而得。
-- **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: 医疗大模型项目 CareGPT，基于 LLaMA2-7B 和 Baichuan-13B 在中文医疗数据上微调而得。
+1. Wang et al. ESRL: Efficient Sampling-based Reinforcement Learning for Sequence Generation. 2023. [[arxiv]](https://arxiv.org/abs/2308.02223)
+1. Yu et al. Open, Closed, or Small Language Models for Text Classification? 2023. [[arxiv]](https://arxiv.org/abs/2308.10092)
+1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
+1. Luceri et al. Leveraging Large Language Models to Detect Influence Campaigns in Social Media. 2023. [[arxiv]](https://arxiv.org/abs/2311.07816)
+1. Zhang et al. Alleviating Hallucinations of Large Language Models through Induced Hallucinations. 2023. [[arxiv]](https://arxiv.org/abs/2312.15710)
+1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
+1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
+1. Choi et al. FACT-GPT: Fact-Checking Augmentation via Claim Matching with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2402.05904)
+1. Zhang et al. AutoMathText: Autonomous Data Selection with Language Models for Mathematical Texts. 2024. [[arxiv]](https://arxiv.org/abs/2402.07625)
+1. Lyu et al. KnowTuning: Knowledge-aware Fine-tuning for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11176)
+1. Yang et al. LaCo: Large Language Model Pruning via Layer Collaps. 2024. [[arxiv]](https://arxiv.org/abs/2402.11187)
+1. Bhardwaj et al. Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic. 2024. [[arxiv]](https://arxiv.org/abs/2402.11746)
+1. Yang et al. Enhancing Empathetic Response Generation by Augmenting LLMs with Small-scale Empathetic Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11801)
+1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. 2024. [[arxiv]](https://arxiv.org/abs/2402.11809)
+1. Cao et al. Head-wise Shareable Attention for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11819)
+1. Zhang et al. Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languages. 2024. [[arxiv]](https://arxiv.org/abs/2402.12204)
+1. Kim et al. Efficient and Effective Vocabulary Expansion Towards Multilingual Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.14714)
+1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
+1. Huang et al. Key-Point-Driven Data Synthesis with its Enhancement on Mathematical Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2403.02333)
+1. Duan et al. Negating Negatives: Alignment without Human Positive Samples via Distributional Dispreference Optimization. 2024. [[arxiv]](https://arxiv.org/abs/2403.03419)
+1. Xie and Schwertfeger. Empowering Robotics with Large Language Models: osmAG Map Comprehension with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2403.08228)
+1. **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: 天文大模型 StarWhisper，基于 ChatGLM2-6B 和 Qwen-14B 在天文数据上微调而得。
+1. **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: 中文法律领域大模型 DISC-LawLLM，基于 Baichuan-13B 微调而得，具有法律推理和知识检索能力。
+1. **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: 孙思邈中文医疗大模型 Sumsimiao，基于 Baichuan-7B 和 ChatGLM-6B 在中文医疗数据上微调而得。
+1. **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: 医疗大模型项目 CareGPT，基于 LLaMA2-7B 和 Baichuan-13B 在中文医疗数据上微调而得。
+1. **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**：MBTI性格大模型项目，根据数据集与训练方式让任意 LLM 拥有 16 个不同的性格类型。
+
+> [!TIP]
+> 如果您有项目希望添加至上述列表，请通过邮件联系或者创建一个 PR。
 
 ## 协议
 
 本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源。
 
-使用模型权重时，请遵循对应的模型协议：[Baichuan](https://huggingface.co/baichuan-inc/Baichuan-13B-Base/resolve/main/Community%20License%20for%20Baichuan-13B%20Model.pdf) / [Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/resolve/main/Community%20License%20for%20Baichuan2%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [InternLM](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [Phi-1.5](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/LICENSE) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf)
+使用模型权重时，请遵循对应的模型协议：[Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## 引用
 
 如果您觉得此项目有帮助，请考虑以下列格式引用
 
 ```bibtex
-@Misc{llama-factory,
-  title = {LLaMA Factory},
-  author = {hiyouga},
-  howpublished = {\url{https://github.com/hiyouga/LLaMA-Factory}},
-  year = {2023}
+@article{zheng2024llamafactory,
+  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models}, 
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Yongqiang Ma},
+  journal={arXiv preprint arXiv:2403.13372},
+  year={2024},
+  url={http://arxiv.org/abs/2403.13372}
 }
 ```
 

data/README.md

@@ -2,21 +2,32 @@ If you are using a custom dataset, please provide your dataset definition in the
 
 ```json
 "dataset_name": {
-  "hf_hub_url": "the name of the dataset repository on the Hugging Face hub. (if specified, ignore below 3 arguments)",
-  "script_url": "the name of the directory containing a dataset loading script. (if specified, ignore below 2 arguments)",
-  "file_name": "the name of the dataset file in the this directory. (required if above are not specified)",
+  "hf_hub_url": "the name of the dataset repository on the Hugging Face hub. (if specified, ignore script_url and file_name)",
+  "ms_hub_url": "the name of the dataset repository on the ModelScope hub. (if specified, ignore script_url and file_name)",
+  "script_url": "the name of the directory containing a dataset loading script. (if specified, ignore file_name)",
+  "file_name": "the name of the dataset file in this directory. (required if above are not specified)",
   "file_sha1": "the SHA-1 hash value of the dataset file. (optional, does not affect training)",
   "subset": "the name of the subset. (optional, default: None)",
+  "folder": "the name of the folder of the dataset repository on the Hugging Face hub. (optional, default: None)",
   "ranking": "whether the dataset is a preference dataset or not. (default: false)",
   "formatting": "the format of the dataset. (optional, default: alpaca, can be chosen from {alpaca, sharegpt})",
-  "columns": {
-    "prompt": "the column name in the dataset containing the prompts. (default: instruction, for alpaca)",
-    "query": "the column name in the dataset containing the queries. (default: input, for alpaca)",
-    "response": "the column name in the dataset containing the responses. (default: output, for alpaca)",
-    "history": "the column name in the dataset containing the histories. (default: None, for alpaca)",
-    "messages": "the column name in the dataset containing the messages. (default: conversations, for sharegpt)",
-    "role": "the key in the message represents the identity. (default: from, for sharegpt)",
-    "content": "the key in the message represents the content. (default: value, for sharegpt)"
+  "columns (optional)": {
+    "prompt": "the column name in the dataset containing the prompts. (default: instruction)",
+    "query": "the column name in the dataset containing the queries. (default: input)",
+    "response": "the column name in the dataset containing the responses. (default: output)",
+    "history": "the column name in the dataset containing the histories. (default: None)",
+    "messages": "the column name in the dataset containing the messages. (default: conversations)",
+    "system": "the column name in the dataset containing the system prompts. (default: None)",
+    "tools": "the column name in the dataset containing the tool description. (default: None)"
+  },
+  "tags (optional, used for the sharegpt format)": {
+    "role_tag": "the key in the message represents the identity. (default: from)",
+    "content_tag": "the key in the message represents the content. (default: value)",
+    "user_tag": "the value of the role_tag represents the user. (default: human)",
+    "assistant_tag": "the value of the role_tag represents the assistant. (default: gpt)",
+    "observation_tag": "the value of the role_tag represents the tool results. (default: observation)",
+    "function_tag": "the value of the role_tag represents the function call. (default: function_call)",
+    "system_tag": "the value of the role_tag represents the system prompt. (default: system, can override system column)"
   }
 }
 ```
@@ -31,6 +42,7 @@ Currently we support dataset in **alpaca** or **sharegpt** format, the dataset i
     "instruction": "user instruction (required)",
     "input": "user input (optional)",
     "output": "model response (required)",
+    "system": "system prompt (optional)",
     "history": [
       ["user instruction in the first round (optional)", "model response in the first round (optional)"],
       ["user instruction in the second round (optional)", "model response in the second round (optional)"]
@@ -47,14 +59,15 @@ Regarding the above dataset, the `columns` in `dataset_info.json` should be:
     "prompt": "instruction",
     "query": "input",
     "response": "output",
+    "system": "system",
     "history": "history"
   }
 }
 ```
 
-where the `prompt` and `response` columns should contain non-empty values, represent instruction and response respectively. The `query` column will be concatenated with the `prompt` column and used as input for the model.
+The `query` column will be concatenated with the `prompt` column and used as the user prompt, then the user prompt would be `prompt\nquery`. The `response` column represents the model response.
 
-The `history` column is a list consisting string tuples representing query-response pairs in history. Note that the responses **in each round will be used for training**.
+The `system` column will be used as the system prompt. The `history` column is a list consisting string tuples representing prompt-response pairs in the history. Note that the responses in the history **will also be used for training**.
 
 For the pre-training datasets, only the `prompt` column will be used for training.
 
@@ -85,7 +98,9 @@ The dataset in sharegpt format should follow the below format:
         "from": "gpt",
         "value": "model response"
       }
-    ]
+    ],
+    "system": "system prompt (optional)",
+    "tools": "tool description (optional)"
   }
 ]
 ```
@@ -96,12 +111,18 @@ Regarding the above dataset, the `columns` in `dataset_info.json` should be:
 "dataset_name": {
   "columns": {
     "messages": "conversations",
-    "role": "from",
-    "content": "value"
+    "system": "system",
+    "tools": "tools"
+  },
+  "tags": {
+    "role_tag": "from",
+    "content_tag": "value",
+    "user_tag": "human",
+    "assistant_tag": "gpt"
   }
 }
 ```
 
-where the `messages` column should be a list whose length is even, and follow the `u/a/u/a/u/a` order.
+where the `messages` column should be a list following the `u/a/u/a/u/a` order.
 
 Pre-training datasets and preference datasets are incompatible with the sharegpt format yet.

data/README_zh.md

@@ -2,21 +2,32 @@
 
 ```json
 "数据集名称": {
-  "hf_hub_url": "Hugging Face 上的项目地址（若指定，则忽略下列三个参数）",
-  "script_url": "包含数据加载脚本的本地文件夹名称（若指定，则忽略下列两个参数）",
+  "hf_hub_url": "Hugging Face 的数据集仓库地址（若指定，则忽略 script_url 和 file_name）",
+  "ms_hub_url": "ModelScope 的数据集仓库地址（若指定，则忽略 script_url 和 file_name）",
+  "script_url": "包含数据加载脚本的本地文件夹名称（若指定，则忽略 file_name）",
   "file_name": "该目录下数据集文件的名称（若上述参数未指定，则此项必需）",
-  "file_sha1": "数据集文件的SHA-1哈希值（可选，留空不影响训练）",
+  "file_sha1": "数据集文件的 SHA-1 哈希值（可选，留空不影响训练）",
   "subset": "数据集子集的名称（可选，默认：None）",
+  "folder": "Hugging Face 仓库的文件夹名称（可选，默认：None）",
   "ranking": "是否为偏好数据集（可选，默认：False）",
   "formatting": "数据集格式（可选，默认：alpaca，可以为 alpaca 或 sharegpt）",
-  "columns": {
-    "prompt": "数据集代表提示词的表头名称（默认：instruction，用于 alpaca 格式）",
-    "query": "数据集代表请求的表头名称（默认：input，用于 alpaca 格式）",
-    "response": "数据集代表回答的表头名称（默认：output，用于 alpaca 格式）",
-    "history": "数据集代表历史对话的表头名称（默认：None，用于 alpaca 格式）",
-    "messages": "数据集代表消息列表的表头名称（默认：conversations，用于 sharegpt 格式）",
-    "role": "消息中代表发送者身份的键名（默认：from，用于 sharegpt 格式）",
-    "content": "消息中代表文本内容的键名（默认：value，用于 sharegpt 格式）"
+  "columns（可选）": {
+    "prompt": "数据集代表提示词的表头名称（默认：instruction）",
+    "query": "数据集代表请求的表头名称（默认：input）",
+    "response": "数据集代表回答的表头名称（默认：output）",
+    "history": "数据集代表历史对话的表头名称（默认：None）",
+    "messages": "数据集代表消息列表的表头名称（默认：conversations）",
+    "system": "数据集代表系统提示的表头名称（默认：None）",
+    "tools": "数据集代表工具描述的表头名称（默认：None）"
+  },
+  "tags（可选，用于 sharegpt 格式）": {
+    "role_tag": "消息中代表发送者身份的键名（默认：from）",
+    "content_tag": "消息中代表文本内容的键名（默认：value）",
+    "user_tag": "消息中代表用户的 role_tag（默认：human）",
+    "assistant_tag": "消息中代表助手的 role_tag（默认：gpt）",
+    "observation_tag": "消息中代表工具返回结果的 role_tag（默认：observation）",
+    "function_tag": "消息中代表工具调用的 role_tag（默认：function_call）",
+    "system_tag": "消息中代表系统提示的 role_tag（默认：system，会覆盖 system 列）"
   }
 }
 ```
@@ -31,6 +42,7 @@
     "instruction": "用户指令（必填）",
     "input": "用户输入（选填）",
     "output": "模型回答（必填）",
+    "system": "系统提示词（选填）",
     "history": [
       ["第一轮指令（选填）", "第一轮回答（选填）"],
       ["第二轮指令（选填）", "第二轮回答（选填）"]
@@ -47,14 +59,15 @@
     "prompt": "instruction",
     "query": "input",
     "response": "output",
+    "system": "system",
     "history": "history"
   }
 }
 ```
 
-其中 `prompt` 和 `response` 列应当是非空的字符串，分别代表用户指令和模型回答。`query` 列的内容将会和 `prompt` 列拼接作为模型输入。
+其中 `query` 列对应的内容会与 `prompt` 列对应的内容拼接后作为用户指令，即用户指令为 `prompt\nquery`。`response` 列对应的内容为模型回答。
 
-`history` 列是由多个字符串二元组构成的列表，分别代表历史消息中每轮的指令和回答。注意每轮的模型回答**均会被用于训练**。
+`system` 列对应的内容将被作为系统提示词。`history` 列是由多个字符串二元组构成的列表，分别代表历史消息中每轮的指令和回答。注意历史消息中的回答**也会被用于训练**。
 
 对于预训练数据集，仅 `prompt` 列中的内容会用于模型训练。
 
@@ -85,7 +98,9 @@
         "from": "gpt",
         "value": "模型回答"
       }
-    ]
+    ],
+    "system": "系统提示词（选填）",
+    "tools": "工具描述（选填）"
   }
 ]
 ```
@@ -96,12 +111,18 @@
 "数据集名称": {
   "columns": {
     "messages": "conversations",
-    "role": "from",
-    "content": "value"
+    "system": "system",
+    "tools": "tools"
+  },
+  "tags": {
+    "role_tag": "from",
+    "content_tag": "value",
+    "user_tag": "human",
+    "assistant_tag": "gpt"
   }
 }
 ```
 
-其中 `messages` 列必须为偶数长度的列表，且符合 `用户/模型/用户/模型/用户/模型` 的顺序。
+其中 `messages` 列应当是一个列表，且符合 `用户/模型/用户/模型/用户/模型` 的顺序。
 
 预训练数据集和偏好数据集尚不支持 sharegpt 格式。

data/alpaca_data_zh_51k.json.REMOVED.git-id

@@ -1 +1 @@
-fc9a6a3458caca2af8dafc6181773fe10c6d8657
+34c723573fbc2d7601f6d9c882ccf5aa4f9bcc4b

data/belle_multiturn/belle_multiturn.py

@@ -1,7 +1,10 @@
+import os
 import json
 import datasets
 
 
+_HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
+
 _DESCRIPTION = "BELLE multiturn chat dataset."
 
 _CITATION = """\
@@ -13,9 +16,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M"
+_HOMEPAGE = "{}/datasets/BelleGroup/multiturn_chat_0.8M".format(_HF_ENDPOINT)
 _LICENSE = "gpl-3.0"
-_URL = "https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json"
+_URL = "{}/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json".format(_HF_ENDPOINT)
 
 
 class BelleMultiturn(datasets.GeneratorBasedBuilder):
@@ -24,9 +27,7 @@ class BelleMultiturn(datasets.GeneratorBasedBuilder):
 
     def _info(self):
         features = datasets.Features({
-            "instruction": datasets.Value("string"),
-            "output": datasets.Value("string"),
-            "history": datasets.Sequence(datasets.Sequence(datasets.Value("string")))
+            "conversations": [{"from": datasets.Value("string"), "value": datasets.Value("string")}]
         })
         return datasets.DatasetInfo(
             description=_DESCRIPTION,
@@ -51,6 +52,7 @@ class BelleMultiturn(datasets.GeneratorBasedBuilder):
         with open(filepath, "r", encoding="utf-8") as f:
             for key, row in enumerate(f):
                 data = json.loads(row)
+                conversations = []
                 prompt = data["instruction"].strip()
                 response = data["output"].strip()
 
@@ -58,7 +60,8 @@ class BelleMultiturn(datasets.GeneratorBasedBuilder):
                 human_idx = prompt.rfind("Human:")
                 query = prompt[human_idx+6:assist_idx].strip()
                 prompt = prompt[:human_idx].strip()
-                history = []
+                conversations.insert(0, {"from": "gpt", "value": response})
+                conversations.insert(0, {"from": "human", "value": query})
 
                 while prompt.rfind("Assistant:") != -1:
                     assist_idx = prompt.rfind("Assistant:")
@@ -66,13 +69,10 @@ class BelleMultiturn(datasets.GeneratorBasedBuilder):
                     if human_idx != -1:
                         old_query = prompt[human_idx+6:assist_idx].strip()
                         old_resp = prompt[assist_idx+10:].strip()
-                        history.insert(0, (old_query, old_resp))
+                        conversations.insert(0, {"from": "gpt", "value": old_resp})
+                        conversations.insert(0, {"from": "human", "value": old_query})
                     else:
                         break
                     prompt = prompt[:human_idx].strip()
 
-                yield key, {
-                    "instruction": query,
-                    "output": response,
-                    "history": history
-                }
+                yield key, {"conversations": conversations}

data/example_dataset/example_dataset.py

@@ -1,6 +1,6 @@
 import json
 import datasets
-from typing import Any, Dict, List
+from typing import Any, Dict, Generator, List, Tuple
 
 
 _DESCRIPTION = "An example of dataset."
@@ -40,7 +40,7 @@ class ExampleDataset(datasets.GeneratorBasedBuilder):
             )
         ]
 
-    def _generate_examples(self, filepath: str) -> Dict[int, Dict[str, Any]]:
+    def _generate_examples(self, filepath: str) -> Generator[Tuple[int, Dict[str, Any]], None, None]:
         example_dataset = json.load(open(filepath, "r", encoding="utf-8"))
         for key, example in enumerate(example_dataset):
             yield key, example

data/glaive_toolcall_10k.json.REMOVED.git-id

@@ -0,0 +1 @@
+4748dff00d1dc42768a5b6cc772143c313017812

data/hh_rlhf_en/hh_rlhf_en.py

@@ -1,13 +1,14 @@
+import os
 import json
 import datasets
 from typing import List
 
-
+_HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
 _DESCRIPTION = "Human preference data about helpfulness and harmlessness."
 _CITATION = ""
-_HOMEPAGE = "https://huggingface.co/datasets/Anthropic/hh-rlhf"
+_HOMEPAGE = "{}/datasets/Anthropic/hh-rlhf".format(_HF_ENDPOINT)
 _LICENSE = "mit"
-_URL = "https://huggingface.co/datasets/Anthropic/hh-rlhf/resolve/main/"
+_URL = "{}/datasets/Anthropic/hh-rlhf/resolve/main/".format(_HF_ENDPOINT)
 _URLS = {
     "train": [
         _URL + "harmless-base/train.jsonl.gz",

data/orca_rlhf.json.REMOVED.git-id

@@ -0,0 +1 @@
+736bcedea2b24a1414765c6d69cbdafaea839f3c

data/sharegpt_zh_27k.json.REMOVED.git-id

@@ -1 +0,0 @@
-38c89869c6aeca2a3af9ea1e09afe460f9b46810

data/ultra_chat/ultra_chat.py

@@ -1,7 +1,9 @@
+import os
 import json
 import datasets
 from typing import List
 
+_HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
 
 _DESCRIPTION = "UltraChat: Large-scale, Informative, and Diverse Multi-round Dialogue Data."
 
@@ -16,9 +18,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "https://huggingface.co/datasets/stingning/ultrachat"
+_HOMEPAGE = "{}/datasets/stingning/ultrachat".format(_HF_ENDPOINT)
 _LICENSE = "cc-by-nc-4.0"
-_BASE_DATA_URL = "https://huggingface.co/datasets/stingning/ultrachat/resolve/main/train_{idx}.jsonl"
+_BASE_DATA_URL = "{}/datasets/stingning/ultrachat/resolve/main/train_{{idx}}.jsonl".format(_HF_ENDPOINT)
 
 
 class UltraChat(datasets.GeneratorBasedBuilder):
@@ -66,6 +68,4 @@ class UltraChat(datasets.GeneratorBasedBuilder):
                         "from": "human" if i % 2 == 0 else "gpt",
                         "value": content[i]
                     } for i in range(len(content))]
-                    yield key, {
-                        "conversations": conversations
-                    }
+                    yield key, {"conversations": conversations}

data/wiki_demo.txt

@@ -1,50 +0,0 @@
-Machine learning (ML) is a field devoted to understanding and building methods that let machines "learn" – that is, methods that leverage data to improve computer performance on some set of tasks.
-Machine learning algorithms build a model based on sample data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so. Machine learning algorithms are used in a wide variety of applications, such as in medicine, email filtering, speech recognition, agriculture, and computer vision, where it is difficult or unfeasible to develop conventional algorithms to perform the needed tasks.
-A subset of machine learning is closely related to computational statistics, which focuses on making predictions using computers, but not all machine learning is statistical learning. The study of mathematical optimization delivers methods, theory and application domains to the field of machine learning. Data mining is a related field of study, focusing on exploratory data analysis through unsupervised learning.
-Some implementations of machine learning use data and neural networks in a way that mimics the working of a biological brain.
-In its application across business problems, machine learning is also referred to as predictive analytics.
-Learning algorithms work on the basis that strategies, algorithms, and inferences that worked well in the past are likely to continue working well in the future. These inferences can sometimes be obvious, such as "since the sun rose every morning for the last 10,000 days, it will probably rise tomorrow morning as well". Other times, they can be more nuanced, such as "X% of families have geographically separate species with color variants, so there is a Y% chance that undiscovered black swans exist".
-Machine learning programs can perform tasks without being explicitly programmed to do so. It involves computers learning from data provided so that they carry out certain tasks. For simple tasks assigned to computers, it is possible to program algorithms telling the machine how to execute all steps required to solve the problem at hand; on the computer's part, no learning is needed. For more advanced tasks, it can be challenging for a human to manually create the needed algorithms. In practice, it can turn out to be more effective to help the machine develop its own algorithm, rather than having human programmers specify every needed step.
-The discipline of machine learning employs various approaches to teach computers to accomplish tasks where no fully satisfactory algorithm is available. In cases where vast numbers of potential answers exist, one approach is to label some of the correct answers as valid. This can then be used as training data for the computer to improve the algorithm(s) it uses to determine correct answers. For example, to train a system for the task of digital character recognition, the MNIST dataset of handwritten digits has often been used.
-The term machine learning was coined in 1959 by Arthur Samuel, an IBM employee and pioneer in the field of computer gaming and artificial intelligence. The synonym self-teaching computers was also used in this time period.
-By the early 1960s an experimental "learning machine" with punched tape memory, called Cybertron, had been developed by Raytheon Company to analyze sonar signals, electrocardiograms, and speech patterns using rudimentary reinforcement learning. It was repetitively "trained" by a human operator/teacher to recognize patterns and equipped with a "goof" button to cause it to re-evaluate incorrect decisions. A representative book on research into machine learning during the 1960s was Nilsson's book on Learning Machines, dealing mostly with machine learning for pattern classification. Interest related to pattern recognition continued into the 1970s, as described by Duda and Hart in 1973. In 1981 a report was given on using teaching strategies so that a neural network learns to recognize 40 characters (26 letters, 10 digits, and 4 special symbols) from a computer terminal.
-Tom M. Mitchell provided a widely quoted, more formal definition of the algorithms studied in the machine learning field: "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E." This definition of the tasks in which machine learning is concerned offers a fundamentally operational definition rather than defining the field in cognitive terms. This follows Alan Turing's proposal in his paper "Computing Machinery and Intelligence", in which the question "Can machines think?" is replaced with the question "Can machines do what we (as thinking entities) can do?".
-Modern-day machine learning has two objectives, one is to classify data based on models which have been developed, the other purpose is to make predictions for future outcomes based on these models. A hypothetical algorithm specific to classifying data may use computer vision of moles coupled with supervised learning in order to train it to classify the cancerous moles. A machine learning algorithm for stock trading may inform the trader of future potential predictions.
-As a scientific endeavor, machine learning grew out of the quest for artificial intelligence (AI). In the early days of AI as an academic discipline, some researchers were interested in having machines learn from data. They attempted to approach the problem with various symbolic methods, as well as what were then termed "neural networks"; these were mostly perceptrons and other models that were later found to be reinventions of the generalized linear models of statistics. Probabilistic reasoning was also employed, especially in automated medical diagnosis.: 488 
-However, an increasing emphasis on the logical, knowledge-based approach caused a rift between AI and machine learning. Probabilistic systems were plagued by theoretical and practical problems of data acquisition and representation.: 488  By 1980, expert systems had come to dominate AI, and statistics was out of favor. Work on symbolic/knowledge-based learning did continue within AI, leading to inductive logic programming, but the more statistical line of research was now outside the field of AI proper, in pattern recognition and information retrieval.: 708–710, 755  Neural networks research had been abandoned by AI and computer science around the same time. This line, too, was continued outside the AI/CS field, as "connectionism", by researchers from other disciplines including Hopfield, Rumelhart, and Hinton. Their main success came in the mid-1980s with the reinvention of backpropagation.: 25 
-Machine learning (ML), reorganized and recognized as its own field, started to flourish in the 1990s. The field changed its goal from achieving artificial intelligence to tackling solvable problems of a practical nature. It shifted focus away from the symbolic approaches it had inherited from AI, and toward methods and models borrowed from statistics, fuzzy logic, and probability theory.
-Machine learning and data mining often employ the same methods and overlap significantly, but while machine learning focuses on prediction, based on known properties learned from the training data, data mining focuses on the discovery of (previously) unknown properties in the data (this is the analysis step of knowledge discovery in databases). Data mining uses many machine learning methods, but with different goals; on the other hand, machine learning also employs data mining methods as "unsupervised learning" or as a preprocessing step to improve learner accuracy. Much of the confusion between these two research communities (which do often have separate conferences and separate journals, ECML PKDD being a major exception) comes from the basic assumptions they work with: in machine learning, performance is usually evaluated with respect to the ability to reproduce known knowledge, while in knowledge discovery and data mining (KDD) the key task is the discovery of previously unknown knowledge. Evaluated with respect to known knowledge, an uninformed (unsupervised) method will easily be outperformed by other supervised methods, while in a typical KDD task, supervised methods cannot be used due to the unavailability of training data.
-Machine learning also has intimate ties to optimization: many learning problems are formulated as minimization of some loss function on a training set of examples. Loss functions express the discrepancy between the predictions of the model being trained and the actual problem instances (for example, in classification, one wants to assign a label to instances, and models are trained to correctly predict the pre-assigned labels of a set of examples).
-The difference between optimization and machine learning arises from the goal of generalization: while optimization algorithms can minimize the loss on a training set, machine learning is concerned with minimizing the loss on unseen samples. Characterizing the generalization of various learning algorithms is an active topic of current research, especially for deep learning algorithms.
-Machine learning and statistics are closely related fields in terms of methods, but distinct in their principal goal: statistics draws population inferences from a sample, while machine learning finds generalizable predictive patterns. According to Michael I. Jordan, the ideas of machine learning, from methodological principles to theoretical tools, have had a long pre-history in statistics. He also suggested the term data science as a placeholder to call the overall field.
-Leo Breiman distinguished two statistical modeling paradigms: data model and algorithmic model, wherein "algorithmic model" means more or less the machine learning algorithms like Random Forest.
-Some statisticians have adopted methods from machine learning, leading to a combined field that they call statistical learning.
-Analytical and computational techniques derived from deep-rooted physics of disordered systems can be extended to large-scale problems, including machine learning, e.g., to analyze the weight space of deep neural networks. Statistical physics is thus finding applications in the area of medical diagnostics.
-A core objective of a learner is to generalize from its experience. Generalization in this context is the ability of a learning machine to perform accurately on new, unseen examples/tasks after having experienced a learning data set. The training examples come from some generally unknown probability distribution (considered representative of the space of occurrences) and the learner has to build a general model about this space that enables it to produce sufficiently accurate predictions in new cases.
-The computational analysis of machine learning algorithms and their performance is a branch of theoretical computer science known as computational learning theory via the Probably Approximately Correct Learning (PAC) model. Because training sets are finite and the future is uncertain, learning theory usually does not yield guarantees of the performance of algorithms. Instead, probabilistic bounds on the performance are quite common. The bias–variance decomposition is one way to quantify generalization error.
-For the best performance in the context of generalization, the complexity of the hypothesis should match the complexity of the function underlying the data. If the hypothesis is less complex than the function, then the model has under fitted the data. If the complexity of the model is increased in response, then the training error decreases. But if the hypothesis is too complex, then the model is subject to overfitting and generalization will be poorer.
-In addition to performance bounds, learning theorists study the time complexity and feasibility of learning. In computational learning theory, a computation is considered feasible if it can be done in polynomial time. There are two kinds of time complexity results: Positive results show that a certain class of functions can be learned in polynomial time. Negative results show that certain classes cannot be learned in polynomial time.
-Machine learning approaches are traditionally divided into three broad categories, which correspond to learning paradigms, depending on the nature of the "signal" or "feedback" available to the learning system:
-Supervised learning: The computer is presented with example inputs and their desired outputs, given by a "teacher", and the goal is to learn a general rule that maps inputs to outputs.
-Unsupervised learning: No labels are given to the learning algorithm, leaving it on its own to find structure in its input. Unsupervised learning can be a goal in itself (discovering hidden patterns in data) or a means towards an end (feature learning).
-Reinforcement learning: A computer program interacts with a dynamic environment in which it must perform a certain goal (such as driving a vehicle or playing a game against an opponent). As it navigates its problem space, the program is provided feedback that's analogous to rewards, which it tries to maximize. Although each algorithm has advantages and limitations, no single algorithm works for all problems.
-Supervised learning algorithms build a mathematical model of a set of data that contains both the inputs and the desired outputs. The data is known as training data, and consists of a set of training examples. Each training example has one or more inputs and the desired output, also known as a supervisory signal. In the mathematical model, each training example is represented by an array or vector, sometimes called a feature vector, and the training data is represented by a matrix. Through iterative optimization of an objective function, supervised learning algorithms learn a function that can be used to predict the output associated with new inputs. An optimal function will allow the algorithm to correctly determine the output for inputs that were not a part of the training data. An algorithm that improves the accuracy of its outputs or predictions over time is said to have learned to perform that task.
-Types of supervised-learning algorithms include active learning, classification and regression. Classification algorithms are used when the outputs are restricted to a limited set of values, and regression algorithms are used when the outputs may have any numerical value within a range. As an example, for a classification algorithm that filters emails, the input would be an incoming email, and the output would be the name of the folder in which to file the email.
-Similarity learning is an area of supervised machine learning closely related to regression and classification, but the goal is to learn from examples using a similarity function that measures how similar or related two objects are. It has applications in ranking, recommendation systems, visual identity tracking, face verification, and speaker verification.
-Unsupervised learning algorithms take a set of data that contains only inputs, and find structure in the data, like grouping or clustering of data points. The algorithms, therefore, learn from test data that has not been labeled, classified or categorized. Instead of responding to feedback, unsupervised learning algorithms identify commonalities in the data and react based on the presence or absence of such commonalities in each new piece of data. A central application of unsupervised learning is in the field of density estimation in statistics, such as finding the probability density function. Though unsupervised learning encompasses other domains involving summarizing and explaining data features. Unsupervised learning algorithms streamlined the process of survey and graph large indel based haplotypes of a gene of interest from pan-genome. 
-Cluster analysis is the assignment of a set of observations into subsets (called clusters) so that observations within the same cluster are similar according to one or more predesignated criteria, while observations drawn from different clusters are dissimilar. Different clustering techniques make different assumptions on the structure of the data, often defined by some similarity metric and evaluated, for example, by internal compactness, or the similarity between members of the same cluster, and separation, the difference between clusters. Other methods are based on estimated density and graph connectivity.
-Semi-supervised learning falls between unsupervised learning (without any labeled training data) and supervised learning (with completely labeled training data). Some of the training examples are missing training labels, yet many machine-learning researchers have found that unlabeled data, when used in conjunction with a small amount of labeled data, can produce a considerable improvement in learning accuracy.
-In weakly supervised learning, the training labels are noisy, limited, or imprecise; however, these labels are often cheaper to obtain, resulting in larger effective training sets.
-Reinforcement learning is an area of machine learning concerned with how software agents ought to take actions in an environment so as to maximize some notion of cumulative reward. Due to its generality, the field is studied in many other disciplines, such as game theory, control theory, operations research, information theory, simulation-based optimization, multi-agent systems, swarm intelligence, statistics and genetic algorithms. In machine learning, the environment is typically represented as a Markov decision process (MDP). Many reinforcements learning algorithms use dynamic programming techniques. Reinforcement learning algorithms do not assume knowledge of an exact mathematical model of the MDP and are used when exact models are infeasible. Reinforcement learning algorithms are used in autonomous vehicles or in learning to play a game against a human opponent.
-Dimensionality reduction is a process of reducing the number of random variables under consideration by obtaining a set of principal variables. In other words, it is a process of reducing the dimension of the feature set, also called the "number of features". Most of the dimensionality reduction techniques can be considered as either feature elimination or extraction. One of the popular methods of dimensionality reduction is principal component analysis (PCA). PCA involves changing higher-dimensional data (e.g., 3D) to a smaller space (e.g., 2D). This results in a smaller dimension of data (2D instead of 3D), while keeping all original variables in the model without changing the data. The manifold hypothesis proposes that high-dimensional data sets lie along low-dimensional manifolds, and many dimensionality reduction techniques make this assumption, leading to the area of manifold learning and manifold regularization.
-Although machine learning has been transformative in some fields, machine-learning programs often fail to deliver expected results. Reasons for this are numerous: lack of (suitable) data, lack of access to the data, data bias, privacy problems, badly chosen tasks and algorithms, wrong tools and people, lack of resources, and evaluation problems.
-In 2018, a self-driving car from Uber failed to detect a pedestrian, who was killed after a collision. Attempts to use machine learning in healthcare with the IBM Watson system failed to deliver even after years of time and billions of dollars invested.
-Machine learning has been used as a strategy to update the evidence related to a systematic review and increased reviewer burden related to the growth of biomedical literature. While it has improved with training sets, it has not yet developed sufficiently to reduce the workload burden without limiting the necessary sensitivity for the findings research themselves.
-Machine learning approaches in particular can suffer from different data biases. A machine learning system trained specifically on current customers may not be able to predict the needs of new customer groups that are not represented in the training data. When trained on human-made data, machine learning is likely to pick up the constitutional and unconscious biases already present in society. Language models learned from data have been shown to contain human-like biases. Machine learning systems used for criminal risk assessment have been found to be biased against black people. In 2015, Google photos would often tag black people as gorillas, and in 2018 this still was not well resolved, but Google reportedly was still using the workaround to remove all gorillas from the training data, and thus was not able to recognize real gorillas at all. Similar issues with recognizing non-white people have been found in many other systems. In 2016, Microsoft tested a chatbot that learned from Twitter, and it quickly picked up racist and sexist language. Because of such challenges, the effective use of machine learning may take longer to be adopted in other domains. Concern for fairness in machine learning, that is, reducing bias in machine learning and propelling its use for human good is increasingly expressed by artificial intelligence scientists, including Fei-Fei Li, who reminds engineers that "There's nothing artificial about AI...It's inspired by people, it's created by people, and—most importantly—it impacts people. It is a powerful tool we are only just beginning to understand, and that is a profound responsibility."
-Learners can also disappoint by "learning the wrong lesson". A toy example is that an image classifier trained only on pictures of brown horses and black cats might conclude that all brown patches are likely to be horses. A real-world example is that, unlike humans, current image classifiers often do not primarily make judgments from the spatial relationship between components of the picture, and they learn relationships between pixels that humans are oblivious to, but that still correlate with images of certain types of real objects. Modifying these patterns on a legitimate image can result in "adversarial" images that the system misclassifies.
-Adversarial vulnerabilities can also result in nonlinear systems, or from non-pattern perturbations. Some systems are so brittle that changing a single adversarial pixel predictably induces misclassification.[citation needed] Machine learning models are often vulnerable to manipulation and/or evasion via adversarial machine learning.
-Researchers have demonstrated how backdoors can be placed undetectably into classifying (e.g., for categories "spam" and well-visible "not spam" of posts) machine learning models which are often developed and/or trained by third parties. Parties can change the classification of any input, including in cases for which a type of data/software transparency is provided, possibly including white-box access.
-Machine learning poses a host of ethical questions. Systems that are trained on datasets collected with biases may exhibit these biases upon use (algorithmic bias), thus digitizing cultural prejudices. For example, in 1988, the UK's Commission for Racial Equality found that St. George's Medical School had been using a computer program trained from data of previous admissions staff and this program had denied nearly 60 candidates who were found to be either women or had non-European sounding names. Using job hiring data from a firm with racist hiring policies may lead to a machine learning system duplicating the bias by scoring job applicants by similarity to previous successful applicants. Responsible collection of data and documentation of algorithmic rules used by a system thus is a critical part of machine learning.
-AI can be well-equipped to make decisions in technical fields, which rely heavily on data and historical information. These decisions rely on the objectivity and logical reasoning. Because human languages contain biases, machines trained on language corpora will necessarily also learn these biases.
-Other forms of ethical challenges, not related to personal biases, are seen in health care. There are concerns among health care professionals that these systems might not be designed in the public's interest but as income-generating machines. This is especially true in the United States where there is a long-standing ethical dilemma of improving health care, but also increase profits. For example, the algorithms could be designed to provide patients with unnecessary tests or medication in which the algorithm's proprietary owners hold stakes. There is potential for machine learning in health care to provide professionals an additional tool to diagnose, medicate, and plan recovery paths for patients, but this requires these biases to be mitigated.
-Since the 2010s, advances in both machine learning algorithms and computer hardware have led to more efficient methods for training deep neural networks (a particular narrow subdomain of machine learning) that contain many layers of non-linear hidden units. By 2019, graphic processing units (GPUs), often with AI-specific enhancements, had displaced CPUs as the dominant method of training large-scale commercial cloud AI. OpenAI estimated the hardware computing used in the largest deep learning projects from AlexNet (2012) to AlphaZero (2017), and found a 300,000-fold increase in the amount of compute required, with a doubling-time trendline of 3.4 months.

data/wiki_demo.txt.REMOVED.git-id

@@ -0,0 +1 @@
+c9cf509b7fdac5490cfd6dae72c2d7b8a60af6cb

docker-compose.yml

@@ -0,0 +1,23 @@
+version: '3.8'
+
+services:
+  llama-factory:
+    build:
+      dockerfile: Dockerfile
+      context: .
+    container_name: llama_factory
+    volumes:
+      - ./hf_cache:/root/.cache/huggingface/
+      - ./data:/app/data
+      - ./output:/app/output
+    ports:
+      - "7860:7860"
+    ipc: host
+    deploy:
+      resources:
+        reservations:
+          devices:
+          - driver: nvidia
+            count: "all"
+            capabilities: [gpu]
+    restart: unless-stopped

examples/accelerate/fsdp_config.yaml

@@ -0,0 +1,25 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_backward_prefetch: BACKWARD_PRE
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: fp16
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate/master_config.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+gpu_ids: all
+machine_rank: 0
+main_process_ip: 192.168.0.1
+main_process_port: 29555
+main_training_function: main
+mixed_precision: fp16
+num_machines: 2
+num_processes: 16
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate/single_config.yaml

@@ -0,0 +1,16 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+gpu_ids: all
+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

examples/accelerate/slave_config.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+gpu_ids: all
+machine_rank: 1
+main_process_ip: 192.168.0.1
+main_process_port: 29555
+main_training_function: main
+mixed_precision: fp16
+num_machines: 2
+num_processes: 16
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/extras/galore/adamw.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/extras/galore/adamw_8bit_bf16.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --optim adamw_8bit \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --pure_bf16

examples/extras/galore/galore_adamw.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --use_galore \
+    --galore_layerwise \
+    --galore_target mlp,self_attn \
+    --galore_rank 128 \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/extras/galore/galore_adamw_8bit_bf16.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --optim adamw_8bit \
+    --use_galore \
+    --galore_layerwise \
+    --galore_target mlp,self_attn \
+    --galore_rank 128 \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --pure_bf16

examples/extras/llama_pro/expand.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+python ../../../scripts/llama_pro.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --output_dir ../../../models/llama2-7b-pro \
+    --num_expand 8

examples/extras/llama_pro/sft.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path ../../../models/llama2-7b-pro \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type freeze \
+    --name_module_trainable all \
+    --num_layer_trainable 8 \
+    --use_llama_pro \
+    --output_dir ../../../saves/LLaMA2-7B-Pro/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/extras/loraplus/sft.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/loraplus/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16 \
+    --loraplus_lr_ratio 16.0

examples/fsdp_qlora/README.md

@@ -0,0 +1,5 @@
+```bash
+pip install git+https://github.com/huggingface/transformers.git
+pip install "accelerate>=0.28.0"
+pip install "bitsandbytes>=0.43.0"
+```

examples/fsdp_qlora/fsdp.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0,1 accelerate launch \
+    --config_file ../accelerate/fsdp_config.yaml \
+    ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-70b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-70B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --quantization_bit 4 \
+    --plot_loss \
+    --fp16

examples/full_multi_gpu/multi_node.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+
+python -m torch.distributed.run \
+    --nproc_per_node $NPROC_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT \
+    ../../src/train_bash.py \
+    --deepspeed ../deepspeed/ds_z3_config.json \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type full \
+    --output_dir ../../saves/LLaMA2-7B/full/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/full_multi_gpu/single_node.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+deepspeed --num_gpus 4 ../../src/train_bash.py \
+    --deepspeed ../deepspeed/ds_z3_config.json \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type full \
+    --output_dir ../../saves/LLaMA2-7B/full/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/lora_multi_gpu/multi_node.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch \
+    --config_file ../accelerate/master_config.yaml \
+    ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/lora_multi_gpu/single_node.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch \
+    --config_file ../accelerate/single_config.yaml \
+    ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/README.md

@@ -0,0 +1,8 @@
+Usage:
+
+- `pretrain.sh`: do pre-train (optional)
+- `sft.sh`: do supervised fine-tune
+- `reward.sh`: do reward modeling (must after sft.sh)
+- `ppo.sh`: do PPO training (must after sft.sh and reward.sh)
+- `dpo.sh`: do DPO training (must after sft.sh)
+- `predict.sh`: do predict (must after sft.sh and dpo.sh)

examples/lora_single_gpu/dpo.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage dpo \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --create_new_adapter \
+    --dataset comparison_gpt4_en \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/dpo \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 1000 \
+    --val_size 0.1 \
+    --dpo_ftx 1.0 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/ppo.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage ppo \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --create_new_adapter \
+    --dataset alpaca_gpt4_en \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --reward_model ../../saves/LLaMA2-7B/lora/reward \
+    --output_dir ../../saves/LLaMA2-7B/lora/ppo \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 512 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 1000 \
+    --top_k 0 \
+    --top_p 0.9 \
+    --max_new_tokens 256 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/predict.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_predict \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft,../../saves/LLaMA2-7B/lora/dpo \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --output_dir ../../saves/LLaMA2-7B/lora/predict \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_eval_batch_size 1 \
+    --max_samples 20 \
+    --predict_with_generate

examples/lora_single_gpu/pretrain.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage pt \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset c4_demo \
+    --dataset_dir ../../data \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/pretrain \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 10000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/reward.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage rm \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --create_new_adapter \
+    --dataset comparison_gpt4_en \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/reward \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 5000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/sft.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/merge_lora/README.md

@@ -0,0 +1,4 @@
+Usage:
+
+- `merge.sh`: merge the lora weights
+- `quantize.sh`: quantize the model with AutoGPTQ (must after merge.sh, optional)

examples/merge_lora/merge.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/export_model.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --template default \
+    --finetuning_type lora \
+    --export_dir ../../models/llama2-7b-sft \
+    --export_size 2 \
+    --export_legacy_format False

examples/merge_lora/quantize.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/export_model.py \
+    --model_name_or_path ../../models/llama2-7b-sft \
+    --template default \
+    --export_dir ../../models/llama2-7b-sft-int4 \
+    --export_quantization_bit 4 \
+    --export_quantization_dataset ../../data/c4_demo.json \
+    --export_size 2 \
+    --export_legacy_format False

examples/qlora_single_gpu/aqlm.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path BlackSamorez/Llama-2-7b-AQLM-2Bit-1x16-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/qlora_single_gpu/awq.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path TheBloke/Llama-2-7B-AWQ \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/qlora_single_gpu/bitsandbytes.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --quantization_bit 4 \
+    --plot_loss \
+    --fp16

examples/qlora_single_gpu/gptq.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path TheBloke/Llama-2-7B-GPTQ \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

pyproject.toml

@@ -1,3 +1,33 @@
 [build-system]
 requires = ["setuptools>=61.0"]
 build-backend = "setuptools.build_meta"
+
+[tool.ruff]
+target-version = "py38"
+line-length = 119
+indent-width = 4
+
+[tool.ruff.lint]
+ignore = ["C408", "C901", "E501", "E731", "E741", "W605"]
+select = ["C", "E", "F", "I", "W"]
+
+[tool.ruff.lint.isort]
+lines-after-imports = 2
+known-first-party = ["llmtuner"]
+known-third-party = [
+    "accelerate",
+    "datasets",
+    "gradio",
+    "numpy",
+    "peft",
+    "torch",
+    "transformers",
+    "trl"
+]
+
+[tool.ruff.format]
+quote-style = "double"
+indent-style = "space"
+docstring-code-format = true
+skip-magic-trailing-comma = false
+line-ending = "auto"

requirements.txt

@@ -1,20 +1,18 @@
 torch>=1.13.1
-transformers>=4.31.0,<4.35.0
-datasets>=2.14.0
-accelerate>=0.21.0
-peft>=0.6.0
-trl==0.7.2
+transformers>=4.37.2
+datasets>=2.14.3
+accelerate>=0.27.2
+peft>=0.9.0
+trl>=0.8.1
 gradio>=3.38.0,<4.0.0
 scipy
+einops
 sentencepiece
 protobuf
-tiktoken
-fire
-jieba
-rouge-chinese
-nltk
 uvicorn
 pydantic
 fastapi
 sse-starlette
 matplotlib
+fire
+galore-torch

scripts/cal_flops.py

@@ -3,42 +3,31 @@
 # Usage: python cal_flops.py --model_name_or_path path_to_model --batch_size 1 --seq_length 512
 # Inspired by: https://www.deepspeed.ai/tutorials/flops-profiler/
 
+from typing import Optional
+
 import fire
 import torch
-from typing import Optional
-from deepspeed.accelerator import get_accelerator # type: ignore
-from deepspeed.profiling.flops_profiler import get_model_profile # type: ignore
+from deepspeed.accelerator import get_accelerator  # type: ignore
+from deepspeed.profiling.flops_profiler import get_model_profile  # type: ignore
 
 from llmtuner import ChatModel
 
 
-def calculate(
+def calculate_flops(
     model_name_or_path: str,
     batch_size: Optional[int] = 1,
     seq_length: Optional[int] = 256,
-    flash_attn: Optional[bool] = False
+    flash_attn: Optional[bool] = False,
 ):
     with get_accelerator().device(0):
-        chat_model = ChatModel(dict(
-            model_name_or_path=model_name_or_path,
-            template="vanilla",
-            flash_attn=flash_attn
-        ))
+        chat_model = ChatModel(dict(model_name_or_path=model_name_or_path, template="vanilla", flash_attn=flash_attn))
         fake_input = torch.ones((batch_size, seq_length), dtype=torch.long, device=chat_model.model.device)
-        input_dict = {
-            "input_ids": fake_input,
-            "labels": fake_input.clone()
-        }
-        flops, macs, params = get_model_profile(
-            chat_model.model,
-            kwargs=input_dict,
-            print_profile=True,
-            detailed=True
-        )
+        input_dict = {"input_ids": fake_input, "labels": fake_input.clone()}
+        flops, macs, params = get_model_profile(chat_model.model, kwargs=input_dict, print_profile=True, detailed=True)
         print("FLOPs:", flops)
         print("MACs:", macs)
         print("Params:", params)
 
 
 if __name__ == "__main__":
-    fire.Fire(calculate)
+    fire.Fire(calculate_flops)

scripts/cal_lr.py

@@ -0,0 +1,77 @@
+# coding=utf-8
+# Calculates the optimal learning rate for 7B/13B models using LLaMA's hyper-parameters.
+# Usage: python cal_lr.py --model_name_or_path path_to_model --dataset alpaca_en --cutoff_len 1024 --batch_size 16
+# Inspired by: https://github.com/imoneoi/openchat/blob/master/ochat/training_deepspeed/train.py
+
+import math
+from typing import Optional
+
+import fire
+import torch
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+from transformers import DataCollatorForLanguageModeling, DataCollatorForSeq2Seq
+
+from llmtuner.data import get_dataset
+from llmtuner.extras.constants import IGNORE_INDEX
+from llmtuner.hparams import get_train_args
+from llmtuner.model import load_model_and_tokenizer
+
+
+BASE_LR = 3e-4  # 1.5e-4 for 30B-70B models
+BASE_BS = 4_000_000  # from llama paper
+
+
+def calculate_lr(
+    model_name_or_path: str,
+    batch_size: int,  # total batch size, namely (batch size * gradient accumulation * world size)
+    stage: Optional[str] = "sft",
+    dataset: Optional[str] = "alpaca_en",
+    dataset_dir: Optional[str] = "data",
+    template: Optional[str] = "default",
+    cutoff_len: Optional[int] = 1024,  # i.e. maximum input length during training
+    is_mistral: Optional[bool] = False,  # mistral model uses a smaller learning rate,
+):
+    model_args, data_args, training_args, finetuning_args, _ = get_train_args(
+        dict(
+            stage=stage,
+            model_name_or_path=model_name_or_path,
+            dataset=dataset,
+            dataset_dir=dataset_dir,
+            template=template,
+            cutoff_len=cutoff_len,
+            output_dir="dummy_dir",
+            overwrite_cache=True,
+        )
+    )
+    _, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, is_trainable=False, add_valuehead=False)
+    trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage=stage)
+    if stage == "pt":
+        data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+    elif stage == "sft":
+        data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX)
+    else:
+        raise NotImplementedError
+
+    dataloader = DataLoader(
+        dataset=trainset, batch_size=batch_size, shuffle=True, collate_fn=data_collator, pin_memory=True
+    )
+    valid_tokens, total_tokens = 0, 0
+    for batch in tqdm(dataloader):
+        valid_tokens += torch.sum(batch["labels"] != IGNORE_INDEX).item()
+        total_tokens += torch.numel(batch["labels"])
+
+    batch_max_len = cutoff_len * batch_size  # max tokens in a batch
+    valid_ratio = valid_tokens / total_tokens
+    batch_valid_len = batch_max_len * valid_ratio
+    lr = BASE_LR * math.sqrt(batch_valid_len / BASE_BS)  # lr ~ sqrt(batch_size)
+    lr = lr / 6.0 if is_mistral else lr
+    print(
+        "Optimal learning rate is {:.2e} for valid ratio% {:.2f} and effective batch size {:.2f}".format(
+            lr, valid_ratio * 100, batch_valid_len
+        )
+    )
+
+
+if __name__ == "__main__":
+    fire.Fire(calculate_lr)

scripts/length_cdf.py

@@ -0,0 +1,52 @@
+# coding=utf-8
+# Calculates the distribution of the input lengths in the dataset.
+# Usage: python length_cdf.py --model_name_or_path path_to_model --dataset alpaca_en --template default
+
+from collections import defaultdict
+from typing import Optional
+
+import fire
+from tqdm import tqdm
+
+from llmtuner.data import get_dataset
+from llmtuner.hparams import get_train_args
+from llmtuner.model import load_model_and_tokenizer
+
+
+def length_cdf(
+    model_name_or_path: str,
+    dataset: Optional[str] = "alpaca_en",
+    dataset_dir: Optional[str] = "data",
+    template: Optional[str] = "default",
+    interval: Optional[int] = 1000,
+):
+    model_args, data_args, training_args, finetuning_args, _ = get_train_args(
+        dict(
+            stage="sft",
+            model_name_or_path=model_name_or_path,
+            dataset=dataset,
+            dataset_dir=dataset_dir,
+            template=template,
+            cutoff_len=1_000_000,
+            output_dir="dummy_dir",
+            overwrite_cache=True,
+        )
+    )
+    _, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, is_trainable=False, add_valuehead=False)
+    trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage="sft")
+    total_num = len(trainset)
+    length_dict = defaultdict(int)
+    for sample in tqdm(trainset["input_ids"]):
+        length_dict[len(sample) // interval * interval] += 1
+
+    length_tuples = list(length_dict.items())
+    length_tuples.sort()
+    count_accu, prob_accu = 0, 0
+    for length, count in length_tuples:
+        count_accu += count
+        prob_accu += count / total_num * 100
+        print("{:d} ({:.2f}%) samples have length < {}.".format(count_accu, prob_accu, length + interval))
+
+
+if __name__ == "__main__":
+    fire.Fire(length_cdf)

scripts/llama_pro.py

@@ -0,0 +1,115 @@
+# coding=utf-8
+# Performs block expansion for LLaMA, Mistral or Qwen1.5 models.
+# Usage: python llama_pro.py --model_name_or_path meta-llama/Llama-2-7b-hf --output_dir llama2_pro --num_expand 8
+# Inspired by: https://github.com/TencentARC/LLaMA-Pro/blob/main/scripts/block_expansion.py
+
+import json
+import os
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Optional
+
+import fire
+import torch
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+from transformers.modeling_utils import (
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_INDEX_NAME,
+    WEIGHTS_NAME,
+    shard_checkpoint,
+)
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedModel
+
+
+def change_name(name: str, old_index: int, new_index: int) -> str:
+    return name.replace(".{:d}.".format(old_index), ".{:d}.".format(new_index))
+
+
+def block_expansion(
+    model_name_or_path: str,
+    output_dir: str,
+    num_expand: int,
+    shard_size: Optional[str] = "2GB",
+    save_safetensors: Optional[bool] = False,
+):
+    config: "PretrainedConfig" = AutoConfig.from_pretrained(model_name_or_path)
+    num_layers = getattr(config, "num_hidden_layers")
+    setattr(config, "num_hidden_layers", num_layers + num_expand)
+    config.save_pretrained(output_dir)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
+    tokenizer.save_pretrained(output_dir)
+
+    config: "PretrainedConfig" = AutoConfig.from_pretrained(model_name_or_path)  # load the original one
+    if save_safetensors:
+        setattr(config, "tie_word_embeddings", False)  # safetensors does not allow shared weights
+
+    model: "PreTrainedModel" = AutoModelForCausalLM.from_pretrained(
+        model_name_or_path,
+        config=config,
+        torch_dtype="auto",
+        trust_remote_code=True,
+        low_cpu_mem_usage=True,
+    )
+    state_dict = model.state_dict()
+
+    if num_layers % num_expand != 0:
+        raise ValueError("`num_layers` {} should be divisible by `num_expand` {}.".format(num_layers, num_expand))
+
+    split = num_layers // num_expand
+    layer_cnt = 0
+    output_state_dict = OrderedDict()
+    for i in range(num_layers):
+        for key, value in state_dict.items():
+            if ".{:d}.".format(i) in key:
+                output_state_dict[change_name(key, i, layer_cnt)] = value
+
+        print("Add layer {} copied from layer {}".format(layer_cnt, i))
+        layer_cnt += 1
+        if (i + 1) % split == 0:
+            for key, value in state_dict.items():
+                if ".{:d}.".format(i) in key:
+                    if "down_proj" in key or "o_proj" in key:
+                        output_state_dict[change_name(key, i, layer_cnt)] = torch.zeros_like(value)
+                    else:
+                        output_state_dict[change_name(key, i, layer_cnt)] = torch.clone(value)
+
+            print("Add layer {} expanded from layer {}".format(layer_cnt, i))
+            layer_cnt += 1
+
+    for key, value in state_dict.items():
+        if key not in output_state_dict:
+            output_state_dict[key] = value
+
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    shards, index = shard_checkpoint(output_state_dict, max_shard_size=shard_size, weights_name=weights_name)
+
+    for shard_file, shard in tqdm(shards.items(), desc="Save weights"):
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
+
+    if index is None:
+        print("Model weights saved in {}".format(os.path.join(output_dir, weights_name)))
+    else:
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
+            json.dump(index, f, indent=2, sort_keys=True)
+        print("Model weights saved in {}".format(output_dir))
+
+    print("Fine-tune this model with:")
+    print("  --model_name_or_path {} \\".format(output_dir))
+    print("  --finetuning_type freeze \\")
+    print("  --name_module_trainable all \\")
+    print("  --num_layer_trainable {} \\".format(num_expand))
+    print("  --use_llama_pro")
+
+
+if __name__ == "__main__":
+    fire.Fire(block_expansion)

scripts/llamafy_baichuan2.py

@@ -1,60 +1,68 @@
 # coding=utf-8
 # Converts the Baichuan2-7B model in the same format as LLaMA2-7B.
-# Usage: python llamafy_baichuan2.py --input_dir input --output_dir output --shard_size 10GB
+# Usage: python llamafy_baichuan2.py --input_dir input --output_dir output
 # Inspired by: https://huggingface.co/fireballoon/baichuan-llama-7b/blob/main/convert_baichuan_to_llama.py
 # Converted model: https://huggingface.co/hiyouga/Baichuan2-7B-Base-LLaMAfied
 
-import os
-import fire
 import json
-import torch
+import os
 from collections import OrderedDict
-from transformers.modeling_utils import shard_checkpoint, WEIGHTS_NAME, WEIGHTS_INDEX_NAME
-from typing import Any, Dict
+from typing import Any, Dict, Optional
+
+import fire
+import torch
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers.modeling_utils import (
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_INDEX_NAME,
+    WEIGHTS_NAME,
+    shard_checkpoint,
+)
 
 
 CONFIG_NAME = "config.json"
 
 
-def save_weight(
-    input_dir: str,
-    output_dir: str,
-    shard_size: str
-):
+def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool):
     baichuan2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
-    for filepath in os.listdir(input_dir):
+    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
         if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".bin"):
             shard_weight = torch.load(os.path.join(input_dir, filepath), map_location="cpu")
             baichuan2_state_dict.update(shard_weight)
 
     llama2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
-    for key, value in baichuan2_state_dict.items():
+    for key, value in tqdm(baichuan2_state_dict.items(), desc="Convert format"):
         if "W_pack" in key:
             proj_size = value.size(0) // 3
             llama2_state_dict[key.replace("W_pack", "q_proj")] = value[:proj_size, :]
-            llama2_state_dict[key.replace("W_pack", "k_proj")] = value[proj_size:2*proj_size, :]
-            llama2_state_dict[key.replace("W_pack", "v_proj")] = value[2*proj_size:, :]
+            llama2_state_dict[key.replace("W_pack", "k_proj")] = value[proj_size : 2 * proj_size, :]
+            llama2_state_dict[key.replace("W_pack", "v_proj")] = value[2 * proj_size :, :]
         elif "lm_head" in key:
             llama2_state_dict[key] = torch.nn.functional.normalize(value)
         else:
             llama2_state_dict[key] = value
 
-    shards, index = shard_checkpoint(llama2_state_dict, max_shard_size=shard_size, weights_name=WEIGHTS_NAME)
-    for shard_file, shard in shards.items():
-        torch.save(shard, os.path.join(output_dir, shard_file))
-    
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    shards, index = shard_checkpoint(llama2_state_dict, max_shard_size=shard_size, weights_name=weights_name)
+
+    for shard_file, shard in tqdm(shards.items(), desc="Save weights"):
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
+
     if index is None:
         print("Model weights saved in {}".format(os.path.join(output_dir, WEIGHTS_NAME)))
     else:
-        with open(os.path.join(output_dir, WEIGHTS_INDEX_NAME), "w", encoding="utf-8") as f:
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
             json.dump(index, f, indent=2, sort_keys=True)
         print("Model weights saved in {}".format(output_dir))
 
 
-def save_config(
-    input_dir: str,
-    output_dir: str
-):
+def save_config(input_dir: str, output_dir: str):
     with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
         llama2_config_dict: Dict[str, Any] = json.load(f)
 
@@ -69,17 +77,15 @@ def save_config(
 
 
 def llamafy_baichuan2(
-    input_dir: str,
-    output_dir: str,
-    shard_size: str
+    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
 ):
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:
         raise print("Output dir already exists", e)
 
-    save_weight(input_dir, output_dir, shard_size)
-    save_config(input_dir, output_dir)    
+    save_weight(input_dir, output_dir, shard_size, save_safetensors)
+    save_config(input_dir, output_dir)
 
 
 if __name__ == "__main__":

scripts/llamafy_internlm2.py

@@ -0,0 +1,114 @@
+# coding=utf-8
+# Converts the InternLM2 model in the same format as LLaMA2.
+# Usage: python llamafy_internlm2.py --input_dir input --output_dir output
+# Warning: We have found that the converted model cannot infer correctly. It will be fixed later.
+
+import json
+import os
+from collections import OrderedDict
+from typing import Any, Dict, Optional
+
+import fire
+import torch
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers.modeling_utils import (
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_INDEX_NAME,
+    WEIGHTS_NAME,
+    shard_checkpoint,
+)
+
+
+CONFIG_NAME = "config.json"
+
+
+def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool):
+    with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
+        internlm2_config_dict: Dict[str, Any] = json.load(f)
+
+    internlm2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
+    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
+        if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".bin"):
+            shard_weight = torch.load(os.path.join(input_dir, filepath), map_location="cpu")
+            internlm2_state_dict.update(shard_weight)
+
+    llama2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
+    for key, value in tqdm(internlm2_state_dict.items(), desc="Convert format"):
+        if "output" in key:
+            llama2_state_dict[key.replace("output", "lm_head")] = value
+        elif "tok_embeddings" in key:
+            llama2_state_dict[key.replace("tok_embeddings", "embed_tokens")] = value
+        elif "wqkv" in key:
+            num_q_heads = internlm2_config_dict["num_attention_heads"]
+            num_kv_heads = internlm2_config_dict["num_key_value_heads"]
+            q_size = value.size(0) // (num_q_heads + 2 * num_kv_heads) * num_q_heads
+            kv_size = value.size(0) // (num_q_heads + 2 * num_kv_heads) * num_kv_heads
+            llama2_state_dict[key.replace("attention.wqkv", "self_attn.q_proj")] = value[:q_size, ...]
+            llama2_state_dict[key.replace("attention.wqkv", "self_attn.k_proj")] = value[
+                q_size : q_size + kv_size, ...
+            ]
+            llama2_state_dict[key.replace("attention.wqkv", "self_attn.v_proj")] = value[q_size + kv_size :, ...]
+        elif "wo" in key:
+            llama2_state_dict[key.replace("attention.wo", "self_attn.o_proj")] = value
+        elif "attention_norm" in key:
+            llama2_state_dict[key.replace("attention_norm", "input_layernorm")] = value
+        elif "ffn_norm" in key:
+            llama2_state_dict[key.replace("ffn_norm", "post_attention_layernorm")] = value
+        elif "w1" in key:
+            llama2_state_dict[key.replace("feed_forward.w1", "mlp.gate_proj")] = value
+        elif "w2" in key:
+            llama2_state_dict[key.replace("feed_forward.w2", "mlp.down_proj")] = value
+        elif "w3" in key:
+            llama2_state_dict[key.replace("feed_forward.w3", "mlp.up_proj")] = value
+        else:
+            llama2_state_dict[key] = value
+
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    shards, index = shard_checkpoint(llama2_state_dict, max_shard_size=shard_size, weights_name=weights_name)
+
+    for shard_file, shard in tqdm(shards.items(), desc="Save weights"):
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
+
+    if index is None:
+        print("Model weights saved in {}".format(os.path.join(output_dir, WEIGHTS_NAME)))
+    else:
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
+            json.dump(index, f, indent=2, sort_keys=True)
+        print("Model weights saved in {}".format(output_dir))
+
+
+def save_config(input_dir: str, output_dir: str):
+    with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
+        llama2_config_dict: Dict[str, Any] = json.load(f)
+
+    llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
+    llama2_config_dict.pop("auto_map", None)
+    llama2_config_dict.pop("bias", None)
+    llama2_config_dict.pop("rope_scaling", None)
+    llama2_config_dict["model_type"] = "llama"
+
+    with open(os.path.join(output_dir, CONFIG_NAME), "w", encoding="utf-8") as f:
+        json.dump(llama2_config_dict, f, indent=2)
+    print("Model config saved in {}".format(os.path.join(output_dir, CONFIG_NAME)))
+
+
+def llamafy_internlm2(
+    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
+):
+    try:
+        os.makedirs(output_dir, exist_ok=False)
+    except Exception as e:
+        raise print("Output dir already exists", e)
+
+    save_weight(input_dir, output_dir, shard_size, save_safetensors)
+    save_config(input_dir, output_dir)
+
+
+if __name__ == "__main__":
+    fire.Fire(llamafy_internlm2)

scripts/llamafy_qwen.py

@@ -1,33 +1,40 @@
 # coding=utf-8
 # Converts the Qwen models in the same format as LLaMA2.
-# Usage: python llamafy_qwen.py --input_dir input --output_dir output --shard_size 10GB
+# Usage: python llamafy_qwen.py --input_dir input --output_dir output
+# Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
 
-import os
-import fire
 import json
-import torch
+import os
 from collections import OrderedDict
+from typing import Any, Dict, Optional
+
+import fire
+import torch
 from safetensors import safe_open
-from transformers.modeling_utils import shard_checkpoint, WEIGHTS_NAME, WEIGHTS_INDEX_NAME
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers.modeling_utils import (
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_INDEX_NAME,
+    WEIGHTS_NAME,
+    shard_checkpoint,
+)
 from transformers.utils import check_min_version
-from typing import Any, Dict
+
 
 try:
     check_min_version("4.34.0")
-except:
+except Exception:
     raise ValueError("Please upgrade `transformers` to 4.34.0")
 
 
 CONFIG_NAME = "config.json"
 
 
-def save_weight(
-    input_dir: str,
-    output_dir: str,
-    shard_size: str
-) -> str:
+def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool) -> str:
     qwen_state_dict: Dict[str, torch.Tensor] = OrderedDict()
-    for filepath in os.listdir(input_dir):
+    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
         if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".safetensors"):
             with safe_open(os.path.join(input_dir, filepath), framework="pt", device="cpu") as f:
                 for key in f.keys():
@@ -35,7 +42,7 @@ def save_weight(
 
     llama2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
     torch_dtype = None
-    for key, value in qwen_state_dict.items():
+    for key, value in tqdm(qwen_state_dict.items(), desc="Convert format"):
         if torch_dtype is None:
             torch_dtype = value.dtype
         if "wte" in key:
@@ -47,13 +54,15 @@ def save_weight(
             if "attn.c_attn" in key:
                 proj_size = value.size(0) // 3
                 llama2_state_dict[key.replace("attn.c_attn", "self_attn.q_proj")] = value[:proj_size, ...]
-                llama2_state_dict[key.replace("attn.c_attn", "self_attn.k_proj")] = value[proj_size:2*proj_size, ...]
-                llama2_state_dict[key.replace("attn.c_attn", "self_attn.v_proj")] = value[2*proj_size:, ...]
+                llama2_state_dict[key.replace("attn.c_attn", "self_attn.k_proj")] = value[
+                    proj_size : 2 * proj_size, ...
+                ]
+                llama2_state_dict[key.replace("attn.c_attn", "self_attn.v_proj")] = value[2 * proj_size :, ...]
             elif "attn.c_proj" in key:
                 llama2_state_dict[key.replace("attn.c_proj", "self_attn.o_proj")] = value
-                llama2_state_dict[key.replace("attn.c_proj.weight", "self_attn.o_proj.bias")] = (
-                    torch.zeros_like(value[:, 0]).squeeze()
-                )
+                llama2_state_dict[key.replace("attn.c_proj.weight", "self_attn.o_proj.bias")] = torch.zeros_like(
+                    value[:, 0]
+                ).squeeze()
             elif "ln_1" in key:
                 llama2_state_dict[key.replace("ln_1", "input_layernorm")] = value
             elif "ln_2" in key:
@@ -69,25 +78,27 @@ def save_weight(
             else:
                 raise KeyError("Unable to process key {}".format(key))
 
-    shards, index = shard_checkpoint(llama2_state_dict, max_shard_size=shard_size, weights_name=WEIGHTS_NAME)
-    for shard_file, shard in shards.items():
-        torch.save(shard, os.path.join(output_dir, shard_file))
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    shards, index = shard_checkpoint(llama2_state_dict, max_shard_size=shard_size, weights_name=weights_name)
+
+    for shard_file, shard in tqdm(shards.items(), desc="Save weights"):
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
 
     if index is None:
-        print("Model weights saved in {}".format(os.path.join(output_dir, WEIGHTS_NAME)))
+        print("Model weights saved in {}".format(os.path.join(output_dir, weights_name)))
     else:
-        with open(os.path.join(output_dir, WEIGHTS_INDEX_NAME), "w", encoding="utf-8") as f:
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
             json.dump(index, f, indent=2, sort_keys=True)
         print("Model weights saved in {}".format(output_dir))
 
     return str(torch_dtype).replace("torch.", "")
 
 
-def save_config(
-    input_dir: str,
-    output_dir: str,
-    torch_dtype: str
-):
+def save_config(input_dir: str, output_dir: str, torch_dtype: str):
     with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
         qwen_config_dict: Dict[str, Any] = json.load(f)
 
@@ -118,17 +129,15 @@ def save_config(
 
 
 def llamafy_qwen(
-    input_dir: str,
-    output_dir: str,
-    shard_size: str
+    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
 ):
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:
         raise print("Output dir already exists", e)
 
-    torch_dtype = save_weight(input_dir, output_dir, shard_size)
-    save_config(input_dir, output_dir, torch_dtype)    
+    torch_dtype = save_weight(input_dir, output_dir, shard_size, save_safetensors)
+    save_config(input_dir, output_dir, torch_dtype)
 
 
 if __name__ == "__main__":

scripts/loftq_init.py

@@ -0,0 +1,82 @@
+# coding=utf-8
+# Initializes LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ)
+# Usage: python loftq_init.py --model_name_or_path path_to_model --save_dir output_dir
+# Inspired by: https://github.com/huggingface/peft/blob/main/examples/loftq_finetuning/quantize_save_load.py
+
+import os
+from typing import TYPE_CHECKING, Optional
+
+import fire
+import torch
+import torch.nn as nn
+from peft import LoftQConfig, LoraConfig, TaskType, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+
+class Shell(nn.Module):
+    def __init__(self, weight: torch.Tensor, bias: Optional[torch.Tensor] = None):
+        super().__init__()
+        self.weight = nn.Parameter(weight, requires_grad=False)
+        if bias is not None:
+            self.bias = nn.Parameter(bias, requires_grad=False)
+
+
+def unwrap_model(model: nn.Module, pattern=".base_layer") -> None:
+    for name in {k.split(pattern)[0] for k, _ in model.named_modules() if pattern in k}:
+        parent_name = ".".join(name.split(".")[:-1])
+        child_name = name.split(".")[-1]
+        parent_module = model.get_submodule(parent_name)
+        child_module = getattr(parent_module, child_name)
+        base_layer = getattr(child_module, "base_layer")
+        weight = getattr(base_layer, "weight", None)
+        bias = getattr(base_layer, "bias", None)
+        setattr(parent_module, child_name, Shell(weight, bias))
+
+    print("Model unwrapped.")
+
+
+def quantize_loftq(
+    model_name_or_path: str,
+    save_dir: str,
+    loftq_bits: Optional[int] = 4,
+    loftq_iter: Optional[int] = 1,
+    lora_alpha: Optional[int] = None,
+    lora_rank: Optional[int] = 16,
+    lora_target: Optional[str] = "q_proj,v_proj",
+    save_safetensors: Optional[bool] = False,
+):
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
+    loftq_config = LoftQConfig(loftq_bits=loftq_bits, loftq_iter=loftq_iter)
+    lora_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        inference_mode=True,
+        r=lora_rank,
+        lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
+        lora_dropout=0.1,
+        target_modules=[name.strip() for name in lora_target.split(",")],
+        init_lora_weights="loftq",
+        loftq_config=loftq_config,
+    )
+
+    # Init LoftQ model
+    lora_model = get_peft_model(model, lora_config)
+    base_model: "PreTrainedModel" = lora_model.get_base_model()
+
+    # Save LoftQ model
+    setattr(lora_model.base_model.peft_config["default"], "base_model_name_or_path", save_dir)
+    setattr(lora_model.base_model.peft_config["default"], "init_lora_weights", True)
+    lora_model.save_pretrained(os.path.join(save_dir, "adapters"), safe_serialization=save_safetensors)
+
+    # Save base model
+    unwrap_model(base_model)
+    base_model.save_pretrained(save_dir, safe_serialization=save_safetensors)
+    tokenizer.save_pretrained(save_dir)
+
+
+if __name__ == "__main__":
+    fire.Fire(quantize_loftq)

setup.py

@@ -1,13 +1,14 @@
 import os
 import re
-from setuptools import setup, find_packages
+
+from setuptools import find_packages, setup
 
 
 def get_version():
     with open(os.path.join("src", "llmtuner", "__init__.py"), "r", encoding="utf-8") as f:
         file_content = f.read()
         pattern = r"{0}\W*=\W*\"([^\"]+)\"".format("__version__")
-        version, = re.findall(pattern, file_content)
+        (version,) = re.findall(pattern, file_content)
         return version
 
 
@@ -18,8 +19,21 @@ def get_requires():
         return lines
 
 
-def main():
+extra_require = {
+    "deepspeed": ["deepspeed"],
+    "metrics": ["nltk", "jieba", "rouge-chinese"],
+    "unsloth": ["torch==2.2.0", "unsloth[cu121-ampere-torch220]"],
+    "vllm": ["vllm>=0.3.3"],
+    "bitsandbytes": ["bitsandbytes>=0.39.0"],
+    "gptq": ["optimum>=1.16.0", "auto-gptq>=0.5.0"],
+    "awq": ["autoawq"],
+    "aqlm": ["aqlm[gpu]>=1.1.0"],
+    "qwen": ["tiktoken", "transformers_stream_generator"],
+    "quality": ["ruff"],
+}
 
+
+def main():
     setup(
         name="llmtuner",
         version=get_version(),
@@ -35,8 +49,9 @@ def main():
         packages=find_packages("src"),
         python_requires=">=3.8.0",
         install_requires=get_requires(),
+        extras_require=extra_require,
         classifiers=[
-            "Development Status :: 3 - Alpha",
+            "Development Status :: 4 - Beta",
             "Intended Audience :: Developers",
             "Intended Audience :: Education",
             "Intended Audience :: Science/Research",
@@ -46,8 +61,9 @@ def main():
             "Programming Language :: Python :: 3.8",
             "Programming Language :: Python :: 3.9",
             "Programming Language :: Python :: 3.10",
+            "Programming Language :: Python :: 3.11",
             "Topic :: Scientific/Engineering :: Artificial Intelligence",
-        ]
+        ],
     )
 
 

src/api_demo.py

@@ -1,3 +1,5 @@
+import os
+
 import uvicorn
 
 from llmtuner import ChatModel, create_app
@@ -6,8 +8,8 @@ from llmtuner import ChatModel, create_app
 def main():
     chat_model = ChatModel()
     app = create_app(chat_model)
-    print("Visit http://localhost:8000/docs for API document.")
-    uvicorn.run(app, host="0.0.0.0", port=8000, workers=1)
+    print("Visit http://localhost:{}/docs for API document.".format(os.environ.get("API_PORT", 8000)))
+    uvicorn.run(app, host="0.0.0.0", port=int(os.environ.get("API_PORT", 8000)), workers=1)
 
 
 if __name__ == "__main__":

src/cli_demo.py

@@ -1,10 +1,19 @@
-import readline
 from llmtuner import ChatModel
+from llmtuner.extras.misc import torch_gc
+
+
+try:
+    import platform
+
+    if platform.system() != "Windows":
+        import readline  # noqa: F401
+except ImportError:
+    print("Install `readline` for a better experience.")
 
 
 def main():
     chat_model = ChatModel()
-    history = []
+    messages = []
     print("Welcome to the CLI application, use `clear` to remove the history, use `exit` to exit the application.")
 
     while True:
@@ -20,19 +29,20 @@ def main():
             break
 
         if query.strip() == "clear":
-            history = []
+            messages = []
+            torch_gc()
             print("History has been removed.")
             continue
 
+        messages.append({"role": "user", "content": query})
         print("Assistant: ", end="", flush=True)
 
         response = ""
-        for new_text in chat_model.stream_chat(query, history):
+        for new_text in chat_model.stream_chat(messages):
             print(new_text, end="", flush=True)
             response += new_text
         print()
-
-        history = history + [(query, response)]
+        messages.append({"role": "assistant", "content": response})
 
 
 if __name__ == "__main__":

src/evaluate.py

@@ -1,190 +1,10 @@
-# coding=utf-8
-# Evaluates the performance of pre-trained models.
-# Usage: python evaluate.py --model_name_or_path path_to_model --checkpoint_dir path_to_ckpt --template vanilla
-#                           --task ceval --split validation --lang zh --n_shot 5 --batch_size 4 --save_name result
-# Inspired by: https://github.com/hendrycks/test/blob/master/evaluate_flan.py
-
-import os
-import fire
-import json
-import torch
-import numpy as np
-import transformers
-from collections import Counter
-from datasets import load_dataset
-from dataclasses import dataclass
-from tqdm import tqdm, trange
-from typing import TYPE_CHECKING, Dict, List, Literal, Optional, Tuple
-
-from llmtuner import ChatModel
-
-if TYPE_CHECKING:
-    from datasets import Dataset
+from llmtuner import Evaluator
 
 
-choices = ["A", "B", "C", "D"]
-
-
-@dataclass
-class EvalTemplate:
-
-    system: str
-    choice: str
-    answer: str
-    prefix: str
-
-    def parse_example(
-        self,
-        example: Dict[str, str]
-    ) -> Tuple[str, str]:
-        candidates = [self.choice.format(choice=ch, content=example[ch]) for ch in choices if ch in example]
-        return "".join([example["question"]] + candidates + [self.answer]), example["answer"]
-
-    def format_example(
-        self,
-        target_data: Dict[str, str],
-        support_set: "Dataset",
-        subject_name: str,
-        use_history: bool
-    ) -> Tuple[str, str, List[Tuple[str, str]]]:
-        query, resp = self.parse_example(target_data)
-        history = [self.parse_example(support_set[k]) for k in range(len(support_set))]
-
-        if len(history):
-            temp = history.pop(0)
-            history.insert(0, (self.system.format(subject=subject_name) + temp[0], temp[1]))
-        else:
-            query = self.system.format(subject=subject_name) + query
-
-        if not use_history:
-            query = "\n\n".join(["".join(item) for item in history] + [query])
-            history = []
-        return query.strip(), resp, history
-
-
-eval_templates = {
-    "en": EvalTemplate(
-        system="The following are multiple choice questions (with answers) about {subject}.\n\n",
-        choice="\n{choice}. {content}",
-        answer="\nAnswer: ",
-        prefix=" "
-    ),
-    "zh": EvalTemplate(
-        system="以下是中国关于{subject}考试的单项选择题，请选出其中的正确答案。\n\n",
-        choice="\n{choice}. {content}",
-        answer="\n答案：",
-        prefix="\n"
-    )
-}
-
-
-@torch.inference_mode()
-def batch_inference(
-    chat_model: ChatModel,
-    batch_input: Dict[str, torch.Tensor],
-    prefix_char: str
-) -> List[str]:
-    logits = chat_model.model(**batch_input).logits
-    lengths = torch.sum(batch_input["attention_mask"], dim=-1)
-    nextword_logits = torch.stack([logits[i, lengths[i] - 1] for i in range(len(lengths))], dim=0)
-    probs = torch.nn.functional.softmax(
-        torch.stack(
-            [
-                nextword_logits[:, chat_model.tokenizer.encode(prefix_char + choice, add_special_tokens=False)[-1]]
-                for choice in choices
-            ],
-            dim=-1
-        ),
-        dim=-1
-    ).detach()
-    return [chr(ord("A") + offset.item()) for offset in torch.argmax(probs, dim=-1)]
-
-
-def evaluate(
-    model_name_or_path: str,
-    finetuning_type: Optional[str] = "lora",
-    checkpoint_dir: Optional[str] = None,
-    template: Optional[str] = "vanilla",
-    task: Optional[str] = "ceval",
-    dataset_dir: Optional[str] = "evaluation",
-    split: Optional[Literal["validation", "test"]] = "validation",
-    lang: Optional[Literal["zh", "en"]] = "zh",
-    n_shot: Optional[int] = 5,
-    n_avg: Optional[int] = 1,
-    batch_size: Optional[int] = 4,
-    save_name: Optional[str] = None,
-    seed: Optional[int] = 42
-):
-    with open(os.path.join(dataset_dir, task, "mapping.json"), "r", encoding="utf-8") as f:
-        categorys: Dict[str, Dict[str, str]] = json.load(f)
-
-    transformers.set_seed(seed)
-    chat_model = ChatModel(dict(
-        model_name_or_path=model_name_or_path,
-        finetuning_type=finetuning_type,
-        checkpoint_dir=checkpoint_dir,
-        template=template
-    ))
-    chat_model.tokenizer.padding_side = "right" # avoid overflow issue in batched inference for llama2
-    eval_template = eval_templates[lang]
-
-    category_corrects: Dict[str, np.ndarray] = {
-        subj: np.array([], dtype="bool") for subj in ["Average", "STEM", "Social Sciences", "Humanities", "Other"]
-    }
-    pbar = tqdm(categorys.keys(), desc="Processing subjects", position=0)
-    results = {}
-    for subject in pbar:
-        dataset = load_dataset(os.path.join(dataset_dir, task), subject)
-        labels, answers, all_outputs = [], [], []
-        for epoch in range(n_avg):
-            pbar.set_postfix_str("{} Trial: {}".format(categorys[subject]["name"], epoch))
-            inputs, outputs = [], []
-            for i in trange(len(dataset[split]), desc="Formatting batches", position=1, leave=False):
-                support_set = dataset["train"].shuffle().select(range(min(n_shot, len(dataset["train"]))))
-                query, resp, history = eval_template.format_example(
-                    target_data=dataset[split][i],
-                    support_set=support_set,
-                    subject_name=categorys[subject]["name"],
-                    use_history=chat_model.template.use_history
-                )
-                input_ids, _ = chat_model.template.encode_oneturn(
-                    tokenizer=chat_model.tokenizer, query=query, resp=resp, history=history
-                )
-                inputs.append({"input_ids": input_ids, "attention_mask": [1] * len(input_ids)})
-                if epoch == 0:
-                    labels.append(resp)
-
-            for i in trange(0, len(inputs), batch_size, desc="Predicting batches", position=1, leave=False):
-                batch_input = chat_model.tokenizer.pad(
-                    inputs[i : i + batch_size], return_attention_mask=True, return_tensors="pt"
-                ).to(chat_model.model.device)
-                preds = batch_inference(chat_model, batch_input, eval_template.prefix)
-                outputs += preds
-            all_outputs.append(outputs)
-
-        for i in range(len(all_outputs[0])):
-            count = Counter([all_outputs[epoch][i] for epoch in range(n_avg)])
-            answers.append(count.most_common(1)[0][0])
-
-        corrects = (np.array(answers) == np.array(labels))
-        category_name = categorys[subject]["category"]
-        category_corrects[category_name] = np.concatenate([category_corrects[category_name], corrects], axis=0)
-        category_corrects["Average"] = np.concatenate([category_corrects["Average"], corrects], axis=0)
-        results[subject] = {str(i): answers[i] for i in range(len(answers))}
-
-    score_info = "\n".join([
-        "{:>15}: {:.2f}".format(category_name, 100 * np.mean(category_correct))
-        for category_name, category_correct in category_corrects.items() if len(category_correct)
-    ])
-
-    print(score_info)
-    if save_name is not None:
-        with open(save_name + ".json", "w", encoding="utf-8", newline="\n") as f:
-            json.dump(results, f, indent=2)
-
-        with open(save_name + ".log", "w", encoding="utf-8", newline="\n") as f:
-            f.write(score_info)
+def main():
+    evaluator = Evaluator()
+    evaluator.eval()
 
 
 if __name__ == "__main__":
-    fire.Fire(evaluate)
+    main()

src/llmtuner/__init__.py

@@ -1,9 +1,11 @@
-# Level: api, webui > chat > tuner > dsets > extras, hparams
+# Level: api, webui > chat, eval, train > data, model > extras, hparams
 
-from llmtuner.api import create_app
-from llmtuner.chat import ChatModel
-from llmtuner.tuner import export_model, run_exp
-from llmtuner.webui import create_ui, create_web_demo
+from .api import create_app
+from .chat import ChatModel
+from .eval import Evaluator
+from .train import export_model, run_exp
+from .webui import create_ui, create_web_demo
 
 
-__version__ = "0.2.1"
+__version__ = "0.6.0"
+__all__ = ["create_app", "ChatModel", "Evaluator", "export_model", "run_exp", "create_ui", "create_web_demo"]

src/llmtuner/api/__init__.py

@@ -1 +1,4 @@
-from llmtuner.api.app import create_app
+from .app import create_app
+
+
+__all__ = ["create_app"]

src/llmtuner/api/app.py

@@ -1,44 +1,67 @@
 import json
-import uvicorn
-from fastapi import FastAPI, HTTPException, status
-from fastapi.middleware.cors import CORSMiddleware
+import os
 from contextlib import asynccontextmanager
-from sse_starlette import EventSourceResponse
-from typing import List, Tuple
+from typing import Any, Dict, Sequence
+
 from pydantic import BaseModel
 
-from llmtuner.extras.misc import torch_gc
-from llmtuner.chat import ChatModel
-from llmtuner.api.protocol import (
-    Role,
-    Finish,
-    ModelCard,
-    ModelList,
-    ChatMessage,
-    DeltaMessage,
+from ..chat import ChatModel
+from ..data import Role as DataRole
+from ..extras.misc import torch_gc
+from ..extras.packages import is_fastapi_availble, is_starlette_available, is_uvicorn_available
+from .protocol import (
+    ChatCompletionMessage,
     ChatCompletionRequest,
     ChatCompletionResponse,
-    ChatCompletionStreamResponse,
     ChatCompletionResponseChoice,
     ChatCompletionResponseStreamChoice,
-    ChatCompletionResponseUsage
+    ChatCompletionResponseUsage,
+    ChatCompletionStreamResponse,
+    Finish,
+    Function,
+    FunctionCall,
+    ModelCard,
+    ModelList,
+    Role,
+    ScoreEvaluationRequest,
+    ScoreEvaluationResponse,
 )
 
 
+if is_fastapi_availble():
+    from fastapi import FastAPI, HTTPException, status
+    from fastapi.middleware.cors import CORSMiddleware
+
+
+if is_starlette_available():
+    from sse_starlette import EventSourceResponse
+
+
+if is_uvicorn_available():
+    import uvicorn
+
+
 @asynccontextmanager
-async def lifespan(app: FastAPI): # collects GPU memory
+async def lifespan(app: "FastAPI"):  # collects GPU memory
     yield
     torch_gc()
 
 
-def to_json(data: BaseModel) -> str:
-    try: # pydantic v2
+def dictify(data: "BaseModel") -> Dict[str, Any]:
+    try:  # pydantic v2
+        return data.model_dump(exclude_unset=True)
+    except AttributeError:  # pydantic v1
+        return data.dict(exclude_unset=True)
+
+
+def jsonify(data: "BaseModel") -> str:
+    try:  # pydantic v2
         return json.dumps(data.model_dump(exclude_unset=True), ensure_ascii=False)
-    except: # pydantic v1
+    except AttributeError:  # pydantic v1
         return data.json(exclude_unset=True, ensure_ascii=False)
 
 
-def create_app(chat_model: ChatModel) -> FastAPI:
+def create_app(chat_model: "ChatModel") -> "FastAPI":
     app = FastAPI(lifespan=lifespan)
 
     app.add_middleware(
@@ -49,6 +72,14 @@ def create_app(chat_model: ChatModel) -> FastAPI:
         allow_headers=["*"],
     )
 
+    role_mapping = {
+        Role.USER: DataRole.USER.value,
+        Role.ASSISTANT: DataRole.ASSISTANT.value,
+        Role.SYSTEM: DataRole.SYSTEM.value,
+        Role.FUNCTION: DataRole.FUNCTION.value,
+        Role.TOOL: DataRole.OBSERVATION.value,
+    }
+
     @app.get("/v1/models", response_model=ModelList)
     async def list_models():
         model_card = ModelCard(id="gpt-3.5-turbo")
@@ -56,91 +87,138 @@ def create_app(chat_model: ChatModel) -> FastAPI:
 
     @app.post("/v1/chat/completions", response_model=ChatCompletionResponse, status_code=status.HTTP_200_OK)
     async def create_chat_completion(request: ChatCompletionRequest):
-        if len(request.messages) < 1 or request.messages[-1].role != Role.USER:
-            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid request")
+        if not chat_model.engine.can_generate:
+            raise HTTPException(status_code=status.HTTP_405_METHOD_NOT_ALLOWED, detail="Not allowed")
 
-        query = request.messages[-1].content
-        prev_messages = request.messages[:-1]
-        if len(prev_messages) > 0 and prev_messages[0].role == Role.SYSTEM:
-            system = prev_messages.pop(0).content
+        if len(request.messages) == 0:
+            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid length")
+
+        if request.messages[0].role == Role.SYSTEM:
+            system = request.messages.pop(0).content
         else:
-            system = None
+            system = ""
 
-        history = []
-        if len(prev_messages) % 2 == 0:
-            for i in range(0, len(prev_messages), 2):
-                if prev_messages[i].role == Role.USER and prev_messages[i+1].role == Role.ASSISTANT:
-                    history.append([prev_messages[i].content, prev_messages[i+1].content])
-                else:
-                    raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Only supports u/a/u/a/u...")
+        if len(request.messages) % 2 == 0:
+            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Only supports u/a/u/a/u...")
+
+        input_messages = []
+        for i, message in enumerate(request.messages):
+            if i % 2 == 0 and message.role not in [Role.USER, Role.TOOL]:
+                raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
+            elif i % 2 == 1 and message.role not in [Role.ASSISTANT, Role.FUNCTION]:
+                raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
+
+            input_messages.append({"role": role_mapping[message.role], "content": message.content})
+
+        tool_list = request.tools
+        if isinstance(tool_list, list) and len(tool_list):
+            try:
+                tools = json.dumps([tool["function"] for tool in tool_list], ensure_ascii=False)
+            except Exception:
+                raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid tools")
+        else:
+            tools = ""
 
         if request.stream:
-            generate = predict(query, history, system, request)
+            if tools:
+                raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Cannot stream function calls.")
+
+            generate = stream_chat_completion(input_messages, system, tools, request)
             return EventSourceResponse(generate, media_type="text/event-stream")
 
-        response, (prompt_length, response_length) = chat_model.chat(
-            query, history, system,
+        responses = await chat_model.achat(
+            input_messages,
+            system,
+            tools,
             do_sample=request.do_sample,
             temperature=request.temperature,
             top_p=request.top_p,
             max_new_tokens=request.max_tokens,
-            num_return_sequences=request.n
+            num_return_sequences=request.n,
         )
 
+        prompt_length, response_length = 0, 0
+        choices = []
+        for i, response in enumerate(responses):
+            if tools:
+                result = chat_model.engine.template.format_tools.extract(response.response_text)
+            else:
+                result = response.response_text
+
+            if isinstance(result, tuple):
+                name, arguments = result
+                function = Function(name=name, arguments=arguments)
+                response_message = ChatCompletionMessage(
+                    role=Role.ASSISTANT, tool_calls=[FunctionCall(function=function)]
+                )
+                finish_reason = Finish.TOOL
+            else:
+                response_message = ChatCompletionMessage(role=Role.ASSISTANT, content=result)
+                finish_reason = Finish.STOP if response.finish_reason == "stop" else Finish.LENGTH
+
+            choices.append(
+                ChatCompletionResponseChoice(index=i, message=response_message, finish_reason=finish_reason)
+            )
+            prompt_length = response.prompt_length
+            response_length += response.response_length
+
         usage = ChatCompletionResponseUsage(
             prompt_tokens=prompt_length,
             completion_tokens=response_length,
-            total_tokens=prompt_length+response_length
+            total_tokens=prompt_length + response_length,
         )
 
-        choices = [ChatCompletionResponseChoice(
-            index=i,
-            message=ChatMessage(role=Role.ASSISTANT, content=choice),
-            finish_reason=Finish.STOP
-        ) for i, choice in enumerate(response)]
-
         return ChatCompletionResponse(model=request.model, choices=choices, usage=usage)
 
-    async def predict(query: str, history: List[Tuple[str, str]], system: str, request: ChatCompletionRequest):
+    async def stream_chat_completion(
+        messages: Sequence[Dict[str, str]], system: str, tools: str, request: ChatCompletionRequest
+    ):
         choice_data = ChatCompletionResponseStreamChoice(
-            index=0,
-            delta=DeltaMessage(role=Role.ASSISTANT),
-            finish_reason=None
+            index=0, delta=ChatCompletionMessage(role=Role.ASSISTANT, content=""), finish_reason=None
         )
         chunk = ChatCompletionStreamResponse(model=request.model, choices=[choice_data])
-        yield to_json(chunk)
+        yield jsonify(chunk)
 
-        for new_text in chat_model.stream_chat(
-            query, history, system,
+        async for new_token in chat_model.astream_chat(
+            messages,
+            system,
+            tools,
             do_sample=request.do_sample,
             temperature=request.temperature,
             top_p=request.top_p,
-            max_new_tokens=request.max_tokens
+            max_new_tokens=request.max_tokens,
         ):
-            if len(new_text) == 0:
+            if len(new_token) == 0:
                 continue
 
             choice_data = ChatCompletionResponseStreamChoice(
-                index=0,
-                delta=DeltaMessage(content=new_text),
-                finish_reason=None
+                index=0, delta=ChatCompletionMessage(content=new_token), finish_reason=None
             )
             chunk = ChatCompletionStreamResponse(model=request.model, choices=[choice_data])
-            yield to_json(chunk)
+            yield jsonify(chunk)
 
         choice_data = ChatCompletionResponseStreamChoice(
-            index=0,
-            delta=DeltaMessage(),
-            finish_reason=Finish.STOP
+            index=0, delta=ChatCompletionMessage(), finish_reason=Finish.STOP
         )
         chunk = ChatCompletionStreamResponse(model=request.model, choices=[choice_data])
-        yield to_json(chunk)
+        yield jsonify(chunk)
         yield "[DONE]"
 
+    @app.post("/v1/score/evaluation", response_model=ScoreEvaluationResponse, status_code=status.HTTP_200_OK)
+    async def create_score_evaluation(request: ScoreEvaluationRequest):
+        if chat_model.engine.can_generate:
+            raise HTTPException(status_code=status.HTTP_405_METHOD_NOT_ALLOWED, detail="Not allowed")
+
+        if len(request.messages) == 0:
+            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid request")
+
+        scores = await chat_model.aget_scores(request.messages, max_length=request.max_length)
+        return ScoreEvaluationResponse(model=request.model, scores=scores)
+
     return app
 
 
 if __name__ == "__main__":
     chat_model = ChatModel()
     app = create_app(chat_model)
-    uvicorn.run(app, host="0.0.0.0", port=8000, workers=1)
+    uvicorn.run(app, host="0.0.0.0", port=int(os.environ.get("API_PORT", 8000)), workers=1)

src/llmtuner/api/protocol.py

@@ -1,30 +1,48 @@
 import time
-from enum import Enum
-from pydantic import BaseModel, Field
+from enum import Enum, unique
 from typing import List, Optional
 
+from pydantic import BaseModel, Field
+from typing_extensions import Literal
 
+
+@unique
 class Role(str, Enum):
     USER = "user"
     ASSISTANT = "assistant"
     SYSTEM = "system"
+    FUNCTION = "function"
+    TOOL = "tool"
 
 
+@unique
 class Finish(str, Enum):
     STOP = "stop"
     LENGTH = "length"
+    TOOL = "tool_calls"
 
 
 class ModelCard(BaseModel):
     id: str
-    object: Optional[str] = "model"
-    created: Optional[int] = Field(default_factory=lambda: int(time.time()))
-    owned_by: Optional[str] = "owner"
+    object: Literal["model"] = "model"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    owned_by: Literal["owner"] = "owner"
 
 
 class ModelList(BaseModel):
-    object: Optional[str] = "list"
-    data: Optional[List[ModelCard]] = []
+    object: Literal["list"] = "list"
+    data: List[ModelCard] = []
+
+
+class Function(BaseModel):
+    name: str
+    arguments: str
+
+
+class FunctionCall(BaseModel):
+    id: Literal["call_default"] = "call_default"
+    type: Literal["function"] = "function"
+    function: Function
 
 
 class ChatMessage(BaseModel):
@@ -32,31 +50,33 @@ class ChatMessage(BaseModel):
     content: str
 
 
-class DeltaMessage(BaseModel):
+class ChatCompletionMessage(BaseModel):
     role: Optional[Role] = None
     content: Optional[str] = None
+    tool_calls: Optional[List[FunctionCall]] = None
 
 
 class ChatCompletionRequest(BaseModel):
     model: str
     messages: List[ChatMessage]
-    do_sample: Optional[bool] = True
+    tools: list = []
+    do_sample: bool = True
     temperature: Optional[float] = None
     top_p: Optional[float] = None
-    n: Optional[int] = 1
+    n: int = 1
     max_tokens: Optional[int] = None
-    stream: Optional[bool] = False
+    stream: bool = False
 
 
 class ChatCompletionResponseChoice(BaseModel):
     index: int
-    message: ChatMessage
+    message: ChatCompletionMessage
     finish_reason: Finish
 
 
 class ChatCompletionResponseStreamChoice(BaseModel):
     index: int
-    delta: DeltaMessage
+    delta: ChatCompletionMessage
     finish_reason: Optional[Finish] = None
 
 
@@ -67,17 +87,30 @@ class ChatCompletionResponseUsage(BaseModel):
 
 
 class ChatCompletionResponse(BaseModel):
-    id: Optional[str] = "chatcmpl-default"
-    object: Optional[str] = "chat.completion"
-    created: Optional[int] = Field(default_factory=lambda: int(time.time()))
+    id: Literal["chatcmpl-default"] = "chatcmpl-default"
+    object: Literal["chat.completion"] = "chat.completion"
+    created: int = Field(default_factory=lambda: int(time.time()))
     model: str
     choices: List[ChatCompletionResponseChoice]
     usage: ChatCompletionResponseUsage
 
 
 class ChatCompletionStreamResponse(BaseModel):
-    id: Optional[str] = "chatcmpl-default"
-    object: Optional[str] = "chat.completion.chunk"
-    created: Optional[int] = Field(default_factory=lambda: int(time.time()))
+    id: Literal["chatcmpl-default"] = "chatcmpl-default"
+    object: Literal["chat.completion.chunk"] = "chat.completion.chunk"
+    created: int = Field(default_factory=lambda: int(time.time()))
     model: str
     choices: List[ChatCompletionResponseStreamChoice]
+
+
+class ScoreEvaluationRequest(BaseModel):
+    model: str
+    messages: List[str]
+    max_length: Optional[int] = None
+
+
+class ScoreEvaluationResponse(BaseModel):
+    id: Literal["scoreeval-default"] = "scoreeval-default"
+    object: Literal["score.evaluation"] = "score.evaluation"
+    model: str
+    scores: List[float]

src/llmtuner/chat/__init__.py

@@ -1 +1,5 @@
-from llmtuner.chat.stream_chat import ChatModel
+from .base_engine import BaseEngine
+from .chat_model import ChatModel
+
+
+__all__ = ["BaseEngine", "ChatModel"]

src/llmtuner/chat/base_engine.py

@@ -0,0 +1,69 @@
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, List, Literal, Optional, Sequence, Union
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+    from ..data import Template
+    from ..extras.packages import is_vllm_available
+    from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+    if is_vllm_available():
+        from vllm import AsyncLLMEngine
+
+
+@dataclass
+class Response:
+    response_text: str
+    response_length: int
+    prompt_length: int
+    finish_reason: Literal["stop", "length"]
+
+
+class BaseEngine(ABC):
+    model: Union["PreTrainedModel", "AsyncLLMEngine"]
+    tokenizer: "PreTrainedTokenizer"
+    can_generate: bool
+    template: "Template"
+    generating_args: Dict[str, Any]
+
+    @abstractmethod
+    def __init__(
+        self,
+        model_args: "ModelArguments",
+        data_args: "DataArguments",
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+    ) -> None: ...
+
+    @abstractmethod
+    async def start(
+        self,
+    ) -> None: ...
+
+    @abstractmethod
+    async def chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]: ...
+
+    @abstractmethod
+    async def stream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]: ...
+
+    @abstractmethod
+    async def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]: ...

src/llmtuner/chat/chat_model.py

@@ -0,0 +1,91 @@
+import asyncio
+from threading import Thread
+from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, Generator, List, Optional, Sequence
+
+from ..hparams import get_infer_args
+from .hf_engine import HuggingfaceEngine
+from .vllm_engine import VllmEngine
+
+
+if TYPE_CHECKING:
+    from .base_engine import BaseEngine, Response
+
+
+def _start_background_loop(loop: asyncio.AbstractEventLoop) -> None:
+    asyncio.set_event_loop(loop)
+    loop.run_forever()
+
+
+class ChatModel:
+    def __init__(self, args: Optional[Dict[str, Any]] = None) -> None:
+        model_args, data_args, finetuning_args, generating_args = get_infer_args(args)
+        if model_args.infer_backend == "huggingface":
+            self.engine: "BaseEngine" = HuggingfaceEngine(model_args, data_args, finetuning_args, generating_args)
+        elif model_args.infer_backend == "vllm":
+            self.engine: "BaseEngine" = VllmEngine(model_args, data_args, finetuning_args, generating_args)
+        else:
+            raise NotImplementedError("Unknown backend: {}".format(model_args.infer_backend))
+
+        self._loop = asyncio.new_event_loop()
+        self._thread = Thread(target=_start_background_loop, args=(self._loop,), daemon=True)
+        self._thread.start()
+        asyncio.run_coroutine_threadsafe(self.engine.start(), self._loop)
+
+    def chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]:
+        task = asyncio.run_coroutine_threadsafe(self.achat(messages, system, tools, **input_kwargs), self._loop)
+        return task.result()
+
+    async def achat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]:
+        return await self.engine.chat(messages, system, tools, **input_kwargs)
+
+    def stream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> Generator[str, None, None]:
+        generator = self.astream_chat(messages, system, tools, **input_kwargs)
+        while True:
+            try:
+                task = asyncio.run_coroutine_threadsafe(generator.__anext__(), self._loop)
+                yield task.result()
+            except StopAsyncIteration:
+                break
+
+    async def astream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]:
+        async for new_token in self.engine.stream_chat(messages, system, tools, **input_kwargs):
+            yield new_token
+
+    def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        task = asyncio.run_coroutine_threadsafe(self.aget_scores(batch_input, **input_kwargs), self._loop)
+        return task.result()
+
+    async def aget_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        return await self.engine.get_scores(batch_input, **input_kwargs)

src/llmtuner/chat/hf_engine.py

@@ -0,0 +1,263 @@
+import asyncio
+import concurrent.futures
+import os
+from threading import Thread
+from typing import TYPE_CHECKING, Any, AsyncGenerator, Callable, Dict, List, Optional, Sequence, Tuple
+
+import torch
+from transformers import GenerationConfig, TextIteratorStreamer
+
+from ..data import get_template_and_fix_tokenizer
+from ..extras.misc import get_logits_processor
+from ..model import load_model_and_tokenizer
+from .base_engine import BaseEngine, Response
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+    from trl import PreTrainedModelWrapper
+
+    from ..data import Template
+    from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+
+class HuggingfaceEngine(BaseEngine):
+    def __init__(
+        self,
+        model_args: "ModelArguments",
+        data_args: "DataArguments",
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+    ) -> None:
+        self.can_generate = finetuning_args.stage == "sft"
+        self.model, self.tokenizer = load_model_and_tokenizer(
+            model_args, finetuning_args, is_trainable=False, add_valuehead=(not self.can_generate)
+        )
+        self.tokenizer.padding_side = "left" if self.can_generate else "right"
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.generating_args = generating_args.to_dict()
+
+    @staticmethod
+    def _process_args(
+        model: "PreTrainedModel",
+        tokenizer: "PreTrainedTokenizer",
+        template: "Template",
+        generating_args: Dict[str, Any],
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> Tuple[Dict[str, Any], int]:
+        paired_messages = messages + [{"role": "assistant", "content": ""}]
+        prompt_ids, _ = template.encode_oneturn(
+            tokenizer=tokenizer, messages=paired_messages, system=system, tools=tools
+        )
+        prompt_length = len(prompt_ids)
+        inputs = torch.tensor([prompt_ids], device=model.device)
+
+        do_sample = input_kwargs.pop("do_sample", None)
+        temperature = input_kwargs.pop("temperature", None)
+        top_p = input_kwargs.pop("top_p", None)
+        top_k = input_kwargs.pop("top_k", None)
+        num_return_sequences = input_kwargs.pop("num_return_sequences", None)
+        repetition_penalty = input_kwargs.pop("repetition_penalty", None)
+        max_length = input_kwargs.pop("max_length", None)
+        max_new_tokens = input_kwargs.pop("max_new_tokens", None)
+
+        generating_args.update(
+            dict(
+                do_sample=do_sample if do_sample is not None else generating_args["do_sample"],
+                temperature=temperature or generating_args["temperature"],
+                top_p=top_p or generating_args["top_p"],
+                top_k=top_k or generating_args["top_k"],
+                num_return_sequences=num_return_sequences or 1,
+                repetition_penalty=repetition_penalty or generating_args["repetition_penalty"],
+                eos_token_id=[tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids,
+                pad_token_id=tokenizer.pad_token_id,
+            )
+        )
+
+        if isinstance(num_return_sequences, int) and num_return_sequences > 1:
+            generating_args["do_sample"] = True
+
+        if max_length:
+            generating_args.pop("max_new_tokens", None)
+            generating_args["max_length"] = max_length
+
+        if max_new_tokens:
+            generating_args.pop("max_length", None)
+            generating_args["max_new_tokens"] = max_new_tokens
+
+        gen_kwargs = dict(
+            inputs=inputs,
+            generation_config=GenerationConfig(**generating_args),
+            logits_processor=get_logits_processor(),
+        )
+
+        return gen_kwargs, prompt_length
+
+    @staticmethod
+    @torch.inference_mode()
+    def _chat(
+        model: "PreTrainedModel",
+        tokenizer: "PreTrainedTokenizer",
+        template: "Template",
+        generating_args: Dict[str, Any],
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> List["Response"]:
+        gen_kwargs, prompt_length = HuggingfaceEngine._process_args(
+            model, tokenizer, template, generating_args, messages, system, tools, input_kwargs
+        )
+        generate_output = model.generate(**gen_kwargs)
+        response_ids = generate_output[:, prompt_length:]
+        response = tokenizer.batch_decode(response_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+        results = []
+        for i in range(len(response)):
+            eos_index = (response_ids[i] == tokenizer.eos_token_id).nonzero()
+            response_length = (eos_index[0].item() + 1) if len(eos_index) else len(response_ids[i])
+            results.append(
+                Response(
+                    response_text=response[i],
+                    response_length=response_length,
+                    prompt_length=prompt_length,
+                    finish_reason="stop" if len(eos_index) else "length",
+                )
+            )
+
+        return results
+
+    @staticmethod
+    @torch.inference_mode()
+    def _stream_chat(
+        model: "PreTrainedModel",
+        tokenizer: "PreTrainedTokenizer",
+        template: "Template",
+        generating_args: Dict[str, Any],
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> Callable[[], str]:
+        gen_kwargs, _ = HuggingfaceEngine._process_args(
+            model, tokenizer, template, generating_args, messages, system, tools, input_kwargs
+        )
+        streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+        gen_kwargs["streamer"] = streamer
+        thread = Thread(target=model.generate, kwargs=gen_kwargs, daemon=True)
+        thread.start()
+
+        def stream():
+            try:
+                return streamer.__next__()
+            except StopIteration:
+                raise StopAsyncIteration()
+
+        return stream
+
+    @staticmethod
+    @torch.inference_mode()
+    def _get_scores(
+        model: "PreTrainedModelWrapper",
+        tokenizer: "PreTrainedTokenizer",
+        batch_input: List[str],
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> List[float]:
+        max_length = input_kwargs.pop("max_length", None)
+        device = getattr(model.pretrained_model, "device", "cuda")
+        inputs = tokenizer(
+            batch_input,
+            padding=True,
+            truncation=True,
+            max_length=max_length or getattr(model.config, "max_position_embeddings", 1024),
+            return_tensors="pt",
+            add_special_tokens=True,
+        ).to(device)
+
+        input_ids: torch.Tensor = inputs["input_ids"]
+        _, _, values = model(**inputs, output_hidden_states=True, return_dict=True)
+
+        if getattr(model.config, "model_type", None) == "chatglm":
+            values = torch.transpose(values, 0, 1)
+
+        scores = []
+        for i in range(input_ids.size(0)):
+            end_indexes = (input_ids[i] != tokenizer.pad_token_id).nonzero()
+            end_index = end_indexes[-1].item() if len(end_indexes) else 0
+            scores.append(values[i, end_index].nan_to_num().item())
+
+        return scores
+
+    async def start(self) -> None:
+        self._semaphore = asyncio.Semaphore(int(os.environ.get("MAX_CONCURRENT", 1)))
+
+    async def chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]:
+        if not self.can_generate:
+            raise ValueError("The current model does not support `chat`.")
+
+        loop = asyncio.get_running_loop()
+        input_args = (
+            self.model,
+            self.tokenizer,
+            self.template,
+            self.generating_args,
+            messages,
+            system,
+            tools,
+            input_kwargs,
+        )
+        async with self._semaphore:
+            with concurrent.futures.ThreadPoolExecutor() as pool:
+                return await loop.run_in_executor(pool, self._chat, *input_args)
+
+    async def stream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]:
+        if not self.can_generate:
+            raise ValueError("The current model does not support `stream_chat`.")
+
+        loop = asyncio.get_running_loop()
+        input_args = (
+            self.model,
+            self.tokenizer,
+            self.template,
+            self.generating_args,
+            messages,
+            system,
+            tools,
+            input_kwargs,
+        )
+        async with self._semaphore:
+            with concurrent.futures.ThreadPoolExecutor() as pool:
+                stream = self._stream_chat(*input_args)
+                while True:
+                    try:
+                        yield await loop.run_in_executor(pool, stream)
+                    except StopAsyncIteration:
+                        break
+
+    async def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        if self.can_generate:
+            raise ValueError("Cannot get scores using an auto-regressive model.")
+
+        loop = asyncio.get_running_loop()
+        input_args = (self.model, self.tokenizer, batch_input, input_kwargs)
+        async with self._semaphore:
+            with concurrent.futures.ThreadPoolExecutor() as pool:
+                return await loop.run_in_executor(pool, self._get_scores, *input_args)

src/llmtuner/chat/stream_chat.py

@@ -1,109 +0,0 @@
-import torch
-from typing import Any, Dict, Generator, List, Optional, Tuple
-from threading import Thread
-from transformers import GenerationConfig, TextIteratorStreamer
-
-from llmtuner.extras.misc import dispatch_model, get_logits_processor
-from llmtuner.extras.template import get_template_and_fix_tokenizer
-from llmtuner.tuner.core import get_infer_args, load_model_and_tokenizer
-
-
-class ChatModel:
-
-    def __init__(self, args: Optional[Dict[str, Any]] = None) -> None:
-        model_args, data_args, finetuning_args, self.generating_args = get_infer_args(args)
-        self.model, self.tokenizer = load_model_and_tokenizer(model_args, finetuning_args)
-        self.tokenizer.padding_side = "left"
-        self.model = dispatch_model(self.model)
-        self.template = get_template_and_fix_tokenizer(data_args.template, self.tokenizer)
-        self.system_prompt = data_args.system_prompt
-
-    def process_args(
-        self,
-        query: str,
-        history: Optional[List[Tuple[str, str]]] = None,
-        system: Optional[str] = None,
-        **input_kwargs
-    ) -> Tuple[Dict[str, Any], int]:
-        system = system or self.system_prompt
-        prompt, _ = self.template.encode_oneturn(
-            tokenizer=self.tokenizer, query=query, resp="", history=history, system=system
-        )
-        prompt_length = len(prompt)
-        input_ids = torch.tensor([prompt], device=self.model.device)
-
-        do_sample = input_kwargs.pop("do_sample", None)
-        temperature = input_kwargs.pop("temperature", None)
-        top_p = input_kwargs.pop("top_p", None)
-        top_k = input_kwargs.pop("top_k", None)
-        num_return_sequences = input_kwargs.pop("num_return_sequences", None)
-        repetition_penalty = input_kwargs.pop("repetition_penalty", None)
-        max_length = input_kwargs.pop("max_length", None)
-        max_new_tokens = input_kwargs.pop("max_new_tokens", None)
-
-        generating_args = self.generating_args.to_dict()
-        generating_args.update(dict(
-            do_sample=do_sample if do_sample is not None else generating_args["do_sample"],
-            temperature=temperature or generating_args["temperature"],
-            top_p=top_p or generating_args["top_p"],
-            top_k=top_k or generating_args["top_k"],
-            num_return_sequences=num_return_sequences or 1,
-            repetition_penalty=repetition_penalty or generating_args["repetition_penalty"],
-            eos_token_id=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
-            pad_token_id=self.tokenizer.pad_token_id
-        ))
-
-        if isinstance(num_return_sequences, int) and num_return_sequences > 1:
-            generating_args["do_sample"] = True
-
-        if max_length:
-            generating_args.pop("max_new_tokens", None)
-            generating_args["max_length"] = max_length
-
-        if max_new_tokens:
-            generating_args.pop("max_length", None)
-            generating_args["max_new_tokens"] = max_new_tokens
-
-        gen_kwargs = dict(
-            inputs=input_ids,
-            generation_config=GenerationConfig(**generating_args),
-            logits_processor=get_logits_processor()
-        )
-
-        return gen_kwargs, prompt_length
-
-    @torch.inference_mode()
-    def chat(
-        self,
-        query: str,
-        history: Optional[List[Tuple[str, str]]] = None,
-        system: Optional[str] = None,
-        **input_kwargs
-    ) -> Tuple[List[str], Tuple[int, int]]:
-        gen_kwargs, prompt_length = self.process_args(query, history, system, **input_kwargs)
-        generate_output = self.model.generate(**gen_kwargs)
-        response_ids = generate_output[:, prompt_length:]
-        response = self.tokenizer.batch_decode(response_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
-        response_length = 0
-        for i in range(len(response_ids)):
-            eos_index = (response_ids[i] == self.tokenizer.eos_token_id).nonzero()
-            response_length += eos_index[0].item() if len(eos_index) else len(response_ids[i])
-
-        return response, (prompt_length, response_length)
-
-    @torch.inference_mode()
-    def stream_chat(
-        self,
-        query: str,
-        history: Optional[List[Tuple[str, str]]] = None,
-        system: Optional[str] = None,
-        **input_kwargs
-    ) -> Generator[str, None, None]:
-        gen_kwargs, _ = self.process_args(query, history, system, **input_kwargs)
-        streamer = TextIteratorStreamer(self.tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
-        gen_kwargs["streamer"] = streamer
-
-        thread = Thread(target=self.model.generate, kwargs=gen_kwargs)
-        thread.start()
-
-        yield from streamer

src/llmtuner/chat/vllm_engine.py

@@ -0,0 +1,149 @@
+import uuid
+from typing import TYPE_CHECKING, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence
+
+from transformers.utils.versions import require_version
+
+from ..data import get_template_and_fix_tokenizer
+from ..extras.misc import get_device_count
+from ..extras.packages import is_vllm_available
+from ..model import load_tokenizer
+from .base_engine import BaseEngine, Response
+
+
+if is_vllm_available():
+    from vllm import AsyncEngineArgs, AsyncLLMEngine, RequestOutput, SamplingParams
+
+if TYPE_CHECKING:
+    from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+
+class VllmEngine(BaseEngine):
+    def __init__(
+        self,
+        model_args: "ModelArguments",
+        data_args: "DataArguments",
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+    ) -> None:
+        require_version("vllm>=0.3.3", "To fix: pip install vllm>=0.3.3")
+        self.can_generate = finetuning_args.stage == "sft"
+        engine_args = AsyncEngineArgs(
+            model=model_args.model_name_or_path,
+            trust_remote_code=True,
+            max_model_len=model_args.vllm_maxlen,
+            tensor_parallel_size=get_device_count() or 1,
+            gpu_memory_utilization=model_args.vllm_gpu_util,
+            disable_log_stats=True,
+            disable_log_requests=True,
+            enforce_eager=model_args.vllm_enforce_eager,
+        )
+        self.model = AsyncLLMEngine.from_engine_args(engine_args)
+        self.tokenizer = load_tokenizer(model_args)
+        self.tokenizer.padding_side = "left"
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.generating_args = generating_args.to_dict()
+
+    async def _generate(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncIterator["RequestOutput"]:
+        request_id = "chatcmpl-{}".format(uuid.uuid4().hex)
+        paired_messages = messages + [{"role": "assistant", "content": ""}]
+        prompt_ids, _ = self.template.encode_oneturn(
+            tokenizer=self.tokenizer, messages=paired_messages, system=system, tools=tools
+        )
+        prompt_length = len(prompt_ids)
+
+        temperature = input_kwargs.pop("temperature", None)
+        top_p = input_kwargs.pop("top_p", None)
+        top_k = input_kwargs.pop("top_k", None)
+        num_return_sequences = input_kwargs.pop("num_return_sequences", None)
+        repetition_penalty = input_kwargs.pop("repetition_penalty", None)
+        max_length = input_kwargs.pop("max_length", None)
+        max_new_tokens = input_kwargs.pop("max_new_tokens", None)
+
+        generating_args = self.generating_args.copy()
+        generating_args.update(
+            dict(
+                temperature=temperature or generating_args["temperature"],
+                top_p=top_p or generating_args["top_p"],
+                top_k=top_k or generating_args["top_k"],
+                num_return_sequences=num_return_sequences or 1,
+                repetition_penalty=repetition_penalty or generating_args["repetition_penalty"],
+            )
+        )
+
+        if max_length:
+            generating_args["max_new_tokens"] = max_length - prompt_length
+
+        if max_new_tokens:
+            generating_args["max_new_tokens"] = max_new_tokens
+
+        sampling_params = SamplingParams(
+            n=generating_args["num_return_sequences"],
+            repetition_penalty=generating_args["repetition_penalty"],
+            temperature=generating_args["temperature"],
+            top_p=generating_args["top_p"],
+            top_k=generating_args["top_k"],
+            use_beam_search=generating_args["num_beams"] > 1,
+            length_penalty=generating_args["length_penalty"],
+            stop_token_ids=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
+            max_tokens=generating_args["max_new_tokens"],
+            skip_special_tokens=True,
+        )
+        result_generator = self.model.generate(
+            prompt=None, sampling_params=sampling_params, request_id=request_id, prompt_token_ids=prompt_ids
+        )
+        return result_generator
+
+    async def start(self) -> None:
+        pass
+
+    async def chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]:
+        final_output = None
+        generator = await self._generate(messages, system, tools, **input_kwargs)
+        async for request_output in generator:
+            final_output = request_output
+
+        results = []
+        for output in final_output.outputs:
+            results.append(
+                Response(
+                    response_text=output.text,
+                    response_length=len(output.token_ids),
+                    prompt_length=len(final_output.prompt_token_ids),
+                    finish_reason=output.finish_reason,
+                )
+            )
+
+        return results
+
+    async def stream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]:
+        generated_text = ""
+        generator = await self._generate(messages, system, tools, **input_kwargs)
+        async for result in generator:
+            delta_text = result.outputs[0].text[len(generated_text) :]
+            generated_text = result.outputs[0].text
+            yield delta_text
+
+    async def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        raise NotImplementedError("vLLM engine does not support get_scores.")

src/llmtuner/data/__init__.py

@@ -0,0 +1,6 @@
+from .loader import get_dataset
+from .template import Template, get_template_and_fix_tokenizer, templates
+from .utils import Role, split_dataset
+
+
+__all__ = ["get_dataset", "Template", "get_template_and_fix_tokenizer", "templates", "Role", "split_dataset"]

src/llmtuner/data/aligner.py

@@ -0,0 +1,133 @@
+from functools import partial
+from typing import TYPE_CHECKING, Any, Dict, List, Union
+
+from datasets import Features
+
+from .utils import Role
+
+
+if TYPE_CHECKING:
+    from datasets import Dataset, IterableDataset
+
+    from ..hparams import DataArguments
+    from .parser import DatasetAttr
+
+
+def convert_alpaca(examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr") -> Dict[str, List[Any]]:
+    outputs = {"prompt": [], "response": [], "system": [], "tools": []}
+    for i in range(len(examples[dataset_attr.prompt])):
+        prompt = []
+        if dataset_attr.history and isinstance(examples[dataset_attr.history][i], list):
+            for old_prompt, old_response in examples[dataset_attr.history][i]:
+                prompt.append({"role": Role.USER.value, "content": old_prompt})
+                prompt.append({"role": Role.ASSISTANT.value, "content": old_response})
+
+        content = []
+        if dataset_attr.prompt and examples[dataset_attr.prompt][i]:
+            content.append(examples[dataset_attr.prompt][i])
+
+        if dataset_attr.query and examples[dataset_attr.query][i]:
+            content.append(examples[dataset_attr.query][i])
+
+        prompt.append({"role": Role.USER.value, "content": "\n".join(content)})
+
+        if dataset_attr.response and isinstance(examples[dataset_attr.response][i], list):
+            response = [
+                {"role": Role.ASSISTANT.value, "content": content} for content in examples[dataset_attr.response][i]
+            ]
+        elif dataset_attr.response and isinstance(examples[dataset_attr.response][i], str):
+            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
+        else:
+            response = []
+
+        outputs["prompt"].append(prompt)
+        outputs["response"].append(response)
+        outputs["system"].append(examples[dataset_attr.system][i] if dataset_attr.system else "")
+        outputs["tools"].append("")
+
+    return outputs
+
+
+def convert_sharegpt(examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr") -> Dict[str, List[Any]]:
+    outputs = {"prompt": [], "response": [], "system": [], "tools": []}
+    tag_mapping = {
+        dataset_attr.user_tag: Role.USER.value,
+        dataset_attr.assistant_tag: Role.ASSISTANT.value,
+        dataset_attr.observation_tag: Role.OBSERVATION.value,
+        dataset_attr.function_tag: Role.FUNCTION.value,
+        dataset_attr.system_tag: Role.SYSTEM.value,
+    }
+    odd_tags = (dataset_attr.user_tag, dataset_attr.observation_tag)
+    even_tags = (dataset_attr.assistant_tag, dataset_attr.function_tag)
+    accept_tags = (odd_tags, even_tags)
+    for i, messages in enumerate(examples[dataset_attr.messages]):
+        if dataset_attr.system_tag and messages[0][dataset_attr.role_tag] == dataset_attr.system_tag:
+            system = messages[0][dataset_attr.content_tag]
+            messages = messages[1:]
+        else:
+            system = examples[dataset_attr.system][i] if dataset_attr.system else ""
+
+        messages = messages[: len(messages) // 2 * 2]  # should be multiples of 2
+        if len(messages) == 0:
+            continue
+
+        aligned_messages = []
+        for turn_idx, message in enumerate(messages):
+            if message[dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
+                raise ValueError("Invalid role tag in {}.".format(messages))
+
+            aligned_messages.append(
+                {"role": tag_mapping[message[dataset_attr.role_tag]], "content": message[dataset_attr.content_tag]}
+            )
+
+        outputs["prompt"].append(aligned_messages[:-1])
+        outputs["response"].append(aligned_messages[-1:])
+        outputs["system"].append(system)
+        outputs["tools"].append(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
+
+    return outputs
+
+
+def align_dataset(
+    dataset: Union["Dataset", "IterableDataset"], dataset_attr: "DatasetAttr", data_args: "DataArguments"
+) -> Union["Dataset", "IterableDataset"]:
+    r"""
+    Aligned dataset:
+        prompt: [{"role": "user", "content": "..."}] * (2T - 1)
+        response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
+        system: "..."
+        tools: "..."
+    """
+    if dataset_attr.formatting == "alpaca":
+        convert_func = partial(convert_alpaca, dataset_attr=dataset_attr)
+    else:
+        convert_func = partial(convert_sharegpt, dataset_attr=dataset_attr)
+
+    column_names = list(next(iter(dataset)).keys())
+    features = Features.from_dict(
+        {
+            "prompt": [
+                {"role": {"dtype": "string", "_type": "Value"}, "content": {"dtype": "string", "_type": "Value"}}
+            ],
+            "response": [
+                {"role": {"dtype": "string", "_type": "Value"}, "content": {"dtype": "string", "_type": "Value"}}
+            ],
+            "system": {"dtype": "string", "_type": "Value"},
+            "tools": {"dtype": "string", "_type": "Value"},
+        }
+    )
+    kwargs = {}
+    if not data_args.streaming:
+        kwargs = dict(
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=(not data_args.overwrite_cache),
+            desc="Converting format of dataset",
+        )
+
+    return dataset.map(
+        convert_func,
+        batched=True,
+        remove_columns=column_names,
+        features=features,
+        **kwargs,
+    )

src/llmtuner/data/formatter.py

@@ -0,0 +1,187 @@
+import json
+import re
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Literal, Optional, Sequence, Set, Tuple, Union
+
+
+SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
+
+
+JSON_FORMAT_PROMPT = (
+    """, in a JSON format representing the kwargs (e.g. ```{"input": "hello world", "num_beams": 5}```)"""
+)
+
+
+TOOL_SYSTEM_PROMPT = (
+    "You have access to the following tools:\n{tool_text}"
+    "Use the following format if using a tool:\n"
+    "```\n"
+    "Action: tool name (one of [{tool_names}]).\n"
+    "Action Input: the input to the tool{format_prompt}.\n"
+    "```\n"
+)
+
+
+def default_tool_formatter(tools: List[Dict[str, Any]]) -> str:
+    tool_text = ""
+    tool_names = []
+    for tool in tools:
+        param_text = ""
+        for name, param in tool["parameters"]["properties"].items():
+            required = ", required" if name in tool["parameters"].get("required", []) else ""
+            enum = ", should be one of [{}]".format(", ".join(param["enum"])) if param.get("enum", None) else ""
+            items = (
+                ", where each item should be {}".format(param["items"].get("type", "")) if param.get("items") else ""
+            )
+            param_text += "  - {name} ({type}{required}): {desc}{enum}{items}\n".format(
+                name=name,
+                type=param.get("type", ""),
+                required=required,
+                desc=param.get("description", ""),
+                enum=enum,
+                items=items,
+            )
+
+        tool_text += "> Tool Name: {name}\nTool Description: {desc}\nTool Args:\n{args}\n".format(
+            name=tool["name"], desc=tool.get("description", ""), args=param_text
+        )
+        tool_names.append(tool["name"])
+
+    return TOOL_SYSTEM_PROMPT.format(
+        tool_text=tool_text, tool_names=", ".join(tool_names), format_prompt=JSON_FORMAT_PROMPT
+    )
+
+
+def default_tool_extractor(content: str) -> Union[str, Tuple[str, str]]:
+    regex = re.compile(r"Action:\s*([a-zA-Z0-9_]+).*?Action Input:\s*(.*)", re.DOTALL)
+    action_match = re.search(regex, content)
+    if not action_match:
+        return content
+
+    tool_name = action_match.group(1).strip()
+    tool_input = action_match.group(2).strip().strip('"').strip("```")
+    try:
+        arguments = json.loads(tool_input)
+    except json.JSONDecodeError:
+        return content
+
+    return tool_name, json.dumps(arguments, ensure_ascii=False)
+
+
+@dataclass
+class Formatter(ABC):
+    slots: SLOTS = field(default_factory=list)
+    tool_format: Optional[Literal["default"]] = None
+
+    @abstractmethod
+    def apply(self, **kwargs) -> SLOTS: ...
+
+    def extract(self, content: str) -> Union[str, Tuple[str, str]]:
+        raise NotImplementedError
+
+
+@dataclass
+class EmptyFormatter(Formatter):
+    def __post_init__(self):
+        has_placeholder = False
+        for slot in filter(lambda s: isinstance(s, str), self.slots):
+            if re.search(r"\{\{[a-zA-Z_][a-zA-Z0-9_]*\}\}", slot):
+                has_placeholder = True
+
+        if has_placeholder:
+            raise ValueError("Empty formatter should not contain any placeholder.")
+
+    def apply(self, **kwargs) -> SLOTS:
+        return self.slots
+
+
+@dataclass
+class StringFormatter(Formatter):
+    def __post_init__(self):
+        has_placeholder = False
+        for slot in filter(lambda s: isinstance(s, str), self.slots):
+            if re.search(r"\{\{[a-zA-Z_][a-zA-Z0-9_]*\}\}", slot):
+                has_placeholder = True
+
+        if not has_placeholder:
+            raise ValueError("A placeholder is required in the string formatter.")
+
+    def apply(self, **kwargs) -> SLOTS:
+        elements = []
+        for slot in self.slots:
+            if isinstance(slot, str):
+                for name, value in kwargs.items():
+                    if not isinstance(value, str):
+                        raise RuntimeError("Expected a string, got {}".format(value))
+
+                    slot = slot.replace("{{" + name + "}}", value, 1)
+                elements.append(slot)
+            elif isinstance(slot, (dict, set)):
+                elements.append(slot)
+            else:
+                raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+
+        return elements
+
+
+@dataclass
+class FunctionFormatter(Formatter):
+    def __post_init__(self):
+        has_name, has_args = False, False
+        for slot in filter(lambda s: isinstance(s, str), self.slots):
+            if "{{name}}" in slot:
+                has_name = True
+            if "{{arguments}}" in slot:
+                has_args = True
+
+        if not has_name or not has_args:
+            raise ValueError("Name and arguments placeholders are required in the function formatter.")
+
+    def apply(self, **kwargs) -> SLOTS:
+        content = kwargs.pop("content")
+        try:
+            function = json.loads(content)
+            name = function["name"]
+            arguments = json.dumps(function["arguments"], ensure_ascii=False)
+        except Exception:
+            name, arguments = "", ""
+
+        elements = []
+        for slot in self.slots:
+            if isinstance(slot, str):
+                slot = slot.replace("{{name}}", name).replace("{{arguments}}", arguments)
+                elements.append(slot)
+            elif isinstance(slot, (dict, set)):
+                elements.append(slot)
+            else:
+                raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+
+        return elements
+
+
+@dataclass
+class ToolFormatter(Formatter):
+    def __post_init__(self):
+        if self.tool_format is None:
+            raise ValueError("Tool format was not found.")
+
+    def apply(self, **kwargs) -> SLOTS:
+        content = kwargs.pop("content")
+        try:
+            tools = json.loads(content)
+            if not len(tools):
+                return [""]
+
+            if self.tool_format == "default":
+                return [default_tool_formatter(tools)]
+            else:
+                raise NotImplementedError
+        except Exception:
+            return [""]
+
+    def extract(self, content: str) -> Union[str, Tuple[str, str]]:
+        if self.tool_format == "default":
+            return default_tool_extractor(content)
+        else:
+            raise NotImplementedError

src/llmtuner/data/loader.py

@@ -0,0 +1,170 @@
+import inspect
+import os
+from typing import TYPE_CHECKING, Literal, Union
+
+from datasets import load_dataset, load_from_disk
+
+from ..extras.constants import FILEEXT2TYPE
+from ..extras.logging import get_logger
+from .aligner import align_dataset
+from .parser import get_dataset_list
+from .preprocess import get_preprocess_and_print_func
+from .template import get_template_and_fix_tokenizer
+from .utils import checksum, merge_dataset
+
+
+if TYPE_CHECKING:
+    from datasets import Dataset, IterableDataset
+    from transformers import Seq2SeqTrainingArguments
+    from transformers.tokenization_utils import PreTrainedTokenizer
+
+    from ..hparams import DataArguments, ModelArguments
+    from .parser import DatasetAttr
+
+
+logger = get_logger(__name__)
+
+
+def load_single_dataset(
+    dataset_attr: "DatasetAttr",
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+) -> Union["Dataset", "IterableDataset"]:
+    logger.info("Loading dataset {}...".format(dataset_attr))
+    data_path, data_name, data_dir, data_files = None, None, None, None
+    if dataset_attr.load_from in ["hf_hub", "ms_hub"]:
+        data_path = dataset_attr.dataset_name
+        data_name = dataset_attr.subset
+        data_dir = dataset_attr.folder
+
+    elif dataset_attr.load_from == "script":
+        data_path = os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)
+        data_name = dataset_attr.subset
+        data_dir = dataset_attr.folder
+
+    elif dataset_attr.load_from == "file":
+        data_files = []
+        local_path = os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)
+        if os.path.isdir(local_path):  # is directory
+            for file_name in os.listdir(local_path):
+                data_files.append(os.path.join(local_path, file_name))
+                if data_path is None:
+                    data_path = FILEEXT2TYPE.get(file_name.split(".")[-1], None)
+                elif data_path != FILEEXT2TYPE.get(file_name.split(".")[-1], None):
+                    raise ValueError("File types should be identical.")
+        elif os.path.isfile(local_path):  # is file
+            data_files.append(local_path)
+            data_path = FILEEXT2TYPE.get(local_path.split(".")[-1], None)
+        else:
+            raise ValueError("File not found.")
+
+        if data_path is None:
+            raise ValueError("File extension must be txt, csv, json or jsonl.")
+
+        checksum(data_files, dataset_attr.file_sha1)
+    else:
+        raise NotImplementedError
+
+    if dataset_attr.load_from == "ms_hub":
+        try:
+            from modelscope import MsDataset
+            from modelscope.utils.config_ds import MS_DATASETS_CACHE
+
+            cache_dir = model_args.cache_dir or MS_DATASETS_CACHE
+            dataset = MsDataset.load(
+                dataset_name=data_path,
+                subset_name=data_name,
+                data_dir=data_dir,
+                data_files=data_files,
+                split=data_args.split,
+                cache_dir=cache_dir,
+                token=model_args.ms_hub_token,
+                use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
+            ).to_hf_dataset()
+        except ImportError:
+            raise ImportError("Please install modelscope via `pip install modelscope -U`")
+    else:
+        if "trust_remote_code" in inspect.signature(load_dataset).parameters:  # for datasets==2.16.0
+            kwargs = {"trust_remote_code": True}
+        else:
+            kwargs = {}
+
+        dataset = load_dataset(
+            path=data_path,
+            name=data_name,
+            data_dir=data_dir,
+            data_files=data_files,
+            split=data_args.split,
+            cache_dir=model_args.cache_dir,
+            token=model_args.hf_hub_token,
+            streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
+            **kwargs,
+        )
+
+    if data_args.streaming and (dataset_attr.load_from == "file"):  # faster than specifying streaming=True
+        dataset = dataset.to_iterable_dataset()  # TODO: add num shards parameter
+
+    if data_args.max_samples is not None:  # truncate dataset
+        num_samples = min(data_args.max_samples, len(dataset))
+        dataset = dataset.select(range(num_samples))
+
+    return align_dataset(dataset, dataset_attr, data_args)
+
+
+def get_dataset(
+    tokenizer: "PreTrainedTokenizer",
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    stage: Literal["pt", "sft", "rm", "ppo"],
+    # split: Optional[str] = "train", # TODO: add split
+) -> Union["Dataset", "IterableDataset"]:
+    template = get_template_and_fix_tokenizer(tokenizer, data_args.template)
+    if data_args.train_on_prompt and template.efficient_eos:
+        raise ValueError("Current template does not support `train_on_prompt`.")
+
+    # Load from cache
+    if data_args.cache_path is not None:
+        if os.path.exists(data_args.cache_path):
+            logger.warning("Loading dataset from disk will ignore other data arguments.")
+            dataset = load_from_disk(data_args.cache_path)
+            if data_args.streaming:
+                dataset = dataset.to_iterable_dataset()
+            return dataset
+
+        if data_args.streaming:
+            raise ValueError("Turn off `streaming` when saving dataset to disk.")
+
+    with training_args.main_process_first(desc="load dataset"):
+        all_datasets = []
+        for dataset_attr in get_dataset_list(data_args):
+            all_datasets.append(load_single_dataset(dataset_attr, model_args, data_args))
+        dataset = merge_dataset(all_datasets, data_args, training_args)
+
+    with training_args.main_process_first(desc="pre-process dataset"):
+        preprocess_func, print_function = get_preprocess_and_print_func(
+            tokenizer, template, data_args, training_args, stage
+        )
+        column_names = list(next(iter(dataset)).keys())
+        kwargs = {}
+        if not data_args.streaming:
+            kwargs = dict(
+                num_proc=data_args.preprocessing_num_workers,
+                load_from_cache_file=(not data_args.overwrite_cache),
+                desc="Running tokenizer on dataset",
+            )
+
+        dataset = dataset.map(preprocess_func, batched=True, remove_columns=column_names, **kwargs)
+
+        if data_args.cache_path is not None and not os.path.exists(data_args.cache_path):
+            if training_args.should_save:
+                dataset.save_to_disk(data_args.cache_path)
+                logger.info("Dataset cache saved at {}.".format(data_args.cache_path))
+
+        if training_args.should_log:
+            try:
+                print_function(next(iter(dataset)))
+            except StopIteration:
+                raise RuntimeError("Cannot find valid samples, check `data/README.md` for the data format.")
+
+        return dataset

src/llmtuner/data/parser.py

@@ -0,0 +1,119 @@
+import json
+import os
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional
+
+from ..extras.constants import DATA_CONFIG
+from ..extras.misc import use_modelscope
+
+
+if TYPE_CHECKING:
+    from ..hparams import DataArguments
+
+
+@dataclass
+class DatasetAttr:
+    r"""
+    Dataset attributes.
+    """
+
+    """ basic configs """
+    load_from: Literal["hf_hub", "ms_hub", "script", "file"]
+    dataset_name: str
+    """ extra configs """
+    file_sha1: Optional[str] = None
+    subset: Optional[str] = None
+    folder: Optional[str] = None
+    ranking: bool = False
+    formatting: Literal["alpaca", "sharegpt"] = "alpaca"
+    """ columns """
+    system: Optional[str] = None
+    """ columns for the alpaca format """
+    prompt: Optional[str] = "instruction"
+    query: Optional[str] = "input"
+    response: Optional[str] = "output"
+    history: Optional[str] = None
+    """ columns for the sharegpt format """
+    messages: Optional[str] = "conversations"
+    tools: Optional[str] = None
+    """ tags for the sharegpt format """
+    role_tag: Optional[str] = "from"
+    content_tag: Optional[str] = "value"
+    user_tag: Optional[str] = "human"
+    assistant_tag: Optional[str] = "gpt"
+    observation_tag: Optional[str] = "observation"
+    function_tag: Optional[str] = "function_call"
+    system_tag: Optional[str] = "system"
+
+    def __repr__(self) -> str:
+        return self.dataset_name
+
+    def set_attr(self, key: str, obj: Dict[str, Any], default: Optional[Any] = None) -> None:
+        setattr(self, key, obj.get(key, default))
+
+
+def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
+    dataset_names = [ds.strip() for ds in data_args.dataset.split(",")] if data_args.dataset is not None else []
+    try:
+        with open(os.path.join(data_args.dataset_dir, DATA_CONFIG), "r") as f:
+            dataset_info = json.load(f)
+    except Exception as err:
+        if data_args.dataset is not None:
+            raise ValueError(
+                "Cannot open {} due to {}.".format(os.path.join(data_args.dataset_dir, DATA_CONFIG), str(err))
+            )
+        dataset_info = None
+
+    if data_args.interleave_probs is not None:
+        data_args.interleave_probs = [float(prob.strip()) for prob in data_args.interleave_probs.split(",")]
+
+    dataset_list: List[DatasetAttr] = []
+    for name in dataset_names:
+        if name not in dataset_info:
+            raise ValueError("Undefined dataset {} in {}.".format(name, DATA_CONFIG))
+
+        has_hf_url = "hf_hub_url" in dataset_info[name]
+        has_ms_url = "ms_hub_url" in dataset_info[name]
+
+        if has_hf_url or has_ms_url:
+            if (use_modelscope() and has_ms_url) or (not has_hf_url):
+                dataset_attr = DatasetAttr("ms_hub", dataset_name=dataset_info[name]["ms_hub_url"])
+            else:
+                dataset_attr = DatasetAttr("hf_hub", dataset_name=dataset_info[name]["hf_hub_url"])
+        elif "script_url" in dataset_info[name]:
+            dataset_attr = DatasetAttr("script", dataset_name=dataset_info[name]["script_url"])
+        else:
+            dataset_attr = DatasetAttr("file", dataset_name=dataset_info[name]["file_name"])
+
+        dataset_attr.set_attr("file_sha1", dataset_info[name])
+        dataset_attr.set_attr("subset", dataset_info[name])
+        dataset_attr.set_attr("folder", dataset_info[name])
+        dataset_attr.set_attr("ranking", dataset_info[name], default=False)
+        dataset_attr.set_attr("formatting", dataset_info[name], default="alpaca")
+
+        if "columns" in dataset_info[name]:
+            column_names = ["system"]
+            if dataset_attr.formatting == "alpaca":
+                column_names.extend(["prompt", "query", "response", "history"])
+            else:
+                column_names.extend(["messages", "tools"])
+
+            for column_name in column_names:
+                dataset_attr.set_attr(column_name, dataset_info[name]["columns"])
+
+        if dataset_attr.formatting == "sharegpt" and "tags" in dataset_info[name]:
+            tag_names = (
+                "role_tag",
+                "content_tag",
+                "user_tag",
+                "assistant_tag",
+                "observation_tag",
+                "function_tag",
+                "system_tag",
+            )
+            for tag in tag_names:
+                dataset_attr.set_attr(tag, dataset_info[name]["tags"])
+
+        dataset_list.append(dataset_attr)
+
+    return dataset_list

src/llmtuner/data/preprocess.py

@@ -0,0 +1,276 @@
+from functools import partial
+from itertools import chain
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Literal, Tuple
+
+from ..extras.constants import IGNORE_INDEX
+from ..extras.logging import get_logger
+from .utils import Role
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments
+    from transformers.tokenization_utils import PreTrainedTokenizer
+
+    from ..hparams import DataArguments
+    from .template import Template
+
+
+logger = get_logger(__name__)
+
+
+def preprocess_pretrain_dataset(
+    examples: Dict[str, List[Any]], tokenizer: "PreTrainedTokenizer", data_args: "DataArguments"
+) -> Dict[str, List[List[int]]]:
+    # build grouped texts with format `X1 X2 X3 ...` if packing is enabled
+    text_examples = [messages[0]["content"] + tokenizer.eos_token for messages in examples["prompt"]]
+    if not data_args.packing:
+        return tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len)
+
+    tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
+    concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}
+    total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
+    block_size = data_args.cutoff_len
+    # we drop the small remainder, and if the total_length < block_size, we exclude this batch
+    total_length = (total_length // block_size) * block_size
+    # split by chunks of cutoff_len
+    result = {
+        k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
+        for k, t in concatenated_examples.items()
+    }
+    if data_args.template == "gemma":
+        for i in range(len(result["input_ids"])):
+            result["input_ids"][i][0] = tokenizer.bos_token_id
+
+    return result
+
+
+def preprocess_supervised_dataset(
+    examples: Dict[str, List[Any]],
+    tokenizer: "PreTrainedTokenizer",
+    template: "Template",
+    data_args: "DataArguments",
+) -> Dict[str, List[List[int]]]:
+    # build inputs with format `<bos> X Y <eos>` and labels with format `<ignore> ... <ignore> Y <eos>`
+    # for multiturn examples, we only mask the prompt part in each prompt-response pair.
+    model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
+
+    for i in range(len(examples["prompt"])):
+        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
+            continue
+
+        messages = examples["prompt"][i] + examples["response"][i]
+        input_ids, labels = [], []
+        for turn_idx, (source_ids, target_ids) in enumerate(
+            template.encode_multiturn(
+                tokenizer,
+                messages,
+                examples["system"][i],
+                examples["tools"][i],
+                data_args.cutoff_len,
+                data_args.reserved_label_len,
+            )
+        ):
+            if data_args.train_on_prompt:
+                source_mask = source_ids
+            elif turn_idx != 0 and template.efficient_eos:
+                source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
+            else:
+                source_mask = [IGNORE_INDEX] * len(source_ids)
+
+            input_ids += source_ids + target_ids
+            labels += source_mask + target_ids
+
+        if template.efficient_eos:
+            input_ids += [tokenizer.eos_token_id]
+            labels += [tokenizer.eos_token_id]
+
+        model_inputs["input_ids"].append(input_ids)
+        model_inputs["attention_mask"].append([1] * len(input_ids))
+        model_inputs["labels"].append(labels)
+
+    return model_inputs
+
+
+def preprocess_packed_supervised_dataset(
+    examples: Dict[str, List[Any]],
+    tokenizer: "PreTrainedTokenizer",
+    template: "Template",
+    data_args: "DataArguments",
+) -> Dict[str, List[List[int]]]:
+    # build inputs with format `<bos> X1 Y1 <eos> <bos> X2 Y2 <eos>`
+    # and labels with format `<ignore> ... <ignore> Y1 <eos> <ignore> ... <ignore> Y2 <eos>`
+    model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
+    input_ids, labels = [], []
+    for i in range(len(examples["prompt"])):
+        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
+            continue
+
+        messages = examples["prompt"][i] + examples["response"][i]
+        for source_ids, target_ids in template.encode_multiturn(
+            tokenizer, messages, examples["system"][i], examples["tools"][i]
+        ):
+            if data_args.train_on_prompt:
+                source_mask = source_ids
+            elif len(input_ids) != 0 and template.efficient_eos:
+                source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
+            else:
+                source_mask = [IGNORE_INDEX] * len(source_ids)
+
+            input_ids += source_ids + target_ids
+            labels += source_mask + target_ids
+
+    if template.efficient_eos:
+        input_ids += [tokenizer.eos_token_id]
+        labels += [tokenizer.eos_token_id]
+
+    total_length = len(input_ids)
+    block_size = data_args.cutoff_len
+    # we drop the small remainder, and if the total_length < block_size, we exclude this batch
+    total_length = (total_length // block_size) * block_size
+    # split by chunks of cutoff_len
+    for i in range(0, total_length, block_size):
+        if not all(label == IGNORE_INDEX for label in labels[i : i + block_size]):
+            model_inputs["input_ids"].append(input_ids[i : i + block_size])
+            model_inputs["attention_mask"].append([1] * block_size)
+            model_inputs["labels"].append(labels[i : i + block_size])
+
+    return model_inputs
+
+
+def preprocess_unsupervised_dataset(
+    examples: Dict[str, List[Any]],
+    tokenizer: "PreTrainedTokenizer",
+    template: "Template",
+    data_args: "DataArguments",
+) -> Dict[str, List[List[int]]]:
+    # build inputs with format `<bos> X` and labels with format `Y <eos>`
+    model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
+
+    for i in range(len(examples["prompt"])):
+        if len(examples["prompt"][i]) % 2 != 1:
+            continue
+
+        if len(examples["response"][i]) == 1:
+            messages = examples["prompt"][i] + examples["response"][i]
+        else:
+            messages = examples["prompt"][i] + [{"role": Role.ASSISTANT.value, "content": ""}]
+
+        input_ids, labels = template.encode_oneturn(
+            tokenizer,
+            messages,
+            examples["system"][i],
+            examples["tools"][i],
+            data_args.cutoff_len,
+            data_args.reserved_label_len,
+        )
+
+        if template.efficient_eos:
+            labels += [tokenizer.eos_token_id]
+
+        model_inputs["input_ids"].append(input_ids)
+        model_inputs["attention_mask"].append([1] * len(input_ids))
+        model_inputs["labels"].append(labels)
+
+    return model_inputs
+
+
+def preprocess_pairwise_dataset(
+    examples: Dict[str, List[Any]],
+    tokenizer: "PreTrainedTokenizer",
+    template: "Template",
+    data_args: "DataArguments",
+) -> Dict[str, List[List[int]]]:
+    # build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
+    model_inputs = {"prompt_ids": [], "chosen_ids": [], "rejected_ids": []}
+    for i in range(len(examples["prompt"])):
+        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
+            continue
+
+        chosen_messages = examples["prompt"][i] + [examples["response"][i][0]]
+        rejected_messages = examples["prompt"][i] + [examples["response"][i][1]]
+        prompt_ids, chosen_ids = template.encode_oneturn(
+            tokenizer,
+            chosen_messages,
+            examples["system"][i],
+            examples["tools"][i],
+            data_args.cutoff_len,
+            data_args.reserved_label_len,
+        )
+        _, rejected_ids = template.encode_oneturn(
+            tokenizer,
+            rejected_messages,
+            examples["system"][i],
+            examples["tools"][i],
+            data_args.cutoff_len,
+            data_args.reserved_label_len,
+        )
+
+        if template.efficient_eos:
+            chosen_ids += [tokenizer.eos_token_id]
+            rejected_ids += [tokenizer.eos_token_id]
+
+        model_inputs["prompt_ids"].append(prompt_ids)
+        model_inputs["chosen_ids"].append(chosen_ids)
+        model_inputs["rejected_ids"].append(rejected_ids)
+
+    return model_inputs
+
+
+def print_supervised_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None:
+    print("input_ids:\n{}".format(example["input_ids"]))
+    print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
+    print("label_ids:\n{}".format(example["labels"]))
+    print(
+        "labels:\n{}".format(
+            tokenizer.decode(list(filter(lambda x: x != IGNORE_INDEX, example["labels"])), skip_special_tokens=False)
+        )
+    )
+
+
+def print_pairwise_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None:
+    print("prompt_ids:\n{}".format(example["prompt_ids"]))
+    print("prompt:\n{}".format(tokenizer.decode(example["prompt_ids"], skip_special_tokens=False)))
+    print("chosen_ids:\n{}".format(example["chosen_ids"]))
+    print("chosen:\n{}".format(tokenizer.decode(example["chosen_ids"], skip_special_tokens=False)))
+    print("rejected_ids:\n{}".format(example["rejected_ids"]))
+    print("rejected:\n{}".format(tokenizer.decode(example["rejected_ids"], skip_special_tokens=False)))
+
+
+def print_unsupervised_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None:
+    print("input_ids:\n{}".format(example["input_ids"]))
+    print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
+
+
+def get_preprocess_and_print_func(
+    tokenizer: "PreTrainedTokenizer",
+    template: "Template",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    stage: Literal["pt", "sft", "rm", "ppo"],
+) -> Tuple[Callable, Callable]:
+    if stage == "pt":
+        preprocess_func = partial(preprocess_pretrain_dataset, tokenizer=tokenizer, data_args=data_args)
+        print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
+    elif stage == "sft" and not training_args.predict_with_generate:
+        if data_args.packing:
+            preprocess_func = partial(
+                preprocess_packed_supervised_dataset, tokenizer=tokenizer, template=template, data_args=data_args
+            )
+        else:
+            preprocess_func = partial(
+                preprocess_supervised_dataset, tokenizer=tokenizer, template=template, data_args=data_args
+            )
+
+        print_function = partial(print_supervised_dataset_example, tokenizer=tokenizer)
+    elif stage == "rm":
+        preprocess_func = partial(
+            preprocess_pairwise_dataset, tokenizer=tokenizer, template=template, data_args=data_args
+        )
+        print_function = partial(print_pairwise_dataset_example, tokenizer=tokenizer)
+    else:
+        preprocess_func = partial(
+            preprocess_unsupervised_dataset, tokenizer=tokenizer, template=template, data_args=data_args
+        )
+        print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
+
+    return preprocess_func, print_function

src/llmtuner/data/template.py

@@ -0,0 +1,773 @@
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Tuple, Union
+
+from ..extras.logging import get_logger
+from .formatter import EmptyFormatter, FunctionFormatter, StringFormatter, ToolFormatter
+from .utils import Role, infer_max_len
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedTokenizer
+
+    from .formatter import SLOTS, Formatter
+
+
+logger = get_logger(__name__)
+
+
+@dataclass
+class Template:
+    format_user: "Formatter"
+    format_assistant: "Formatter"
+    format_system: "Formatter"
+    format_function: "Formatter"
+    format_observation: "Formatter"
+    format_tools: "Formatter"
+    format_separator: "Formatter"
+    default_system: str
+    stop_words: List[str]
+    efficient_eos: bool
+    replace_eos: bool
+    force_system: bool
+
+    def encode_oneturn(
+        self,
+        tokenizer: "PreTrainedTokenizer",
+        messages: List[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        cutoff_len: int = 1_000_000,
+        reserved_label_len: int = 1,
+    ) -> Tuple[List[int], List[int]]:
+        r"""
+        Returns a single pair of token ids representing prompt and response respectively.
+        """
+        encoded_pairs = self._encode(tokenizer, messages, system, tools, cutoff_len, reserved_label_len)
+        prompt_ids = []
+        for query_ids, resp_ids in encoded_pairs[:-1]:
+            prompt_ids += query_ids + resp_ids
+        prompt_ids = prompt_ids + encoded_pairs[-1][0]
+        answer_ids = encoded_pairs[-1][1]
+        return prompt_ids, answer_ids
+
+    def encode_multiturn(
+        self,
+        tokenizer: "PreTrainedTokenizer",
+        messages: List[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        cutoff_len: int = 1_000_000,
+        reserved_label_len: int = 1,
+    ) -> Sequence[Tuple[List[int], List[int]]]:
+        r"""
+        Returns multiple pairs of token ids representing prompts and responses respectively.
+        """
+        return self._encode(tokenizer, messages, system, tools, cutoff_len, reserved_label_len)
+
+    def _encode(
+        self,
+        tokenizer: "PreTrainedTokenizer",
+        messages: List[Dict[str, str]],
+        system: str,
+        tools: str,
+        cutoff_len: int,
+        reserved_label_len: int,
+    ) -> Sequence[Tuple[List[int], List[int]]]:
+        r"""
+        Encodes formatted inputs to pairs of token ids.
+        Turn 0: system + query        resp
+        Turn t: sep + query           resp
+        """
+        system = system or self.default_system
+        encoded_messages = []
+        for i, message in enumerate(messages):
+            elements = []
+            if i == 0 and (system or tools or self.force_system):
+                tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
+                elements += self.format_system.apply(content=(system + tool_text))
+            elif i > 0 and i % 2 == 0:
+                elements += self.format_separator.apply()
+
+            if message["role"] == Role.USER.value:
+                elements += self.format_user.apply(content=message["content"], idx=str(i // 2))
+            elif message["role"] == Role.ASSISTANT.value:
+                elements += self.format_assistant.apply(content=message["content"])
+            elif message["role"] == Role.OBSERVATION.value:
+                elements += self.format_observation.apply(content=message["content"])
+            elif message["role"] == Role.FUNCTION.value:
+                elements += self.format_function.apply(content=message["content"])
+            else:
+                raise NotImplementedError("Unexpected role: {}".format(message["role"]))
+
+            encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
+
+        return self._make_pairs(encoded_messages, cutoff_len, reserved_label_len)
+
+    def _convert_elements_to_ids(
+        self, tokenizer: "PreTrainedTokenizer", elements: List[Union[str, Dict[str, str]]]
+    ) -> List[int]:
+        r"""
+        Converts elements to token ids.
+        """
+        token_ids = []
+        for elem in elements:
+            if isinstance(elem, str):
+                if len(elem) != 0:
+                    token_ids += tokenizer.encode(elem, add_special_tokens=False)
+            elif isinstance(elem, dict):
+                token_ids += [tokenizer.convert_tokens_to_ids(elem.get("token"))]
+            elif isinstance(elem, set):
+                if "bos_token" in elem and tokenizer.bos_token_id is not None:
+                    token_ids += [tokenizer.bos_token_id]
+                elif "eos_token" in elem and tokenizer.eos_token_id is not None:
+                    token_ids += [tokenizer.eos_token_id]
+            else:
+                raise ValueError("Input must be string, set[str] or dict[str, str], got {}".format(type(elem)))
+
+        return token_ids
+
+    def _make_pairs(
+        self,
+        encoded_messages: Sequence[List[int]],
+        cutoff_len: int,
+        reserved_label_len: int,
+    ) -> Sequence[Tuple[List[int], List[int]]]:
+        encoded_pairs = []
+        total_length = 0
+        for i in range(0, len(encoded_messages), 2):
+            if total_length >= cutoff_len:
+                break
+
+            max_source_len, max_target_len = infer_max_len(
+                source_len=len(encoded_messages[i]),
+                target_len=len(encoded_messages[i + 1]),
+                max_len=(cutoff_len - total_length),
+                reserved_label_len=reserved_label_len,
+            )
+            source_ids = encoded_messages[i][:max_source_len]
+            target_ids = encoded_messages[i + 1][:max_target_len]
+            total_length += len(source_ids) + len(target_ids)
+            encoded_pairs.append((source_ids, target_ids))
+
+        return encoded_pairs
+
+
+@dataclass
+class Llama2Template(Template):
+    def _encode(
+        self,
+        tokenizer: "PreTrainedTokenizer",
+        messages: List[Dict[str, str]],
+        system: str,
+        tools: str,
+        cutoff_len: int,
+        reserved_label_len: int,
+    ) -> Sequence[Tuple[List[int], List[int]]]:
+        r"""
+        Encodes formatted inputs to pairs of token ids.
+        Turn 0: system + query        resp
+        Turn t: sep + query           resp
+        """
+        system = system or self.default_system
+        encoded_messages = []
+        for i, message in enumerate(messages):
+            elements = []
+            system_text = ""
+            if i == 0 and (system or tools or self.force_system):
+                tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
+                system_text = self.format_system.apply(content=(system + tool_text))[0]
+            elif i > 0 and i % 2 == 0:
+                elements += self.format_separator.apply()
+
+            if message["role"] == Role.USER.value:
+                elements += self.format_user.apply(content=system_text + message["content"])
+            elif message["role"] == Role.ASSISTANT.value:
+                elements += self.format_assistant.apply(content=message["content"])
+            elif message["role"] == Role.OBSERVATION.value:
+                elements += self.format_observation.apply(content=message["content"])
+            elif message["role"] == Role.FUNCTION.value:
+                elements += self.format_function.apply(content=message["content"])
+            else:
+                raise NotImplementedError("Unexpected role: {}".format(message["role"]))
+
+            encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
+
+        return self._make_pairs(encoded_messages, cutoff_len, reserved_label_len)
+
+
+templates: Dict[str, Template] = {}
+
+
+def _register_template(
+    name: str,
+    format_user: Optional["Formatter"] = None,
+    format_assistant: Optional["Formatter"] = None,
+    format_system: Optional["Formatter"] = None,
+    format_function: Optional["Formatter"] = None,
+    format_observation: Optional["Formatter"] = None,
+    format_tools: Optional["Formatter"] = None,
+    format_separator: Optional["Formatter"] = None,
+    default_system: str = "",
+    stop_words: List[str] = [],
+    efficient_eos: bool = False,
+    replace_eos: bool = False,
+    force_system: bool = False,
+) -> None:
+    r"""
+    Registers a chat template.
+
+    To add the following chat template:
+    ```
+    [HUMAN]:
+    user prompt here
+    [AI]:
+    model response here
+
+    [HUMAN]:
+    user prompt here
+    [AI]:
+    model response here
+    ```
+
+    The corresponding code should be:
+    ```
+    _register_template(
+        name="custom",
+        format_user=StringFormatter(slots=["[HUMAN]:\n{{content}}\n[AI]:\n"]),
+        format_separator=EmptyFormatter(slots=["\n\n"]),
+        efficient_eos=True,
+    )
+    ```
+    """
+    eos_slots = [] if efficient_eos else [{"eos_token"}]
+    template_class = Llama2Template if name.startswith("llama2") else Template
+    default_user_formatter = StringFormatter(slots=["{{content}}"])
+    default_assistant_formatter = StringFormatter(slots=["{{content}}"] + eos_slots)
+    default_function_formatter = FunctionFormatter(slots=["Action: {{name}}\nAction Input: {{arguments}}"] + eos_slots)
+    default_tool_formatter = ToolFormatter(tool_format="default")
+    default_separator_formatter = EmptyFormatter()
+    templates[name] = template_class(
+        format_user=format_user or default_user_formatter,
+        format_assistant=format_assistant or default_assistant_formatter,
+        format_system=format_system or default_user_formatter,
+        format_function=format_function or default_function_formatter,
+        format_observation=format_observation or format_user or default_user_formatter,
+        format_tools=format_tools or default_tool_formatter,
+        format_separator=format_separator or default_separator_formatter,
+        default_system=default_system,
+        stop_words=stop_words,
+        efficient_eos=efficient_eos,
+        replace_eos=replace_eos,
+        force_system=force_system,
+    )
+
+
+def _add_or_replace_eos_token(tokenizer: "PreTrainedTokenizer", eos_token: str) -> None:
+    is_added = tokenizer.eos_token_id is None
+    num_added_tokens = tokenizer.add_special_tokens({"eos_token": eos_token})
+
+    if is_added:
+        logger.info("Add eos token: {}".format(tokenizer.eos_token))
+    else:
+        logger.info("Replace eos token: {}".format(tokenizer.eos_token))
+
+    if num_added_tokens > 0:
+        logger.warning("New tokens have been added, make sure `resize_vocab` is True.")
+
+
+def _jinja_escape(content: str) -> str:
+    return content.replace("\n", r"\n").replace("'", r"\'")
+
+
+def _convert_slots_to_jinja(slots: "SLOTS", tokenizer: "PreTrainedTokenizer", placeholder: str = "content") -> str:
+    slot_items = []
+    for slot in slots:
+        if isinstance(slot, str):
+            slot_pieces = slot.split("{{content}}")
+            if slot_pieces[0]:
+                slot_items.append("'" + _jinja_escape(slot_pieces[0]) + "'")
+            if len(slot_pieces) > 1:
+                slot_items.append(placeholder)
+                if slot_pieces[1]:
+                    slot_items.append("'" + _jinja_escape(slot_pieces[1]) + "'")
+        elif isinstance(slot, set):
+            if "bos_token" in slot:
+                slot_items.append("'" + tokenizer.bos_token + "'")
+            elif "eos_token" in slot:  # do not use {{ eos_token }} since it may be replaced
+                slot_items.append("'" + tokenizer.eos_token + "'")
+        elif isinstance(slot, dict):
+            raise ValueError("Dict is not supported.")
+
+    return " + ".join(slot_items)
+
+
+def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer") -> str:
+    jinja_template = ""
+
+    if template.default_system:
+        jinja_template += "{% set system_message = '" + _jinja_escape(template.default_system) + "' %}"
+
+    jinja_template += (
+        "{% if messages[0]['role'] == 'system' %}" "{% set system_message = messages[0]['content'] %}" "{% endif %}"
+    )
+
+    system_message = _convert_slots_to_jinja(template.format_system.apply(), tokenizer, placeholder="system_message")
+    if isinstance(template, Llama2Template):
+        pass
+    elif template.force_system:
+        jinja_template += "{{ " + system_message + " }}"
+    else:
+        jinja_template += "{% if system_message is defined %}{{ " + system_message + " }}{% endif %}"
+
+    jinja_template += "{% for message in messages %}"
+    jinja_template += "{% set content = message['content'] %}"
+    if isinstance(template, Llama2Template):
+        jinja_template += "{% if loop.index0 == 0 and system_message is defined %}"
+        jinja_template += "{% set content = " + system_message + " + message['content'] %}"
+        jinja_template += "{% endif %}"
+    jinja_template += "{% if message['role'] == 'user' %}"
+    user_message = _convert_slots_to_jinja(template.format_user.apply(), tokenizer)
+    jinja_template += "{{ " + user_message + " }}"
+    jinja_template += "{% elif message['role'] == 'assistant' %}"
+    assistant_message = _convert_slots_to_jinja(
+        template.format_assistant.apply() + template.format_separator.apply(), tokenizer
+    )
+    jinja_template += "{{ " + assistant_message + " }}"
+    jinja_template += "{% endif %}"
+    jinja_template += "{% endfor %}"
+    return jinja_template
+
+
+def get_template_and_fix_tokenizer(
+    tokenizer: "PreTrainedTokenizer",
+    name: Optional[str] = None,
+) -> Template:
+    if name is None:
+        template = templates["vanilla"]  # placeholder
+    else:
+        template = templates.get(name, None)
+        if template is None:
+            raise ValueError("Template {} does not exist.".format(name))
+
+    stop_words = template.stop_words
+    if template.replace_eos:
+        if not stop_words:
+            raise ValueError("Stop words are required to replace the EOS token.")
+
+        _add_or_replace_eos_token(tokenizer, eos_token=stop_words[0])
+        stop_words = stop_words[1:]
+
+    if tokenizer.eos_token_id is None:
+        _add_or_replace_eos_token(tokenizer, eos_token="<|endoftext|>")
+
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+        logger.info("Add pad token: {}".format(tokenizer.pad_token))
+
+    if stop_words:
+        num_added_tokens = tokenizer.add_special_tokens(
+            dict(additional_special_tokens=stop_words), replace_additional_special_tokens=False
+        )
+        logger.info("Add {} to stop words.".format(",".join(stop_words)))
+        if num_added_tokens > 0:
+            logger.warning("New tokens have been added, make sure `resize_vocab` is True.")
+
+    try:
+        tokenizer.chat_template = _get_jinja_template(template, tokenizer)
+    except ValueError:
+        logger.info("Cannot add this chat template to tokenizer.")
+
+    return template
+
+
+_register_template(
+    name="alpaca",
+    format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n\n### Response:\n"]),
+    format_separator=EmptyFormatter(slots=["\n\n"]),
+    default_system=(
+        "Below is an instruction that describes a task. " "Write a response that appropriately completes the request."
+    ),
+)
+
+
+_register_template(
+    name="aquila",
+    format_user=StringFormatter(slots=["Human: {{content}}###Assistant:"]),
+    format_separator=EmptyFormatter(slots=["###"]),
+    default_system=(
+        "A chat between a curious human and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the human's questions."
+    ),
+    stop_words=["</s>"],
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="atom",
+    format_user=StringFormatter(
+        slots=[{"bos_token"}, "Human: {{content}}\n", {"eos_token"}, {"bos_token"}, "Assistant:"]
+    ),
+    format_assistant=StringFormatter(slots=["{{content}}\n", {"eos_token"}]),
+)
+
+
+_register_template(
+    name="baichuan",
+    format_user=StringFormatter(slots=["<reserved_102>{{content}}<reserved_103>"]),
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="baichuan2",
+    format_user=StringFormatter(slots=["<reserved_106>{{content}}<reserved_107>"]),
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="belle",
+    format_user=StringFormatter(slots=["Human: {{content}}\n\nBelle: "]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    format_separator=EmptyFormatter(slots=["\n\n"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="bluelm",
+    format_user=StringFormatter(slots=[{"token": "[|Human|]:"}, "{{content}}", {"token": "[|AI|]:"}]),
+)
+
+
+_register_template(
+    name="chatglm2",
+    format_user=StringFormatter(slots=["[Round {{idx}}]\n\n问：{{content}}\n\n答："]),
+    format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, "{{content}}"]),
+    format_separator=EmptyFormatter(slots=["\n\n"]),
+    efficient_eos=True,
+    force_system=True,
+)
+
+
+_register_template(
+    name="chatglm3",
+    format_user=StringFormatter(slots=[{"token": "<|user|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]),
+    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
+    format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, "{{content}}"]),
+    format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
+    format_observation=StringFormatter(
+        slots=[{"token": "<|observation|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]
+    ),
+    stop_words=["<|user|>", "<|observation|>"],
+    efficient_eos=True,
+    force_system=True,
+)
+
+
+_register_template(
+    name="chatglm3_system",
+    format_user=StringFormatter(slots=[{"token": "<|user|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]),
+    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
+    format_system=StringFormatter(
+        slots=[{"token": "[gMASK]"}, {"token": "sop"}, {"token": "<|system|>"}, "\n", "{{content}}"]
+    ),
+    format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
+    format_observation=StringFormatter(
+        slots=[{"token": "<|observation|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]
+    ),
+    default_system=(
+        "You are ChatGLM3, a large language model trained by Zhipu.AI. "
+        "Follow the user's instructions carefully. Respond using markdown."
+    ),
+    stop_words=["<|user|>", "<|observation|>"],
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="chatml",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    stop_words=["<|im_end|>", "<|im_start|>"],
+    replace_eos=True,
+)
+
+
+_register_template(
+    name="chatml_de",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system="Du bist ein freundlicher und hilfsbereiter KI-Assistent.",
+    stop_words=["<|im_end|>", "<|im_start|>"],
+    replace_eos=True,
+)
+
+
+_register_template(
+    name="codegeex2",
+    format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="cpm",
+    format_user=StringFormatter(slots=["<用户>{{content}}<AI>"]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="deepseek",
+    format_user=StringFormatter(slots=["User: {{content}}\n\nAssistant:"]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="deepseekcoder",
+    format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n### Response:"]),
+    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
+    format_separator=EmptyFormatter(slots=["\n<|EOT|>\n"]),
+    default_system=(
+        "You are an AI programming assistant, utilizing the Deepseek Coder model, "
+        "developed by Deepseek Company, and you only answer questions related to computer science. "
+        "For politically sensitive questions, security and privacy issues, "
+        "and other non-computer science questions, you will refuse to answer\n"
+    ),
+    stop_words=["<|EOT|>"],
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="default",
+    format_user=StringFormatter(slots=["Human: {{content}}\nAssistant: "]),
+    format_system=StringFormatter(slots=["{{content}}\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+)
+
+
+_register_template(
+    name="falcon",
+    format_user=StringFormatter(slots=["User: {{content}}\nFalcon:"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="gemma",
+    format_user=StringFormatter(slots=["<start_of_turn>user\n{{content}}<end_of_turn>\n<start_of_turn>model\n"]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    format_separator=EmptyFormatter(slots=["<end_of_turn>\n"]),
+    efficient_eos=True,
+    force_system=True,
+)
+
+
+_register_template(
+    name="intern",
+    format_user=StringFormatter(slots=["<|User|>:{{content}}", {"token": "<eoh>"}, "\n<|Bot|>:"]),
+    format_separator=EmptyFormatter(slots=[{"token": "<eoa>"}, "\n"]),
+    stop_words=["<eoa>"],
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="intern2",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system=(
+        "You are an AI assistant whose name is InternLM (书生·浦语).\n"
+        "- InternLM (书生·浦语) is a conversational language model that is developed "
+        "by Shanghai AI Laboratory (上海人工智能实验室). It is designed to be helpful, honest, and harmless.\n"
+        "- InternLM (书生·浦语) can understand and communicate fluently in the language chosen "
+        "by the user such as English and 中文."
+    ),
+    stop_words=["<|im_end|>"],
+    efficient_eos=True,  # internlm2 tokenizer cannot set eos_token_id
+)
+
+
+_register_template(
+    name="llama2",
+    format_user=StringFormatter(slots=[{"bos_token"}, "[INST] {{content}} [/INST]"]),
+    format_system=StringFormatter(slots=["<<SYS>>\n{{content}}\n<</SYS>>\n\n"]),
+    default_system=(
+        "You are a helpful, respectful and honest assistant. "
+        "Always answer as helpfully as possible, while being safe. "
+        "Your answers should not include any harmful, unethical, "
+        "racist, sexist, toxic, dangerous, or illegal content. "
+        "Please ensure that your responses are socially unbiased and positive in nature.\n\n"
+        "If a question does not make any sense, or is not factually coherent, "
+        "explain why instead of answering something not correct. "
+        "If you don't know the answer to a question, please don't share false information."
+    ),
+)
+
+
+_register_template(
+    name="llama2_zh",
+    format_user=StringFormatter(slots=[{"bos_token"}, "[INST] {{content}} [/INST]"]),
+    format_system=StringFormatter(slots=["<<SYS>>\n{{content}}\n<</SYS>>\n\n"]),
+    default_system="You are a helpful assistant. 你是一个乐于助人的助手。",
+)
+
+
+_register_template(
+    name="mistral",
+    format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="olmo",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>"]),
+    format_assistant=StringFormatter(slots=["{{content}}", {"eos_token"}]),
+    format_system=StringFormatter(slots=[{"eos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="openchat",
+    format_user=StringFormatter(slots=["GPT4 Correct User: {{content}}", {"eos_token"}, "GPT4 Correct Assistant:"]),
+    format_assistant=StringFormatter(slots=["{{content}}", {"eos_token"}]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="orion",
+    format_user=StringFormatter(slots=["Human: {{content}}\n\nAssistant: ", {"eos_token"}]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
+    name="qwen",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system="You are a helpful assistant.",
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+)
+
+
+_register_template(
+    name="solar",
+    format_user=StringFormatter(slots=["### User:\n{{content}}\n\n### Assistant:\n"]),
+    format_system=StringFormatter(slots=["### System:\n{{content}}\n\n"]),
+    efficient_eos=True,
+)
+
+
+_register_template(
+    name="starchat",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|end|>\n<|assistant|>"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}<|end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    stop_words=["<|end|>"],
+    replace_eos=True,
+    force_system=True,
+)
+
+
+_register_template(
+    name="vanilla",
+)
+
+
+_register_template(
+    name="vicuna",
+    format_user=StringFormatter(slots=["USER: {{content}} ASSISTANT:"]),
+    default_system=(
+        "A chat between a curious user and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the user's questions."
+    ),
+)
+
+
+_register_template(
+    name="xuanyuan",
+    format_user=StringFormatter(slots=["Human: {{content}} Assistant:"]),
+    default_system=(
+        "以下是用户和人工智能助手之间的对话。用户以Human开头，人工智能助手以Assistant开头，"
+        "会对人类提出的问题给出有帮助、高质量、详细和礼貌的回答，并且总是拒绝参与与不道德、"
+        "不安全、有争议、政治敏感等相关的话题、问题和指示。\n"
+    ),
+)
+
+
+_register_template(
+    name="xverse",
+    format_user=StringFormatter(slots=["Human: {{content}}\n\nAssistant: "]),
+)
+
+
+_register_template(
+    name="yayi",
+    format_user=StringFormatter(slots=[{"token": "<|Human|>"}, ":\n{{content}}\n\n", {"token": "<|YaYi|>"}, ":"]),
+    format_system=StringFormatter(slots=[{"token": "<|System|>"}, ":\n{{content}}\n\n"]),
+    format_separator=EmptyFormatter(slots=["\n\n"]),
+    default_system=(
+        "You are a helpful, respectful and honest assistant named YaYi "
+        "developed by Beijing Wenge Technology Co.,Ltd. "
+        "Always answer as helpfully as possible, while being safe.  "
+        "Your answers should not include any harmful, unethical, "
+        "racist, sexist, toxic, dangerous, or illegal content. "
+        "Please ensure that your responses are socially unbiased and positive in nature.\n\n"
+        "If a question does not make any sense, or is not factually coherent, "
+        "explain why instead of answering something not correct. "
+        "If you don't know the answer to a question, please don't share false information."
+    ),
+    stop_words=["<|End|>"],
+)
+
+
+_register_template(
+    name="yi",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+)
+
+
+_register_template(
+    name="yuan",
+    format_user=StringFormatter(slots=["{{content}}", {"token": "<sep>"}]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    stop_words=["<eod>"],
+    replace_eos=True,
+)
+
+
+_register_template(
+    name="zephyr",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}", {"eos_token"}, "<|assistant|>"]),
+    format_assistant=StringFormatter(slots=["\n{{content}}", {"eos_token"}]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}", {"eos_token"}]),
+    default_system="You are a friendly chatbot who always responds in the style of a pirate",
+)
+
+
+_register_template(
+    name="ziya",
+    format_user=StringFormatter(slots=["<human>:{{content}}\n<bot>:"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+)

src/llmtuner/data/utils.py

@@ -0,0 +1,94 @@
+import hashlib
+from enum import Enum, unique
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+from datasets import concatenate_datasets, interleave_datasets
+
+from ..extras.logging import get_logger
+
+
+if TYPE_CHECKING:
+    from datasets import Dataset, IterableDataset
+    from transformers import Seq2SeqTrainingArguments
+
+    from llmtuner.hparams import DataArguments
+
+
+logger = get_logger(__name__)
+
+
+@unique
+class Role(str, Enum):
+    USER = "user"
+    ASSISTANT = "assistant"
+    SYSTEM = "system"
+    FUNCTION = "function"
+    OBSERVATION = "observation"
+
+
+def checksum(data_files: List[str], file_sha1: Optional[str] = None) -> None:
+    if file_sha1 is None:
+        logger.warning("Checksum failed: missing SHA-1 hash value in dataset_info.json.")
+        return
+
+    if len(data_files) != 1:
+        logger.warning("Checksum failed: too many files.")
+        return
+
+    with open(data_files[0], "rb") as f:
+        sha1 = hashlib.sha1(f.read()).hexdigest()
+        if sha1 != file_sha1:
+            logger.warning("Checksum failed: mismatched SHA-1 hash value at {}.".format(data_files[0]))
+
+
+def infer_max_len(source_len: int, target_len: int, max_len: int, reserved_label_len: int) -> Tuple[int, int]:
+    max_target_len = int(max_len * (target_len / (source_len + target_len)))
+    max_target_len = max(max_target_len, reserved_label_len)
+    max_source_len = max_len - max_target_len
+    return max_source_len, max_target_len
+
+
+def merge_dataset(
+    all_datasets: List[Union["Dataset", "IterableDataset"]],
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+) -> Union["Dataset", "IterableDataset"]:
+    if len(all_datasets) == 1:
+        return all_datasets[0]
+    elif data_args.mix_strategy == "concat":
+        if data_args.streaming:
+            logger.warning("The samples between different datasets will not be mixed in streaming mode.")
+        return concatenate_datasets(all_datasets)
+    elif data_args.mix_strategy.startswith("interleave"):
+        if not data_args.streaming:
+            logger.warning("We recommend using `mix_strategy=concat` in non-streaming mode.")
+        return interleave_datasets(
+            datasets=all_datasets,
+            probabilities=data_args.interleave_probs,
+            seed=training_args.seed,
+            stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
+        )
+    else:
+        raise ValueError("Unknown mixing strategy.")
+
+
+def split_dataset(
+    dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", training_args: "Seq2SeqTrainingArguments"
+) -> Dict[str, "Dataset"]:
+    if training_args.do_train:
+        if data_args.val_size > 1e-6:  # Split the dataset
+            if data_args.streaming:
+                val_set = dataset.take(int(data_args.val_size))
+                train_set = dataset.skip(int(data_args.val_size))
+                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
+                return {"train_dataset": train_set, "eval_dataset": val_set}
+            else:
+                val_size = int(data_args.val_size) if data_args.val_size > 1 else data_args.val_size
+                dataset = dataset.train_test_split(test_size=val_size, seed=training_args.seed)
+                return {"train_dataset": dataset["train"], "eval_dataset": dataset["test"]}
+        else:
+            if data_args.streaming:
+                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
+            return {"train_dataset": dataset}
+    else:  # do_eval or do_predict
+        return {"eval_dataset": dataset}

src/llmtuner/dsets/__init__.py

@@ -1,3 +0,0 @@
-from llmtuner.dsets.loader import get_dataset
-from llmtuner.dsets.preprocess import preprocess_dataset
-from llmtuner.dsets.utils import split_dataset

src/llmtuner/dsets/loader.py

@@ -1,145 +0,0 @@
-import os
-from typing import TYPE_CHECKING, Any, Dict, List, Union
-
-from datasets import concatenate_datasets, interleave_datasets, load_dataset
-
-from llmtuner.dsets.utils import checksum, EXT2TYPE
-from llmtuner.extras.logging import get_logger
-
-if TYPE_CHECKING:
-    from datasets import Dataset, IterableDataset
-    from llmtuner.hparams import ModelArguments, DataArguments
-
-
-logger = get_logger(__name__)
-
-
-def get_dataset(
-    model_args: "ModelArguments",
-    data_args: "DataArguments"
-) -> Union["Dataset", "IterableDataset"]:
-    max_samples = data_args.max_samples
-    all_datasets: List[Union["Dataset", "IterableDataset"]] = [] # support multiple datasets
-
-    for dataset_attr in data_args.dataset_list:
-        logger.info("Loading dataset {}...".format(dataset_attr))
-
-        if dataset_attr.load_from == "hf_hub":
-            data_path = dataset_attr.dataset_name
-            data_name = dataset_attr.subset
-            data_files = None
-        elif dataset_attr.load_from == "script":
-            data_path = os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)
-            data_name = dataset_attr.subset
-            data_files = None
-        elif dataset_attr.load_from == "file":
-            data_path, data_name = None, None
-            data_files: List[str] = []
-            if os.path.isdir(os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)): # is directory
-                for file_name in os.listdir(os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)):
-                    data_files.append(os.path.join(data_args.dataset_dir, dataset_attr.dataset_name, file_name))
-                    if data_path is None:
-                        data_path = EXT2TYPE.get(file_name.split(".")[-1], None)
-                    else:
-                        assert data_path == EXT2TYPE.get(file_name.split(".")[-1], None), "file types are not identical."
-            elif os.path.isfile(os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)): # is file
-                data_files.append(os.path.join(data_args.dataset_dir, dataset_attr.dataset_name))
-                data_path = EXT2TYPE.get(dataset_attr.dataset_name.split(".")[-1], None)
-            else:
-                raise ValueError("File not found.")
-
-            assert data_path, "File extension must be txt, csv, json or jsonl."
-            checksum(data_files, dataset_attr.dataset_sha1)
-        else:
-            raise NotImplementedError
-
-        dataset = load_dataset(
-            path=data_path,
-            name=data_name,
-            data_files=data_files,
-            split=data_args.split,
-            cache_dir=model_args.cache_dir,
-            token=model_args.hf_hub_token,
-            streaming=data_args.streaming
-        )
-
-        if max_samples is not None: # truncate dataset
-            dataset = dataset.select(range(min(len(dataset), max_samples)))
-
-        def convert_format(examples: Dict[str, List[Any]]) -> Dict[str, List[Any]]:
-            # convert dataset from sharegpt format to alpaca format
-            outputs = {"prompt": [], "query": [], "response": [], "history": []}
-            for msg_list in examples[dataset_attr.messages]:
-                msg_list = msg_list[:len(msg_list) // 2 * 2] # should be multiples of 2
-                if len(msg_list) == 0:
-                    continue
-
-                msg_pairs = []
-                user_role, assistant_role = None, None
-                for idx in range(0, len(msg_list), 2):
-                    if user_role is None and assistant_role is None:
-                        user_role = msg_list[idx][dataset_attr.role]
-                        assistant_role = msg_list[idx + 1][dataset_attr.role]
-                    else:
-                        if (
-                            msg_list[idx][dataset_attr.role] != user_role
-                            or msg_list[idx+1][dataset_attr.role] != assistant_role
-                        ):
-                            raise ValueError("Only accepts conversation in u/a/u/a/u/a order.")
-                    msg_pairs.append((msg_list[idx][dataset_attr.content], msg_list[idx + 1][dataset_attr.content]))
-
-                if len(msg_pairs) != 0:
-                    outputs["prompt"].append(msg_pairs[-1][0])
-                    outputs["query"].append("")
-                    outputs["response"].append(msg_pairs[-1][1])
-                    outputs["history"].append(msg_pairs[:-1])
-
-            return outputs
-
-        if dataset_attr.formatting == "sharegpt": # convert format
-            column_names = list(next(iter(dataset)).keys())
-            kwargs = {}
-            if not data_args.streaming:
-                kwargs = dict(
-                    num_proc=data_args.preprocessing_num_workers,
-                    load_from_cache_file=(not data_args.overwrite_cache),
-                    desc="Converting format of dataset"
-                )
-
-            dataset = dataset.map(
-                convert_format,
-                batched=True,
-                remove_columns=column_names,
-                **kwargs
-            )
-        else:
-            for column_name in ["prompt", "query", "response", "history"]: # align dataset
-                if getattr(dataset_attr, column_name) and getattr(dataset_attr, column_name) != column_name:
-                    dataset = dataset.rename_column(getattr(dataset_attr, column_name), column_name)
-
-        if dataset_attr.system_prompt: # add system prompt
-            system_prompt = dataset_attr.system_prompt
-            if data_args.streaming:
-                dataset = dataset.map(lambda _: {"system": system_prompt})
-            else:
-                dataset = dataset.add_column("system", [system_prompt] * len(dataset))
-
-        all_datasets.append(dataset)
-
-    if len(data_args.dataset_list) == 1:
-        return all_datasets[0]
-    elif data_args.mix_strategy == "concat":
-        if data_args.streaming:
-            logger.warning("The samples between different datasets will not be mixed in streaming mode.")
-        return concatenate_datasets(all_datasets)
-    elif data_args.mix_strategy.startswith("interleave"):
-        if not data_args.streaming:
-            logger.warning("We recommend using `mix_strategy=concat` in non-streaming mode.")
-        return interleave_datasets(
-            datasets=all_datasets,
-            probabilities=data_args.interleave_probs,
-            seed=data_args.seed,
-            stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted"
-        )
-    else:
-        raise ValueError("Unknown mixing strategy.")

src/llmtuner/dsets/preprocess.py

@@ -1,272 +0,0 @@
-import os
-import tiktoken
-from itertools import chain
-from typing import TYPE_CHECKING, Any, Dict, Generator, List, Literal, Union
-
-from datasets import load_from_disk
-
-from llmtuner.extras.constants import IGNORE_INDEX
-from llmtuner.extras.logging import get_logger
-from llmtuner.extras.template import get_template_and_fix_tokenizer
-
-if TYPE_CHECKING:
-    from datasets import Dataset, IterableDataset
-    from transformers import Seq2SeqTrainingArguments
-    from transformers.tokenization_utils import PreTrainedTokenizer
-    from llmtuner.hparams import DataArguments
-
-
-logger = get_logger(__name__)
-
-
-def preprocess_dataset(
-    dataset: Union["Dataset", "IterableDataset"],
-    tokenizer: "PreTrainedTokenizer",
-    data_args: "DataArguments",
-    training_args: "Seq2SeqTrainingArguments",
-    stage: Literal["pt", "sft", "rm", "ppo"]
-) -> Union["Dataset", "IterableDataset"]:
-    template = get_template_and_fix_tokenizer(data_args.template, tokenizer)
-
-    if data_args.train_on_prompt and template.efficient_eos:
-        raise ValueError("Current template does not support `train_on_prompt`.")
-
-    def construct_example(examples: Dict[str, List[Any]]) -> Generator[Any, None, None]:
-        for i in range(len(examples["prompt"])):
-            query, response = examples["prompt"][i], examples["response"][i]
-            query = query + "\n" + examples["query"][i] if "query" in examples and examples["query"][i] else query
-            history = examples["history"][i] if "history" in examples else None
-            system = examples["system"][i] if "system" in examples else None
-            yield query, response, history, system
-
-    def preprocess_pretrain_dataset(examples: Dict[str, List[Any]]) -> Dict[str, List[List[int]]]:
-        # build grouped texts with format `X1 X2 X3 ...`
-        if isinstance(getattr(tokenizer, "tokenizer", None), tiktoken.Encoding): # for tiktoken tokenizer (Qwen)
-            kwargs = dict(allowed_special="all")
-        else:
-            kwargs = dict(add_special_tokens=True)
-
-        if hasattr(tokenizer, "add_eos_token"): # for LLaMA tokenizer
-            add_eos_token_flag = getattr(tokenizer, "add_eos_token")
-            setattr(tokenizer, "add_eos_token", True)
-
-        tokenized_examples = tokenizer(examples["prompt"], **kwargs)
-        concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}
-        total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
-        block_size = data_args.cutoff_len
-        # we drop the small remainder, and if the total_length < block_size, we exclude this batch
-        total_length = (total_length // block_size) * block_size
-        # split by chunks of cutoff_len
-        result = {
-            k: [t[i: i + block_size] for i in range(0, total_length, block_size)]
-            for k, t in concatenated_examples.items()
-        }
-        # make sure the saved tokenizer is the same as the original one
-        if hasattr(tokenizer, "add_eos_token"):
-            setattr(tokenizer, "add_eos_token", add_eos_token_flag)
-        return result
-
-    def preprocess_supervised_dataset(examples: Dict[str, List[Any]]) -> Dict[str, List[List[int]]]:
-        # build inputs with format `<bos> X Y <eos>` and labels with format `<ignore> ... <ignore> Y <eos>`
-        # for multiturn examples, we only mask the prompt part in each prompt-response pair.
-        model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
-
-        for query, response, history, system in construct_example(examples):
-            if not (isinstance(query, str) and isinstance(response, str) and query != "" and response != ""):
-                continue
-
-            input_ids, labels = [], []
-            for turn_idx, (source_ids, target_ids) in enumerate(template.encode_multiturn(
-                tokenizer, query, response, history, system
-            )):
-                total_len = len(source_ids) + len(target_ids)
-                max_source_len = int(data_args.cutoff_len * (len(source_ids) / total_len))
-                max_target_len = int(data_args.cutoff_len * (len(target_ids) / total_len))
-
-                if len(source_ids) > max_source_len:
-                    source_ids = source_ids[:max_source_len]
-                if len(target_ids) > max_target_len:
-                    target_ids = target_ids[:max_target_len]
-
-                if data_args.train_on_prompt:
-                    source_mask = source_ids
-                elif turn_idx != 0 and template.efficient_eos:
-                    source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
-                else:
-                    source_mask = [IGNORE_INDEX] * len(source_ids)
-
-                input_ids += source_ids + target_ids
-                labels += source_mask + target_ids
-
-            if template.efficient_eos:
-                input_ids += [tokenizer.eos_token_id]
-                labels += [tokenizer.eos_token_id]
-
-            if len(input_ids) > data_args.cutoff_len:
-                input_ids = input_ids[:data_args.cutoff_len]
-                labels = labels[:data_args.cutoff_len]
-
-            model_inputs["input_ids"].append(input_ids)
-            model_inputs["attention_mask"].append([1] * len(input_ids))
-            model_inputs["labels"].append(labels)
-
-        return model_inputs
-
-    def preprocess_packed_supervised_dataset(examples: Dict[str, List[Any]]) -> Dict[str, List[List[int]]]:
-        # build inputs with format `<bos> X1 Y1 <eos> <bos> X2 Y2 <eos>`
-        # and labels with format `<ignore> ... <ignore> Y1 <eos> <ignore> ... <ignore> Y2 <eos>`
-        model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
-        input_ids, labels = [], []
-        for query, response, history, system in construct_example(examples):
-            if not (isinstance(query, str) and isinstance(response, str) and query != "" and response != ""):
-                continue
-
-            for turn_idx, (source_ids, target_ids) in enumerate(template.encode_multiturn(
-                tokenizer, query, response, history, system
-            )):
-                if data_args.train_on_prompt:
-                    source_mask = source_ids
-                elif turn_idx != 0 and template.efficient_eos:
-                    source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
-                else:
-                    source_mask = [IGNORE_INDEX] * len(source_ids)
-                input_ids += source_ids + target_ids
-                labels += source_mask + target_ids
-
-        if template.efficient_eos:
-            input_ids += [tokenizer.eos_token_id]
-            labels += [tokenizer.eos_token_id]
-
-        total_length = len(input_ids)
-        block_size = data_args.cutoff_len
-        # we drop the small remainder, and if the total_length < block_size, we exclude this batch
-        total_length = (total_length // block_size) * block_size
-        # split by chunks of cutoff_len
-        for i in range(0, total_length, block_size):
-            model_inputs["input_ids"].append(input_ids[i: i + block_size])
-            model_inputs["attention_mask"].append([1] * block_size)
-            model_inputs["labels"].append(labels[i: i + block_size])
-
-        return model_inputs
-
-    def preprocess_unsupervised_dataset(examples: Dict[str, List[Any]]) -> Dict[str, List[List[int]]]:
-        # build inputs with format `<bos> X` and labels with format `Y <eos>`
-        model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
-
-        for query, response, history, system in construct_example(examples):
-            if not (isinstance(query, str) and query != ""):
-                continue
-
-            input_ids, labels = template.encode_oneturn(tokenizer, query, response, history, system)
-
-            if template.efficient_eos:
-                labels += [tokenizer.eos_token_id]
-
-            if len(input_ids) > data_args.cutoff_len:
-                input_ids = input_ids[:data_args.cutoff_len]
-            if len(labels) > data_args.cutoff_len:
-                labels = labels[:data_args.cutoff_len]
-
-            model_inputs["input_ids"].append(input_ids)
-            model_inputs["attention_mask"].append([1] * len(input_ids))
-            model_inputs["labels"].append(labels)
-
-        return model_inputs
-
-    def preprocess_pairwise_dataset(examples: Dict[str, List[Any]]) -> Dict[str, List[List[int]]]:
-        # build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
-        model_inputs = {"prompt_ids": [], "chosen_ids": [], "rejected_ids": []}
-        for query, response, history, system in construct_example(examples):
-            if not (isinstance(query, str) and isinstance(response, list) and query != "" and len(response) > 1):
-                continue
-
-            prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, query, response[0], history, system)
-            _, rejected_ids = template.encode_oneturn(tokenizer, query, response[1], history, system)
-
-            if template.efficient_eos:
-                chosen_ids += [tokenizer.eos_token_id]
-                rejected_ids += [tokenizer.eos_token_id]
-
-            total_len = len(prompt_ids) + max(len(chosen_ids), len(rejected_ids))
-            max_source_len = int(data_args.cutoff_len * (len(prompt_ids) / total_len))
-            max_target_len = int(data_args.cutoff_len * (max(len(chosen_ids), len(rejected_ids)) / total_len))
-
-            if len(prompt_ids) > max_source_len:
-                prompt_ids = prompt_ids[:max_source_len]
-            if len(chosen_ids) > max_target_len:
-                chosen_ids = chosen_ids[:max_target_len]
-            if len(rejected_ids) > max_target_len:
-                rejected_ids = rejected_ids[:max_target_len]
-
-            model_inputs["prompt_ids"].append(prompt_ids)
-            model_inputs["chosen_ids"].append(chosen_ids)
-            model_inputs["rejected_ids"].append(rejected_ids)
-
-        return model_inputs
-
-    def print_supervised_dataset_example(example: Dict[str, List[int]]) -> None:
-        print("input_ids:\n{}".format(example["input_ids"]))
-        print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
-        print("label_ids:\n{}".format(example["labels"]))
-        print("labels:\n{}".format(
-            tokenizer.decode(list(filter(lambda x: x != IGNORE_INDEX, example["labels"])), skip_special_tokens=False)
-        ))
-
-    def print_pairwise_dataset_example(example: Dict[str, List[int]]) -> None:
-        print("prompt_ids:\n{}".format(example["prompt_ids"]))
-        print("prompt:\n{}".format(tokenizer.decode(example["prompt_ids"], skip_special_tokens=False)))
-        print("chosen_ids:\n{}".format(example["chosen_ids"]))
-        print("chosen:\n{}".format(tokenizer.decode(example["chosen_ids"], skip_special_tokens=False)))
-        print("rejected_ids:\n{}".format(example["rejected_ids"]))
-        print("rejected:\n{}".format(tokenizer.decode(example["rejected_ids"], skip_special_tokens=False)))
-
-    def print_unsupervised_dataset_example(example: Dict[str, List[int]]) -> None:
-        print("input_ids:\n{}".format(example["input_ids"]))
-        print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
-
-    if stage == "pt":
-        preprocess_func = preprocess_pretrain_dataset
-        print_function = print_unsupervised_dataset_example
-    elif stage == "sft" and not training_args.predict_with_generate:
-        preprocess_func = preprocess_packed_supervised_dataset if data_args.sft_packing else preprocess_supervised_dataset
-        print_function = print_supervised_dataset_example
-    elif stage == "rm":
-        preprocess_func = preprocess_pairwise_dataset
-        print_function = print_pairwise_dataset_example
-    else:
-        preprocess_func = preprocess_unsupervised_dataset
-        print_function = print_unsupervised_dataset_example
-
-    if data_args.cache_path is not None and os.path.exists(data_args.cache_path):
-        logger.warning("Loading dataset from disk will ignore other data arguments.")
-        return load_from_disk(data_args.cache_path)
-
-    with training_args.main_process_first(desc="dataset map pre-processing"):
-        column_names = list(next(iter(dataset)).keys())
-        kwargs = {}
-        if not data_args.streaming:
-            kwargs = dict(
-                num_proc=data_args.preprocessing_num_workers,
-                load_from_cache_file=(not data_args.overwrite_cache),
-                desc="Running tokenizer on dataset"
-            )
-
-        dataset = dataset.map(
-            preprocess_func,
-            batched=True,
-            remove_columns=column_names,
-            **kwargs
-        )
-
-        if data_args.cache_path is not None and not os.path.exists(data_args.cache_path):
-            if training_args.should_save:
-                dataset.save_to_disk(data_args.cache_path)
-            raise SystemExit("Dataset saved, rerun this script with the same `--cache_path`.")
-
-        if training_args.should_log:
-            try:
-                print_function(next(iter(dataset)))
-            except StopIteration:
-                raise RuntimeError("Empty dataset!")
-
-        return dataset

src/llmtuner/dsets/utils.py

@@ -1,61 +0,0 @@
-import hashlib
-from typing import TYPE_CHECKING, Dict, List, Optional, Union
-
-from llmtuner.extras.logging import get_logger
-
-if TYPE_CHECKING:
-    from datasets import Dataset, IterableDataset
-    from transformers import TrainingArguments
-    from llmtuner.hparams import DataArguments
-
-
-logger = get_logger(__name__)
-
-
-EXT2TYPE = {
-    "arrow": "arrow",
-    "csv": "csv",
-    "json": "json",
-    "jsonl": "json",
-    "parquet": "parquet",
-    "txt": "text"
-}
-
-
-def checksum(data_files: List[str], file_sha1: Optional[str] = None) -> None:
-    if file_sha1 is None:
-        logger.warning("Checksum failed: missing SHA-1 hash value in dataset_info.json.")
-        return
-
-    if len(data_files) != 1:
-        logger.warning("Checksum failed: too many files.")
-        return
-
-    with open(data_files[0], "rb") as f:
-        sha1 = hashlib.sha1(f.read()).hexdigest()
-        if sha1 != file_sha1:
-            logger.warning("Checksum failed: mismatched SHA-1 hash value at {}.".format(data_files[0]))
-
-
-def split_dataset(
-    dataset: Union["Dataset", "IterableDataset"],
-    data_args: "DataArguments",
-    training_args: "TrainingArguments"
-) -> Dict[str, "Dataset"]:
-    if training_args.do_train:
-        if data_args.val_size > 1e-6: # Split the dataset
-            if data_args.streaming:
-                val_set = dataset.take(int(data_args.val_size))
-                train_set = dataset.skip(int(data_args.val_size))
-                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
-                return {"train_dataset": train_set, "eval_dataset": val_set}
-            else:
-                val_size = int(data_args.val_size) if data_args.val_size > 1 else data_args.val_size
-                dataset = dataset.train_test_split(test_size=val_size, seed=training_args.seed)
-                return {"train_dataset": dataset["train"], "eval_dataset": dataset["test"]}
-        else:
-            if data_args.streaming:
-                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
-            return {"train_dataset": dataset}
-    else: # do_eval or do_predict
-        return {"eval_dataset": dataset}

src/llmtuner/eval/__init__.py

@@ -0,0 +1,4 @@
+from .evaluator import Evaluator
+
+
+__all__ = ["Evaluator"]

src/llmtuner/eval/evaluator.py

@@ -0,0 +1,122 @@
+# Inspired by: https://github.com/hendrycks/test/blob/master/evaluate_flan.py
+
+import inspect
+import json
+import os
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+import torch
+from datasets import load_dataset
+from tqdm import tqdm, trange
+from transformers.utils import cached_file
+
+from ..data import get_template_and_fix_tokenizer
+from ..extras.constants import CHOICES, SUBJECTS
+from ..hparams import get_eval_args
+from ..model import load_model_and_tokenizer
+from .template import get_eval_template
+
+
+class Evaluator:
+    def __init__(self, args: Optional[Dict[str, Any]] = None) -> None:
+        self.model_args, self.data_args, self.eval_args, finetuning_args = get_eval_args(args)
+        self.model, self.tokenizer = load_model_and_tokenizer(self.model_args, finetuning_args)
+        self.tokenizer.padding_side = "right"  # avoid overflow issue in batched inference for llama2
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, self.data_args.template)
+        self.eval_template = get_eval_template(self.eval_args.lang)
+        self.choice_inputs = [
+            self.tokenizer.encode(self.eval_template.prefix + ch, add_special_tokens=False)[-1] for ch in CHOICES
+        ]
+
+    @torch.inference_mode()
+    def batch_inference(self, batch_input: Dict[str, torch.Tensor]) -> List[str]:
+        logits = self.model(**batch_input).logits
+        lengths = torch.sum(batch_input["attention_mask"], dim=-1)
+        word_probs = torch.stack([logits[i, lengths[i] - 1] for i in range(len(lengths))], dim=0)
+        choice_probs = torch.nn.functional.softmax(word_probs[:, self.choice_inputs], dim=-1).detach()
+        return [chr(ord("A") + offset.item()) for offset in torch.argmax(choice_probs, dim=-1)]
+
+    def eval(self) -> None:
+        mapping = cached_file(
+            path_or_repo_id=os.path.join(self.eval_args.task_dir, self.eval_args.task),
+            filename="mapping.json",
+            cache_dir=self.model_args.cache_dir,
+            token=self.model_args.hf_hub_token,
+        )
+
+        with open(mapping, "r", encoding="utf-8") as f:
+            categorys: Dict[str, Dict[str, str]] = json.load(f)
+
+        category_corrects = {subj: np.array([], dtype="bool") for subj in SUBJECTS}
+        pbar = tqdm(categorys.keys(), desc="Processing subjects", position=0)
+        results = {}
+        for subject in pbar:
+            if "trust_remote_code" in inspect.signature(load_dataset).parameters:  # for datasets==2.16.0
+                kwargs = {"trust_remote_code": True}
+            else:
+                kwargs = {}
+
+            dataset = load_dataset(
+                path=os.path.join(self.eval_args.task_dir, self.eval_args.task),
+                name=subject,
+                cache_dir=self.model_args.cache_dir,
+                download_mode=self.eval_args.download_mode,
+                token=self.model_args.hf_hub_token,
+                **kwargs,
+            )
+            pbar.set_postfix_str(categorys[subject]["name"])
+            inputs, outputs, labels = [], [], []
+            for i in trange(len(dataset[self.data_args.split]), desc="Formatting batches", position=1, leave=False):
+                support_set = (
+                    dataset["train"].shuffle().select(range(min(self.eval_args.n_shot, len(dataset["train"]))))
+                )
+                messages = self.eval_template.format_example(
+                    target_data=dataset[self.data_args.split][i],
+                    support_set=support_set,
+                    subject_name=categorys[subject]["name"],
+                )
+
+                input_ids, _ = self.template.encode_oneturn(tokenizer=self.tokenizer, messages=messages)
+                inputs.append({"input_ids": input_ids, "attention_mask": [1] * len(input_ids)})
+                labels.append(messages[-1]["content"])
+
+            for i in trange(
+                0, len(inputs), self.eval_args.batch_size, desc="Predicting batches", position=1, leave=False
+            ):
+                batch_input = self.tokenizer.pad(
+                    inputs[i : i + self.eval_args.batch_size], return_attention_mask=True, return_tensors="pt"
+                ).to(self.model.device)
+                preds = self.batch_inference(batch_input)
+                outputs += preds
+
+            corrects = np.array(outputs) == np.array(labels)
+            category_name = categorys[subject]["category"]
+            category_corrects[category_name] = np.concatenate([category_corrects[category_name], corrects], axis=0)
+            category_corrects["Average"] = np.concatenate([category_corrects["Average"], corrects], axis=0)
+            results[subject] = {str(i): outputs[i] for i in range(len(outputs))}
+
+        pbar.close()
+        self._save_results(category_corrects, results)
+
+    def _save_results(self, category_corrects: Dict[str, np.ndarray], results: Dict[str, Dict[int, str]]) -> None:
+        score_info = "\n".join(
+            [
+                "{:>15}: {:.2f}".format(category_name, 100 * np.mean(category_correct))
+                for category_name, category_correct in category_corrects.items()
+                if len(category_correct)
+            ]
+        )
+        print(score_info)
+        if self.eval_args.save_dir is not None:
+            os.makedirs(self.eval_args.save_dir, exist_ok=False)
+            with open(os.path.join(self.eval_args.save_dir, "results.json"), "w", encoding="utf-8", newline="\n") as f:
+                json.dump(results, f, indent=2)
+
+            with open(os.path.join(self.eval_args.save_dir, "results.log"), "w", encoding="utf-8", newline="\n") as f:
+                f.write(score_info)
+
+
+if __name__ == "__main__":
+    evaluator = Evaluator()
+    evaluator.eval()

src/llmtuner/eval/template.py

@@ -0,0 +1,67 @@
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, List, Tuple
+
+from ..data import Role
+from ..extras.constants import CHOICES
+
+
+if TYPE_CHECKING:
+    from datasets import Dataset
+
+
+@dataclass
+class EvalTemplate:
+    system: str
+    choice: str
+    answer: str
+    prefix: str
+
+    def parse_example(self, example: Dict[str, str]) -> Tuple[str, str]:
+        candidates = [self.choice.format(choice=ch, content=example[ch]) for ch in CHOICES if ch in example]
+        return "".join([example["question"]] + candidates + [self.answer]), example["answer"]
+
+    def format_example(
+        self, target_data: Dict[str, str], support_set: "Dataset", subject_name: str
+    ) -> List[Dict[str, str]]:
+        messages = []
+        for k in range(len(support_set)):
+            prompt, response = self.parse_example(support_set[k])
+            messages.append({"role": Role.USER, "content": prompt})
+            messages.append({"role": Role.ASSISTANT, "content": response})
+
+        prompt, response = self.parse_example(target_data)
+        messages.append({"role": Role.USER, "content": prompt})
+        messages.append({"role": Role.ASSISTANT, "content": response})
+        messages[0]["content"] = self.system.format(subject=subject_name) + messages[0]["content"]
+        return messages
+
+
+eval_templates: Dict[str, "EvalTemplate"] = {}
+
+
+def register_eval_template(name: str, system: str, choice: str, answer: str, prefix: str) -> None:
+    eval_templates[name] = EvalTemplate(system=system, choice=choice, answer=answer, prefix=prefix)
+
+
+def get_eval_template(name: str) -> "EvalTemplate":
+    eval_template = eval_templates.get(name, None)
+    assert eval_template is not None, "Template {} does not exist.".format(name)
+    return eval_template
+
+
+register_eval_template(
+    name="en",
+    system="The following are multiple choice questions (with answers) about {subject}.\n\n",
+    choice="\n{choice}. {content}",
+    answer="\nAnswer: ",
+    prefix=" ",
+)
+
+
+register_eval_template(
+    name="zh",
+    system="以下是中国关于{subject}考试的单项选择题，请选出其中的正确答案。\n\n",
+    choice="\n{choice}. {content}",
+    answer="\n答案：",
+    prefix="\n",
+)

src/llmtuner/extras/callbacks.py

@@ -1,46 +1,38 @@
-import os
 import json
+import os
 import time
-from typing import TYPE_CHECKING
 from datetime import timedelta
+from typing import TYPE_CHECKING
 
 from transformers import TrainerCallback
-from transformers.trainer_utils import has_length, PREFIX_CHECKPOINT_DIR
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, has_length
+
+from .constants import LOG_FILE_NAME
+from .logging import get_logger
+from .misc import fix_valuehead_checkpoint
 
-from llmtuner.extras.constants import LOG_FILE_NAME
-from llmtuner.extras.logging import get_logger
 
 if TYPE_CHECKING:
-    from transformers import TrainingArguments, TrainerState, TrainerControl
+    from transformers import TrainerControl, TrainerState, TrainingArguments
 
 
 logger = get_logger(__name__)
 
 
-class SavePeftModelCallback(TrainerCallback):
-
+class FixValueHeadModelCallback(TrainerCallback):
     def on_save(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
         r"""
         Event called after a checkpoint save.
         """
         if args.should_save:
-            output_dir = os.path.join(args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, state.global_step))
-            model = kwargs.pop("model")
-            if getattr(model, "is_peft_model", False):
-                getattr(model, "pretrained_model").save_pretrained(output_dir)
-
-    def on_train_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
-        r"""
-        Event called at the end of training.
-        """
-        if args.should_save:
-            model = kwargs.pop("model")
-            if getattr(model, "is_peft_model", False):
-                getattr(model, "pretrained_model").save_pretrained(args.output_dir)
+            fix_valuehead_checkpoint(
+                model=kwargs.pop("model"),
+                output_dir=os.path.join(args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, state.global_step)),
+                safe_serialization=args.save_safetensors,
+            )
 
 
 class LogCallback(TrainerCallback):
-
     def __init__(self, runner=None):
         self.runner = runner
         self.in_training = False
@@ -106,7 +98,9 @@ class LogCallback(TrainerCallback):
             self.cur_steps = 0
             self.max_steps = 0
 
-    def on_predict(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", *other, **kwargs):
+    def on_predict(
+        self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", *other, **kwargs
+    ):
         r"""
         Event called after a successful prediction.
         """
@@ -132,18 +126,22 @@ class LogCallback(TrainerCallback):
             epoch=state.log_history[-1].get("epoch", None),
             percentage=round(self.cur_steps / self.max_steps * 100, 2) if self.max_steps != 0 else 100,
             elapsed_time=self.elapsed_time,
-            remaining_time=self.remaining_time
+            remaining_time=self.remaining_time,
         )
         if self.runner is not None:
-            logger.info("{{'loss': {:.4f}, 'learning_rate': {:2.4e}, 'epoch': {:.2f}}}".format(
-                logs["loss"] or 0, logs["learning_rate"] or 0, logs["epoch"] or 0
-            ))
+            logger.info(
+                "{{'loss': {:.4f}, 'learning_rate': {:2.4e}, 'epoch': {:.2f}}}".format(
+                    logs["loss"] or 0, logs["learning_rate"] or 0, logs["epoch"] or 0
+                )
+            )
 
         os.makedirs(args.output_dir, exist_ok=True)
         with open(os.path.join(args.output_dir, "trainer_log.jsonl"), "a", encoding="utf-8") as f:
             f.write(json.dumps(logs) + "\n")
 
-    def on_prediction_step(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+    def on_prediction_step(
+        self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs
+    ):
         r"""
         Event called after a prediction step.
         """

src/llmtuner/extras/logging.py

@@ -1,8 +1,11 @@
-import sys
 import logging
+import sys
 
 
 class LoggerHandler(logging.Handler):
+    r"""
+    Logger handler used in Web UI.
+    """
 
     def __init__(self):
         super().__init__()
@@ -19,19 +22,12 @@ class LoggerHandler(logging.Handler):
         self.log += "\n\n"
 
 
-def reset_logging():
-    r"""
-    Removes basic config of root logger
-    """
-    root = logging.getLogger()
-    list(map(root.removeHandler, root.handlers))
-    list(map(root.removeFilter, root.filters))
-
-
 def get_logger(name: str) -> logging.Logger:
+    r"""
+    Gets a standard logger with a stream hander to stdout.
+    """
     formatter = logging.Formatter(
-        fmt="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S"
+        fmt="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S"
     )
     handler = logging.StreamHandler(sys.stdout)
     handler.setFormatter(formatter)
@@ -41,3 +37,12 @@ def get_logger(name: str) -> logging.Logger:
     logger.addHandler(handler)
 
     return logger
+
+
+def reset_logging() -> None:
+    r"""
+    Removes basic config of root logger. (unused in script)
+    """
+    root = logging.getLogger()
+    list(map(root.removeHandler, root.handlers))
+    list(map(root.removeFilter, root.filters))