Merge pull request #6124 from hiyouga/hiyouga/release

[release] release v0.9.1

Former-commit-id: f61cdd99fd282612884c92d36e111ad46b4e0d00

.dockerignore

@@ -4,10 +4,12 @@
 .venv
 cache
 data
+docker
+saves
 hf_cache
+ms_cache
+om_cache
 output
-examples
 .dockerignore
 .gitattributes
 .gitignore
-Dockerfile

.env.local

@@ -0,0 +1,37 @@
+# Note: actually we do not support .env, just for reference
+# api
+API_HOST=
+API_PORT=
+API_KEY=
+API_MODEL_NAME=
+FASTAPI_ROOT_PATH=
+MAX_CONCURRENT=
+# general
+DISABLE_VERSION_CHECK=
+FORCE_CHECK_IMPORTS=
+LLAMAFACTORY_VERBOSITY=
+USE_MODELSCOPE_HUB=
+USE_OPENMIND_HUB=
+RECORD_VRAM=
+# torchrun
+FORCE_TORCHRUN=
+MASTER_ADDR=
+MASTER_PORT=
+NNODES=
+NODE_RANK=
+NPROC_PER_NODE=
+# wandb
+WANDB_DISABLED=
+WANDB_PROJECT=
+WANDB_API_KEY=
+# gradio ui
+GRADIO_SHARE=
+GRADIO_SERVER_NAME=
+GRADIO_SERVER_PORT=
+GRADIO_ROOT_PATH=
+GRADIO_IPV6=
+# setup
+ENABLE_SHORT_CONSOLE=1
+# reserved (do not use)
+LLAMABOARD_ENABLED=
+LLAMABOARD_WORKDIR=

.github/CONTRIBUTING.md

@@ -19,3 +19,49 @@ There are several ways you can contribute to LLaMA Factory:
 ### Style guide
 
 LLaMA Factory follows the [Google Python Style Guide](https://google.github.io/styleguide/pyguide.html), check it for details.
+
+### Create a Pull Request
+
+1. Fork the [repository](https://github.com/hiyouga/LLaMA-Factory) by clicking on the [Fork](https://github.com/hiyouga/LLaMA-Factory/fork) button on the repository's page. This creates a copy of the code under your GitHub user account.
+
+2. Clone your fork to your local disk, and add the base repository as a remote:
+
+```bash
+git clone git@github.com:[username]/LLaMA-Factory.git
+cd LLaMA-Factory
+git remote add upstream https://github.com/hiyouga/LLaMA-Factory.git
+```
+
+3. Create a new branch to hold your development changes:
+
+```bash
+git checkout -b dev_your_branch
+```
+
+4. Set up a development environment by running the following command in a virtual environment:
+
+```bash
+pip install -e ".[dev]"
+```
+
+If LLaMA Factory was already installed in the virtual environment, remove it with `pip uninstall llamafactory` before reinstalling it in editable mode with the -e flag.
+
+5. Check code before commit:
+
+```bash
+make commit
+make style && make quality
+make test
+```
+
+6. Submit changes:
+
+```bash
+git add .
+git commit -m "commit message"
+git fetch upstream
+git rebase upstream/main
+git push -u origin dev_your_branch
+```
+
+7. Create a merge request from your branch `dev_your_branch` at [origin repo](https://github.com/hiyouga/LLaMA-Factory).

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,13 +1,19 @@
 name: "\U0001F41B Bug / Help"
 description: Create a report to help us improve the LLaMA Factory
 body:
+  - type: markdown
+    attributes:
+      value: |
+        Issues included in **FAQs** or those with **insufficient** information may be closed without a response.
+        包含在**常见问题**内或提供信息**不完整**的 issues 可能不会被回复。
+
   - 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。
+        Please ensure you have read the README carefully and searched the existing issues (including FAQs).
+        请确保您已经认真阅读了 README 并且搜索过现有的 issues（包括常见问题）。
 
       options:
         - label: I have read the README and searched the existing issues.
@@ -38,7 +44,9 @@ body:
         请合理使用 Markdown 标签来格式化您的文本。
 
       placeholder: |
+        ```bash
         llamafactory-cli train ...
+        ```
 
   - type: textarea
     id: expected-behavior

.github/PULL_REQUEST_TEMPLATE.md

@@ -5,3 +5,4 @@ Fixes # (issue)
 ## Before submitting
 
 - [ ] Did you read the [contributor guideline](https://github.com/hiyouga/LLaMA-Factory/blob/main/.github/CONTRIBUTING.md)?
+- [ ] Did you write any new necessary tests?

.github/workflows/label_issue.yml

@@ -0,0 +1,30 @@
+name: label_issue
+
+on:
+  issues:
+    types:
+      - opened
+
+jobs:
+  label_issue:
+    runs-on: ubuntu-latest
+
+    permissions:
+      issues: write
+
+    steps:
+      - env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          ISSUE_URL: ${{ github.event.issue.html_url }}
+          ISSUE_TITLE: ${{ github.event.issue.title }}
+        run: |
+          LABEL=pending
+          NPU_KEYWORDS=(npu huawei ascend 华为 昇腾)
+          ISSUE_TITLE_LOWER=$(echo $ISSUE_TITLE | tr '[:upper:]' '[:lower:]')
+          for KEYWORD in ${NPU_KEYWORDS[@]}; do
+            if [[ $ISSUE_TITLE_LOWER == *$KEYWORD* ]] && [[ $ISSUE_TITLE_LOWER != *input* ]]; then
+              LABEL=pending,npu
+              break
+            fi
+          done
+          gh issue edit $ISSUE_URL --add-label $LABEL

.github/workflows/publish.yml

@@ -0,0 +1,40 @@
+name: publish
+
+on:
+  release:
+    types:
+      - published
+
+jobs:
+  publish:
+    name: Upload release to PyPI
+
+    runs-on: ubuntu-latest
+
+    environment:
+      name: release
+      url: https://pypi.org/p/llamafactory
+
+    permissions:
+      id-token: write
+
+    steps:
+      - name: Checkout
+        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 build
+
+      - name: Build package
+        run: |
+          python -m build
+
+      - name: Publish package
+        uses: pypa/gh-action-pypi-publish@release/v1

.github/workflows/tests.yml

@@ -3,14 +3,14 @@ name: tests
 on:
   push:
     branches:
-      - main
+      - "main"
     paths:
       - "**.py"
       - "requirements.txt"
       - ".github/workflows/*.yml"
   pull_request:
     branches:
-      - main
+      - "main"
     paths:
       - "**.py"
       - "requirements.txt"
@@ -18,22 +18,48 @@ on:
 
 jobs:
   tests:
-    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version:
+          - "3.8"  # TODO: remove py38 in next transformers release
+          - "3.9"
+          - "3.10"
+          - "3.11"
+        os:
+          - "ubuntu-latest"
+          - "windows-latest"
+          - "macos-13"
+
+    runs-on: ${{ matrix.os }}
+
+    environment:
+      name: tests
+
+    env:
+      HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      OS_NAME: ${{ matrix.os }}
+
     steps:
-      - uses: actions/checkout@v4
+      - name: Checkout
+        uses: actions/checkout@v4
+
       - name: Set up Python
         uses: actions/setup-python@v5
         with:
-          python-version: "3.8"
+          python-version: ${{ matrix.python-version }}
           cache: "pip"
           cache-dependency-path: "setup.py"
+
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install .[torch,dev]
+          python -m pip install ".[torch,dev]"
+
       - name: Check quality
         run: |
           make style && make quality
+
       - name: Test with pytest
         run: |
           make test

.gitignore

@@ -159,7 +159,15 @@ cython_debug/
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 .idea/
 
+# vscode
+.vscode/
+
 # custom .gitignore
-user.config
-saves/
+ms_cache/
+hf_cache/
+om_cache/
 cache/
+config/
+saves/
+output/
+wandb/

.pre-commit-config.yaml

@@ -0,0 +1,28 @@
+repos:
+-   repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v5.0.0
+    hooks:
+    -   id: check-ast
+    -   id: check-added-large-files
+        args: ['--maxkb=25000']
+    -   id: check-merge-conflict
+    -   id: check-yaml
+    -   id: debug-statements
+    -   id: end-of-file-fixer
+    -   id: trailing-whitespace
+        args: [--markdown-linebreak-ext=md]
+    -   id: no-commit-to-branch
+        args: ['--branch', 'main']
+
+-   repo: https://github.com/asottile/pyupgrade
+    rev: v3.17.0
+    hooks:
+    -   id: pyupgrade
+        args: [--py38-plus]
+
+-   repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.6.9
+    hooks:
+    -   id: ruff
+        args: [--fix]
+    -   id: ruff-format

CITATION.cff

@@ -12,12 +12,16 @@ authors:
   given-names: "Yanhan"
 - family-names: "Luo"
   given-names: "Zheyan"
+- family-names: "Feng"
+  given-names: "Zhangchi"
 - 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
+  type: conference-paper
+  conference:
+    name: "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)"
   authors:
     - family-names: "Zheng"
       given-names: "Yaowei"
@@ -29,9 +33,12 @@ preferred-citation:
       given-names: "Yanhan"
     - family-names: "Luo"
       given-names: "Zheyan"
+    - family-names: "Feng"
+      given-names: "Zhangchi"
     - 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
+  publisher: "Association for Computational Linguistics"
+  address: "Bangkok, Thailand"

Dockerfile

@@ -1,14 +0,0 @@
-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 .[metrics,bitsandbytes,qwen]
-
-VOLUME [ "/root/.cache/huggingface/", "/app/data", "/app/output" ]
-EXPOSE 7860
-
-CMD [ "llamafactory-cli", "webui" ]

MANIFEST.in

@@ -0,0 +1 @@
+include LICENSE requirements.txt

Makefile

@@ -1,6 +1,13 @@
-.PHONY: quality style test
+.PHONY: build commit quality style test
 
-check_dirs := scripts src tests
+check_dirs := scripts src tests setup.py
+
+build:
+	pip install build && python -m build
+
+commit:
+	pre-commit install
+	pre-commit run --all-files
 
 quality:
 	ruff check $(check_dirs)
@@ -11,4 +18,4 @@ style:
 	ruff format $(check_dirs)
 
 test:
-	pytest tests/
+	CUDA_VISIBLE_DEVICES= WANDB_DISABLED=true pytest -vv tests/

README.md

@@ -4,7 +4,7 @@
 [![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/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-44-green)](#projects-using-llama-factory)
+[![Citation](https://img.shields.io/badge/citation-93-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/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
 [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
@@ -12,22 +12,32 @@
 [![Open in DSW](https://gallery.pai-ml.com/assets/open-in-dsw.svg)](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory)
 [![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)
+[![SageMaker](https://img.shields.io/badge/SageMaker-Open%20in%20AWS-blue)](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
 
 [![GitHub Tread](https://trendshift.io/api/badge/repositories/4535)](https://trendshift.io/repositories/4535)
 
-👋 Join our [WeChat](assets/wechat.jpg).
+👋 Join our [WeChat](assets/wechat.jpg) or [NPU user group](assets/wechat_npu.jpg).
 
 \[ English | [中文](README_zh.md) \]
 
 **Fine-tuning a large language model can be easy as...**
 
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/9840a653-7e9c-41c8-ae89-7ace5698baf6
+https://github.com/user-attachments/assets/7c96b465-9df7-45f4-8053-bf03e58386d3
 
 Choose your path:
 
+- **Documentation (WIP)**: https://llamafactory.readthedocs.io/zh-cn/latest/
 - **Colab**: https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing
-- **PAI-DSW**: https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory
 - **Local machine**: Please refer to [usage](#getting-started)
+- **PAI-DSW**: [Llama3 Example](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory) | [Qwen2-VL Example](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_qwen2vl)
+- **Amazon SageMaker**: [Blog](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
+
+Recent activities:
+
+- **2024/10/18-2024/11/30**: Build a personal tour guide bot using PAI+LLaMA Factory. [[website]](https://developer.aliyun.com/topic/llamafactory2)
+
+> [!NOTE]
+> Except for the above links, all other websites are unauthorized third-party websites. Please carefully use them.
 
 ## Table of Contents
 
@@ -46,11 +56,11 @@ Choose your path:
 
 ## Features
 
-- **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
+- **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Qwen2-VL, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
 - **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc.
-- **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, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ and Agent tuning.
-- **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA.
+- **Scalable resources**: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ.
+- **Advanced algorithms**: [GaLore](https://github.com/jiaweizzhao/GaLore), [BAdam](https://github.com/Ledzy/BAdam), [Adam-mini](https://github.com/zyushun/Adam-mini), DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA and Agent tuning.
+- **Practical tricks**: [FlashAttention-2](https://github.com/Dao-AILab/flash-attention), [Unsloth](https://github.com/unslothai/unsloth), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), RoPE scaling, NEFTune and rsLoRA.
 - **Experiment monitors**: LlamaBoard, TensorBoard, Wandb, MLflow, etc.
 - **Faster inference**: OpenAI-style API, Gradio UI and CLI with vLLM worker.
 
@@ -71,15 +81,27 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Changelog
 
-[24/06/07] We supported fine-tuning the **[Qwen-2](https://qwenlm.github.io/blog/qwen2/)** series models.
+[24/10/09] We supported downloading pre-trained models and datasets from the **[Modelers Hub](https://modelers.cn/models)**. See [this tutorial](#download-from-modelers-hub) for usage.
 
-[24/06/05] We supported fine-tuning the **[GLM-4-9B/GLM-4-9B-Chat](https://github.com/THUDM/GLM-4)** models.
+[24/09/19] We support fine-tuning the **[Qwen2.5](https://qwenlm.github.io/blog/qwen2.5/)** models.
 
-[24/05/26] We supported **[SimPO](https://arxiv.org/abs/2405.14734)** algorithm for preference learning. See [examples](examples/README.md) for usage.
+[24/08/30] We support fine-tuning the **[Qwen2-VL](https://qwenlm.github.io/blog/qwen2-vl/)** models. Thank [@simonJJJ](https://github.com/simonJJJ)'s PR.
+
+[24/08/27] We support **[Liger Kernel](https://github.com/linkedin/Liger-Kernel)**. Try `enable_liger_kernel: true` for efficient training.
+
+[24/08/09] We support **[Adam-mini](https://github.com/zyushun/Adam-mini)** optimizer. See [examples](examples/README.md) for usage. Thank [@relic-yuexi](https://github.com/relic-yuexi)'s PR.
 
 <details><summary>Full Changelog</summary>
 
-[24/05/20] We supported fine-tuning the **PaliGemma** series models. Note that the PaliGemma models are pre-trained models, you need to fine-tune them with `gemma` template for chat completion.
+[24/07/04] We support [contamination-free packed training](https://github.com/MeetKai/functionary/tree/main/functionary/train/packing). Use `neat_packing: true` to activate it. Thank [@chuan298](https://github.com/chuan298)'s PR.
+
+[24/06/16] We support **[PiSSA](https://arxiv.org/abs/2404.02948)** algorithm. See [examples](examples/README.md) for usage.
+
+[24/06/07] We supported fine-tuning the **[Qwen2](https://qwenlm.github.io/blog/qwen2/)** and **[GLM-4](https://github.com/THUDM/GLM-4)** models.
+
+[24/05/26] We supported **[SimPO](https://arxiv.org/abs/2405.14734)** algorithm for preference learning. See [examples](examples/README.md) for usage.
+
+[24/05/20] We supported fine-tuning the **PaliGemma** series models. Note that the PaliGemma models are pre-trained models, you need to fine-tune them with `paligemma` template for chat completion.
 
 [24/05/18] We supported **[KTO](https://arxiv.org/abs/2402.01306)** algorithm for preference learning. See [examples](examples/README.md) for usage.
 
@@ -91,7 +113,7 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 [24/04/21] We supported **[Mixture-of-Depths](https://arxiv.org/abs/2404.02258)** according to [AstraMindAI's implementation](https://github.com/astramind-ai/Mixture-of-depths). See [examples](examples/README.md) for usage.
 
-[24/04/16] We supported **[BAdam](https://arxiv.org/abs/2404.02827)**. See [examples](examples/README.md) for usage.
+[24/04/16] We supported **[BAdam](https://arxiv.org/abs/2404.02827)** optimizer. See [examples](examples/README.md) for usage.
 
 [24/04/16] We supported **[unsloth](https://github.com/unslothai/unsloth)**'s long-sequence training (Llama-2-7B-56k within 24GB). It achieves **117%** speed and **50%** memory compared with FlashAttention-2, more benchmarks can be found in [this page](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison).
 
@@ -103,7 +125,7 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 [24/03/13] We supported **[LoRA+](https://arxiv.org/abs/2402.12354)**. See [examples](examples/README.md) for usage.
 
-[24/03/07] We supported gradient low-rank projection (**[GaLore](https://arxiv.org/abs/2403.03507)**) algorithm. See [examples](examples/README.md) for usage.
+[24/03/07] We supported **[GaLore](https://arxiv.org/abs/2403.03507)** optimizer. See [examples](examples/README.md) for usage.
 
 [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.
 
@@ -119,7 +141,7 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 [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](#download-from-modelscope-hub) for usage.
+[23/12/01] We supported downloading pre-trained models and datasets from the **[ModelScope Hub](https://modelscope.cn/models)**. See [this tutorial](#download-from-modelscope-hub) for usage.
 
 [23/10/21] We supported **[NEFTune](https://arxiv.org/abs/2310.05914)** trick for fine-tuning. Try `neftune_noise_alpha: 5` argument to activate NEFTune.
 
@@ -151,35 +173,40 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Supported Models
 
-| Model                                                    | Model size                       | Template  |
-| -------------------------------------------------------- | -------------------------------- | --------- |
-| [Baichuan2](https://huggingface.co/baichuan-inc)         | 7B/13B                           | baichuan2 |
-| [BLOOM](https://huggingface.co/bigscience)               | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [BLOOMZ](https://huggingface.co/bigscience)              | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [ChatGLM3](https://huggingface.co/THUDM)                 | 6B                               | chatglm3  |
-| [Command-R](https://huggingface.co/CohereForAI)          | 35B/104B                         | cohere    |
-| [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B/236B                  | deepseek  |
-| [Falcon](https://huggingface.co/tiiuae)                  | 7B/11B/40B/180B                  | falcon    |
-| [Gemma/CodeGemma](https://huggingface.co/google)         | 2B/7B                            | gemma     |
-| [GLM4](https://huggingface.co/THUDM)                     | 9B                               | glm4      |
-| [InternLM2](https://huggingface.co/internlm)             | 7B/20B                           | intern2   |
-| [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B                   | -         |
-| [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                       | llama2    |
-| [LLaMA-3](https://huggingface.co/meta-llama)             | 8B/70B                           | llama3    |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)             | 7B/13B                           | vicuna    |
-| [Mistral/Mixtral](https://huggingface.co/mistralai)      | 7B/8x7B/8x22B                    | mistral   |
-| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                            | -         |
-| [PaliGemma](https://huggingface.co/google)               | 3B                               | gemma     |
-| [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                        | -         |
-| [Phi-3](https://huggingface.co/microsoft)                | 4B/7B/14B                        | phi       |
-| [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B                  | qwen      |
-| [Qwen1.5 (Code/MoE)](https://huggingface.co/Qwen)        | 0.5B/1.8B/4B/7B/14B/32B/72B/110B | qwen      |
-| [Qwen2 (MoE)](https://huggingface.co/Qwen)               | 0.5B/1.5B/7B/57B/72B             | qwen      |
-| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                        | -         |
-| [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                       | xverse    |
-| [Yi (1/1.5)](https://huggingface.co/01-ai)               | 6B/9B/34B                        | yi        |
-| [Yi-VL](https://huggingface.co/01-ai)                    | 6B/34B                           | yi_vl     |
-| [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                      | yuan      |
+| Model                                                             | Model size                       | Template         |
+| ----------------------------------------------------------------- | -------------------------------- | ---------------- |
+| [Baichuan 2](https://huggingface.co/baichuan-inc)                 | 7B/13B                           | baichuan2        |
+| [BLOOM/BLOOMZ](https://huggingface.co/bigscience)                 | 560M/1.1B/1.7B/3B/7.1B/176B      | -                |
+| [ChatGLM3](https://huggingface.co/THUDM)                          | 6B                               | chatglm3         |
+| [Command R](https://huggingface.co/CohereForAI)                   | 35B/104B                         | cohere           |
+| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)         | 7B/16B/67B/236B                  | deepseek         |
+| [Falcon](https://huggingface.co/tiiuae)                           | 7B/11B/40B/180B                  | falcon           |
+| [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google)          | 2B/7B/9B/27B                     | gemma            |
+| [GLM-4](https://huggingface.co/THUDM)                             | 9B                               | glm4             |
+| [Index](https://huggingface.co/IndexTeam)                         | 1.9B                             | index            |
+| [InternLM2/InternLM2.5](https://huggingface.co/internlm)          | 7B/20B                           | intern2          |
+| [Llama](https://github.com/facebookresearch/llama)                | 7B/13B/33B/65B                   | -                |
+| [Llama 2](https://huggingface.co/meta-llama)                      | 7B/13B/70B                       | llama2           |
+| [Llama 3-3.2](https://huggingface.co/meta-llama)                  | 1B/3B/8B/70B                     | llama3           |
+| [Llama 3.2 Vision](https://huggingface.co/meta-llama)             | 11B/90B                          | mllama           |
+| [LLaVA-1.5](https://huggingface.co/llava-hf)                      | 7B/13B                           | llava            |
+| [LLaVA-NeXT](https://huggingface.co/llava-hf)                     | 7B/8B/13B/34B/72B/110B           | llava_next       |
+| [LLaVA-NeXT-Video](https://huggingface.co/llava-hf)               | 7B/34B                           | llava_next_video |
+| [MiniCPM](https://huggingface.co/openbmb)                         | 1B/2B/4B                         | cpm/cpm3         |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)               | 7B/8x7B/8x22B                    | mistral          |
+| [OLMo](https://huggingface.co/allenai)                            | 1B/7B                            | -                |
+| [PaliGemma](https://huggingface.co/google)                        | 3B                               | paligemma        |
+| [Phi-1.5/Phi-2](https://huggingface.co/microsoft)                 | 1.3B/2.7B                        | -                |
+| [Phi-3](https://huggingface.co/microsoft)                         | 4B/14B                           | phi              |
+| [Phi-3-small](https://huggingface.co/microsoft)                   | 7B                               | phi_small        |
+| [Pixtral](https://huggingface.co/mistralai)                       | 12B                              | pixtral          |
+| [Qwen (1-2.5) (Code/Math/MoE)](https://huggingface.co/Qwen)       | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen             |
+| [Qwen2-VL](https://huggingface.co/Qwen)                           | 2B/7B/72B                        | qwen2_vl         |
+| [StarCoder 2](https://huggingface.co/bigcode)                     | 3B/7B/15B                        | -                |
+| [XVERSE](https://huggingface.co/xverse)                           | 7B/13B/65B                       | xverse           |
+| [Yi/Yi-1.5 (Code)](https://huggingface.co/01-ai)                  | 1.5B/6B/9B/34B                   | yi               |
+| [Yi-VL](https://huggingface.co/01-ai)                             | 6B/34B                           | yi_vl            |
+| [Yuan 2](https://huggingface.co/IEITYuan)                         | 2B/51B/102B                      | yuan             |
 
 > [!NOTE]
 > 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 "instruct/chat" models.
@@ -203,6 +230,9 @@ You also can add a custom chat template to [template.py](src/llamafactory/data/t
 | ORPO Training          | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | SimPO Training         | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
+> [!TIP]
+> The implementation details of PPO can be found in [this blog](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html).
+
 ## Provided Datasets
 
 <details><summary>Pre-training datasets</summary>
@@ -259,7 +289,12 @@ You also can add a custom chat template to [template.py](src/llamafactory/data/t
 - [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
 - [STEM (zh)](https://huggingface.co/datasets/hfl/stem_zh_instruction)
 - [Ruozhiba (zh)](https://huggingface.co/datasets/hfl/ruozhiba_gpt4_turbo)
+- [Neo-sft (zh)](https://huggingface.co/datasets/m-a-p/neo_sft_phase2)
+- [WebInstructSub (en)](https://huggingface.co/datasets/TIGER-Lab/WebInstructSub)
+- [Magpie-Pro-300K-Filtered (en)](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered)
+- [Magpie-ultra-v0.1 (en)](https://huggingface.co/datasets/argilla/magpie-ultra-v0.1)
 - [LLaVA mixed (en&zh)](https://huggingface.co/datasets/BUAADreamer/llava-en-zh-300k)
+- [Pokemon-gpt4o-captions (en&zh)](https://huggingface.co/datasets/jugg1024/pokemon-gpt4o-captions)
 - [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)
@@ -276,6 +311,8 @@ You also can add a custom chat template to [template.py](src/llamafactory/data/t
 
 - [DPO mixed (en&zh)](https://huggingface.co/datasets/hiyouga/DPO-En-Zh-20k)
 - [UltraFeedback (en)](https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized)
+- [RLHF-V (en)](https://huggingface.co/datasets/openbmb/RLHF-V-Dataset)
+- [VLFeedback (en)](https://huggingface.co/datasets/Zhihui/VLFeedback)
 - [Orca DPO Pairs (en)](https://huggingface.co/datasets/Intel/orca_dpo_pairs)
 - [HH-RLHF (en)](https://huggingface.co/datasets/Anthropic/hh-rlhf)
 - [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
@@ -296,20 +333,20 @@ huggingface-cli login
 | Mandatory    | Minimum | Recommend |
 | ------------ | ------- | --------- |
 | python       | 3.8     | 3.11      |
-| torch        | 1.13.1  | 2.3.0     |
-| transformers | 4.41.2  | 4.41.2    |
-| datasets     | 2.16.0  | 2.19.2    |
-| accelerate   | 0.30.1  | 0.30.1    |
-| peft         | 0.11.1  | 0.11.1    |
-| trl          | 0.8.6   | 0.9.4     |
+| torch        | 1.13.1  | 2.4.0     |
+| transformers | 4.41.2  | 4.43.4    |
+| datasets     | 2.16.0  | 2.20.0    |
+| accelerate   | 0.30.1  | 0.32.0    |
+| peft         | 0.11.1  | 0.12.0    |
+| trl          | 0.8.6   | 0.9.6     |
 
 | Optional     | Minimum | Recommend |
 | ------------ | ------- | --------- |
 | CUDA         | 11.6    | 12.2      |
 | deepspeed    | 0.10.0  | 0.14.0    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.4.3     |
-| flash-attn   | 2.3.0   | 2.5.9     |
+| vllm         | 0.4.3   | 0.5.0     |
+| flash-attn   | 2.3.0   | 2.6.3     |
 
 ### Hardware Requirement
 
@@ -335,10 +372,10 @@ huggingface-cli login
 ```bash
 git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
 cd LLaMA-Factory
-pip install -e '.[torch,metrics]'
+pip install -e ".[torch,metrics]"
 ```
 
-Extra dependencies available: torch, torch_npu, metrics, deepspeed, bitsandbytes, vllm, galore, badam, gptq, awq, aqlm, qwen, modelscope, quality
+Extra dependencies available: torch, torch-npu, metrics, deepspeed, liger-kernel, bitsandbytes, hqq, eetq, gptq, awq, aqlm, vllm, galore, badam, adam-mini, qwen, modelscope, openmind, quality
 
 > [!TIP]
 > Use `pip install --no-deps -e .` to resolve package conflicts.
@@ -357,9 +394,7 @@ To enable FlashAttention-2 on the Windows platform, you need to install the prec
 
 <details><summary>For Ascend NPU users</summary>
 
-Join [NPU user group](assets/wechat_npu.jpg).
-
-To install LLaMA Factory on Ascend NPU devices, please specify extra dependencies: `pip install -e '.[torch-npu,metrics]'`. Additionally, you need to install the **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**. Please follow the [installation tutorial](https://www.hiascend.com/document/detail/en/CANNCommunityEdition/600alphaX/softwareinstall/instg/atlasdeploy_03_0031.html) or use the following commands:
+To install LLaMA Factory on Ascend NPU devices, please specify extra dependencies: `pip install -e ".[torch-npu,metrics]"`. Additionally, you need to install the **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**. Please follow the [installation tutorial](https://www.hiascend.com/document/detail/en/CANNCommunityEdition/600alphaX/softwareinstall/instg/atlasdeploy_03_0031.html) or use the following commands:
 
 ```bash
 # replace the url according to your CANN version and devices
@@ -382,20 +417,17 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch-npu    | 2.1.0   | 2.1.0.post3 |
 | deepspeed    | 0.13.2  | 0.13.2      |
 
-Docker image:
-
-- 32GB: [Download page](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html)
-- 64GB: [Download page](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
-
 Remember to use `ASCEND_RT_VISIBLE_DEVICES` instead of `CUDA_VISIBLE_DEVICES` to specify the device to use.
 
 If you cannot infer model on NPU devices, try setting `do_sample: false` in the configurations.
 
+Download the pre-built Docker images: [32GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html) | [64GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
+
 </details>
 
 ### Data Preparation
 
-Please refer to [data/README.md](data/README.md) for checking the details about the format of dataset files. You can either use datasets on HuggingFace / ModelScope hub or load the dataset in local disk.
+Please refer to [data/README.md](data/README.md) for checking the details about the format of dataset files. You can either use datasets on HuggingFace / ModelScope / Modelers hub or load the dataset in local disk.
 
 > [!NOTE]
 > Please update `data/dataset_info.json` to use your custom dataset.
@@ -405,9 +437,9 @@ Please refer to [data/README.md](data/README.md) for checking the details about
 Use the following 3 commands to run LoRA **fine-tuning**, **inference** and **merging** of the Llama3-8B-Instruct model, respectively.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 See [examples/README.md](examples/README.md) for advanced usage (including distributed training).
@@ -417,50 +449,148 @@ See [examples/README.md](examples/README.md) for advanced usage (including distr
 
 ### Fine-Tuning with LLaMA Board GUI (powered by [Gradio](https://github.com/gradio-app/gradio))
 
-#### Use local environment
+```bash
+llamafactory-cli webui
+```
+
+### Build Docker
+
+For CUDA users:
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 GRADIO_SHARE=1 llamafactory-cli webui
+cd docker/docker-cuda/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+For Ascend NPU users:
+
+```bash
+cd docker/docker-npu/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+For AMD ROCm users:
+
+```bash
+cd docker/docker-rocm/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+<details><summary>Build without Docker Compose</summary>
+
+For CUDA users:
+
+```bash
+docker build -f ./docker/docker-cuda/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit --gpus=all \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./om_cache:/root/.cache/openmind \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
+```
+
+For Ascend NPU users:
+
+```bash
+# Choose docker image upon your environment
+docker build -f ./docker/docker-npu/Dockerfile \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+# Change `device` upon your resources
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./om_cache:/root/.cache/openmind \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
+    -v /etc/ascend_install.info:/etc/ascend_install.info \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/davinci0 \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
+```
+
+For AMD ROCm users:
+
+```bash
+docker build -f ./docker/docker-rocm/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./om_cache:/root/.cache/openmind \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v ./saves:/app/saves \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/kfd \
+    --device /dev/dri \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
 </details>
 
-#### Use Docker
-
-```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
-```
-
-#### Use Docker Compose
-
-```bash
-docker compose -f ./docker-compose.yml up -d
-```
-
 <details><summary>Details about volume</summary>
 
-- hf_cache: Utilize Hugging Face 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.
+- `hf_cache`: Utilize Hugging Face cache on the host machine. Reassignable if a cache already exists in a different directory.
+- `ms_cache`: Similar to Hugging Face cache but for ModelScope users.
+- `om_cache`: Similar to Hugging Face cache but for Modelers users.
+- `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.
 
 </details>
 
 ### Deploy with OpenAI-style API and vLLM
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1 API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
+API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
 ```
 
 > [!TIP]
-> Visit https://platform.openai.com/docs/api-reference/chat/create for API document.
+> Visit [this page](https://platform.openai.com/docs/api-reference/chat/create) for API document.
+>
+> Examples: [Image understanding](scripts/test_image.py) | [Function calling](scripts/test_toolcall.py)
 
 ### Download from ModelScope Hub
 
@@ -472,9 +602,19 @@ export USE_MODELSCOPE_HUB=1 # `set USE_MODELSCOPE_HUB=1` for Windows
 
 Train the model by specifying a model ID of the ModelScope Hub as the `model_name_or_path`. You can find a full list of model IDs at [ModelScope Hub](https://modelscope.cn/models), e.g., `LLM-Research/Meta-Llama-3-8B-Instruct`.
 
+### Download from Modelers Hub
+
+You can also use Modelers Hub to download models and datasets.
+
+```bash
+export USE_OPENMIND_HUB=1 # `set USE_OPENMIND_HUB=1` for Windows
+```
+
+Train the model by specifying a model ID of the Modelers Hub as the `model_name_or_path`. You can find a full list of model IDs at [Modelers Hub](https://modelers.cn/models), e.g., `TeleAI/TeleChat-7B-pt`.
+
 ### Use W&B Logger
 
-To use [Weights & Biases](https://wandb.ai) for logging experimental results, you need to add the following arguments.
+To use [Weights & Biases](https://wandb.ai) for logging experimental results, you need to add the following arguments to yaml files.
 
 ```yaml
 report_to: wandb
@@ -494,45 +634,95 @@ If you have a project that should be incorporated, please contact via email or c
 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. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. KDD 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. ACL 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. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. ACL 2024 Findings. [[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. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. ACL 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. Wu et al. Large Language Models are Parallel Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2403.09073)
 1. Zhang et al. EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling. 2024. [[arxiv]](https://arxiv.org/abs/2403.14541)
 1. Weller et al. FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2403.15246)
-1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
+1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. COLING 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
 1. Zan et al. CodeS: Natural Language to Code Repository via Multi-Layer Sketch. 2024. [[arxiv]](https://arxiv.org/abs/2403.16443)
 1. Liu et al. Extensive Self-Contrast Enables Feedback-Free Language Model Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2404.00604)
 1. Luo et al. BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.02827)
 1. Du et al. Chinese Tiny LLM: Pretraining a Chinese-Centric Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2404.04167)
-1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
+1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. ICML 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
 1. Liu et al. Dynamic Generation of Personalities with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.07084)
 1. Shang et al. How Far Have We Gone in Stripped Binary Code Understanding Using Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.09836)
 1. Huang et al. LLMTune: Accelerate Database Knob Tuning with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.11581)
 1. Deng et al. Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction. 2024. [[arxiv]](https://arxiv.org/abs/2404.14215)
 1. Acikgoz et al. Hippocrates: An Open-Source Framework for Advancing Large Language Models in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2404.16621)
-1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2404.17140)
-1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2404.17140)
+1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. NAACL 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Xu et al. Large Language Models for Cyber Security: A Systematic Literature Review. 2024. [[arxiv]](https://arxiv.org/abs/2405.04760)
+1. Dammu et al. "They are uncultured": Unveiling Covert Harms and Social Threats in LLM Generated Conversations. 2024. [[arxiv]](https://arxiv.org/abs/2405.05378)
+1. Yi et al. A safety realignment framework via subspace-oriented model fusion for large language models. 2024. [[arxiv]](https://arxiv.org/abs/2405.09055)
+1. Lou et al. SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling. 2024. [[arxiv]](https://arxiv.org/abs/2405.12739)
+1. Zhang et al. Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2405.13816)
+1. Zhang et al. TS-Align: A Teacher-Student Collaborative Framework for Scalable Iterative Finetuning of Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2405.20215)
+1. Zihong Chen. Sentence Segmentation and Sentence Punctuation Based on XunziALLM. 2024. [[paper]](https://aclanthology.org/2024.lt4hala-1.30)
+1. Gao et al. The Best of Both Worlds: Toward an Honest and Helpful Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2406.00380)
+1. Wang and Song. MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset. 2024. [[arxiv]](https://arxiv.org/abs/2406.02106)
+1. Hu et al. Computational Limits of Low-Rank Adaptation (LoRA) for Transformer-Based Models. 2024. [[arxiv]](https://arxiv.org/abs/2406.03136)
+1. Ge et al. Time Sensitive Knowledge Editing through Efficient Finetuning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2406.04496)
+1. Tan et al. Peer Review as A Multi-Turn and Long-Context Dialogue with Role-Based Interactions. 2024. [[arxiv]](https://arxiv.org/abs/2406.05688)
+1. Song et al. Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters. 2024. [[arxiv]](https://arxiv.org/abs/2406.05955)
+1. Gu et al. RWKV-CLIP: A Robust Vision-Language Representation Learner. 2024. [[arxiv]](https://arxiv.org/abs/2406.06973)
+1. Chen et al. Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees. 2024. [[arxiv]](https://arxiv.org/abs/2406.07115)
+1. Zhu et al. Are Large Language Models Good Statisticians?. 2024. [[arxiv]](https://arxiv.org/abs/2406.07815)
+1. Li et al. Know the Unknown: An Uncertainty-Sensitive Method for LLM Instruction Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2406.10099)
+1. Ding et al. IntentionQA: A Benchmark for Evaluating Purchase Intention Comprehension Abilities of Language Models in E-commerce. 2024. [[arxiv]](https://arxiv.org/abs/2406.10173)
+1. He et al. COMMUNITY-CROSS-INSTRUCT: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities. 2024. [[arxiv]](https://arxiv.org/abs/2406.12074)
+1. Lin et al. FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving. 2024. [[arxiv]](https://arxiv.org/abs/2406.14408)
+1. Treutlein et al. Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data. 2024. [[arxiv]](https://arxiv.org/abs/2406.14546)
+1. Feng et al. SS-Bench: A Benchmark for Social Story Generation and Evaluation. 2024. [[arxiv]](https://arxiv.org/abs/2406.15695)
+1. Feng et al. Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement. 2024. [[arxiv]](https://arxiv.org/abs/2406.17233)
+1. Liu et al. Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals. 2024. [[arxiv]](https://arxiv.org/abs/2406.18069)
+1. Iyer et al. Exploring Very Low-Resource Translation with LLMs: The University of Edinburgh's Submission to AmericasNLP 2024 Translation Task. AmericasNLP 2024. [[paper]](https://aclanthology.org/2024.americasnlp-1.25)
+1. Li et al. Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring. 2024. [[arxiv]](https://arxiv.org/abs/2406.19949)
+1. Yang et al. Financial Knowledge Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2407.00365)
+1. Lin et al. DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging. 2024. [[arxiv]](https://arxiv.org/abs/2407.01470)
+1. Bako et al. Evaluating the Semantic Profiling Abilities of LLMs for Natural Language Utterances in Data Visualization. 2024. [[arxiv]](https://arxiv.org/abs/2407.06129)
+1. Huang et al. RoLoRA: Fine-tuning Rotated Outlier-free LLMs for Effective Weight-Activation Quantization. 2024. [[arxiv]](https://arxiv.org/abs/2407.08044)
+1. Jiang et al. LLM-Collaboration on Automatic Science Journalism for the General Audience. 2024. [[arxiv]](https://arxiv.org/abs/2407.09756)
+1. Inouye et al. Applied Auto-tuning on LoRA Hyperparameters. 2024. [[paper]](https://scholarcommons.scu.edu/cseng_senior/272/)
+1. Qi et al. Research on Tibetan Tourism Viewpoints information generation system based on LLM. 2024. [[arxiv]](https://arxiv.org/abs/2407.13561)
+1. Xu et al. Course-Correction: Safety Alignment Using Synthetic Preferences. 2024. [[arxiv]](https://arxiv.org/abs/2407.16637)
+1. Sun et al. LAMBDA: A Large Model Based Data Agent. 2024. [[arxiv]](https://arxiv.org/abs/2407.17535)
+1. Zhu et al. CollectiveSFT: Scaling Large Language Models for Chinese Medical Benchmark with Collective Instructions in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2407.19705)
+1. Yu et al. Correcting Negative Bias in Large Language Models through Negative Attention Score Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2408.00137)
+1. Xie et al. The Power of Personalized Datasets: Advancing Chinese Composition Writing for Elementary School through Targeted Model Fine-Tuning. IALP 2024. [[paper]](https://www.asianlp.sg/conferences/ialp2024/proceedings/papers/IALP2024_P055.pdf)
+1. Liu et al. Instruct-Code-Llama: Improving Capabilities of Language Model in Competition Level Code Generation by Online Judge Feedback. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_11)
+1. Wang et al. Cybernetic Sentinels: Unveiling the Impact of Safety Data Selection on Model Security in Supervised Fine-Tuning. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_23)
+1. Xia et al. Understanding the Performance and Estimating the Cost of LLM Fine-Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2408.04693)
+1. Zeng et al. Perceive, Reflect, and Plan: Designing LLM Agent for Goal-Directed City Navigation without Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2408.04168)
+1. Xia et al. Using Pre-trained Language Model for Accurate ESG Prediction. FinNLP 2024. [[paper]](https://aclanthology.org/2024.finnlp-2.1/)
+1. Liang et al. I-SHEEP: Self-Alignment of LLM from Scratch through an Iterative Self-Enhancement Paradigm. 2024. [[arxiv]](https://arxiv.org/abs/2408.08072)
+1. Bai et al. Aligning Large Language Model with Direct Multi-Preference Optimization for Recommendation. CIKM 2024. [[paper]](https://dl.acm.org/doi/10.1145/3627673.3679611)
 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/X-D-Lab/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.
-1. **[Luminia-13B-v3](https://huggingface.co/Nekochu/Luminia-13B-v3)**: A large language model specialized in generate metadata for stable diffusion. [[🤗Demo]](https://huggingface.co/spaces/Nekochu/Luminia-13B_SD_Prompt)
+1. **[Luminia-13B-v3](https://huggingface.co/Nekochu/Luminia-13B-v3)**: A large language model specialized in generate metadata for stable diffusion. [[demo]](https://huggingface.co/spaces/Nekochu/Luminia-13B_SD_Prompt)
 1. **[Chinese-LLaVA-Med](https://github.com/BUAADreamer/Chinese-LLaVA-Med)**: A multimodal large language model specialized in Chinese medical domain, based on LLaVA-1.5-7B.
+1. **[AutoRE](https://github.com/THUDM/AutoRE)**: A document-level relation extraction system based on large language models.
+1. **[NVIDIA RTX AI Toolkit](https://github.com/NVIDIA/RTX-AI-Toolkit)**: SDKs for fine-tuning LLMs on Windows PC for NVIDIA RTX.
+1. **[LazyLLM](https://github.com/LazyAGI/LazyLLM)**: An easy and lazy way for building multi-agent LLMs applications and supports model fine-tuning via LLaMA Factory.
+1. **[RAG-Retrieval](https://github.com/NLPJCL/RAG-Retrieval)**: A full pipeline for RAG retrieval model fine-tuning, inference, and distillation. [[blog]](https://zhuanlan.zhihu.com/p/987727357)
+
 
 </details>
 
@@ -540,17 +730,19 @@ If you have a project that should be incorporated, please contact via email or c
 
 This repository is licensed under the [Apache-2.0 License](LICENSE).
 
-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) / [Command-R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [LLaMA-3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [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) / [Yi-1.5](LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+Please follow the model licenses to use the corresponding model weights: [Baichuan 2](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) / [Command R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [Index](https://huggingface.co/IndexTeam/Index-1.9B/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [MiniCPM](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) / [Mistral/Mixtral/Pixtral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder 2](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) / [Yi-1.5](LICENSE) / [Yuan 2](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## Citation
 
 If this work is helpful, please kindly cite as:
 
 ```bibtex
-@article{zheng2024llamafactory,
+@inproceedings{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},
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Zhangchi Feng and Yongqiang Ma},
+  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
+  address={Bangkok, Thailand},
+  publisher={Association for Computational Linguistics},
   year={2024},
   url={http://arxiv.org/abs/2403.13372}
 }

README_zh.md

@@ -4,7 +4,7 @@
 [![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/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-44-green)](#使用了-llama-factory-的项目)
+[![Citation](https://img.shields.io/badge/citation-93-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/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
 [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
@@ -12,22 +12,33 @@
 [![Open in DSW](https://gallery.pai-ml.com/assets/open-in-dsw.svg)](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory)
 [![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)
+[![SageMaker](https://img.shields.io/badge/SageMaker-Open%20in%20AWS-blue)](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
 
 [![GitHub Tread](https://trendshift.io/api/badge/repositories/4535)](https://trendshift.io/repositories/4535)
 
-👋 加入我们的[微信群](assets/wechat.jpg)。
+👋 加入我们的[微信群](assets/wechat.jpg)或 [NPU 用户群](assets/wechat_npu.jpg)。
 
 \[ [English](README.md) | 中文 \]
 
 **微调大模型可以像这样轻松…**
 
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd-d76c6d0a6594
+https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 
 选择你的打开方式：
 
+- **入门教程**：https://zhuanlan.zhihu.com/p/695287607
+- **框架文档**：https://llamafactory.readthedocs.io/zh-cn/latest/
 - **Colab**：https://colab.research.google.com/drive/1d5KQtbemerlSDSxZIfAaWXhKr30QypiK?usp=sharing
-- **PAI-DSW**: https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory
 - **本地机器**：请见[如何使用](#如何使用)
+- **PAI-DSW**：[Llama3 案例](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory) | [Qwen2-VL 案例](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_qwen2vl)
+- **Amazon SageMaker**：[博客](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
+
+近期活动：
+
+- **2024/10/18-2024/11/30**：使用 PAI+LLaMA Factory 构建个性化导游机器人。[[活动页面]](https://developer.aliyun.com/topic/llamafactory2)
+
+> [!NOTE]
+> 除上述链接以外的其他网站均为未经许可的第三方网站，请小心甄别。
 
 ## 目录
 
@@ -46,11 +57,11 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 项目特色
 
-- **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
+- **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Qwen2-VL、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
 - **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO 训练、DPO 训练、KTO 训练、ORPO 训练等等。
-- **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、BAdam、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ 和 Agent 微调。
-- **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。
+- **多种精度**：16 比特全参数微调、冻结微调、LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ 的 2/3/4/5/6/8 比特 QLoRA 微调。
+- **先进算法**：[GaLore](https://github.com/jiaweizzhao/GaLore)、[BAdam](https://github.com/Ledzy/BAdam)、[Adam-mini](https://github.com/zyushun/Adam-mini)、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ、PiSSA 和 Agent 微调。
+- **实用技巧**：[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)、[Unsloth](https://github.com/unslothai/unsloth)、[Liger Kernel](https://github.com/linkedin/Liger-Kernel)、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。
 
@@ -71,15 +82,27 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 更新日志
 
-[24/06/07] 我们支持了 **[Qwen-2](https://qwenlm.github.io/blog/qwen2/)** 系列模型的微调。
+[24/10/09] 我们支持了从 **[魔乐社区](https://modelers.cn/models)** 下载预训练模型和数据集。详细用法请参照 [此教程](#从魔乐社区下载)。
 
-[24/06/05] 我们支持了 **[GLM-4-9B/GLM-4-9B-Chat](https://github.com/THUDM/GLM-4)** 模型的微调。
+[24/09/19] 我们支持了 **[Qwen2.5](https://qwenlm.github.io/blog/qwen2.5/)** 模型的微调。
 
-[24/05/26] 我们支持了 **[SimPO](https://arxiv.org/abs/2405.14734)** 偏好对齐算法。详细用法请参照 [examples](examples/README_zh.md)。
+[24/08/30] 我们支持了 **[Qwen2-VL](https://qwenlm.github.io/blog/qwen2-vl/)** 模型的微调。感谢 [@simonJJJ](https://github.com/simonJJJ) 的 PR。
+
+[24/08/27] 我们支持了 **[Liger Kernel](https://github.com/linkedin/Liger-Kernel)**。请使用 `enable_liger_kernel: true` 来加速训练。
+
+[24/08/09] 我们支持了 **[Adam-mini](https://github.com/zyushun/Adam-mini)** 优化器。详细用法请参照 [examples](examples/README_zh.md)。感谢 [@relic-yuexi](https://github.com/relic-yuexi) 的 PR。
 
 <details><summary>展开日志</summary>
 
-[24/05/20] 我们支持了 **PaliGemma** 系列模型的微调。注意 PaliGemma 是预训练模型，你需要使用 `gemma` 模板进行微调使其获得对话能力。
+[24/07/04] 我们支持了[无污染打包训练](https://github.com/MeetKai/functionary/tree/main/functionary/train/packing)。请使用 `neat_packing: true` 参数。感谢 [@chuan298](https://github.com/chuan298) 的 PR。
+
+[24/06/16] 我们支持了 **[PiSSA](https://arxiv.org/abs/2404.02948)** 算法。详细用法请参照 [examples](examples/README_zh.md)。
+
+[24/06/07] 我们支持了 **[Qwen2](https://qwenlm.github.io/blog/qwen2/)** 和 **[GLM-4](https://github.com/THUDM/GLM-4)** 模型的微调。
+
+[24/05/26] 我们支持了 **[SimPO](https://arxiv.org/abs/2405.14734)** 偏好对齐算法。详细用法请参照 [examples](examples/README_zh.md)。
+
+[24/05/20] 我们支持了 **PaliGemma** 系列模型的微调。注意 PaliGemma 是预训练模型，你需要使用 `paligemma` 模板进行微调使其获得对话能力。
 
 [24/05/18] 我们支持了 **[KTO](https://arxiv.org/abs/2402.01306)** 偏好对齐算法。详细用法请参照 [examples](examples/README_zh.md)。
 
@@ -91,7 +114,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 [24/04/21] 我们基于 [AstraMindAI 的仓库](https://github.com/astramind-ai/Mixture-of-depths)支持了 **[混合深度训练](https://arxiv.org/abs/2404.02258)**。详细用法请参照 [examples](examples/README_zh.md)。
 
-[24/04/16] 我们支持了 **[BAdam](https://arxiv.org/abs/2404.02827)**。详细用法请参照 [examples](examples/README_zh.md)。
+[24/04/16] 我们支持了 **[BAdam](https://arxiv.org/abs/2404.02827)** 优化器。详细用法请参照 [examples](examples/README_zh.md)。
 
 [24/04/16] 我们支持了 **[unsloth](https://github.com/unslothai/unsloth)** 的长序列训练（24GB 可训练 Llama-2-7B-56k）。该方法相比 FlashAttention-2 提供了 **117%** 的训练速度和 **50%** 的显存节约。更多数据请见[此页面](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison)。
 
@@ -103,7 +126,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 [24/03/13] 我们支持了 **[LoRA+](https://arxiv.org/abs/2402.12354)**。详细用法请参照 [examples](examples/README_zh.md)。
 
-[24/03/07] 我们支持了梯度低秩投影（**[GaLore](https://arxiv.org/abs/2403.03507)**）算法。详细用法请参照 [examples](examples/README_zh.md)。
+[24/03/07] 我们支持了 **[GaLore](https://arxiv.org/abs/2403.03507)** 优化器。详细用法请参照 [examples](examples/README_zh.md)。
 
 [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。
 
@@ -151,35 +174,39 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 模型
 
-| 模型名                                                   | 模型大小                          | Template  |
-| -------------------------------------------------------- | -------------------------------- | --------- |
-| [Baichuan2](https://huggingface.co/baichuan-inc)         | 7B/13B                           | baichuan2 |
-| [BLOOM](https://huggingface.co/bigscience)               | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [BLOOMZ](https://huggingface.co/bigscience)              | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [ChatGLM3](https://huggingface.co/THUDM)                 | 6B                               | chatglm3  |
-| [Command-R](https://huggingface.co/CohereForAI)          | 35B/104B                         | cohere    |
-| [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B/236B                  | deepseek  |
-| [Falcon](https://huggingface.co/tiiuae)                  | 7B/11B/40B/180B                  | falcon    |
-| [Gemma/CodeGemma](https://huggingface.co/google)         | 2B/7B                            | gemma     |
-| [GLM4](https://huggingface.co/THUDM)                     | 9B                               | glm4      |
-| [InternLM2](https://huggingface.co/internlm)             | 7B/20B                           | intern2   |
-| [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B                   | -         |
-| [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                       | llama2    |
-| [LLaMA-3](https://huggingface.co/meta-llama)             | 8B/70B                           | llama3    |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)             | 7B/13B                           | vicuna    |
-| [Mistral/Mixtral](https://huggingface.co/mistralai)      | 7B/8x7B/8x22B                    | mistral   |
-| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                            | -         |
-| [PaliGemma](https://huggingface.co/google)               | 3B                               | gemma     |
-| [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                        | -         |
-| [Phi-3](https://huggingface.co/microsoft)                | 4B/7B/14B                        | phi       |
-| [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B                  | qwen      |
-| [Qwen1.5 (Code/MoE)](https://huggingface.co/Qwen)        | 0.5B/1.8B/4B/7B/14B/32B/72B/110B | qwen      |
-| [Qwen2 (MoE)](https://huggingface.co/Qwen)               | 0.5B/1.5B/7B/57B/72B             | qwen      |
-| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                        | -         |
-| [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                       | xverse    |
-| [Yi (1/1.5)](https://huggingface.co/01-ai)               | 6B/9B/34B                        | yi        |
-| [Yi-VL](https://huggingface.co/01-ai)                    | 6B/34B                           | yi_vl     |
-| [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                      | yuan      |
+| 模型名                                                            | 模型大小                          | Template         |
+| ----------------------------------------------------------------- | -------------------------------- | ---------------- |
+| [Baichuan 2](https://huggingface.co/baichuan-inc)                 | 7B/13B                           | baichuan2        |
+| [BLOOM/BLOOMZ](https://huggingface.co/bigscience)                 | 560M/1.1B/1.7B/3B/7.1B/176B      | -                |
+| [ChatGLM3](https://huggingface.co/THUDM)                          | 6B                               | chatglm3         |
+| [Command R](https://huggingface.co/CohereForAI)                   | 35B/104B                         | cohere           |
+| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)         | 7B/16B/67B/236B                  | deepseek         |
+| [Falcon](https://huggingface.co/tiiuae)                           | 7B/11B/40B/180B                  | falcon           |
+| [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google)          | 2B/7B/9B/27B                     | gemma            |
+| [GLM-4](https://huggingface.co/THUDM)                             | 9B                               | glm4             |
+| [Index](https://huggingface.co/IndexTeam)                         | 1.9B                             | index            |
+| [InternLM2/InternLM2.5](https://huggingface.co/internlm)          | 7B/20B                           | intern2          |
+| [Llama](https://github.com/facebookresearch/llama)                | 7B/13B/33B/65B                   | -                |
+| [Llama 2](https://huggingface.co/meta-llama)                      | 7B/13B/70B                       | llama2           |
+| [Llama 3-3.2](https://huggingface.co/meta-llama)                  | 1B/3B/8B/70B                     | llama3           |
+| [Llama 3.2 Vision](https://huggingface.co/meta-llama)             | 11B/90B                          | mllama           |
+| [LLaVA-1.5](https://huggingface.co/llava-hf)                      | 7B/13B                           | llava            |
+| [LLaVA-NeXT](https://huggingface.co/llava-hf)                     | 7B/8B/13B/34B/72B/110B           | llava_next       |
+| [LLaVA-NeXT-Video](https://huggingface.co/llava-hf)               | 7B/34B                           | llava_next_video |
+| [MiniCPM](https://huggingface.co/openbmb)                         | 1B/2B/4B                         | cpm/cpm3         |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)               | 7B/8x7B/8x22B                    | mistral          |
+| [OLMo](https://huggingface.co/allenai)                            | 1B/7B                            | -                |
+| [PaliGemma](https://huggingface.co/google)                        | 3B                               | paligemma        |
+| [Phi-1.5/Phi-2](https://huggingface.co/microsoft)                 | 1.3B/2.7B                        | -                |
+| [Phi-3](https://huggingface.co/microsoft)                         | 4B/7B/14B                        | phi              |
+| [Pixtral](https://huggingface.co/mistralai)                       | 12B                              | pixtral          |
+| [Qwen (1-2.5) (Code/Math/MoE)](https://huggingface.co/Qwen)       | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen             |
+| [Qwen2-VL](https://huggingface.co/Qwen)                           | 2B/7B/72B                        | qwen2_vl         |
+| [StarCoder 2](https://huggingface.co/bigcode)                     | 3B/7B/15B                        | -                |
+| [XVERSE](https://huggingface.co/xverse)                           | 7B/13B/65B                       | xverse           |
+| [Yi/Yi-1.5 (Code)](https://huggingface.co/01-ai)                  | 1.5B/6B/9B/34B                   | yi               |
+| [Yi-VL](https://huggingface.co/01-ai)                             | 6B/34B                           | yi_vl            |
+| [Yuan 2](https://huggingface.co/IEITYuan)                         | 2B/51B/102B                      | yuan             |
 
 > [!NOTE]
 > 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”（Instruct/Chat）模型请务必使用**对应的模板**。
@@ -203,6 +230,9 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 | ORPO 训练              | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | SimPO 训练             | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
+> [!TIP]
+> 有关 PPO 的实现细节，请参考[此博客](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html)。
+
 ## 数据集
 
 <details><summary>预训练数据集</summary>
@@ -259,7 +289,12 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 - [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
 - [STEM (zh)](https://huggingface.co/datasets/hfl/stem_zh_instruction)
 - [Ruozhiba (zh)](https://huggingface.co/datasets/hfl/ruozhiba_gpt4_turbo)
+- [Neo-sft (zh)](https://huggingface.co/datasets/m-a-p/neo_sft_phase2)
+- [WebInstructSub (en)](https://huggingface.co/datasets/TIGER-Lab/WebInstructSub)
+- [Magpie-Pro-300K-Filtered (en)](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered)
+- [Magpie-ultra-v0.1 (en)](https://huggingface.co/datasets/argilla/magpie-ultra-v0.1)
 - [LLaVA mixed (en&zh)](https://huggingface.co/datasets/BUAADreamer/llava-en-zh-300k)
+- [Pokemon-gpt4o-captions (en&zh)](https://huggingface.co/datasets/jugg1024/pokemon-gpt4o-captions)
 - [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)
@@ -276,6 +311,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 - [DPO mixed (en&zh)](https://huggingface.co/datasets/hiyouga/DPO-En-Zh-20k)
 - [UltraFeedback (en)](https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized)
+- [RLHF-V (en)](https://huggingface.co/datasets/openbmb/RLHF-V-Dataset)
+- [VLFeedback (en)](https://huggingface.co/datasets/Zhihui/VLFeedback)
 - [Orca DPO Pairs (en)](https://huggingface.co/datasets/Intel/orca_dpo_pairs)
 - [HH-RLHF (en)](https://huggingface.co/datasets/Anthropic/hh-rlhf)
 - [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
@@ -296,20 +333,20 @@ huggingface-cli login
 | 必需项       | 至少     | 推荐      |
 | ------------ | ------- | --------- |
 | python       | 3.8     | 3.11      |
-| torch        | 1.13.1  | 2.3.0     |
-| transformers | 4.41.2  | 4.41.2    |
-| datasets     | 2.16.0  | 2.19.2    |
-| accelerate   | 0.30.1  | 0.30.1    |
-| peft         | 0.11.1  | 0.11.1    |
-| trl          | 0.8.6   | 0.9.4     |
+| torch        | 1.13.1  | 2.4.0     |
+| transformers | 4.41.2  | 4.43.4    |
+| datasets     | 2.16.0  | 2.20.0    |
+| accelerate   | 0.30.1  | 0.32.0    |
+| peft         | 0.11.1  | 0.12.0    |
+| trl          | 0.8.6   | 0.9.6     |
 
 | 可选项       | 至少     | 推荐      |
 | ------------ | ------- | --------- |
 | CUDA         | 11.6    | 12.2      |
 | deepspeed    | 0.10.0  | 0.14.0    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.4.3     |
-| flash-attn   | 2.3.0   | 2.5.9     |
+| vllm         | 0.4.3   | 0.5.0     |
+| flash-attn   | 2.3.0   | 2.6.3     |
 
 ### 硬件依赖
 
@@ -335,10 +372,10 @@ huggingface-cli login
 ```bash
 git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
 cd LLaMA-Factory
-pip install -e '.[torch,metrics]'
+pip install -e ".[torch,metrics]"
 ```
 
-可选的额外依赖项：torch、torch_npu、metrics、deepspeed、bitsandbytes、vllm、galore、badam、gptq、awq、aqlm、qwen、modelscope、quality
+可选的额外依赖项：torch、torch-npu、metrics、deepspeed、liger-kernel、bitsandbytes、hqq、eetq、gptq、awq、aqlm、vllm、galore、badam、adam-mini、qwen、modelscope、openmind、quality
 
 > [!TIP]
 > 遇到包冲突时，可使用 `pip install --no-deps -e .` 解决。
@@ -357,9 +394,7 @@ pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/downl
 
 <details><summary>昇腾 NPU 用户指南</summary>
 
-加入 [NPU 用户群](assets/wechat_npu.jpg)。
-
-在昇腾 NPU 设备上安装 LLaMA Factory 时，需要指定额外依赖项，使用 `pip install -e '.[torch-npu,metrics]'` 命令安装。此外，还需要安装 **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**，安装方法请参考[安装教程](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/quickstart/quickstart/quickstart_18_0004.html)或使用以下命令：
+在昇腾 NPU 设备上安装 LLaMA Factory 时，需要指定额外依赖项，使用 `pip install -e ".[torch-npu,metrics]"` 命令安装。此外，还需要安装 **[Ascend CANN Toolkit 与 Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**，安装方法请参考[安装教程](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/quickstart/quickstart/quickstart_18_0004.html)或使用以下命令：
 
 ```bash
 # 请替换 URL 为 CANN 版本和设备型号对应的 URL
@@ -382,20 +417,17 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch-npu    | 2.1.0   | 2.1.0.post3 |
 | deepspeed    | 0.13.2  | 0.13.2      |
 
-Docker 镜像：
-
-- 32GB：[下载地址](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html)
-- 64GB：[下载地址](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
-
 请使用 `ASCEND_RT_VISIBLE_DEVICES` 而非 `CUDA_VISIBLE_DEVICES` 来指定运算设备。
 
 如果遇到无法正常推理的情况，请尝试设置 `do_sample: false`。
 
+下载预构建 Docker 镜像：[32GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html) | [64GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
+
 </details>
 
 ### 数据准备
 
-关于数据集文件的格式，请参考 [data/README_zh.md](data/README_zh.md) 的内容。你可以使用 HuggingFace / ModelScope 上的数据集或加载本地数据集。
+关于数据集文件的格式，请参考 [data/README_zh.md](data/README_zh.md) 的内容。你可以使用 HuggingFace / ModelScope / Modelers 上的数据集或加载本地数据集。
 
 > [!NOTE]
 > 使用自定义数据集时，请更新 `data/dataset_info.json` 文件。
@@ -405,9 +437,9 @@ Docker 镜像：
 下面三行命令分别对 Llama3-8B-Instruct 模型进行 LoRA **微调**、**推理**和**合并**。
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 高级用法请参考 [examples/README_zh.md](examples/README_zh.md)（包括多 GPU 微调）。
@@ -417,48 +449,148 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_s
 
 ### LLaMA Board 可视化微调（由 [Gradio](https://github.com/gradio-app/gradio) 驱动）
 
-#### 使用本地环境
-
 ```bash
-CUDA_VISIBLE_DEVICES=0 GRADIO_SHARE=1 llamafactory-cli webui
+llamafactory-cli webui
 ```
 
-#### 使用 Docker
+### 构建 Docker
+
+CUDA 用户：
 
 ```bash
-docker build -f ./Dockerfile -t llama-factory:latest .
-docker run --gpus=all \
-    -v ./hf_cache:/root/.cache/huggingface/ \
+cd docker/docker-cuda/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+昇腾 NPU 用户：
+
+```bash
+cd docker/docker-npu/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+AMD ROCm 用户：
+
+```bash
+cd docker/docker-rocm/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+<details><summary>不使用 Docker Compose 构建</summary>
+
+CUDA 用户：
+
+```bash
+docker build -f ./docker/docker-cuda/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit --gpus=all \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./om_cache:/root/.cache/openmind \
     -v ./data:/app/data \
     -v ./output:/app/output \
     -p 7860:7860 \
+    -p 8000:8000 \
     --shm-size 16G \
-    --name llama_factory \
-    -d llama-factory:latest
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
-#### 使用 Docker Compose
+昇腾 NPU 用户：
 
 ```bash
-docker compose -f ./docker-compose.yml up -d
+# 根据您的环境选择镜像
+docker build -f ./docker/docker-npu/Dockerfile \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+# 根据您的资源更改 `device`
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./om_cache:/root/.cache/openmind \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
+    -v /etc/ascend_install.info:/etc/ascend_install.info \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/davinci0 \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
+AMD ROCm 用户：
+
+```bash
+docker build -f ./docker/docker-rocm/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./om_cache:/root/.cache/openmind \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v ./saves:/app/saves \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/kfd \
+    --device /dev/dri \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
+```
+
+</details>
+
 <details><summary>数据卷详情</summary>
 
-- hf_cache：使用宿主机的 Hugging Face 缓存文件夹，允许更改为新的目录。
-- data：宿主机中存放数据集的文件夹路径。
-- output：将导出目录设置为该路径后，即可在宿主机中访问导出后的模型。
+- `hf_cache`：使用宿主机的 Hugging Face 缓存文件夹，允许更改为新的目录。
+- `ms_cache`：类似 Hugging Face 缓存文件夹，为 ModelScope 用户提供。
+- `om_cache`：类似 Hugging Face 缓存文件夹，为 Modelers 用户提供。
+- `data`：宿主机中存放数据集的文件夹路径。
+- `output`：将导出目录设置为该路径后，即可在宿主机中访问导出后的模型。
 
 </details>
 
 ### 利用 vLLM 部署 OpenAI API
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1 API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
+API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
 ```
 
 > [!TIP]
-> API 文档请查阅 https://platform.openai.com/docs/api-reference/chat/create。
+> API 文档请查阅[这里](https://platform.openai.com/docs/api-reference/chat/create)。
+>
+> 示例：[图像理解](scripts/test_image.py) | [工具调用](scripts/test_toolcall.py)
 
 ### 从魔搭社区下载
 
@@ -470,9 +602,19 @@ export USE_MODELSCOPE_HUB=1 # Windows 使用 `set USE_MODELSCOPE_HUB=1`
 
 将 `model_name_or_path` 设置为模型 ID 来加载对应的模型。在[魔搭社区](https://modelscope.cn/models)查看所有可用的模型，例如 `LLM-Research/Meta-Llama-3-8B-Instruct`。
 
+### 从魔乐社区下载
+
+您也可以通过下述方法，使用魔乐社区下载数据集和模型。
+
+```bash
+export USE_OPENMIND_HUB=1 # Windows 使用 `set USE_OPENMIND_HUB=1`
+```
+
+将 `model_name_or_path` 设置为模型 ID 来加载对应的模型。在[魔乐社区](https://modelers.cn/models)查看所有可用的模型，例如 `TeleAI/TeleChat-7B-pt`。
+
 ### 使用 W&B 面板
 
-若要使用 [Weights & Biases](https://wandb.ai) 记录实验数据，请添加下面的参数。
+若要使用 [Weights & Biases](https://wandb.ai) 记录实验数据，请在 yaml 文件中添加下面的参数。
 
 ```yaml
 report_to: wandb
@@ -492,45 +634,94 @@ run_name: test_run # 可选
 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. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. KDD 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. ACL 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. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. ACL 2024 Findings. [[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. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. ACL 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. Wu et al. Large Language Models are Parallel Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2403.09073)
 1. Zhang et al. EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling. 2024. [[arxiv]](https://arxiv.org/abs/2403.14541)
 1. Weller et al. FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2403.15246)
-1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
+1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. COLING 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
 1. Zan et al. CodeS: Natural Language to Code Repository via Multi-Layer Sketch. 2024. [[arxiv]](https://arxiv.org/abs/2403.16443)
 1. Liu et al. Extensive Self-Contrast Enables Feedback-Free Language Model Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2404.00604)
 1. Luo et al. BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.02827)
 1. Du et al. Chinese Tiny LLM: Pretraining a Chinese-Centric Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2404.04167)
-1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
+1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. ICML 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
 1. Liu et al. Dynamic Generation of Personalities with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.07084)
 1. Shang et al. How Far Have We Gone in Stripped Binary Code Understanding Using Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.09836)
 1. Huang et al. LLMTune: Accelerate Database Knob Tuning with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.11581)
 1. Deng et al. Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction. 2024. [[arxiv]](https://arxiv.org/abs/2404.14215)
 1. Acikgoz et al. Hippocrates: An Open-Source Framework for Advancing Large Language Models in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2404.16621)
-1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2404.17140)
-1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2404.17140)
+1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. NAACL 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Xu et al. Large Language Models for Cyber Security: A Systematic Literature Review. 2024. [[arxiv]](https://arxiv.org/abs/2405.04760)
+1. Dammu et al. "They are uncultured": Unveiling Covert Harms and Social Threats in LLM Generated Conversations. 2024. [[arxiv]](https://arxiv.org/abs/2405.05378)
+1. Yi et al. A safety realignment framework via subspace-oriented model fusion for large language models. 2024. [[arxiv]](https://arxiv.org/abs/2405.09055)
+1. Lou et al. SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling. 2024. [[arxiv]](https://arxiv.org/abs/2405.12739)
+1. Zhang et al. Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2405.13816)
+1. Zhang et al. TS-Align: A Teacher-Student Collaborative Framework for Scalable Iterative Finetuning of Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2405.20215)
+1. Zihong Chen. Sentence Segmentation and Sentence Punctuation Based on XunziALLM. 2024. [[paper]](https://aclanthology.org/2024.lt4hala-1.30)
+1. Gao et al. The Best of Both Worlds: Toward an Honest and Helpful Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2406.00380)
+1. Wang and Song. MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset. 2024. [[arxiv]](https://arxiv.org/abs/2406.02106)
+1. Hu et al. Computational Limits of Low-Rank Adaptation (LoRA) for Transformer-Based Models. 2024. [[arxiv]](https://arxiv.org/abs/2406.03136)
+1. Ge et al. Time Sensitive Knowledge Editing through Efficient Finetuning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2406.04496)
+1. Tan et al. Peer Review as A Multi-Turn and Long-Context Dialogue with Role-Based Interactions. 2024. [[arxiv]](https://arxiv.org/abs/2406.05688)
+1. Song et al. Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters. 2024. [[arxiv]](https://arxiv.org/abs/2406.05955)
+1. Gu et al. RWKV-CLIP: A Robust Vision-Language Representation Learner. 2024. [[arxiv]](https://arxiv.org/abs/2406.06973)
+1. Chen et al. Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees. 2024. [[arxiv]](https://arxiv.org/abs/2406.07115)
+1. Zhu et al. Are Large Language Models Good Statisticians?. 2024. [[arxiv]](https://arxiv.org/abs/2406.07815)
+1. Li et al. Know the Unknown: An Uncertainty-Sensitive Method for LLM Instruction Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2406.10099)
+1. Ding et al. IntentionQA: A Benchmark for Evaluating Purchase Intention Comprehension Abilities of Language Models in E-commerce. 2024. [[arxiv]](https://arxiv.org/abs/2406.10173)
+1. He et al. COMMUNITY-CROSS-INSTRUCT: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities. 2024. [[arxiv]](https://arxiv.org/abs/2406.12074)
+1. Lin et al. FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving. 2024. [[arxiv]](https://arxiv.org/abs/2406.14408)
+1. Treutlein et al. Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data. 2024. [[arxiv]](https://arxiv.org/abs/2406.14546)
+1. Feng et al. SS-Bench: A Benchmark for Social Story Generation and Evaluation. 2024. [[arxiv]](https://arxiv.org/abs/2406.15695)
+1. Feng et al. Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement. 2024. [[arxiv]](https://arxiv.org/abs/2406.17233)
+1. Liu et al. Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals. 2024. [[arxiv]](https://arxiv.org/abs/2406.18069)
+1. Iyer et al. Exploring Very Low-Resource Translation with LLMs: The University of Edinburgh's Submission to AmericasNLP 2024 Translation Task. AmericasNLP 2024. [[paper]](https://aclanthology.org/2024.americasnlp-1.25)
+1. Li et al. Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring. 2024. [[arxiv]](https://arxiv.org/abs/2406.19949)
+1. Yang et al. Financial Knowledge Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2407.00365)
+1. Lin et al. DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging. 2024. [[arxiv]](https://arxiv.org/abs/2407.01470)
+1. Bako et al. Evaluating the Semantic Profiling Abilities of LLMs for Natural Language Utterances in Data Visualization. 2024. [[arxiv]](https://arxiv.org/abs/2407.06129)
+1. Huang et al. RoLoRA: Fine-tuning Rotated Outlier-free LLMs for Effective Weight-Activation Quantization. 2024. [[arxiv]](https://arxiv.org/abs/2407.08044)
+1. Jiang et al. LLM-Collaboration on Automatic Science Journalism for the General Audience. 2024. [[arxiv]](https://arxiv.org/abs/2407.09756)
+1. Inouye et al. Applied Auto-tuning on LoRA Hyperparameters. 2024. [[paper]](https://scholarcommons.scu.edu/cseng_senior/272/)
+1. Qi et al. Research on Tibetan Tourism Viewpoints information generation system based on LLM. 2024. [[arxiv]](https://arxiv.org/abs/2407.13561)
+1. Xu et al. Course-Correction: Safety Alignment Using Synthetic Preferences. 2024. [[arxiv]](https://arxiv.org/abs/2407.16637)
+1. Sun et al. LAMBDA: A Large Model Based Data Agent. 2024. [[arxiv]](https://arxiv.org/abs/2407.17535)
+1. Zhu et al. CollectiveSFT: Scaling Large Language Models for Chinese Medical Benchmark with Collective Instructions in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2407.19705)
+1. Yu et al. Correcting Negative Bias in Large Language Models through Negative Attention Score Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2408.00137)
+1. Xie et al. The Power of Personalized Datasets: Advancing Chinese Composition Writing for Elementary School through Targeted Model Fine-Tuning. IALP 2024. [[paper]](https://www.asianlp.sg/conferences/ialp2024/proceedings/papers/IALP2024_P055.pdf)
+1. Liu et al. Instruct-Code-Llama: Improving Capabilities of Language Model in Competition Level Code Generation by Online Judge Feedback. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_11)
+1. Wang et al. Cybernetic Sentinels: Unveiling the Impact of Safety Data Selection on Model Security in Supervised Fine-Tuning. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_23)
+1. Xia et al. Understanding the Performance and Estimating the Cost of LLM Fine-Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2408.04693)
+1. Zeng et al. Perceive, Reflect, and Plan: Designing LLM Agent for Goal-Directed City Navigation without Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2408.04168)
+1. Xia et al. Using Pre-trained Language Model for Accurate ESG Prediction. FinNLP 2024. [[paper]](https://aclanthology.org/2024.finnlp-2.1/)
+1. Liang et al. I-SHEEP: Self-Alignment of LLM from Scratch through an Iterative Self-Enhancement Paradigm. 2024. [[arxiv]](https://arxiv.org/abs/2408.08072)
+1. Bai et al. Aligning Large Language Model with Direct Multi-Preference Optimization for Recommendation. CIKM 2024. [[paper]](https://dl.acm.org/doi/10.1145/3627673.3679611)
 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/X-D-Lab/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。
-1. **[Luminia-13B-v3](https://huggingface.co/Nekochu/Luminia-13B-v3)**：一个用于生成 Stable Diffusion 提示词的大型语言模型。[[🤗Demo]](https://huggingface.co/spaces/Nekochu/Luminia-13B_SD_Prompt)
+1. **[Luminia-13B-v3](https://huggingface.co/Nekochu/Luminia-13B-v3)**：一个用于生成 Stable Diffusion 提示词的大型语言模型。[[demo]](https://huggingface.co/spaces/Nekochu/Luminia-13B_SD_Prompt)
 1. **[Chinese-LLaVA-Med](https://github.com/BUAADreamer/Chinese-LLaVA-Med)**：中文多模态医学大模型，基于 LLaVA-1.5-7B 在中文多模态医疗数据上微调而得。
+1. **[AutoRE](https://github.com/THUDM/AutoRE)**：基于大语言模型的文档级关系抽取系统。
+1. **[NVIDIA RTX AI Toolkit](https://github.com/NVIDIA/RTX-AI-Toolkit)**：在 Windows 主机上利用英伟达 RTX 设备进行大型语言模型微调的开发包。
+1. **[LazyLLM](https://github.com/LazyAGI/LazyLLM)**：一个低代码构建多 Agent 大模型应用的开发工具，支持基于 LLaMA Factory 的模型微调.
+1. **[RAG-Retrieval](https://github.com/NLPJCL/RAG-Retrieval)**：一个全链路 RAG 检索模型微调、推理和蒸馏代码库。[[blog]](https://zhuanlan.zhihu.com/p/987727357)
 
 </details>
 
@@ -538,17 +729,19 @@ run_name: test_run # 可选
 
 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：[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) / [Command-R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [LLaMA-3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [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) / [Yi-1.5](LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+使用模型权重时，请遵循对应的模型协议：[Baichuan 2](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) / [Command R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [Index](https://huggingface.co/IndexTeam/Index-1.9B/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [MiniCPM](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) / [Mistral/Mixtral/Pixtral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder 2](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) / [Yi-1.5](LICENSE) / [Yuan 2](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
-@article{zheng2024llamafactory,
+@inproceedings{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},
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Zhangchi Feng and Yongqiang Ma},
+  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
+  address={Bangkok, Thailand},
+  publisher={Association for Computational Linguistics},
   year={2024},
   url={http://arxiv.org/abs/2403.13372}
 }

data/README.md

@@ -11,8 +11,9 @@ Currently we support datasets in **alpaca** and **sharegpt** format.
   "formatting": "the format of the dataset. (optional, default: alpaca, can be chosen from {alpaca, sharegpt})",
   "ranking": "whether the dataset is a preference dataset or not. (default: False)",
   "subset": "the name of the subset. (optional, default: None)",
+  "split": "the name of dataset split to be used. (optional, default: train)",
   "folder": "the name of the folder of the dataset repository on the Hugging Face hub. (optional, default: None)",
-  "num_samples": "the number of samples in the dataset used for training. (optional, default: None)",
+  "num_samples": "the number of samples in the dataset to be used. (optional, default: None)",
   "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)",
@@ -22,6 +23,7 @@ Currently we support datasets in **alpaca** and **sharegpt** format.
     "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)",
     "images": "the column name in the dataset containing the image inputs. (default: None)",
+    "videos": "the column name in the dataset containing the videos inputs. (default: None)",
     "chosen": "the column name in the dataset containing the chosen answers. (default: None)",
     "rejected": "the column name in the dataset containing the rejected answers. (default: None)",
     "kto_tag": "the column name in the dataset containing the kto tags. (default: None)"
@@ -106,7 +108,7 @@ Regarding the above dataset, the *dataset description* in `dataset_info.json` sh
 
 ### Preference Dataset
 
-Preference datasets are used for reward modeling, DPO training and ORPO training.
+Preference datasets are used for reward modeling, DPO training, ORPO and SimPO training.
 
 It requires a better response in `chosen` column and a worse response in `rejected` column.
 
@@ -138,67 +140,15 @@ Regarding the above dataset, the *dataset description* in `dataset_info.json` sh
 
 ### KTO Dataset
 
-- [Example dataset](kto_en_demo.json)
+An additional column `kto_tag` is required. Please refer to the [sharegpt](#sharegpt-format) format for details.
 
-KTO datasets require a extra `kto_tag` column containing the boolean human feedback.
+### Multimodal Image Dataset
 
-```json
-[
-  {
-    "instruction": "human instruction (required)",
-    "input": "human input (optional)",
-    "output": "model response (required)",
-    "kto_tag": "human feedback [true/false] (required)"
-  }
-]
-```
+An additional column `images` is required. Please refer to the [sharegpt](#sharegpt-format) format for details.
 
-Regarding the above dataset, the *dataset description* in `dataset_info.json` should be:
+### Multimodal Video Dataset
 
-```json
-"dataset_name": {
-  "file_name": "data.json",
-  "columns": {
-    "prompt": "instruction",
-    "query": "input",
-    "response": "output",
-    "kto_tag": "kto_tag"
-  }
-}
-```
-
-### Multimodal Dataset
-
-- [Example dataset](mllm_demo.json)
-
-Multimodal datasets require a `images` column containing the paths to the input images. Currently we only support one image.
-
-```json
-[
-  {
-    "instruction": "human instruction (required)",
-    "input": "human input (optional)",
-    "output": "model response (required)",
-    "images": [
-      "image path (required)"
-    ]
-  }
-]
-```
-
-Regarding the above dataset, the *dataset description* in `dataset_info.json` should be:
-
-```json
-"dataset_name": {
-  "file_name": "data.json",
-  "columns": {
-    "prompt": "instruction",
-    "query": "input",
-    "response": "output",
-    "images": "images"
-  }
-}
-```
+An additional column `videos` is required. Please refer to the [sharegpt](#sharegpt-format) format for details.
 
 ## Sharegpt Format
 
@@ -251,6 +201,10 @@ Regarding the above dataset, the *dataset description* in `dataset_info.json` sh
 }
 ```
 
+### Pre-training Dataset
+
+Not yet supported, please use the [alpaca](#alpaca-format) format.
+
 ### Preference Dataset
 
 - [Example dataset](dpo_en_demo.json)
@@ -301,6 +255,125 @@ Regarding the above dataset, the *dataset description* in `dataset_info.json` sh
 }
 ```
 
+### KTO Dataset
+
+- [Example dataset](kto_en_demo.json)
+
+KTO datasets require a extra `kto_tag` column containing the boolean human feedback.
+
+```json
+[
+  {
+    "conversations": [
+      {
+        "from": "human",
+        "value": "human instruction"
+      },
+      {
+        "from": "gpt",
+        "value": "model response"
+      }
+    ],
+    "kto_tag": "human feedback [true/false] (required)"
+  }
+]
+```
+
+Regarding the above dataset, the *dataset description* in `dataset_info.json` should be:
+
+```json
+"dataset_name": {
+  "file_name": "data.json",
+  "formatting": "sharegpt",
+  "columns": {
+    "messages": "conversations",
+    "kto_tag": "kto_tag"
+  }
+}
+```
+
+### Multimodal Image Dataset
+
+- [Example dataset](mllm_demo.json)
+
+Multimodal image datasets require a `images` column containing the paths to the input images.
+
+The number of images should be identical to the `<image>` tokens in the conversations.
+
+```json
+[
+  {
+    "conversations": [
+      {
+        "from": "human",
+        "value": "<image>human instruction"
+      },
+      {
+        "from": "gpt",
+        "value": "model response"
+      }
+    ],
+    "images": [
+      "image path (required)"
+    ]
+  }
+]
+```
+
+Regarding the above dataset, the *dataset description* in `dataset_info.json` should be:
+
+```json
+"dataset_name": {
+  "file_name": "data.json",
+  "formatting": "sharegpt",
+  "columns": {
+    "messages": "conversations",
+    "images": "images"
+  }
+}
+```
+
+### Multimodal Video Dataset
+
+- [Example dataset](mllm_video_demo.json)
+
+Multimodal video datasets require a `videos` column containing the paths to the input videos.
+
+The number of videos should be identical to the `<video>` tokens in the conversations.
+
+```json
+[
+  {
+    "conversations": [
+      {
+        "from": "human",
+        "value": "<video>human instruction"
+      },
+      {
+        "from": "gpt",
+        "value": "model response"
+      }
+    ],
+    "videos": [
+      "video path (required)"
+    ]
+  }
+]
+```
+
+Regarding the above dataset, the *dataset description* in `dataset_info.json` should be:
+
+```json
+"dataset_name": {
+  "file_name": "data.json",
+  "formatting": "sharegpt",
+  "columns": {
+    "messages": "conversations",
+    "videos": "videos"
+  }
+}
+```
+
 ### OpenAI Format
 
 The openai format is simply a special case of the sharegpt format, where the first message may be a system prompt.
@@ -344,7 +417,3 @@ Regarding the above dataset, the *dataset description* in `dataset_info.json` sh
   }
 }
 ```
-
-The KTO datasets and multimodal datasets in sharegpt format are similar to the alpaca format.
-
-Pre-training datasets are **incompatible** with the sharegpt format.

data/README_zh.md

@@ -11,8 +11,9 @@
   "formatting": "数据集格式（可选，默认：alpaca，可以为 alpaca 或 sharegpt）",
   "ranking": "是否为偏好数据集（可选，默认：False）",
   "subset": "数据集子集的名称（可选，默认：None）",
+  "split": "所使用的数据集切分（可选，默认：train）",
   "folder": "Hugging Face 仓库的文件夹名称（可选，默认：None）",
-  "num_samples": "该数据集中用于训练的样本数量。（可选，默认：None）",
+  "num_samples": "该数据集所使用的样本数量。（可选，默认：None）",
   "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）",
@@ -22,6 +23,7 @@
     "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）",
     "images": "数据集代表图像输入的表头名称（默认：None）",
+    "videos": "数据集代表视频输入的表头名称（默认：None）",
     "chosen": "数据集代表更优回答的表头名称（默认：None）",
     "rejected": "数据集代表更差回答的表头名称（默认：None）",
     "kto_tag": "数据集代表 KTO 标签的表头名称（默认：None）"
@@ -106,7 +108,7 @@
 
 ### 偏好数据集
 
-偏好数据集用于奖励模型训练、DPO 训练和 ORPO 训练。
+偏好数据集用于奖励模型训练、DPO 训练、ORPO 训练和 SimPO 训练。
 
 它需要在 `chosen` 列中提供更优的回答，并在 `rejected` 列中提供更差的回答。
 
@@ -138,67 +140,15 @@
 
 ### KTO 数据集
 
-- [样例数据集](kto_en_demo.json)
+KTO 数据集需要提供额外的 `kto_tag` 列。详情请参阅 [sharegpt](#sharegpt-格式)。
 
-KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人类反馈。
+### 多模态图像数据集
 
-```json
-[
-  {
-    "instruction": "人类指令（必填）",
-    "input": "人类输入（选填）",
-    "output": "模型回答（必填）",
-    "kto_tag": "人类反馈 [true/false]（必填）"
-  }
-]
-```
+多模态图像数据集需要提供额外的 `images` 列。详情请参阅 [sharegpt](#sharegpt-格式)。
 
-对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
+### 多模态视频数据集
 
-```json
-"数据集名称": {
-  "file_name": "data.json",
-  "columns": {
-    "prompt": "instruction",
-    "query": "input",
-    "response": "output",
-    "kto_tag": "kto_tag"
-  }
-}
-```
-
-### 多模态数据集
-
-- [样例数据集](mllm_demo.json)
-
-多模态数据集需要额外添加一个 `images` 列，包含输入图像的路径。目前我们仅支持单张图像输入。
-
-```json
-[
-  {
-    "instruction": "人类指令（必填）",
-    "input": "人类输入（选填）",
-    "output": "模型回答（必填）",
-    "images": [
-      "图像路径（必填）"
-    ]
-  }
-]
-```
-
-对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
-
-```json
-"数据集名称": {
-  "file_name": "data.json",
-  "columns": {
-    "prompt": "instruction",
-    "query": "input",
-    "response": "output",
-    "images": "images"
-  }
-}
-```
+多模态视频数据集需要提供额外的 `videos` 列。详情请参阅 [sharegpt](#sharegpt-格式)。
 
 ## Sharegpt 格式
 
@@ -251,6 +201,10 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
 }
 ```
 
+### 预训练数据集
+
+尚不支持，请使用 [alpaca](#alpaca-格式) 格式。
+
 ### 偏好数据集
 
 - [样例数据集](dpo_zh_demo.json)
@@ -301,6 +255,125 @@ Sharegpt 格式的偏好数据集同样需要在 `chosen` 列中提供更优的
 }
 ```
 
+### KTO 数据集
+
+- [样例数据集](kto_en_demo.json)
+
+KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人类反馈。
+
+```json
+[
+  {
+    "conversations": [
+      {
+        "from": "human",
+        "value": "人类指令"
+      },
+      {
+        "from": "gpt",
+        "value": "模型回答"
+      }
+    ],
+    "kto_tag": "人类反馈 [true/false]（必填）"
+  }
+]
+```
+
+对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
+
+```json
+"数据集名称": {
+  "file_name": "data.json",
+  "formatting": "sharegpt",
+  "columns": {
+    "messages": "conversations",
+    "kto_tag": "kto_tag"
+  }
+}
+```
+
+### 多模态图像数据集
+
+- [样例数据集](mllm_demo.json)
+
+多模态图像数据集需要额外添加一个 `images` 列，包含输入图像的路径。
+
+注意图片的数量必须与文本中所有 `<image>` 标记的数量严格一致。
+
+```json
+[
+  {
+    "conversations": [
+      {
+        "from": "human",
+        "value": "<image>人类指令"
+      },
+      {
+        "from": "gpt",
+        "value": "模型回答"
+      }
+    ],
+    "images": [
+      "图像路径（必填）"
+    ]
+  }
+]
+```
+
+对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
+
+```json
+"数据集名称": {
+  "file_name": "data.json",
+  "formatting": "sharegpt",
+  "columns": {
+    "messages": "conversations",
+    "images": "images"
+  }
+}
+```
+
+### 多模态视频数据集
+
+- [样例数据集](mllm_video_demo.json)
+
+多模态视频数据集需要额外添加一个 `videos` 列，包含输入视频的路径。
+
+注意视频的数量必须与文本中所有 `<video>` 标记的数量严格一致。
+
+```json
+[
+  {
+    "conversations": [
+      {
+        "from": "human",
+        "value": "<video>人类指令"
+      },
+      {
+        "from": "gpt",
+        "value": "模型回答"
+      }
+    ],
+    "videos": [
+      "视频路径（必填）"
+    ]
+  }
+]
+```
+
+对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
+
+```json
+"数据集名称": {
+  "file_name": "data.json",
+  "formatting": "sharegpt",
+  "columns": {
+    "messages": "conversations",
+    "videos": "videos"
+  }
+}
+```
+
 ### OpenAI 格式
 
 OpenAI 格式仅仅是 sharegpt 格式的一种特殊情况，其中第一条消息可能是系统提示词。
@@ -344,7 +417,3 @@ OpenAI 格式仅仅是 sharegpt 格式的一种特殊情况，其中第一条消
   }
 }
 ```
-
-Sharegpt 格式中的 KTO 数据集和多模态数据集与 alpaca 格式的类似。
-
-预训练数据集**不支持** sharegpt 格式。

data/belle_multiturn/belle_multiturn.py

@@ -17,9 +17,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "{}/datasets/BelleGroup/multiturn_chat_0.8M".format(_HF_ENDPOINT)
+_HOMEPAGE = f"{_HF_ENDPOINT}/datasets/BelleGroup/multiturn_chat_0.8M"
 _LICENSE = "gpl-3.0"
-_URL = "{}/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json".format(_HF_ENDPOINT)
+_URL = f"{_HF_ENDPOINT}/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json"
 
 
 class BelleMultiturn(datasets.GeneratorBasedBuilder):
@@ -38,7 +38,7 @@ class BelleMultiturn(datasets.GeneratorBasedBuilder):
         return [datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": file_path})]
 
     def _generate_examples(self, filepath: str):
-        with open(filepath, "r", encoding="utf-8") as f:
+        with open(filepath, encoding="utf-8") as f:
             for key, row in enumerate(f):
                 data = json.loads(row)
                 conversations = []

data/hh_rlhf_en/hh_rlhf_en.py

@@ -8,9 +8,9 @@ import datasets
 _HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
 _DESCRIPTION = "Human preference data about helpfulness and harmlessness."
 _CITATION = ""
-_HOMEPAGE = "{}/datasets/Anthropic/hh-rlhf".format(_HF_ENDPOINT)
+_HOMEPAGE = f"{_HF_ENDPOINT}/datasets/Anthropic/hh-rlhf"
 _LICENSE = "mit"
-_URL = "{}/datasets/Anthropic/hh-rlhf/resolve/main/".format(_HF_ENDPOINT)
+_URL = f"{_HF_ENDPOINT}/datasets/Anthropic/hh-rlhf/resolve/main/"
 _URLS = {
     "train": [
         _URL + "harmless-base/train.jsonl.gz",
@@ -53,7 +53,7 @@ class HhRlhfEn(datasets.GeneratorBasedBuilder):
     def _generate_examples(self, filepaths: List[str]):
         key = 0
         for filepath in filepaths:
-            with open(filepath, "r", encoding="utf-8") as f:
+            with open(filepath, encoding="utf-8") as f:
                 for row in f:
                     data = json.loads(row)
                     chosen = data["chosen"]

data/ultra_chat/ultra_chat.py

@@ -20,9 +20,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "{}/datasets/stingning/ultrachat".format(_HF_ENDPOINT)
+_HOMEPAGE = f"{_HF_ENDPOINT}/datasets/stingning/ultrachat"
 _LICENSE = "cc-by-nc-4.0"
-_BASE_DATA_URL = "{}/datasets/stingning/ultrachat/resolve/main/train_{{idx}}.jsonl".format(_HF_ENDPOINT)
+_BASE_DATA_URL = f"{_HF_ENDPOINT}/datasets/stingning/ultrachat/resolve/main/train_{{idx}}.jsonl"
 
 
 class UltraChat(datasets.GeneratorBasedBuilder):
@@ -42,7 +42,7 @@ class UltraChat(datasets.GeneratorBasedBuilder):
 
     def _generate_examples(self, filepaths: List[str]):
         for filepath in filepaths:
-            with open(filepath, "r", encoding="utf-8") as f:
+            with open(filepath, encoding="utf-8") as f:
                 for row in f:
                     try:
                         data = json.loads(row)

docker-compose.yml

@@ -1,23 +0,0 @@
-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

docker/docker-cuda/Dockerfile

@@ -0,0 +1,72 @@
+# Default use the NVIDIA official image with PyTorch 2.3.0
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/index.html
+ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:24.02-py3
+FROM ${BASE_IMAGE}
+
+# Define environments
+ENV MAX_JOBS=4
+ENV FLASH_ATTENTION_FORCE_BUILD=TRUE
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+# Define installation arguments
+ARG INSTALL_BNB=false
+ARG INSTALL_VLLM=false
+ARG INSTALL_DEEPSPEED=false
+ARG INSTALL_FLASHATTN=false
+ARG INSTALL_LIGER_KERNEL=false
+ARG INSTALL_HQQ=false
+ARG INSTALL_EETQ=false
+ARG PIP_INDEX=https://pypi.org/simple
+
+# Set the working directory
+WORKDIR /app
+
+# Install the requirements
+COPY requirements.txt /app
+RUN pip config set global.index-url "$PIP_INDEX" && \
+    pip config set global.extra-index-url "$PIP_INDEX" && \
+    python -m pip install --upgrade pip && \
+    python -m pip install -r requirements.txt
+
+# Copy the rest of the application into the image
+COPY . /app
+
+# Install the LLaMA Factory
+RUN EXTRA_PACKAGES="metrics"; \
+    if [ "$INSTALL_BNB" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},bitsandbytes"; \
+    fi; \
+    if [ "$INSTALL_VLLM" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},vllm"; \
+    fi; \
+    if [ "$INSTALL_DEEPSPEED" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},deepspeed"; \
+    fi; \
+    if [ "$INSTALL_LIGER_KERNEL" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},liger-kernel"; \
+    fi; \
+    if [ "$INSTALL_HQQ" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},hqq"; \
+    fi; \
+    if [ "$INSTALL_EETQ" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},eetq"; \
+    fi; \
+    pip install -e ".[$EXTRA_PACKAGES]"
+
+# Rebuild flash attention
+RUN pip uninstall -y transformer-engine flash-attn && \
+    if [ "$INSTALL_FLASHATTN" == "true" ]; then \
+        pip uninstall -y ninja && pip install ninja && \
+        pip install --no-cache-dir flash-attn --no-build-isolation; \
+    fi
+
+# Set up volumes
+VOLUME [ "/root/.cache/huggingface", "/root/.cache/modelscope", "/app/data", "/app/output" ]
+
+# Expose port 7860 for the LLaMA Board
+ENV GRADIO_SERVER_PORT 7860
+EXPOSE 7860
+
+# Expose port 8000 for the API service
+ENV API_PORT 8000
+EXPOSE 8000

docker/docker-cuda/docker-compose.yml

@@ -0,0 +1,37 @@
+services:
+  llamafactory:
+    build:
+      dockerfile: ./docker/docker-cuda/Dockerfile
+      context: ../..
+      args:
+        INSTALL_BNB: false
+        INSTALL_VLLM: false
+        INSTALL_DEEPSPEED: false
+        INSTALL_FLASHATTN: false
+        INSTALL_LIGER_KERNEL: false
+        INSTALL_HQQ: false
+        INSTALL_EETQ: false
+        PIP_INDEX: https://pypi.org/simple
+    container_name: llamafactory
+    volumes:
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../om_cache:/root/.cache/openmind
+      - ../../data:/app/data
+      - ../../output:/app/output
+    ports:
+      - "7860:7860"
+      - "8000:8000"
+    ipc: host
+    tty: true
+    shm_size: '16gb'
+    stdin_open: true
+    command: bash
+    deploy:
+      resources:
+        reservations:
+          devices:
+          - driver: nvidia
+            count: "all"
+            capabilities: [gpu]
+    restart: unless-stopped

docker/docker-npu/Dockerfile

@@ -0,0 +1,45 @@
+# Use the Ubuntu 22.04 image with CANN 8.0.rc1
+# More versions can be found at https://hub.docker.com/r/ascendai/cann/tags
+# FROM ascendai/cann:8.0.rc1-910-ubuntu22.04-py3.8
+FROM ascendai/cann:8.0.rc1-910b-ubuntu22.04-py3.8
+# FROM ascendai/cann:8.0.rc1-910-openeuler22.03-py3.8
+# FROM ascendai/cann:8.0.rc1-910b-openeuler22.03-py3.8
+
+# Define environments
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Define installation arguments
+ARG INSTALL_DEEPSPEED=false
+ARG PIP_INDEX=https://pypi.org/simple
+ARG TORCH_INDEX=https://download.pytorch.org/whl/cpu
+
+# Set the working directory
+WORKDIR /app
+
+# Install the requirements
+COPY requirements.txt /app
+RUN pip config set global.index-url "$PIP_INDEX" && \
+    pip config set global.extra-index-url "$TORCH_INDEX" && \
+    python -m pip install --upgrade pip && \
+    python -m pip install -r requirements.txt
+
+# Copy the rest of the application into the image
+COPY . /app
+
+# Install the LLaMA Factory
+RUN EXTRA_PACKAGES="torch-npu,metrics"; \
+    if [ "$INSTALL_DEEPSPEED" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},deepspeed"; \
+    fi; \
+    pip install -e ".[$EXTRA_PACKAGES]"
+
+# Set up volumes
+VOLUME [ "/root/.cache/huggingface", "/root/.cache/modelscope", "/app/data", "/app/output" ]
+
+# Expose port 7860 for the LLaMA Board
+ENV GRADIO_SERVER_PORT 7860
+EXPOSE 7860
+
+# Expose port 8000 for the API service
+ENV API_PORT 8000
+EXPOSE 8000

docker/docker-npu/docker-compose.yml

@@ -0,0 +1,33 @@
+services:
+  llamafactory:
+    build:
+      dockerfile: ./docker/docker-npu/Dockerfile
+      context: ../..
+      args:
+        INSTALL_DEEPSPEED: false
+        PIP_INDEX: https://pypi.org/simple
+    container_name: llamafactory
+    volumes:
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../om_cache:/root/.cache/openmind
+      - ../../data:/app/data
+      - ../../output:/app/output
+      - /usr/local/dcmi:/usr/local/dcmi
+      - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
+      - /usr/local/Ascend/driver:/usr/local/Ascend/driver
+      - /etc/ascend_install.info:/etc/ascend_install.info
+    ports:
+      - "7860:7860"
+      - "8000:8000"
+    ipc: host
+    tty: true
+    shm_size: '16gb'
+    stdin_open: true
+    command: bash
+    devices:
+      - /dev/davinci0
+      - /dev/davinci_manager
+      - /dev/devmm_svm
+      - /dev/hisi_hdc
+    restart: unless-stopped

docker/docker-rocm/Dockerfile

@@ -0,0 +1,65 @@
+FROM hardandheavy/transformers-rocm:2.2.0
+
+# Define environments
+ENV MAX_JOBS=4
+ENV FLASH_ATTENTION_FORCE_BUILD=TRUE
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+# Define installation arguments
+ARG INSTALL_BNB=false
+ARG INSTALL_VLLM=false
+ARG INSTALL_DEEPSPEED=false
+ARG INSTALL_FLASHATTN=false
+ARG INSTALL_LIGER_KERNEL=false
+ARG INSTALL_HQQ=false
+ARG PIP_INDEX=https://pypi.org/simple
+
+# Set the working directory
+WORKDIR /app
+
+# Install the requirements
+COPY requirements.txt /app
+RUN pip config set global.index-url "$PIP_INDEX" && \
+    pip config set global.extra-index-url "$PIP_INDEX" && \
+    python -m pip install --upgrade pip && \
+    python -m pip install -r requirements.txt
+
+# Copy the rest of the application into the image
+COPY . /app
+
+# Install the LLaMA Factory
+RUN EXTRA_PACKAGES="metrics"; \
+    if [ "$INSTALL_BNB" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},bitsandbytes"; \
+    fi; \
+    if [ "$INSTALL_VLLM" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},vllm"; \
+    fi; \
+    if [ "$INSTALL_DEEPSPEED" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},deepspeed"; \
+    fi; \
+    if [ "$INSTALL_LIGER_KERNEL" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},liger-kernel"; \
+    fi; \
+    if [ "$INSTALL_HQQ" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},hqq"; \
+    fi; \
+    pip install -e ".[$EXTRA_PACKAGES]"
+
+# Rebuild flash attention
+RUN pip uninstall -y transformer-engine flash-attn && \
+    if [ "$INSTALL_FLASHATTN" == "true" ]; then \
+        pip uninstall -y ninja && pip install ninja && \
+        pip install --no-cache-dir flash-attn --no-build-isolation; \
+    fi
+
+# Set up volumes
+VOLUME [ "/root/.cache/huggingface", "/root/.cache/modelscope", "/app/data", "/app/output" ]
+
+# Expose port 7860 for the LLaMA Board
+ENV GRADIO_SERVER_PORT 7860
+EXPOSE 7860
+
+# Expose port 8000 for the API service
+ENV API_PORT 8000
+EXPOSE 8000

docker/docker-rocm/docker-compose.yml

@@ -0,0 +1,33 @@
+services:
+  llamafactory:
+    build:
+      dockerfile: ./docker/docker-rocm/Dockerfile
+      context: ../..
+      args:
+        INSTALL_BNB: false
+        INSTALL_VLLM: false
+        INSTALL_DEEPSPEED: false
+        INSTALL_FLASHATTN: false
+        INSTALL_LIGER_KERNEL: false
+        INSTALL_HQQ: false
+        PIP_INDEX: https://pypi.org/simple
+    container_name: llamafactory
+    volumes:
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../om_cache:/root/.cache/openmind
+      - ../../data:/app/data
+      - ../../output:/app/output
+      - ../../saves:/app/saves
+    ports:
+      - "7860:7860"
+      - "8000:8000"
+    ipc: host
+    tty: true
+    shm_size: '16gb'
+    stdin_open: true
+    command: bash
+    devices:
+      - /dev/kfd:/dev/kfd
+      - /dev/dri:/dev/dri
+    restart: unless-stopped

evaluation/ceval/ceval.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 import os
 
 import datasets

evaluation/cmmlu/cmmlu.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 import os
 
 import datasets

evaluation/mmlu/mmlu.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 import os
 
 import datasets
@@ -157,5 +158,4 @@ class MMLU(datasets.GeneratorBasedBuilder):
         df = pd.read_csv(filepath, header=None)
         df.columns = ["question", "A", "B", "C", "D", "answer"]
 
-        for i, instance in enumerate(df.to_dict(orient="records")):
-            yield i, instance
+        yield from enumerate(df.to_dict(orient="records"))

examples/README.md

@@ -4,59 +4,66 @@ Make sure to execute these commands in the `LLaMA-Factory` directory.
 
 ## Table of Contents
 
-- [LoRA Fine-Tuning on A Single GPU](#lora-fine-tuning-on-a-single-gpu)
-- [QLoRA Fine-Tuning on a Single GPU](#qlora-fine-tuning-on-a-single-gpu)
-- [LoRA Fine-Tuning on Multiple GPUs](#lora-fine-tuning-on-multiple-gpus)
-- [LoRA Fine-Tuning on Multiple NPUs](#lora-fine-tuning-on-multiple-npus)
-- [Full-Parameter Fine-Tuning on Multiple GPUs](#full-parameter-fine-tuning-on-multiple-gpus)
+- [LoRA Fine-Tuning](#lora-fine-tuning)
+- [QLoRA Fine-Tuning](#qlora-fine-tuning)
+- [Full-Parameter Fine-Tuning](#full-parameter-fine-tuning)
 - [Merging LoRA Adapters and Quantization](#merging-lora-adapters-and-quantization)
 - [Inferring LoRA Fine-Tuned Models](#inferring-lora-fine-tuned-models)
 - [Extras](#extras)
 
+Use `CUDA_VISIBLE_DEVICES` (GPU) or `ASCEND_RT_VISIBLE_DEVICES` (NPU) to choose computing devices.
+
 ## Examples
 
-### LoRA Fine-Tuning on A Single GPU
+### LoRA Fine-Tuning
 
 #### (Continuous) Pre-Training
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml
 ```
 
 #### Supervised Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### Multimodal Supervised Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml
-```
-
-#### Reward Modeling
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml
-```
-
-#### PPO Training
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml
+llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
+llamafactory-cli train examples/train_lora/qwen2vl_lora_sft.yaml
 ```
 
 #### DPO/ORPO/SimPO Training
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
+```
+
+#### Multimodal DPO/ORPO/SimPO Training
+
+```bash
+llamafactory-cli train examples/train_lora/qwen2vl_lora_dpo.yaml
+```
+
+#### Reward Modeling
+
+```bash
+llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
+```
+
+#### PPO Training
+
+```bash
+llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
 ```
 
 #### KTO Training
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml
 ```
 
 #### Preprocess Dataset
@@ -64,95 +71,85 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lo
 It is useful for large dataset, use `tokenized_path` in config to load the preprocessed dataset.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml
+llamafactory-cli train examples/train_lora/llama3_preprocess.yaml
 ```
 
 #### Evaluating on MMLU/CMMLU/C-Eval Benchmarks
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml
+llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml
 ```
 
 #### Batch Predicting and Computing BLEU and ROUGE Scores
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml
-```
-
-### QLoRA Fine-Tuning on a Single GPU
-
-#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes Quantization (Recommended)
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml
-```
-
-#### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml
-```
-
-#### Supervised Fine-Tuning with 4-bit AWQ Quantization
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml
-```
-
-#### Supervised Fine-Tuning with 2-bit AQLM Quantization
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml
-```
-
-### LoRA Fine-Tuning on Multiple GPUs
-
-#### Supervised Fine-Tuning on Single Node
-
-```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
 ```
 
 #### Supervised Fine-Tuning on Multiple Nodes
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### Supervised Fine-Tuning with DeepSpeed ZeRO-3 (Weight Sharding)
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.yaml
 ```
 
-### LoRA Fine-Tuning on Multiple NPUs
+### QLoRA Fine-Tuning
 
-#### Supervised Fine-Tuning with DeepSpeed ZeRO-0
+#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes/HQQ/EETQ Quantization (Recommended)
 
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_npu/llama3_lora_sft_ds.yaml
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_otfq.yaml
 ```
 
-### Full-Parameter Fine-Tuning on Multiple GPUs
+#### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_gptq.yaml
+```
+
+#### Supervised Fine-Tuning with 4-bit AWQ Quantization
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_awq.yaml
+```
+
+#### Supervised Fine-Tuning with 2-bit AQLM Quantization
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_aqlm.yaml
+```
+
+### Full-Parameter Fine-Tuning
 
 #### Supervised Fine-Tuning on Single Node
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
 ```
 
 #### Supervised Fine-Tuning on Multiple Nodes
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
+```
+
+#### Multimodal Supervised Fine-Tuning
+
+```bash
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/qwen2vl_full_sft.yaml
 ```
 
 #### Batch Predicting and Computing BLEU and ROUGE Scores
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml
+llamafactory-cli train examples/train_full/llama3_full_predict.yaml
 ```
 
 ### Merging LoRA Adapters and Quantization
@@ -162,35 +159,33 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llam
 Note: DO NOT use quantized model or `quantization_bit` when merging LoRA adapters.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 #### Quantizing Model using AutoGPTQ
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
+llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
 ```
 
 ### Inferring LoRA Fine-Tuned Models
 
-Use `CUDA_VISIBLE_DEVICES=0,1` to infer models on multiple devices.
-
 #### Use CLI
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### Use Web UI
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### Launch OpenAI-style API
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml
+llamafactory-cli api examples/inference/llama3_lora_sft.yaml
 ```
 
 ### Extras
@@ -198,36 +193,48 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.y
 #### Full-Parameter Fine-Tuning using GaLore
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
 ```
 
 #### Full-Parameter Fine-Tuning using BAdam
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
+```
+
+#### Full-Parameter Fine-Tuning using Adam-mini
+
+```bash
+llamafactory-cli train examples/extras/adam_mini/qwen2_full_sft.yaml
 ```
 
 #### LoRA+ Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+```
+
+#### PiSSA Fine-Tuning
+
+```bash
+llamafactory-cli train examples/extras/pissa/llama3_lora_sft.yaml
 ```
 
 #### Mixture-of-Depths Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
 ```
 
 #### LLaMA-Pro Fine-Tuning
 
 ```bash
 bash examples/extras/llama_pro/expand.sh
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
+llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
 ```
 
 #### FSDP+QLoRA Fine-Tuning
 
 ```bash
-bash examples/extras/fsdp_qlora/single_node.sh
+bash examples/extras/fsdp_qlora/train.sh
 ```

examples/README_zh.md

@@ -4,59 +4,66 @@
 
 ## 目录
 
-- [单 GPU LoRA 微调](#单-gpu-lora-微调)
-- [单 GPU QLoRA 微调](#单-gpu-qlora-微调)
-- [多 GPU LoRA 微调](#多-gpu-lora-微调)
-- [多 NPU LoRA 微调](#多-npu-lora-微调)
-- [多 GPU 全参数微调](#多-gpu-全参数微调)
+- [LoRA 微调](#lora-微调)
+- [QLoRA 微调](#qlora-微调)
+- [全参数微调](#全参数微调)
 - [合并 LoRA 适配器与模型量化](#合并-lora-适配器与模型量化)
 - [推理 LoRA 模型](#推理-lora-模型)
 - [杂项](#杂项)
 
+使用 `CUDA_VISIBLE_DEVICES`（GPU）或 `ASCEND_RT_VISIBLE_DEVICES`（NPU）选择计算设备。
+
 ## 示例
 
-### 单 GPU LoRA 微调
+### LoRA 微调
 
 #### （增量）预训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml
 ```
 
 #### 指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### 多模态指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml
-```
-
-#### 奖励模型训练
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml
-```
-
-#### PPO 训练
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml
+llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
+llamafactory-cli train examples/train_lora/qwen2vl_lora_sft.yaml
 ```
 
 #### DPO/ORPO/SimPO 训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
+```
+
+#### 多模态 DPO/ORPO/SimPO 训练
+
+```bash
+llamafactory-cli train examples/train_lora/qwen2vl_lora_dpo.yaml
+```
+
+#### 奖励模型训练
+
+```bash
+llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
+```
+
+#### PPO 训练
+
+```bash
+llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
 ```
 
 #### KTO 训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml
 ```
 
 #### 预处理数据集
@@ -64,95 +71,85 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lo
 对于大数据集有帮助，在配置中使用 `tokenized_path` 以加载预处理后的数据集。
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml
+llamafactory-cli train examples/train_lora/llama3_preprocess.yaml
 ```
 
 #### 在 MMLU/CMMLU/C-Eval 上评估
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml
+llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml
 ```
 
 #### 批量预测并计算 BLEU 和 ROUGE 分数
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
 ```
 
-### 单 GPU QLoRA 微调
-
-#### 基于 4/8 比特 Bitsandbytes 量化进行指令监督微调（推荐）
+#### 多机指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml
-```
-
-#### 基于 4/8 比特 GPTQ 量化进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml
-```
-
-#### 基于 4 比特 AWQ 量化进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml
-```
-
-#### 基于 2 比特 AQLM 量化进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml
-```
-
-### 多 GPU LoRA 微调
-
-#### 在单机上进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
-```
-
-#### 在多机上进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### 使用 DeepSpeed ZeRO-3 平均分配显存
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.yaml
 ```
 
-### 多 NPU LoRA 微调
+### QLoRA 微调
 
-#### 使用 DeepSpeed ZeRO-0 进行指令监督微调
+#### 基于 4/8 比特 Bitsandbytes/HQQ/EETQ 量化进行指令监督微调（推荐）
 
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_npu/llama3_lora_sft_ds.yaml
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_otfq.yaml
 ```
 
-### 多 GPU 全参数微调
+#### 基于 4/8 比特 GPTQ 量化进行指令监督微调
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_gptq.yaml
+```
+
+#### 基于 4 比特 AWQ 量化进行指令监督微调
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_awq.yaml
+```
+
+#### 基于 2 比特 AQLM 量化进行指令监督微调
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_aqlm.yaml
+```
+
+### 全参数微调
 
 #### 在单机上进行指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
 ```
 
 #### 在多机上进行指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
+```
+
+#### 多模态指令监督微调
+
+```bash
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/qwen2vl_full_sft.yaml
 ```
 
 #### 批量预测并计算 BLEU 和 ROUGE 分数
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml
+llamafactory-cli train examples/train_full/llama3_full_predict.yaml
 ```
 
 ### 合并 LoRA 适配器与模型量化
@@ -162,35 +159,33 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llam
 注：请勿使用量化后的模型或 `quantization_bit` 参数来合并 LoRA 适配器。
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 #### 使用 AutoGPTQ 量化模型
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
+llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
 ```
 
 ### 推理 LoRA 模型
 
-使用 `CUDA_VISIBLE_DEVICES=0,1` 进行多卡推理。
-
 #### 使用命令行接口
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### 使用浏览器界面
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### 启动 OpenAI 风格 API
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml
+llamafactory-cli api examples/inference/llama3_lora_sft.yaml
 ```
 
 ### 杂项
@@ -198,36 +193,48 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.y
 #### 使用 GaLore 进行全参数训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
 ```
 
 #### 使用 BAdam 进行全参数训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
+```
+
+#### 使用 Adam-mini 进行全参数训练
+
+```bash
+llamafactory-cli train examples/extras/adam_mini/qwen2_full_sft.yaml
 ```
 
 #### LoRA+ 微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+```
+
+#### PiSSA 微调
+
+```bash
+llamafactory-cli train examples/extras/pissa/llama3_lora_sft.yaml
 ```
 
 #### 深度混合微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
 ```
 
 #### LLaMA-Pro 微调
 
 ```bash
 bash examples/extras/llama_pro/expand.sh
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
+llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
 ```
 
 #### FSDP+QLoRA 微调
 
 ```bash
-bash examples/extras/fsdp_qlora/single_node.sh
+bash examples/extras/fsdp_qlora/train.sh
 ```

examples/extras/adam_mini/qwen2_full_sft.yaml

@@ -0,0 +1,39 @@
+### model
+model_name_or_path: Qwen/Qwen2-1.5B-Instruct
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+use_adam_mini: true
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: qwen
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/qwen2-1_5b/full/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-5
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

examples/extras/badam/llama3_full_sft.yaml

@@ -6,14 +6,16 @@ stage: sft
 do_train: true
 finetuning_type: full
 use_badam: true
+badam_mode: layer
 badam_switch_mode: ascending
 badam_switch_interval: 50
 badam_verbose: 2
+# deepspeed: examples/deepspeed/ds_z3_config.json
 
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -28,11 +30,10 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 8
-learning_rate: 1.0e-4
+learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-pure_bf16: true
 
 ### eval
 val_size: 0.1

examples/extras/fsdp_qlora/llama3_lora_sft.yaml

@@ -8,13 +8,10 @@ do_train: true
 finetuning_type: lora
 lora_target: all
 
-### ddp
-ddp_timeout: 180000000
-
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -33,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/galore/llama3_full_sft.yaml

@@ -14,7 +14,7 @@ galore_scale: 2.0
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -29,11 +29,12 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 1
-learning_rate: 1.0e-4
+learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
 pure_bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/llama_pro/expand.sh

@@ -2,5 +2,5 @@
 
 python scripts/llama_pro.py \
     --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
-    --output_dir models/llama3-8b-instruct-pro \
+    --output_dir models/llama3-8b-pro \
     --num_expand 8

examples/extras/llama_pro/llama3_freeze_sft.yaml

@@ -1,5 +1,5 @@
 ### model
-model_name_or_path: models/llama3-8b-instruct-pro
+model_name_or_path: models/llama3-8b-pro
 
 ### method
 stage: sft
@@ -12,13 +12,13 @@ use_llama_pro: true
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b-instruct-pro/freeze/sft
+output_dir: saves/llama3-8b-pro/freeze/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -31,7 +31,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/loraplus/llama3_lora_sft.yaml

@@ -11,7 +11,7 @@ loraplus_lr_ratio: 16.0
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -30,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/mod/llama3_full_sft.yaml

@@ -10,7 +10,7 @@ mixture_of_depths: convert
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -26,11 +26,12 @@ overwrite_output_dir: true
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 8
 optim: paged_adamw_8bit
-learning_rate: 1.0e-4
+learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
 pure_bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/pissa/init.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+
+python scripts/pissa_init.py \
+    --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
+    --output_dir models/llama3-8b-pissa

examples/extras/pissa/llama3_lora_sft.yaml

@@ -1,17 +1,19 @@
 ### model
 model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
-quantization_bit: 4
 
 ### method
 stage: sft
 do_train: true
 finetuning_type: lora
 lora_target: all
+pissa_init: true
+pissa_iter: 16
+pissa_convert: true
 
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -30,7 +32,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/inference/llava1_5.yaml

@@ -0,0 +1,2 @@
+model_name_or_path: llava-hf/llava-1.5-7b-hf
+template: llava

examples/inference/qwen2_vl.yaml

@@ -0,0 +1,2 @@
+model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
+template: qwen2_vl

examples/merge_lora/qwen2vl_lora_sft.yaml

@@ -0,0 +1,13 @@
+### Note: DO NOT use quantized model or quantization_bit when merging lora adapters
+
+### model
+model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
+adapter_name_or_path: saves/qwen2_vl-7b/lora/sft
+template: qwen2_vl
+finetuning_type: lora
+
+### export
+export_dir: models/qwen2_vl_lora_sft
+export_size: 2
+export_device: cpu
+export_legacy_format: false

examples/train_full/llama3_full_predict.yaml

@@ -7,9 +7,9 @@ do_predict: true
 finetuning_type: full
 
 ### dataset
-dataset: identity,alpaca_en_demo
+eval_dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 50
 overwrite_cache: true
 preprocessing_num_workers: 16

examples/train_full/llama3_full_sft_ds3.yaml

@@ -5,15 +5,12 @@ model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 stage: sft
 do_train: true
 finetuning_type: full
-
-### ddp
-ddp_timeout: 180000000
 deepspeed: examples/deepspeed/ds_z3_config.json
 
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -28,11 +25,12 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 2
-learning_rate: 1.0e-4
+learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_full/qwen2vl_full_sft.yaml

@@ -0,0 +1,39 @@
+### model
+model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z3_config.json
+
+### dataset
+dataset: mllm_demo,identity
+template: qwen2_vl
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/qwen2_vl-7b/full/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 2
+learning_rate: 1.0e-5
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

examples/train_lora/llama3_lora_dpo.yaml

@@ -7,12 +7,12 @@ do_train: true
 finetuning_type: lora
 lora_target: all
 pref_beta: 0.1
-pref_loss: sigmoid  # [sigmoid (dpo), orpo, simpo]
+pref_loss: sigmoid  # choices: [sigmoid (dpo), orpo, simpo]
 
 ### dataset
 dataset: dpo_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -31,7 +31,8 @@ learning_rate: 5.0e-6
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_eval.yaml

@@ -6,8 +6,7 @@ adapter_name_or_path: saves/llama3-8b/lora/sft
 finetuning_type: lora
 
 ### dataset
-task: mmlu
-split: test
+task: mmlu_test  # choices: [mmlu_test, ceval_validation, cmmlu_test]
 template: fewshot
 lang: en
 n_shot: 5

examples/train_lora/llama3_lora_kto.yaml

@@ -6,11 +6,12 @@ stage: kto
 do_train: true
 finetuning_type: lora
 lora_target: all
+pref_beta: 0.1
 
 ### dataset
 dataset: kto_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -29,7 +30,8 @@ learning_rate: 5.0e-6
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_ppo.yaml

@@ -11,7 +11,7 @@ lora_target: all
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -30,7 +30,8 @@ learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### generate
 max_new_tokens: 512

examples/train_lora/llama3_lora_predict.yaml

@@ -8,9 +8,9 @@ do_predict: true
 finetuning_type: lora
 
 ### dataset
-dataset: identity,alpaca_en_demo
+eval_dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 50
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -22,3 +22,4 @@ overwrite_output_dir: true
 ### eval
 per_device_eval_batch_size: 1
 predict_with_generate: true
+ddp_timeout: 180000000

examples/train_lora/llama3_lora_pretrain.yaml

@@ -9,13 +9,13 @@ lora_target: all
 
 ### dataset
 dataset: c4_demo
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/lora/sft
+output_dir: saves/llama3-8b/lora/pretrain
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -28,7 +28,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_reward.yaml

@@ -10,7 +10,7 @@ lora_target: all
 ### dataset
 dataset: dpo_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -25,11 +25,12 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 8
-learning_rate: 1.0e-5
+learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_sft.yaml

@@ -10,7 +10,7 @@ lora_target: all
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_sft_ds0.yaml

@@ -6,15 +6,12 @@ stage: sft
 do_train: true
 finetuning_type: lora
 lora_target: all
-
-### ddp
-ddp_timeout: 180000000
 deepspeed: examples/deepspeed/ds_z0_config.json
 
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -33,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_sft_ds3.yaml

@@ -6,15 +6,12 @@ stage: sft
 do_train: true
 finetuning_type: lora
 lora_target: all
-
-### ddp
-ddp_timeout: 180000000
 deepspeed: examples/deepspeed/ds_z3_config.json
 
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -33,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_preprocess.yaml

@@ -10,7 +10,7 @@ lora_target: all
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16

examples/train_lora/llava1_5_lora_sft.yaml

@@ -1,6 +1,5 @@
 ### model
 model_name_or_path: llava-hf/llava-1.5-7b-hf
-visual_inputs: true
 
 ### method
 stage: sft
@@ -10,8 +9,8 @@ lora_target: all
 
 ### dataset
 dataset: mllm_demo
-template: vicuna
-cutoff_len: 1024
+template: llava
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -30,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/qwen2vl_lora_dpo.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
+
+### method
+stage: dpo
+do_train: true
+finetuning_type: lora
+lora_target: all
+pref_beta: 0.1
+pref_loss: sigmoid  # choices: [sigmoid (dpo), orpo, simpo]
+
+### dataset
+dataset: rlhf_v
+template: qwen2_vl
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/qwen2_vl-7b/lora/dpo
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 5.0e-6
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

examples/train_lora/qwen2vl_lora_sft.yaml

@@ -1,5 +1,5 @@
 ### model
-model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
 
 ### method
 stage: sft
@@ -7,19 +7,16 @@ do_train: true
 finetuning_type: lora
 lora_target: all
 
-### ddp
-ddp_timeout: 180000000
-
 ### dataset
-dataset: identity,alpaca_en_demo
-template: llama3
-cutoff_len: 1024
+dataset: mllm_demo,identity  # video: mllm_video_demo
+template: qwen2_vl
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/lora/sft
+output_dir: saves/qwen2_vl-7b/lora/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -27,12 +24,13 @@ overwrite_output_dir: true
 
 ### train
 per_device_train_batch_size: 1
-gradient_accumulation_steps: 2
+gradient_accumulation_steps: 8
 learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_aqlm.yaml

@@ -10,7 +10,7 @@ lora_target: all
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_awq.yaml

@@ -10,7 +10,7 @@ lora_target: all
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_gptq.yaml

@@ -10,7 +10,7 @@ lora_target: all
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_otfq.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+quantization_bit: 4
+quantization_method: bitsandbytes  # choices: [bitsandbytes (4/8), hqq (2/3/4/5/6/8), eetq (8)]
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

requirements.txt

@@ -1,9 +1,10 @@
-transformers>=4.41.2
-datasets>=2.16.0
-accelerate>=0.30.1
-peft>=0.11.1
-trl>=0.8.6
-gradio>=4.0.0
+transformers>=4.41.2,<=4.46.1
+datasets>=2.16.0,<=3.1.0
+accelerate>=0.34.0,<=1.0.1
+peft>=0.11.1,<=0.12.0
+trl>=0.8.6,<=0.9.6
+gradio>=4.0.0,<5.0.0
+pandas>=2.0.0
 scipy
 einops
 sentencepiece
@@ -17,3 +18,6 @@ matplotlib>=3.7.0
 fire
 packaging
 pyyaml
+numpy<2.0.0
+av
+tyro<0.9.0

scripts/cal_flops.py

@@ -1,7 +1,19 @@
-# coding=utf-8
-# Calculates the flops of pre-trained models.
-# 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/
+# Copyright 2024 Microsoft Corporation and the LlamaFactory team.
+#
+# This code is inspired by the Microsoft's DeepSpeed library.
+# https://www.deepspeed.ai/tutorials/flops-profiler/
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import fire
 import torch
@@ -14,14 +26,20 @@ from llamafactory.chat import ChatModel
 def calculate_flops(
     model_name_or_path: str,
     batch_size: int = 1,
-    seq_length: int = 256,
+    seq_length: int = 512,
     flash_attn: str = "auto",
 ):
+    r"""
+    Calculates the flops of pre-trained models.
+    Usage: python cal_flops.py --model_name_or_path path_to_model --batch_size 1 --seq_length 512
+    """
     with get_accelerator().device(0):
         chat_model = ChatModel(dict(model_name_or_path=model_name_or_path, template="empty", flash_attn=flash_attn))
-        fake_input = torch.ones((batch_size, seq_length), dtype=torch.long, device=chat_model.model.device)
+        fake_input = torch.ones((batch_size, seq_length), dtype=torch.long, device=chat_model.engine.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)
+        flops, macs, params = get_model_profile(
+            chat_model.engine.model, kwargs=input_dict, print_profile=True, detailed=True
+        )
         print("FLOPs:", flops)
         print("MACs:", macs)
         print("Params:", params)

scripts/cal_lr.py

@@ -1,7 +1,19 @@
-# 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
+# Copyright 2024 imoneoi and the LlamaFactory team.
+#
+# This code is inspired by the imoneoi's OpenChat library.
+# https://github.com/imoneoi/openchat/blob/3.6.0/ochat/training_deepspeed/train.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import math
 from typing import Literal
@@ -12,7 +24,7 @@ from torch.utils.data import DataLoader
 from tqdm import tqdm
 from transformers import DataCollatorForLanguageModeling, DataCollatorForSeq2Seq
 
-from llamafactory.data import get_dataset
+from llamafactory.data import get_dataset, get_template_and_fix_tokenizer
 from llamafactory.extras.constants import IGNORE_INDEX
 from llamafactory.hparams import get_train_args
 from llamafactory.model import load_tokenizer
@@ -26,12 +38,18 @@ def calculate_lr(
     model_name_or_path: str,
     batch_size: int,  # total batch size, namely (batch size * gradient accumulation * world size)
     stage: Literal["pt", "sft"] = "sft",
-    dataset: str = "alpaca_en",
+    dataset: str = "alpaca_en_demo",
     dataset_dir: str = "data",
     template: str = "default",
     cutoff_len: int = 1024,  # i.e. maximum input length during training
-    is_mistral: bool = False,  # mistral model uses a smaller learning rate,
+    is_mistral_or_gemma: bool = False,  # mistral and gemma models opt for a smaller learning rate,
+    packing: bool = False,
 ):
+    r"""
+    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_demo --cutoff_len 1024 --batch_size 16
+    """
     model_args, data_args, training_args, _, _ = get_train_args(
         dict(
             stage=stage,
@@ -40,19 +58,22 @@ def calculate_lr(
             dataset_dir=dataset_dir,
             template=template,
             cutoff_len=cutoff_len,
+            packing=packing,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    trainset = get_dataset(template, model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
     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
+        raise NotImplementedError(f"Stage does not supported: {stage}.")
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     valid_tokens, total_tokens = 0, 0
@@ -64,7 +85,7 @@ def calculate_lr(
     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
+    lr = lr / 6.0 if is_mistral_or_gemma 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

scripts/cal_mfu.py

@@ -0,0 +1,163 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+
+import fire
+import torch
+import torch.distributed as dist
+from transformers import AutoConfig
+
+from llamafactory.train.tuner import run_exp
+
+
+BASE = 2  # gemm (add + mul)
+
+
+def compute_model_flops(
+    model_name_or_path: str,
+    total_batch_size: int,
+    seq_length: int,
+    include_backward: bool = True,
+    include_recompute: bool = False,
+    include_flashattn: bool = False,
+) -> int:
+    r"""
+    Calculates the FLOPs of model per forward/backward pass.
+    """
+    config = AutoConfig.from_pretrained(model_name_or_path)
+    hidden_size = getattr(config, "hidden_size", None)
+    vocab_size = getattr(config, "vocab_size", None)
+    intermediate_size = getattr(config, "intermediate_size", None)
+    num_attention_heads = getattr(config, "num_attention_heads", None)
+    num_key_value_heads = getattr(config, "num_key_value_heads", None)
+    num_hidden_layers = getattr(config, "num_hidden_layers", None)
+    tie_word_embeddings = getattr(config, "tie_word_embeddings", False)
+
+    # mlp module
+    mlp_flops_per_token = 3 * BASE * hidden_size * intermediate_size  # up, gate, down
+    mlp_flops = total_batch_size * seq_length * num_hidden_layers * mlp_flops_per_token
+
+    # attn projector module
+    q_flops_per_token = BASE * hidden_size * hidden_size
+    o_flops_per_token = BASE * hidden_size * hidden_size
+    k_flops_per_token = BASE * hidden_size * hidden_size * num_key_value_heads // num_attention_heads
+    v_flops_per_token = BASE * hidden_size * hidden_size * num_key_value_heads // num_attention_heads
+    attn_proj_flops_per_token = q_flops_per_token + o_flops_per_token + k_flops_per_token + v_flops_per_token
+    attn_proj_flops = total_batch_size * seq_length * num_hidden_layers * attn_proj_flops_per_token
+
+    # attn sdpa module
+    sdpa_flops_per_layer = 2 * BASE * hidden_size * seq_length * seq_length  # (q * k^T) * v
+    sdpa_flops = total_batch_size * num_hidden_layers * sdpa_flops_per_layer
+
+    # embedding module
+    embedding_flops_per_token = hidden_size * vocab_size
+    embedding_flops = total_batch_size * seq_length * embedding_flops_per_token
+    if tie_word_embeddings is False:
+        embedding_flops *= 2
+
+    non_embedding_flops = mlp_flops + attn_proj_flops + sdpa_flops
+    non_embedding_coeff, embedding_coeff = 1, 1
+    if include_backward:
+        non_embedding_coeff += 2
+        embedding_coeff += 2
+
+    if include_recompute:
+        non_embedding_coeff += 1
+
+    total_flops = non_embedding_coeff * non_embedding_flops + embedding_coeff * embedding_flops
+
+    if include_flashattn:
+        total_flops += sdpa_flops
+
+    return total_flops
+
+
+def compute_device_flops(world_size: int) -> float:
+    r"""
+    Calculates the FLOPs of the device capability per second.
+    """
+    device_name = torch.cuda.get_device_name()
+    if "H100" in device_name or "H800" in device_name:
+        return 989 * 1e12 * world_size
+    elif "A100" in device_name or "A800" in device_name:
+        return 312 * 1e12 * world_size
+    elif "V100" in device_name:
+        return 125 * 1e12 * world_size
+    elif "4090" in device_name:
+        return 98 * 1e12 * world_size
+    else:
+        raise NotImplementedError(f"Device not supported: {device_name}.")
+
+
+def calculate_mfu(
+    model_name_or_path: str,
+    batch_size: int = 1,
+    seq_length: int = 1024,
+    num_steps: int = 100,
+    finetuning_type: str = "lora",
+    flash_attn: str = "auto",
+    deepspeed_stage: int = 0,
+    disable_gc: bool = False,
+    liger_kernel: bool = False,
+    unsloth_gc: bool = False,
+) -> float:
+    r"""
+    Calculates MFU for given model and hyper-params.
+    Usage: python cal_mfu.py --model_name_or_path path_to_model --batch_size 1 --seq_length 1024
+    """
+    args = {
+        "model_name_or_path": model_name_or_path,
+        "flash_attn": flash_attn,
+        "disable_gradient_checkpointing": disable_gc,
+        "enable_liger_kernel": liger_kernel,
+        "use_unsloth_gc": unsloth_gc,
+        "stage": "pt",
+        "do_train": True,
+        "finetuning_type": finetuning_type,
+        "dataset": "c4_demo",
+        "cutoff_len": seq_length,
+        "output_dir": os.path.join("saves", "test_mfu"),
+        "logging_strategy": "no",
+        "save_strategy": "no",
+        "save_only_model": True,
+        "overwrite_output_dir": True,
+        "per_device_train_batch_size": batch_size,
+        "max_steps": num_steps,
+        "bf16": True,
+    }
+    if deepspeed_stage in [2, 3]:
+        args["deepspeed"] = f"examples/deepspeed/ds_z{deepspeed_stage}_config.json"
+
+    run_exp(args)
+    with open(os.path.join("saves", "test_mfu", "all_results.json"), encoding="utf-8") as f:
+        result = json.load(f)
+
+    if dist.is_initialized():
+        world_size = dist.get_world_size()
+    else:
+        world_size = 1
+
+    total_batch_size = batch_size * world_size
+    mfu_value = (
+        result["train_steps_per_second"]
+        * compute_model_flops(model_name_or_path, total_batch_size, seq_length)
+        / compute_device_flops(world_size)
+    )
+    print(f"MFU: {mfu_value * 100:.2f}%")
+
+
+if __name__ == "__main__":
+    fire.Fire(calculate_mfu)

scripts/cal_ppl.py

@@ -1,6 +1,16 @@
-# coding=utf-8
-# Calculates the ppl on the dataset of the pre-trained models.
-# Usage: python cal_ppl.py --model_name_or_path path_to_model --save_name ppl.json
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 from dataclasses import dataclass
@@ -12,7 +22,7 @@ from torch.utils.data import DataLoader
 from tqdm import tqdm
 from transformers import DataCollatorForLanguageModeling, DataCollatorForSeq2Seq
 
-from llamafactory.data import get_dataset
+from llamafactory.data import get_dataset, get_template_and_fix_tokenizer
 from llamafactory.extras.constants import IGNORE_INDEX
 from llamafactory.hparams import get_train_args
 from llamafactory.model import load_model, load_tokenizer
@@ -44,18 +54,22 @@ class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
         return super().__call__(chosen_features)
 
 
-def cal_ppl(
+def calculate_ppl(
     model_name_or_path: str,
     save_name: str,
     batch_size: int = 4,
     stage: Literal["pt", "sft", "rm"] = "sft",
-    dataset: str = "alpaca_en",
+    dataset: str = "alpaca_en_demo",
     dataset_dir: str = "data",
     template: str = "default",
     cutoff_len: int = 1024,
     max_samples: Optional[int] = None,
     train_on_prompt: bool = False,
 ):
+    r"""
+    Calculates the ppl on the dataset of the pre-trained models.
+    Usage: python cal_ppl.py --model_name_or_path path_to_model --dataset alpaca_en_demo --save_name ppl.json
+    """
     model_args, data_args, training_args, finetuning_args, _ = get_train_args(
         dict(
             stage=stage,
@@ -68,11 +82,13 @@ def cal_ppl(
             train_on_prompt=train_on_prompt,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    trainset = get_dataset(template, model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
     model = load_model(tokenizer, model_args, finetuning_args, is_trainable=False)
     if stage == "pt":
         data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
@@ -83,7 +99,7 @@ def cal_ppl(
             tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX, train_on_prompt=train_on_prompt
         )
     else:
-        raise NotImplementedError
+        raise NotImplementedError(f"Stage does not supported: {stage}.")
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     criterion = torch.nn.CrossEntropyLoss(reduction="none")
@@ -108,9 +124,9 @@ def cal_ppl(
     with open(save_name, "w", encoding="utf-8") as f:
         json.dump(perplexities, f, indent=2)
 
-    print("Average perplexity is {:.2f}".format(total_ppl / len(perplexities)))
-    print("Perplexities have been saved at {}.".format(save_name))
+    print(f"Average perplexity is {total_ppl / len(perplexities):.2f}")
+    print(f"Perplexities have been saved at {save_name}.")
 
 
 if __name__ == "__main__":
-    fire.Fire(cal_ppl)
+    fire.Fire(calculate_ppl)

scripts/length_cdf.py

@@ -1,24 +1,38 @@
-# 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
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 from collections import defaultdict
 
 import fire
 from tqdm import tqdm
 
-from llamafactory.data import get_dataset
+from llamafactory.data import get_dataset, get_template_and_fix_tokenizer
 from llamafactory.hparams import get_train_args
 from llamafactory.model import load_tokenizer
 
 
 def length_cdf(
     model_name_or_path: str,
-    dataset: str = "alpaca_en",
+    dataset: str = "alpaca_en_demo",
     dataset_dir: str = "data",
     template: str = "default",
     interval: int = 1000,
 ):
+    r"""
+    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_demo --template default
+    """
     model_args, data_args, training_args, _, _ = get_train_args(
         dict(
             stage="sft",
@@ -29,10 +43,12 @@ def length_cdf(
             cutoff_len=1_000_000,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
-    trainset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
+    template = get_template_and_fix_tokenizer(tokenizer_module["tokenizer"], data_args)
+    trainset = get_dataset(template, model_args, data_args, training_args, "sft", **tokenizer_module)["train_dataset"]
     total_num = len(trainset)
     length_dict = defaultdict(int)
     for sample in tqdm(trainset["input_ids"]):
@@ -44,7 +60,7 @@ def length_cdf(
     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))
+        print(f"{count_accu:d} ({prob_accu:.2f}%) samples have length < {length + interval}.")
 
 
 if __name__ == "__main__":

scripts/llama_pro.py

@@ -1,12 +1,24 @@
-# coding=utf-8
-# Performs block expansion for LLaMA, Mistral, Qwen1.5 or Yi 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
+# Copyright 2024 Tencent Inc. and the LlamaFactory team.
+#
+# This code is inspired by the Tencent's LLaMA-Pro library.
+# https://github.com/TencentARC/LLaMA-Pro/blob/main/scripts/block_expansion.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 import os
 from collections import OrderedDict
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING
 
 import fire
 import torch
@@ -27,16 +39,20 @@ if TYPE_CHECKING:
 
 
 def change_name(name: str, old_index: int, new_index: int) -> str:
-    return name.replace(".{:d}.".format(old_index), ".{:d}.".format(new_index))
+    return name.replace(f".{old_index:d}.", f".{new_index:d}.")
 
 
 def block_expansion(
     model_name_or_path: str,
     output_dir: str,
     num_expand: int,
-    shard_size: Optional[str] = "2GB",
-    save_safetensors: Optional[bool] = False,
+    shard_size: str = "2GB",
+    save_safetensors: bool = True,
 ):
+    r"""
+    Performs block expansion for LLaMA, Mistral, Qwen1.5 or Yi models.
+    Usage: python llama_pro.py --model_name_or_path meta-llama/Llama-2-7b-hf --output_dir llama2_pro --num_expand 8
+    """
     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)
@@ -59,27 +75,27 @@ def block_expansion(
     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))
+        raise ValueError(f"`num_layers` {num_layers} should be divisible by `num_expand` {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:
+            if f".{i:d}." in key:
                 output_state_dict[change_name(key, i, layer_cnt)] = value
 
-        print("Add layer {} copied from layer {}".format(layer_cnt, i))
+        print(f"Add layer {layer_cnt} copied from layer {i}")
         layer_cnt += 1
         if (i + 1) % split == 0:
             for key, value in state_dict.items():
-                if ".{:d}.".format(i) in key:
+                if f".{i:d}." 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))
+            print(f"Add layer {layer_cnt} expanded from layer {i}")
             layer_cnt += 1
 
     for key, value in state_dict.items():
@@ -96,17 +112,17 @@ def block_expansion(
             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(f"Model weights saved in {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(f"Model weights saved in {output_dir}")
 
-    print("Fine-tune this model with:")
-    print("model_name_or_path: {}".format(output_dir))
+    print("- Fine-tune this model with:")
+    print(f"model_name_or_path: {output_dir}")
     print("finetuning_type: freeze")
-    print("freeze_trainable_layers: {}".format(num_expand))
+    print(f"freeze_trainable_layers: {num_expand}")
     print("use_llama_pro: true")
 
 

scripts/llamafy_baichuan2.py

@@ -1,13 +1,21 @@
-# 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
-# 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
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 import os
 from collections import OrderedDict
-from typing import Any, Dict, Optional
+from typing import Any, Dict
 
 import fire
 import torch
@@ -54,16 +62,16 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso
             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(f"Model weights saved in {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(f"Model weights saved in {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:
+    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
         llama2_config_dict: Dict[str, Any] = json.load(f)
 
     llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
@@ -73,12 +81,20 @@ def save_config(input_dir: str, output_dir: str):
 
     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)))
+    print(f"Model config saved in {os.path.join(output_dir, CONFIG_NAME)}")
 
 
 def llamafy_baichuan2(
-    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
+    input_dir: str,
+    output_dir: str,
+    shard_size: str = "2GB",
+    save_safetensors: bool = True,
 ):
+    r"""
+    Converts the Baichuan2-7B model in the same format as LLaMA2-7B.
+    Usage: python llamafy_baichuan2.py --input_dir input --output_dir output
+    Converted model: https://huggingface.co/hiyouga/Baichuan2-7B-Base-LLaMAfied
+    """
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:

scripts/llamafy_qwen.py

@@ -1,12 +1,21 @@
-# coding=utf-8
-# Converts the Qwen models in the same format as LLaMA2.
-# Usage: python llamafy_qwen.py --input_dir input --output_dir output
-# Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 import os
 from collections import OrderedDict
-from typing import Any, Dict, Optional
+from typing import Any, Dict
 
 import fire
 import torch
@@ -76,7 +85,7 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso
             elif "lm_head" in key:
                 llama2_state_dict[key] = value
             else:
-                raise KeyError("Unable to process key {}".format(key))
+                raise KeyError(f"Unable to process key {key}")
 
     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)
@@ -88,18 +97,18 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso
             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(f"Model weights saved in {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(f"Model weights saved in {output_dir}")
 
     return str(torch_dtype).replace("torch.", "")
 
 
 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:
+    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
         qwen_config_dict: Dict[str, Any] = json.load(f)
 
     llama2_config_dict: Dict[str, Any] = OrderedDict()
@@ -125,12 +134,20 @@ def save_config(input_dir: str, output_dir: str, torch_dtype: str):
 
     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)))
+    print(f"Model config saved in {os.path.join(output_dir, CONFIG_NAME)}")
 
 
 def llamafy_qwen(
-    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
+    input_dir: str,
+    output_dir: str,
+    shard_size: str = "2GB",
+    save_safetensors: bool = False,
 ):
+    r"""
+    Converts the Qwen models in the same format as LLaMA2.
+    Usage: python llamafy_qwen.py --input_dir input --output_dir output
+    Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
+    """
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:

scripts/loftq_init.py

@@ -1,14 +1,24 @@
-# 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
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is based on the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.10.0/examples/loftq_finetuning/quantize_save_load.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import os
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING
 
 import fire
-import torch
-import torch.nn as nn
 from peft import LoftQConfig, LoraConfig, TaskType, get_peft_model
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
@@ -17,65 +27,61 @@ 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,
+    output_dir: str,
+    loftq_bits: int = 4,
+    loftq_iter: int = 4,
+    lora_alpha: int = None,
+    lora_rank: int = 16,
+    lora_dropout: float = 0,
+    lora_target: tuple = ("q_proj", "v_proj"),
+    save_safetensors: bool = True,
 ):
+    r"""
+    Initializes LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ)
+    Usage: python loftq_init.py --model_name_or_path path_to_model --output_dir output_dir
+    """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
     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(",")],
+        lora_dropout=lora_dropout,
+        target_modules=lora_target,
         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()
+    print("Initializing LoftQ weights, it may be take several minutes, wait patiently.")
+    peft_model = get_peft_model(model, lora_config)
+    loftq_dir = os.path.join(output_dir, "loftq_init")
 
     # 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)
+    setattr(peft_model.peft_config["default"], "base_model_name_or_path", os.path.abspath(output_dir))
+    setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply loftq again
+    peft_model.save_pretrained(loftq_dir, safe_serialization=save_safetensors)
+    print(f"Adapter weights saved in {loftq_dir}")
 
     # Save base model
-    unwrap_model(base_model)
-    base_model.save_pretrained(save_dir, safe_serialization=save_safetensors)
-    tokenizer.save_pretrained(save_dir)
+    base_model: "PreTrainedModel" = peft_model.unload()
+    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
+    tokenizer.save_pretrained(output_dir)
+    print(f"Model weights saved in {output_dir}")
+
+    print("- Fine-tune this model with:")
+    print(f"model_name_or_path: {output_dir}")
+    print(f"adapter_name_or_path: {loftq_dir}")
+    print("finetuning_type: lora")
+    print(f"quantization_bit: {loftq_bits}")
 
 
 if __name__ == "__main__":

scripts/pissa_init.py

@@ -0,0 +1,86 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is based on the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.11.0/examples/pissa_finetuning/preprocess.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import TYPE_CHECKING
+
+import fire
+from peft import LoraConfig, TaskType, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+
+def quantize_pissa(
+    model_name_or_path: str,
+    output_dir: str,
+    pissa_iter: int = 16,
+    lora_alpha: int = None,
+    lora_rank: int = 16,
+    lora_dropout: float = 0,
+    lora_target: tuple = ("q_proj", "v_proj"),
+    save_safetensors: bool = True,
+):
+    r"""
+    Initializes LoRA weights with Principal Singular values and Singular vectors Adaptation (PiSSA)
+    Usage: python pissa_init.py --model_name_or_path path_to_model --output_dir output_dir
+    """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
+    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")
+
+    lora_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        r=lora_rank,
+        lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
+        lora_dropout=lora_dropout,
+        target_modules=lora_target,
+        init_lora_weights="pissa" if pissa_iter == -1 else f"pissa_niter_{pissa_iter}",
+    )
+
+    # Init PiSSA model
+    peft_model = get_peft_model(model, lora_config)
+    pissa_dir = os.path.join(output_dir, "pissa_init")
+
+    # Save PiSSA model
+    setattr(peft_model.peft_config["default"], "base_model_name_or_path", os.path.abspath(output_dir))
+    setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply pissa again
+    peft_model.save_pretrained(pissa_dir, safe_serialization=save_safetensors)
+    print(f"Adapter weights saved in {pissa_dir}")
+
+    # Save base model
+    base_model: "PreTrainedModel" = peft_model.unload()
+    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
+    tokenizer.save_pretrained(output_dir)
+    print(f"Model weights saved in {output_dir}")
+
+    print("- Fine-tune this model with:")
+    print(f"model_name_or_path: {output_dir}")
+    print(f"adapter_name_or_path: {pissa_dir}")
+    print("finetuning_type: lora")
+    print("pissa_init: false")
+    print("pissa_convert: true")
+    print("- and optionally with:")
+    print("quantization_bit: 4")
+
+
+if __name__ == "__main__":
+    fire.Fire(quantize_pissa)

scripts/test_image.py

@@ -0,0 +1,65 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+from openai import OpenAI
+from transformers.utils.versions import require_version
+
+
+require_version("openai>=1.5.0", "To fix: pip install openai>=1.5.0")
+
+
+def main():
+    client = OpenAI(
+        api_key="{}".format(os.environ.get("API_KEY", "0")),
+        base_url="http://localhost:{}/v1".format(os.environ.get("API_PORT", 8000)),
+    )
+    messages = []
+    messages.append(
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Output the color and number of each box."},
+                {
+                    "type": "image_url",
+                    "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-VL/boxes.png"},
+                },
+            ],
+        }
+    )
+    result = client.chat.completions.create(messages=messages, model="test")
+    messages.append(result.choices[0].message)
+    print("Round 1:", result.choices[0].message.content)
+    # The image shows a pyramid of colored blocks with numbers on them. Here are the colors and numbers of ...
+    messages.append(
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What kind of flower is this?"},
+                {
+                    "type": "image_url",
+                    "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-VL/flowers.jpg"},
+                },
+            ],
+        }
+    )
+    result = client.chat.completions.create(messages=messages, model="test")
+    messages.append(result.choices[0].message)
+    print("Round 2:", result.choices[0].message.content)
+    # The image shows a cluster of forget-me-not flowers. Forget-me-nots are small ...
+
+
+if __name__ == "__main__":
+    main()

scripts/test_toolcall.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 import os
 from typing import Sequence

setup.py

@@ -1,39 +1,67 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 import re
+from typing import List
 
 from setuptools import find_packages, setup
 
 
-def get_version():
-    with open(os.path.join("src", "llamafactory", "extras", "env.py"), "r", encoding="utf-8") as f:
+def get_version() -> str:
+    with open(os.path.join("src", "llamafactory", "extras", "env.py"), encoding="utf-8") as f:
         file_content = f.read()
         pattern = r"{}\W*=\W*\"([^\"]+)\"".format("VERSION")
         (version,) = re.findall(pattern, file_content)
         return version
 
 
-def get_requires():
-    with open("requirements.txt", "r", encoding="utf-8") as f:
+def get_requires() -> List[str]:
+    with open("requirements.txt", encoding="utf-8") as f:
         file_content = f.read()
         lines = [line.strip() for line in file_content.strip().split("\n") if not line.startswith("#")]
         return lines
 
 
+def get_console_scripts() -> List[str]:
+    console_scripts = ["llamafactory-cli = llamafactory.cli:main"]
+    if os.environ.get("ENABLE_SHORT_CONSOLE", "1").lower() in ["true", "1"]:
+        console_scripts.append("lmf = llamafactory.cli:main")
+
+    return console_scripts
+
+
 extra_require = {
     "torch": ["torch>=1.13.1"],
     "torch-npu": ["torch==2.1.0", "torch-npu==2.1.0.post3", "decorator"],
     "metrics": ["nltk", "jieba", "rouge-chinese"],
-    "deepspeed": ["deepspeed>=0.10.0,<=0.14.0"],
+    "deepspeed": ["deepspeed>=0.10.0,<=0.14.4"],
+    "liger-kernel": ["liger-kernel"],
     "bitsandbytes": ["bitsandbytes>=0.39.0"],
-    "vllm": ["vllm>=0.4.3"],
-    "galore": ["galore-torch"],
-    "badam": ["badam"],
-    "gptq": ["optimum>=1.16.0", "auto-gptq>=0.5.0"],
+    "hqq": ["hqq"],
+    "eetq": ["eetq"],
+    "gptq": ["optimum>=1.17.0", "auto-gptq>=0.5.0"],
     "awq": ["autoawq"],
     "aqlm": ["aqlm[gpu]>=1.1.0"],
+    "vllm": ["vllm>=0.4.3,<0.6.4"],
+    "galore": ["galore-torch"],
+    "badam": ["badam>=1.2.1"],
+    "adam-mini": ["adam-mini"],
     "qwen": ["transformers_stream_generator"],
     "modelscope": ["modelscope"],
-    "dev": ["ruff", "pytest"],
+    "openmind": ["openmind"],
+    "dev": ["pre-commit", "ruff", "pytest"],
 }
 
 
@@ -44,7 +72,7 @@ def main():
         author="hiyouga",
         author_email="hiyouga" "@" "buaa.edu.cn",
         description="Easy-to-use LLM fine-tuning framework",
-        long_description=open("README.md", "r", encoding="utf-8").read(),
+        long_description=open("README.md", encoding="utf-8").read(),
         long_description_content_type="text/markdown",
         keywords=["LLaMA", "BLOOM", "Falcon", "LLM", "ChatGPT", "transformer", "pytorch", "deep learning"],
         license="Apache 2.0 License",
@@ -54,7 +82,7 @@ def main():
         python_requires=">=3.8.0",
         install_requires=get_requires(),
         extras_require=extra_require,
-        entry_points={"console_scripts": ["llamafactory-cli = llamafactory.cli:main"]},
+        entry_points={"console_scripts": get_console_scripts()},
         classifiers=[
             "Development Status :: 4 - Beta",
             "Intended Audience :: Developers",

src/api.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 
 import uvicorn
@@ -9,9 +23,9 @@ from llamafactory.chat import ChatModel
 def main():
     chat_model = ChatModel()
     app = create_app(chat_model)
-    api_host = os.environ.get("API_HOST", "0.0.0.0")
-    api_port = int(os.environ.get("API_PORT", "8000"))
-    print("Visit http://localhost:{}/docs for API document.".format(api_port))
+    api_host = os.getenv("API_HOST", "0.0.0.0")
+    api_port = int(os.getenv("API_PORT", "8000"))
+    print(f"Visit http://localhost:{api_port}/docs for API document.")
     uvicorn.run(app, host=api_host, port=api_port)
 
 

src/llamafactory/__init__.py

@@ -1,6 +1,47 @@
-# Level: api, webui > chat, eval, train > data, model > hparams > extras
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
-from .cli import VERSION
+r"""
+Efficient fine-tuning of large language models.
+
+Level:
+  api, webui > chat, eval, train > data, model > hparams > extras
+
+Dependency graph:
+  main:
+    transformers>=4.41.2,<=4.46.1
+    datasets>=2.16.0,<=3.1.0
+    accelerate>=0.34.0,<=1.0.1
+    peft>=0.11.1,<=0.12.0
+    trl>=0.8.6,<=0.9.6
+  attention:
+    transformers>=4.42.4 (gemma+fa2)
+  longlora:
+    transformers>=4.41.2,<=4.46.1
+  packing:
+    transformers>=4.41.2,<=4.46.1
+
+Disable version checking: DISABLE_VERSION_CHECK=1
+Enable VRAM recording: RECORD_VRAM=1
+Force check imports: FORCE_CHECK_IMPORTS=1
+Force using torchrun: FORCE_TORCHRUN=1
+Set logging verbosity: LLAMAFACTORY_VERBOSITY=WARN
+Use modelscope: USE_MODELSCOPE_HUB=1
+Use openmind: USE_OPENMIND_HUB=1
+"""
+
+from .extras.env import VERSION
 
 
 __version__ = VERSION

src/llamafactory/api/app.py

@@ -1,5 +1,21 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import asyncio
 import os
 from contextlib import asynccontextmanager
+from functools import partial
 from typing import Optional
 
 from typing_extensions import Annotated
@@ -36,14 +52,24 @@ if is_uvicorn_available():
     import uvicorn
 
 
+async def sweeper() -> None:
+    while True:
+        torch_gc()
+        await asyncio.sleep(300)
+
+
 @asynccontextmanager
-async def lifespan(app: "FastAPI"):  # collects GPU memory
+async def lifespan(app: "FastAPI", chat_model: "ChatModel"):  # collects GPU memory
+    if chat_model.engine_type == "huggingface":
+        asyncio.create_task(sweeper())
+
     yield
     torch_gc()
 
 
 def create_app(chat_model: "ChatModel") -> "FastAPI":
-    app = FastAPI(lifespan=lifespan)
+    root_path = os.getenv("FASTAPI_ROOT_PATH", "")
+    app = FastAPI(lifespan=partial(lifespan, chat_model=chat_model), root_path=root_path)
     app.add_middleware(
         CORSMiddleware,
         allow_origins=["*"],
@@ -51,7 +77,7 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
         allow_methods=["*"],
         allow_headers=["*"],
     )
-    api_key = os.environ.get("API_KEY")
+    api_key = os.getenv("API_KEY")
     security = HTTPBearer(auto_error=False)
 
     async def verify_api_key(auth: Annotated[Optional[HTTPAuthorizationCredentials], Depends(security)]):
@@ -65,7 +91,7 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
         dependencies=[Depends(verify_api_key)],
     )
     async def list_models():
-        model_card = ModelCard(id="gpt-3.5-turbo")
+        model_card = ModelCard(id=os.getenv("API_MODEL_NAME", "gpt-3.5-turbo"))
         return ModelList(data=[model_card])
 
     @app.post(
@@ -102,7 +128,7 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
 def run_api() -> None:
     chat_model = ChatModel()
     app = create_app(chat_model)
-    api_host = os.environ.get("API_HOST", "0.0.0.0")
-    api_port = int(os.environ.get("API_PORT", "8000"))
-    print("Visit http://localhost:{}/docs for API document.".format(api_port))
+    api_host = os.getenv("API_HOST", "0.0.0.0")
+    api_port = int(os.getenv("API_PORT", "8000"))
+    print(f"Visit http://localhost:{api_port}/docs for API document.")
     uvicorn.run(app, host=api_host, port=api_port)

src/llamafactory/api/chat.py

@@ -1,12 +1,27 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import base64
 import io
 import json
 import os
+import re
 import uuid
 from typing import TYPE_CHECKING, AsyncGenerator, Dict, List, Optional, Tuple
 
 from ..data import Role as DataRole
-from ..extras.logging import get_logger
+from ..extras import logging
 from ..extras.packages import is_fastapi_available, is_pillow_available, is_requests_available
 from .common import dictify, jsonify
 from .protocol import (
@@ -37,13 +52,12 @@ if is_requests_available():
 
 
 if TYPE_CHECKING:
-    from numpy.typing import NDArray
-
     from ..chat import ChatModel
+    from ..data.mm_plugin import ImageInput
     from .protocol import ChatCompletionRequest, ScoreEvaluationRequest
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 ROLE_MAPPING = {
     Role.USER: DataRole.USER.value,
     Role.ASSISTANT: DataRole.ASSISTANT.value,
@@ -55,8 +69,8 @@ ROLE_MAPPING = {
 
 def _process_request(
     request: "ChatCompletionRequest",
-) -> Tuple[List[Dict[str, str]], Optional[str], Optional[str], Optional["NDArray"]]:
-    logger.info("==== request ====\n{}".format(json.dumps(dictify(request), indent=2, ensure_ascii=False)))
+) -> Tuple[List[Dict[str, str]], Optional[str], Optional[str], Optional[List["ImageInput"]]]:
+    logger.info_rank0(f"==== request ====\n{json.dumps(dictify(request), indent=2, ensure_ascii=False)}")
 
     if len(request.messages) == 0:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid length")
@@ -70,7 +84,7 @@ def _process_request(
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Only supports u/a/u/a/u...")
 
     input_messages = []
-    image = None
+    images = []
     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")
@@ -78,9 +92,11 @@ def _process_request(
             raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
 
         if message.role == Role.ASSISTANT and isinstance(message.tool_calls, list) and len(message.tool_calls):
-            name = message.tool_calls[0].function.name
-            arguments = message.tool_calls[0].function.arguments
-            content = json.dumps({"name": name, "argument": arguments}, ensure_ascii=False)
+            tool_calls = [
+                {"name": tool_call.function.name, "arguments": tool_call.function.arguments}
+                for tool_call in message.tool_calls
+            ]
+            content = json.dumps(tool_calls, ensure_ascii=False)
             input_messages.append({"role": ROLE_MAPPING[Role.FUNCTION], "content": content})
         elif isinstance(message.content, list):
             for input_item in message.content:
@@ -88,15 +104,14 @@ def _process_request(
                     input_messages.append({"role": ROLE_MAPPING[message.role], "content": input_item.text})
                 else:
                     image_url = input_item.image_url.url
-                    if image_url.startswith("data:image"):  # base64 image
-                        image_data = base64.b64decode(image_url.split(",", maxsplit=1)[1])
-                        image_path = io.BytesIO(image_data)
+                    if re.match(r"^data:image\/(png|jpg|jpeg|gif|bmp);base64,(.+)$", image_url):  # base64 image
+                        image_stream = io.BytesIO(base64.b64decode(image_url.split(",", maxsplit=1)[1]))
                     elif os.path.isfile(image_url):  # local file
-                        image_path = open(image_url, "rb")
+                        image_stream = open(image_url, "rb")
                     else:  # web uri
-                        image_path = requests.get(image_url, stream=True).raw
+                        image_stream = requests.get(image_url, stream=True).raw
 
-                    image = Image.open(image_path).convert("RGB")
+                    images.append(Image.open(image_stream).convert("RGB"))
         else:
             input_messages.append({"role": ROLE_MAPPING[message.role], "content": message.content})
 
@@ -104,12 +119,12 @@ def _process_request(
     if isinstance(tool_list, list) and len(tool_list):
         try:
             tools = json.dumps([dictify(tool.function) for tool in tool_list], ensure_ascii=False)
-        except Exception:
+        except json.JSONDecodeError:
             raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid tools")
     else:
         tools = None
 
-    return input_messages, system, tools, image
+    return input_messages, system, tools, images or None
 
 
 def _create_stream_chat_completion_chunk(
@@ -127,13 +142,13 @@ def _create_stream_chat_completion_chunk(
 async def create_chat_completion_response(
     request: "ChatCompletionRequest", chat_model: "ChatModel"
 ) -> "ChatCompletionResponse":
-    completion_id = "chatcmpl-{}".format(uuid.uuid4().hex)
-    input_messages, system, tools, image = _process_request(request)
+    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
+    input_messages, system, tools, images = _process_request(request)
     responses = await chat_model.achat(
         input_messages,
         system,
         tools,
-        image,
+        images,
         do_sample=request.do_sample,
         temperature=request.temperature,
         top_p=request.top_p,
@@ -146,15 +161,17 @@ async def create_chat_completion_response(
     choices = []
     for i, response in enumerate(responses):
         if tools:
-            result = chat_model.engine.template.format_tools.extract(response.response_text)
+            result = chat_model.engine.template.extract_tool(response.response_text)
         else:
             result = response.response_text
 
-        if isinstance(result, tuple):
-            name, arguments = result
-            function = Function(name=name, arguments=arguments)
-            tool_call = FunctionCall(id="call_{}".format(uuid.uuid4().hex), function=function)
-            response_message = ChatCompletionMessage(role=Role.ASSISTANT, tool_calls=[tool_call])
+        if isinstance(result, list):
+            tool_calls = []
+            for tool in result:
+                function = Function(name=tool[0], arguments=tool[1])
+                tool_calls.append(FunctionCall(id=f"call_{uuid.uuid4().hex}", function=function))
+
+            response_message = ChatCompletionMessage(role=Role.ASSISTANT, tool_calls=tool_calls)
             finish_reason = Finish.TOOL
         else:
             response_message = ChatCompletionMessage(role=Role.ASSISTANT, content=result)
@@ -176,8 +193,8 @@ async def create_chat_completion_response(
 async def create_stream_chat_completion_response(
     request: "ChatCompletionRequest", chat_model: "ChatModel"
 ) -> AsyncGenerator[str, None]:
-    completion_id = "chatcmpl-{}".format(uuid.uuid4().hex)
-    input_messages, system, tools, image = _process_request(request)
+    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
+    input_messages, system, tools, images = _process_request(request)
     if tools:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Cannot stream function calls.")
 
@@ -191,7 +208,7 @@ async def create_stream_chat_completion_response(
         input_messages,
         system,
         tools,
-        image,
+        images,
         do_sample=request.do_sample,
         temperature=request.temperature,
         top_p=request.top_p,
@@ -212,8 +229,9 @@ async def create_stream_chat_completion_response(
 async def create_score_evaluation_response(
     request: "ScoreEvaluationRequest", chat_model: "ChatModel"
 ) -> "ScoreEvaluationResponse":
+    score_id = f"scoreval-{uuid.uuid4().hex}"
     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 ScoreEvaluationResponse(id=score_id, model=request.model, scores=scores)

src/llamafactory/api/common.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 from typing import TYPE_CHECKING, Any, Dict
 

src/llamafactory/api/protocol.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import time
 from enum import Enum, unique
 from typing import Any, Dict, List, Optional, Union
@@ -82,7 +96,7 @@ class ChatCompletionRequest(BaseModel):
     model: str
     messages: List[ChatMessage]
     tools: Optional[List[FunctionAvailable]] = None
-    do_sample: bool = True
+    do_sample: Optional[bool] = None
     temperature: Optional[float] = None
     top_p: Optional[float] = None
     n: int = 1

src/llamafactory/chat/__init__.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from .base_engine import BaseEngine
 from .chat_model import ChatModel
 

src/llamafactory/chat/base_engine.py

@@ -1,14 +1,28 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 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 numpy.typing import NDArray
     from transformers import PreTrainedModel, PreTrainedTokenizer
     from vllm import AsyncLLMEngine
 
     from ..data import Template
+    from ..data.mm_plugin import ImageInput, VideoInput
     from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
 
 
@@ -21,6 +35,12 @@ class Response:
 
 
 class BaseEngine(ABC):
+    r"""
+    Base class for inference engine of chat models.
+
+    Must implements async methods: chat(), stream_chat() and get_scores().
+    """
+
     model: Union["PreTrainedModel", "AsyncLLMEngine"]
     tokenizer: "PreTrainedTokenizer"
     can_generate: bool
@@ -34,12 +54,11 @@ class BaseEngine(ABC):
         data_args: "DataArguments",
         finetuning_args: "FinetuningArguments",
         generating_args: "GeneratingArguments",
-    ) -> None: ...
-
-    @abstractmethod
-    async def start(
-        self,
-    ) -> None: ...
+    ) -> None:
+        r"""
+        Initializes an inference engine.
+        """
+        ...
 
     @abstractmethod
     async def chat(
@@ -47,9 +66,14 @@ class BaseEngine(ABC):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
-    ) -> List["Response"]: ...
+    ) -> List["Response"]:
+        r"""
+        Gets a list of responses of the chat model.
+        """
+        ...
 
     @abstractmethod
     async def stream_chat(
@@ -57,13 +81,22 @@ class BaseEngine(ABC):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
-    ) -> AsyncGenerator[str, None]: ...
+    ) -> AsyncGenerator[str, None]:
+        r"""
+        Gets the response token-by-token of the chat model.
+        """
+        ...
 
     @abstractmethod
     async def get_scores(
         self,
         batch_input: List[str],
         **input_kwargs,
-    ) -> List[float]: ...
+    ) -> List[float]:
+        r"""
+        Gets a list of scores of the reward model.
+        """
+        ...

src/llamafactory/chat/chat_model.py

@@ -1,4 +1,22 @@
+# Copyright 2024 THUDM and the LlamaFactory team.
+#
+# This code is inspired by the THUDM's ChatGLM implementation.
+# https://github.com/THUDM/ChatGLM-6B/blob/main/cli_demo.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import asyncio
+import os
 from threading import Thread
 from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, Generator, List, Optional, Sequence
 
@@ -9,40 +27,53 @@ from .vllm_engine import VllmEngine
 
 
 if TYPE_CHECKING:
-    from numpy.typing import NDArray
-
+    from ..data.mm_plugin import ImageInput, VideoInput
     from .base_engine import BaseEngine, Response
 
 
-def _start_background_loop(loop: asyncio.AbstractEventLoop) -> None:
+def _start_background_loop(loop: "asyncio.AbstractEventLoop") -> None:
     asyncio.set_event_loop(loop)
     loop.run_forever()
 
 
 class ChatModel:
+    r"""
+    General class for chat models. Backed by huggingface or vllm engines.
+
+    Supports both sync and async methods.
+    Sync methods: chat(), stream_chat() and get_scores().
+    Async methods: achat(), astream_chat() and aget_scores().
+    """
+
     def __init__(self, args: Optional[Dict[str, Any]] = None) -> None:
         model_args, data_args, finetuning_args, generating_args = get_infer_args(args)
+        self.engine_type = model_args.infer_backend
         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))
+            raise NotImplementedError(f"Unknown backend: {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,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> List["Response"]:
-        task = asyncio.run_coroutine_threadsafe(self.achat(messages, system, tools, image, **input_kwargs), self._loop)
+        r"""
+        Gets a list of responses of the chat model.
+        """
+        task = asyncio.run_coroutine_threadsafe(
+            self.achat(messages, system, tools, images, videos, **input_kwargs), self._loop
+        )
         return task.result()
 
     async def achat(
@@ -50,20 +81,28 @@ class ChatModel:
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> List["Response"]:
-        return await self.engine.chat(messages, system, tools, image, **input_kwargs)
+        r"""
+        Asynchronously gets a list of responses of the chat model.
+        """
+        return await self.engine.chat(messages, system, tools, images, videos, **input_kwargs)
 
     def stream_chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> Generator[str, None, None]:
-        generator = self.astream_chat(messages, system, tools, image, **input_kwargs)
+        r"""
+        Gets the response token-by-token of the chat model.
+        """
+        generator = self.astream_chat(messages, system, tools, images, videos, **input_kwargs)
         while True:
             try:
                 task = asyncio.run_coroutine_threadsafe(generator.__anext__(), self._loop)
@@ -76,10 +115,14 @@ class ChatModel:
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> AsyncGenerator[str, None]:
-        async for new_token in self.engine.stream_chat(messages, system, tools, image, **input_kwargs):
+        r"""
+        Asynchronously gets the response token-by-token of the chat model.
+        """
+        async for new_token in self.engine.stream_chat(messages, system, tools, images, videos, **input_kwargs):
             yield new_token
 
     def get_scores(
@@ -87,6 +130,9 @@ class ChatModel:
         batch_input: List[str],
         **input_kwargs,
     ) -> List[float]:
+        r"""
+        Gets a list of scores of the reward model.
+        """
         task = asyncio.run_coroutine_threadsafe(self.aget_scores(batch_input, **input_kwargs), self._loop)
         return task.result()
 
@@ -95,17 +141,18 @@ class ChatModel:
         batch_input: List[str],
         **input_kwargs,
     ) -> List[float]:
+        r"""
+        Asynchronously gets a list of scores of the reward model.
+        """
         return await self.engine.get_scores(batch_input, **input_kwargs)
 
 
 def run_chat() -> None:
-    try:
-        import platform
-
-        if platform.system() != "Windows":
+    if os.name != "nt":
+        try:
             import readline  # noqa: F401
-    except ImportError:
-        print("Install `readline` for a better experience.")
+        except ImportError:
+            print("Install `readline` for a better experience.")
 
     chat_model = ChatModel()
     messages = []

src/llamafactory/chat/hf_engine.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import asyncio
 import concurrent.futures
 import os
@@ -6,25 +20,26 @@ from typing import TYPE_CHECKING, Any, AsyncGenerator, Callable, Dict, List, Opt
 
 import torch
 from transformers import GenerationConfig, TextIteratorStreamer
+from typing_extensions import override
 
 from ..data import get_template_and_fix_tokenizer
-from ..extras.logging import get_logger
+from ..extras import logging
+from ..extras.constants import IMAGE_PLACEHOLDER, VIDEO_PLACEHOLDER
 from ..extras.misc import get_logits_processor
 from ..model import load_model, load_tokenizer
 from .base_engine import BaseEngine, Response
 
 
 if TYPE_CHECKING:
-    from numpy.typing import NDArray
     from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
-    from transformers.image_processing_utils import BaseImageProcessor
     from trl import PreTrainedModelWrapper
 
     from ..data import Template
+    from ..data.mm_plugin import ImageInput, VideoInput
     from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 class HuggingfaceEngine(BaseEngine):
@@ -40,11 +55,19 @@ class HuggingfaceEngine(BaseEngine):
         self.tokenizer = tokenizer_module["tokenizer"]
         self.processor = tokenizer_module["processor"]
         self.tokenizer.padding_side = "left" if self.can_generate else "right"
-        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args)
         self.model = load_model(
             self.tokenizer, model_args, finetuning_args, is_trainable=False, add_valuehead=(not self.can_generate)
         )  # must after fixing tokenizer to resize vocab
         self.generating_args = generating_args.to_dict()
+        try:
+            asyncio.get_event_loop()
+        except RuntimeError:
+            logger.warning_once("There is no current event loop, creating a new one.")
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
+
+        self.semaphore = asyncio.Semaphore(int(os.getenv("MAX_CONCURRENT", "1")))
 
     @staticmethod
     def _process_args(
@@ -56,31 +79,30 @@ class HuggingfaceEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         input_kwargs: Optional[Dict[str, Any]] = {},
     ) -> Tuple[Dict[str, Any], int]:
-        if (
-            processor is not None
-            and image is not None
-            and not hasattr(processor, "image_seq_length")
-            and template.image_token not in messages[0]["content"]
-        ):  # llava-like models
-            messages[0]["content"] = template.image_token + messages[0]["content"]
+        mm_input_dict = {"images": [], "videos": [], "imglens": [0], "vidlens": [0]}
+        if images is not None:
+            mm_input_dict.update({"images": images, "imglens": [len(images)]})
+            if not any(IMAGE_PLACEHOLDER in message["content"] for message in messages):
+                messages[0]["content"] = IMAGE_PLACEHOLDER * len(images) + messages[0]["content"]
 
+        if videos is not None:
+            mm_input_dict.update({"videos": videos, "vidlens": [len(videos)]})
+            if not any(VIDEO_PLACEHOLDER in message["content"] for message in messages):
+                messages[0]["content"] = VIDEO_PLACEHOLDER * len(videos) + messages[0]["content"]
+
+        messages = template.mm_plugin.process_messages(
+            messages, mm_input_dict["images"], mm_input_dict["videos"], processor
+        )
         paired_messages = messages + [{"role": "assistant", "content": ""}]
         system = system or generating_args["default_system"]
-        pixel_values = None
-        prompt_ids, _ = template.encode_oneturn(
-            tokenizer=tokenizer, messages=paired_messages, system=system, tools=tools
+        prompt_ids, _ = template.encode_oneturn(tokenizer, paired_messages, system, tools)
+        prompt_ids, _ = template.mm_plugin.process_token_ids(
+            prompt_ids, None, mm_input_dict["images"], mm_input_dict["videos"], tokenizer, processor
         )
-        if processor is not None and image is not None:  # add image features
-            image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
-            batch_feature = image_processor(image, return_tensors="pt")
-            pixel_values = batch_feature.to(model.device)["pixel_values"]  # shape (B, C, H, W)
-            if hasattr(processor, "image_seq_length"):  # paligemma models
-                image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
-                prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
-
         prompt_length = len(prompt_ids)
         inputs = torch.tensor([prompt_ids], device=model.device)
         attention_mask = torch.ones_like(inputs, dtype=torch.bool)
@@ -97,7 +119,7 @@ class HuggingfaceEngine(BaseEngine):
         stop: Optional[Union[str, List[str]]] = input_kwargs.pop("stop", None)
 
         if stop is not None:
-            logger.warning("Stop parameter is not supported in Huggingface engine yet.")
+            logger.warning_rank0("Stop parameter is not supported by the huggingface engine yet.")
 
         generating_args = generating_args.copy()
         generating_args.update(
@@ -142,8 +164,14 @@ class HuggingfaceEngine(BaseEngine):
             logits_processor=get_logits_processor(),
         )
 
-        if pixel_values is not None:
-            gen_kwargs["pixel_values"] = pixel_values
+        mm_inputs = template.mm_plugin.get_mm_inputs(**mm_input_dict, batch_ids=[prompt_ids], processor=processor)
+        for key, value in mm_inputs.items():
+            if isinstance(value, list) and all(isinstance(v, torch.Tensor) for v in value):  # for pixtral inputs
+                value = torch.stack(value)  # assume they have same sizes
+            elif not isinstance(value, torch.Tensor):
+                value = torch.tensor(value)
+
+            gen_kwargs[key] = value.to(model.device)
 
         return gen_kwargs, prompt_length
 
@@ -158,11 +186,22 @@ class HuggingfaceEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         input_kwargs: Optional[Dict[str, Any]] = {},
     ) -> List["Response"]:
         gen_kwargs, prompt_length = HuggingfaceEngine._process_args(
-            model, tokenizer, processor, template, generating_args, messages, system, tools, image, input_kwargs
+            model,
+            tokenizer,
+            processor,
+            template,
+            generating_args,
+            messages,
+            system,
+            tools,
+            images,
+            videos,
+            input_kwargs,
         )
         generate_output = model.generate(**gen_kwargs)
         response_ids = generate_output[:, prompt_length:]
@@ -193,11 +232,22 @@ class HuggingfaceEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         input_kwargs: Optional[Dict[str, Any]] = {},
     ) -> Callable[[], str]:
         gen_kwargs, _ = HuggingfaceEngine._process_args(
-            model, tokenizer, processor, template, generating_args, messages, system, tools, image, input_kwargs
+            model,
+            tokenizer,
+            processor,
+            template,
+            generating_args,
+            messages,
+            system,
+            tools,
+            images,
+            videos,
+            input_kwargs,
         )
         streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
         gen_kwargs["streamer"] = streamer
@@ -220,40 +270,28 @@ class HuggingfaceEngine(BaseEngine):
         batch_input: List[str],
         input_kwargs: Optional[Dict[str, Any]] = {},
     ) -> List[float]:
-        max_length = input_kwargs.pop("max_length", None)
+        max_length: Optional[int] = input_kwargs.pop("max_length", None)
         device = getattr(model.pretrained_model, "device", "cuda")
-        inputs = tokenizer(
+        inputs: Dict[str, "torch.Tensor"] = 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,
+            add_special_tokens=False,
         ).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())
-
+        values: "torch.Tensor" = model(**inputs, return_dict=True, use_cache=False)[-1]
+        scores = values.gather(dim=-1, index=(inputs["attention_mask"].sum(dim=-1, keepdim=True) - 1))
         return scores
 
-    async def start(self) -> None:
-        self._semaphore = asyncio.Semaphore(int(os.environ.get("MAX_CONCURRENT", 1)))
-
+    @override
     async def chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> List["Response"]:
         if not self.can_generate:
@@ -269,19 +307,22 @@ class HuggingfaceEngine(BaseEngine):
             messages,
             system,
             tools,
-            image,
+            images,
+            videos,
             input_kwargs,
         )
-        async with self._semaphore:
+        async with self.semaphore:
             with concurrent.futures.ThreadPoolExecutor() as pool:
                 return await loop.run_in_executor(pool, self._chat, *input_args)
 
+    @override
     async def stream_chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> AsyncGenerator[str, None]:
         if not self.can_generate:
@@ -297,10 +338,11 @@ class HuggingfaceEngine(BaseEngine):
             messages,
             system,
             tools,
-            image,
+            images,
+            videos,
             input_kwargs,
         )
-        async with self._semaphore:
+        async with self.semaphore:
             with concurrent.futures.ThreadPoolExecutor() as pool:
                 stream = self._stream_chat(*input_args)
                 while True:
@@ -309,6 +351,7 @@ class HuggingfaceEngine(BaseEngine):
                     except StopAsyncIteration:
                         break
 
+    @override
     async def get_scores(
         self,
         batch_input: List[str],
@@ -319,6 +362,6 @@ class HuggingfaceEngine(BaseEngine):
 
         loop = asyncio.get_running_loop()
         input_args = (self.model, self.tokenizer, batch_input, input_kwargs)
-        async with self._semaphore:
+        async with self.semaphore:
             with concurrent.futures.ThreadPoolExecutor() as pool:
                 return await loop.run_in_executor(pool, self._get_scores, *input_args)

src/llamafactory/chat/vllm_engine.py

@@ -1,29 +1,49 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import uuid
-from typing import TYPE_CHECKING, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence, Union
+from typing import TYPE_CHECKING, Any, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence, Union
+
+from typing_extensions import override
 
 from ..data import get_template_and_fix_tokenizer
-from ..extras.logging import get_logger
+from ..extras import logging
+from ..extras.constants import IMAGE_PLACEHOLDER
 from ..extras.misc import get_device_count
-from ..extras.packages import is_vllm_available
+from ..extras.packages import is_pillow_available, is_vllm_available
 from ..model import load_config, load_tokenizer
+from ..model.model_utils.quantization import QuantizationMethod
 from ..model.model_utils.visual import LlavaMultiModalProjectorForYiVLForVLLM
 from .base_engine import BaseEngine, Response
 
 
+if is_pillow_available():
+    from PIL import Image
+    from PIL.Image import Image as ImageObject
+
+
 if is_vllm_available():
     from vllm import AsyncEngineArgs, AsyncLLMEngine, RequestOutput, SamplingParams
     from vllm.lora.request import LoRARequest
-    from vllm.sequence import MultiModalData
 
 
 if TYPE_CHECKING:
-    from numpy.typing import NDArray
-    from transformers.image_processing_utils import BaseImageProcessor
-
+    from ..data.mm_plugin import ImageInput, VideoInput
     from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 class VllmEngine(BaseEngine):
@@ -35,20 +55,25 @@ class VllmEngine(BaseEngine):
         generating_args: "GeneratingArguments",
     ) -> None:
         config = load_config(model_args)  # may download model from ms hub
+        if getattr(config, "quantization_config", None):  # gptq models should use float16
+            quantization_config: Dict[str, Any] = getattr(config, "quantization_config", None)
+            quant_method = quantization_config.get("quant_method", "")
+            if quant_method == QuantizationMethod.GPTQ and model_args.infer_dtype == "auto":
+                model_args.infer_dtype = "float16"
 
         self.can_generate = finetuning_args.stage == "sft"
         tokenizer_module = load_tokenizer(model_args)
         self.tokenizer = tokenizer_module["tokenizer"]
         self.processor = tokenizer_module["processor"]
         self.tokenizer.padding_side = "left"
-        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args)
         self.generating_args = generating_args.to_dict()
 
         engine_args = {
             "model": model_args.model_name_or_path,
             "trust_remote_code": True,
             "download_dir": model_args.cache_dir,
-            "dtype": model_args.vllm_dtype,
+            "dtype": model_args.infer_dtype,
             "max_model_len": model_args.vllm_maxlen,
             "tensor_parallel_size": get_device_count() or 1,
             "gpu_memory_utilization": model_args.vllm_gpu_util,
@@ -58,20 +83,14 @@ class VllmEngine(BaseEngine):
             "enable_lora": model_args.adapter_name_or_path is not None,
             "max_lora_rank": model_args.vllm_max_lora_rank,
         }
+        if isinstance(model_args.vllm_config, dict):
+            engine_args.update(model_args.vllm_config)
 
-        if model_args.visual_inputs:
-            image_size = config.vision_config.image_size
-            patch_size = config.vision_config.patch_size
-            self.image_feature_size = (image_size // patch_size) ** 2
-            engine_args["image_input_type"] = "pixel_values"
-            engine_args["image_token_id"] = self.tokenizer.convert_tokens_to_ids(self.template.image_token)
-            engine_args["image_input_shape"] = "1,3,{},{}".format(image_size, image_size)
-            engine_args["image_feature_size"] = self.image_feature_size
-            if getattr(config, "is_yi_vl_derived_model", None):
-                import vllm.model_executor.models.llava
+        if getattr(config, "is_yi_vl_derived_model", None):
+            import vllm.model_executor.models.llava
 
-                logger.info("Detected Yi-VL model, applying projector patch.")
-                vllm.model_executor.models.llava.LlavaMultiModalProjector = LlavaMultiModalProjectorForYiVLForVLLM
+            logger.info_rank0("Detected Yi-VL model, applying projector patch.")
+            vllm.model_executor.models.llava.LlavaMultiModalProjector = LlavaMultiModalProjectorForYiVLForVLLM
 
         self.model = AsyncLLMEngine.from_engine_args(AsyncEngineArgs(**engine_args))
         if model_args.adapter_name_or_path is not None:
@@ -84,35 +103,28 @@ class VllmEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> AsyncIterator["RequestOutput"]:
-        request_id = "chatcmpl-{}".format(uuid.uuid4().hex)
+        request_id = f"chatcmpl-{uuid.uuid4().hex}"
+        if images is not None:
+            if not any(IMAGE_PLACEHOLDER in message["content"] for message in messages):
+                messages[0]["content"] = IMAGE_PLACEHOLDER * len(images) + messages[0]["content"]
 
-        if (
-            self.processor is not None
-            and image is not None
-            and not hasattr(self.processor, "image_seq_length")
-            and self.template.image_token not in messages[0]["content"]
-        ):  # llava-like models (TODO: paligemma models)
-            messages[0]["content"] = self.template.image_token * self.image_feature_size + messages[0]["content"]
-
-        paired_messages = messages + [{"role": "assistant", "content": ""}]
-        system = system or self.generating_args["default_system"]
-        prompt_ids, _ = self.template.encode_oneturn(
-            tokenizer=self.tokenizer, messages=paired_messages, system=system, tools=tools
-        )
-
-        if self.processor is not None and image is not None:  # add image features
-            image_processor: "BaseImageProcessor" = getattr(self.processor, "image_processor")
-            pixel_values = image_processor(image, return_tensors="pt")["pixel_values"]
-            multi_modal_data = MultiModalData(type=MultiModalData.Type.IMAGE, data=pixel_values)
+        if self.template.mm_plugin.__class__.__name__ == "Qwen2vlPlugin":  # temporary solution
+            image_str = f"<|vision_start|>{self.template.mm_plugin.image_token}<|vision_end|>"
         else:
-            multi_modal_data = None
+            image_str = self.template.mm_plugin.image_token or ""
 
+        paired_messages = [
+            {"role": message["role"], "content": message["content"].replace(IMAGE_PLACEHOLDER, image_str)}
+            for message in messages
+        ] + [{"role": "assistant", "content": ""}]
+        system = system or self.generating_args["default_system"]
+        prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools)
         prompt_length = len(prompt_ids)
 
-        use_beam_search: bool = self.generating_args["num_beams"] > 1
         temperature: Optional[float] = input_kwargs.pop("temperature", None)
         top_p: Optional[float] = input_kwargs.pop("top_p", None)
         top_k: Optional[float] = input_kwargs.pop("top_k", None)
@@ -123,6 +135,9 @@ class VllmEngine(BaseEngine):
         max_new_tokens: Optional[int] = input_kwargs.pop("max_new_tokens", None)
         stop: Optional[Union[str, List[str]]] = input_kwargs.pop("stop", None)
 
+        if length_penalty is not None:
+            logger.warning_rank0("Length penalty is not supported by the vllm engine yet.")
+
         if "max_new_tokens" in self.generating_args:
             max_tokens = self.generating_args["max_new_tokens"]
         elif "max_length" in self.generating_args:
@@ -146,35 +161,47 @@ class VllmEngine(BaseEngine):
             temperature=temperature if temperature is not None else self.generating_args["temperature"],
             top_p=(top_p if top_p is not None else self.generating_args["top_p"]) or 1.0,  # top_p must > 0
             top_k=top_k if top_k is not None else self.generating_args["top_k"],
-            use_beam_search=use_beam_search,
-            length_penalty=length_penalty if length_penalty is not None else self.generating_args["length_penalty"],
             stop=stop,
             stop_token_ids=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
             max_tokens=max_tokens,
             skip_special_tokens=True,
         )
 
+        if images is not None:  # add image features
+            image_data = []
+            for image in images:
+                if not isinstance(image, (str, ImageObject)):
+                    raise ValueError(f"Expected image input is a path or PIL.Image, but got {type(image)}.")
+
+                if isinstance(image, str):
+                    image = Image.open(image).convert("RGB")
+
+                image_data.append(image)
+
+            multi_modal_data = {"image": image_data}
+        else:
+            multi_modal_data = None
+
         result_generator = self.model.generate(
-            inputs={"prompt_token_ids": prompt_ids, "multi_modal_data": multi_modal_data},
+            {"prompt_token_ids": prompt_ids, "multi_modal_data": multi_modal_data},
             sampling_params=sampling_params,
             request_id=request_id,
             lora_request=self.lora_request,
         )
         return result_generator
 
-    async def start(self) -> None:
-        pass
-
+    @override
     async def chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> List["Response"]:
         final_output = None
-        generator = await self._generate(messages, system, tools, image, **input_kwargs)
+        generator = await self._generate(messages, system, tools, images, videos, **input_kwargs)
         async for request_output in generator:
             final_output = request_output
 
@@ -191,21 +218,24 @@ class VllmEngine(BaseEngine):
 
         return results
 
+    @override
     async def stream_chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        images: Optional[Sequence["ImageInput"]] = None,
+        videos: Optional[Sequence["VideoInput"]] = None,
         **input_kwargs,
     ) -> AsyncGenerator[str, None]:
         generated_text = ""
-        generator = await self._generate(messages, system, tools, image, **input_kwargs)
+        generator = await self._generate(messages, system, tools, images, videos, **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
 
+    @override
     async def get_scores(
         self,
         batch_input: List[str],

src/llamafactory/cli.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 import random
 import subprocess
@@ -8,8 +22,8 @@ from . import launcher
 from .api.app import run_api
 from .chat.chat_model import run_chat
 from .eval.evaluator import run_eval
+from .extras import logging
 from .extras.env import VERSION, print_env
-from .extras.logging import get_logger
 from .extras.misc import get_device_count
 from .train.tuner import export_model, run_exp
 from .webui.interface import run_web_demo, run_web_ui
@@ -33,7 +47,7 @@ USAGE = (
 WELCOME = (
     "-" * 58
     + "\n"
-    + "| Welcome to LLaMA Factory, version {}".format(VERSION)
+    + f"| Welcome to LLaMA Factory, version {VERSION}"
     + " " * (21 - len(VERSION))
     + "|\n|"
     + " " * 56
@@ -42,7 +56,7 @@ WELCOME = (
     + "-" * 58
 )
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 @unique
@@ -60,7 +74,7 @@ class Command(str, Enum):
 
 
 def main():
-    command = sys.argv.pop(1)
+    command = sys.argv.pop(1) if len(sys.argv) != 1 else Command.HELP
     if command == Command.API:
         run_api()
     elif command == Command.CHAT:
@@ -72,26 +86,28 @@ def main():
     elif command == Command.EXPORT:
         export_model()
     elif command == Command.TRAIN:
-        force_torchrun = os.environ.get("FORCE_TORCHRUN", "0").lower() in ["true", "1"]
+        force_torchrun = os.getenv("FORCE_TORCHRUN", "0").lower() in ["true", "1"]
         if force_torchrun or get_device_count() > 1:
-            master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1")
-            master_port = os.environ.get("MASTER_PORT", str(random.randint(20001, 29999)))
-            logger.info("Initializing distributed tasks at: {}:{}".format(master_addr, master_port))
-            subprocess.run(
+            master_addr = os.getenv("MASTER_ADDR", "127.0.0.1")
+            master_port = os.getenv("MASTER_PORT", str(random.randint(20001, 29999)))
+            logger.info_rank0(f"Initializing distributed tasks at: {master_addr}:{master_port}")
+            process = subprocess.run(
                 (
                     "torchrun --nnodes {nnodes} --node_rank {node_rank} --nproc_per_node {nproc_per_node} "
                     "--master_addr {master_addr} --master_port {master_port} {file_name} {args}"
-                ).format(
-                    nnodes=os.environ.get("NNODES", "1"),
-                    node_rank=os.environ.get("RANK", "0"),
-                    nproc_per_node=os.environ.get("NPROC_PER_NODE", str(get_device_count())),
+                )
+                .format(
+                    nnodes=os.getenv("NNODES", "1"),
+                    node_rank=os.getenv("NODE_RANK", "0"),
+                    nproc_per_node=os.getenv("NPROC_PER_NODE", str(get_device_count())),
                     master_addr=master_addr,
                     master_port=master_port,
                     file_name=launcher.__file__,
                     args=" ".join(sys.argv[1:]),
-                ),
-                shell=True,
+                )
+                .split()
             )
+            sys.exit(process.returncode)
         else:
             run_exp()
     elif command == Command.WEBDEMO:
@@ -103,4 +119,4 @@ def main():
     elif command == Command.HELP:
         print(USAGE)
     else:
-        raise NotImplementedError("Unknown command: {}".format(command))
+        raise NotImplementedError(f"Unknown command: {command}.")

src/llamafactory/data/__init__.py

@@ -1,4 +1,23 @@
-from .collator import KTODataCollatorWithPadding, PairwiseDataCollatorWithPadding
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .collator import (
+    KTODataCollatorWithPadding,
+    MultiModalDataCollatorForSeq2Seq,
+    PairwiseDataCollatorWithPadding,
+    SFTDataCollatorWith4DAttentionMask,
+)
 from .data_utils import Role, split_dataset
 from .loader import get_dataset
 from .template import TEMPLATES, Template, get_template_and_fix_tokenizer
@@ -6,7 +25,9 @@ from .template import TEMPLATES, Template, get_template_and_fix_tokenizer
 
 __all__ = [
     "KTODataCollatorWithPadding",
+    "MultiModalDataCollatorForSeq2Seq",
     "PairwiseDataCollatorWithPadding",
+    "SFTDataCollatorWith4DAttentionMask",
     "Role",
     "split_dataset",
     "get_dataset",

src/llamafactory/data/aligner.py

@@ -1,99 +1,147 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 from functools import partial
-from typing import TYPE_CHECKING, Any, Dict, List, Union
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Union
 
-from datasets import Features
-
-from ..extras.logging import get_logger
+from ..extras import logging
 from .data_utils import Role
 
 
 if TYPE_CHECKING:
     from datasets import Dataset, IterableDataset
+    from transformers import Seq2SeqTrainingArguments
 
     from ..hparams import DataArguments
+    from .mm_plugin import ImageInput, VideoInput
     from .parser import DatasetAttr
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
-def _convert_images(images: List[Any], dataset_attr: "DatasetAttr", data_args: "DataArguments") -> List[Any]:
+def _convert_images(
+    images: Union["ImageInput", Sequence["ImageInput"]],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+) -> Optional[List["ImageInput"]]:
     r"""
     Optionally concatenates image path to dataset dir when loading from local disk.
     """
-    outputs = []
-    if dataset_attr.load_from in ["script", "file"]:
-        for image in images:
-            if isinstance(image, str) and os.path.isfile(os.path.join(data_args.dataset_dir, image)):
-                outputs.append(os.path.join(data_args.dataset_dir, image))
-            else:
-                outputs.append(image)
+    if not isinstance(images, list):
+        images = [images]
+    elif len(images) == 0:
+        return None
+    else:
+        images = images[:]
 
-    return outputs
+    if dataset_attr.load_from in ["script", "file"]:
+        for i in range(len(images)):
+            if isinstance(images[i], str) and os.path.isfile(os.path.join(data_args.image_dir, images[i])):
+                images[i] = os.path.join(data_args.image_dir, images[i])
+
+    return images
+
+
+def _convert_videos(
+    videos: Union["VideoInput", Sequence["VideoInput"]],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+) -> Optional[List["VideoInput"]]:
+    r"""
+    Optionally concatenates video path to dataset dir when loading from local disk.
+    """
+    if not isinstance(videos, list):
+        videos = [videos]
+    elif len(videos) == 0:
+        return None
+    else:
+        videos = videos[:]
+
+    if dataset_attr.load_from in ["script", "file"]:
+        for i in range(len(videos)):
+            if isinstance(videos[i], str) and os.path.isfile(os.path.join(data_args.image_dir, videos[i])):
+                videos[i] = os.path.join(data_args.image_dir, videos[i])
+
+    return videos
 
 
 def convert_alpaca(
-    examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr", data_args: "DataArguments"
-) -> Dict[str, List[Any]]:
+    example: Dict[str, Any],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+) -> Dict[str, Any]:
     r"""
     Converts alpaca format dataset to the standard format.
     """
-    outputs = {"prompt": [], "response": [], "system": [], "tools": [], "images": []}
+    prompt = []
+    if dataset_attr.history and isinstance(example[dataset_attr.history], list):
+        for old_prompt, old_response in example[dataset_attr.history]:
+            prompt.append({"role": Role.USER.value, "content": old_prompt})
+            prompt.append({"role": Role.ASSISTANT.value, "content": old_response})
+
+    query = []
+    if dataset_attr.prompt and example[dataset_attr.prompt]:
+        query.append(example[dataset_attr.prompt])
+
+    if dataset_attr.query and example[dataset_attr.query]:
+        query.append(example[dataset_attr.query])
+
+    prompt.append({"role": Role.USER.value, "content": "\n".join(query)})  # "prompt\nquery"
+
+    if dataset_attr.kto_tag and isinstance(example[dataset_attr.kto_tag], bool):  # kto example
+        response = [{"role": Role.ASSISTANT.value, "content": example[dataset_attr.response]}]
+        if example[dataset_attr.kto_tag]:
+            response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
+        else:
+            response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
+    elif (
+        dataset_attr.ranking
+        and isinstance(example[dataset_attr.chosen], str)
+        and isinstance(example[dataset_attr.rejected], str)
+    ):  # pairwise example
+        response = [
+            {"role": Role.ASSISTANT.value, "content": example[dataset_attr.chosen]},
+            {"role": Role.ASSISTANT.value, "content": example[dataset_attr.rejected]},
+        ]
+    elif dataset_attr.response and isinstance(example[dataset_attr.response], str):  # normal example
+        response = [{"role": Role.ASSISTANT.value, "content": example[dataset_attr.response]}]
+    else:  # unsupervised
+        response = []
+
     convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
-    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)})  # "prompt\nquery"
-
-        if dataset_attr.kto_tag and isinstance(examples[dataset_attr.kto_tag][i], bool):  # kto example
-            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
-            if examples[dataset_attr.kto_tag][i]:
-                response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
-            else:
-                response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
-        elif (
-            dataset_attr.ranking
-            and isinstance(examples[dataset_attr.chosen][i], str)
-            and isinstance(examples[dataset_attr.rejected][i], str)
-        ):  # pairwise example
-            response = [
-                {"role": Role.ASSISTANT.value, "content": examples[dataset_attr.chosen][i]},
-                {"role": Role.ASSISTANT.value, "content": examples[dataset_attr.rejected][i]},
-            ]
-        elif dataset_attr.response and isinstance(examples[dataset_attr.response][i], str):  # normal example
-            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
-        else:  # unsupervised
-            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(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
-        outputs["images"].append(convert_images(examples[dataset_attr.images][i]) if dataset_attr.images else [])
-
-    return outputs
+    convert_videos = partial(_convert_videos, dataset_attr=dataset_attr, data_args=data_args)
+    output = {
+        "_prompt": prompt,
+        "_response": response,
+        "_system": example[dataset_attr.system] if dataset_attr.system else "",
+        "_tools": example[dataset_attr.tools] if dataset_attr.tools else "",
+        "_images": convert_images(example[dataset_attr.images]) if dataset_attr.images else None,
+        "_videos": convert_videos(example[dataset_attr.videos]) if dataset_attr.videos else None,
+    }
+    return output
 
 
 def convert_sharegpt(
-    examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr", data_args: "DataArguments"
-) -> Dict[str, List[Any]]:
+    example: Dict[str, Any],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+) -> Dict[str, Any]:
     r"""
     Converts sharegpt format dataset to the standard format.
     """
-    outputs = {"prompt": [], "response": [], "system": [], "tools": [], "images": []}
-    convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
     tag_mapping = {
         dataset_attr.user_tag: Role.USER.value,
         dataset_attr.assistant_tag: Role.ASSISTANT.value,
@@ -104,86 +152,95 @@ def convert_sharegpt(
     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 = example[dataset_attr.messages]
+    if (
+        dataset_attr.system_tag
+        and len(messages) != 0
+        and messages[0][dataset_attr.role_tag] == dataset_attr.system_tag
+    ):
+        system = messages[0][dataset_attr.content_tag]
+        messages = messages[1:]
+    else:
+        system = example[dataset_attr.system] if dataset_attr.system else ""
 
-        if len(messages) == 0:
-            continue
-
-        aligned_messages = []
-        broken_data = False
-        for turn_idx, message in enumerate(messages):
-            if message[dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
-                logger.warning("Invalid role tag in {}.".format(messages))
-                broken_data = True
-
-            aligned_messages.append(
-                {"role": tag_mapping[message[dataset_attr.role_tag]], "content": message[dataset_attr.content_tag]}
-            )
-
-        if (not dataset_attr.ranking and len(aligned_messages) % 2 != 0) or (
-            dataset_attr.ranking and len(aligned_messages) % 2 == 0
-        ):
-            logger.warning("Invalid message count in {}.".format(messages))
+    aligned_messages = []
+    broken_data = False
+    for turn_idx, message in enumerate(messages):
+        if message[dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
+            logger.warning_rank0(f"Invalid role tag in {messages}.")
             broken_data = True
 
-        if dataset_attr.kto_tag and isinstance(examples[dataset_attr.kto_tag][i], bool):  # kto example
-            prompt = aligned_messages[:-1]
-            response = aligned_messages[-1:]
-            if examples[dataset_attr.kto_tag][i]:
-                response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
-            else:
-                response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
-        elif (
-            dataset_attr.ranking
-            and isinstance(examples[dataset_attr.chosen][i], dict)
-            and isinstance(examples[dataset_attr.rejected][i], dict)
-        ):  # pairwise example
-            chosen = examples[dataset_attr.chosen][i]
-            rejected = examples[dataset_attr.rejected][i]
-            if (
-                chosen[dataset_attr.role_tag] not in accept_tags[-1]
-                or rejected[dataset_attr.role_tag] not in accept_tags[-1]
-            ):
-                logger.warning("Invalid role tag in {}.".format([chosen, rejected]))
-                broken_data = True
+        aligned_messages.append(
+            {"role": tag_mapping[message[dataset_attr.role_tag]], "content": message[dataset_attr.content_tag]}
+        )
 
-            prompt = aligned_messages
-            response = [
-                {"role": tag_mapping[chosen[dataset_attr.role_tag]], "content": chosen[dataset_attr.content_tag]},
-                {"role": tag_mapping[rejected[dataset_attr.role_tag]], "content": rejected[dataset_attr.content_tag]},
-            ]
-        else:  # normal example
-            prompt = aligned_messages[:-1]
-            response = aligned_messages[-1:]
+    if (not dataset_attr.ranking and len(aligned_messages) % 2 != 0) or (
+        dataset_attr.ranking and len(aligned_messages) % 2 == 0
+    ):
+        logger.warning_rank0(f"Invalid message count in {messages}.")
+        broken_data = True
 
-        if broken_data:
-            logger.warning("Skipping this abnormal example.")
-            continue
+    if dataset_attr.kto_tag and isinstance(example[dataset_attr.kto_tag], bool):  # kto example
+        prompt = aligned_messages[:-1]
+        response = aligned_messages[-1:]
+        if example[dataset_attr.kto_tag]:
+            response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
+        else:
+            response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
+    elif (
+        dataset_attr.ranking
+        and isinstance(example[dataset_attr.chosen], dict)
+        and isinstance(example[dataset_attr.rejected], dict)
+    ):  # pairwise example
+        chosen = example[dataset_attr.chosen]
+        rejected = example[dataset_attr.rejected]
+        if (
+            chosen[dataset_attr.role_tag] not in accept_tags[-1]
+            or rejected[dataset_attr.role_tag] not in accept_tags[-1]
+        ):
+            logger.warning_rank0(f"Invalid role tag in {[chosen, rejected]}.")
+            broken_data = True
 
-        outputs["prompt"].append(prompt)
-        outputs["response"].append(response)
-        outputs["system"].append(system)
-        outputs["tools"].append(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
-        outputs["images"].append(convert_images(examples[dataset_attr.images][i]) if dataset_attr.images else [])
+        prompt = aligned_messages
+        response = [
+            {"role": tag_mapping[chosen[dataset_attr.role_tag]], "content": chosen[dataset_attr.content_tag]},
+            {"role": tag_mapping[rejected[dataset_attr.role_tag]], "content": rejected[dataset_attr.content_tag]},
+        ]
+    else:  # normal example
+        prompt = aligned_messages[:-1]
+        response = aligned_messages[-1:]
 
-    return outputs
+    if broken_data:
+        logger.warning_rank0("Skipping this abnormal example.")
+        prompt, response = [], []
+
+    convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
+    convert_videos = partial(_convert_videos, dataset_attr=dataset_attr, data_args=data_args)
+    output = {
+        "_prompt": prompt,
+        "_response": response,
+        "_system": system,
+        "_tools": example[dataset_attr.tools] if dataset_attr.tools else "",
+        "_images": convert_images(example[dataset_attr.images]) if dataset_attr.images else None,
+        "_videos": convert_videos(example[dataset_attr.videos]) if dataset_attr.videos else None,
+    }
+    return output
 
 
 def align_dataset(
-    dataset: Union["Dataset", "IterableDataset"], dataset_attr: "DatasetAttr", data_args: "DataArguments"
+    dataset: Union["Dataset", "IterableDataset"],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
 ) -> Union["Dataset", "IterableDataset"]:
     r"""
     Aligned dataset:
-        prompt: [{"role": "user", "content": "..."}] * (2T - 1)
-        response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
-        system: "..."
-        tools: "...",
-        images: [],
+        _prompt: [{"role": "user", "content": "..."}] * (2T - 1)
+        _response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
+        _system: "..."
+        _tools: "...",
+        _images: [],
+        _videos: [],
     """
     if dataset_attr.formatting == "alpaca":
         convert_func = partial(convert_alpaca, dataset_attr=dataset_attr, data_args=data_args)
@@ -191,31 +248,17 @@ def align_dataset(
         convert_func = partial(convert_sharegpt, dataset_attr=dataset_attr, data_args=data_args)
 
     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"},
-            "images": [{"_type": "Image"}],
-        }
-    )
     kwargs = {}
     if not data_args.streaming:
         kwargs = dict(
             num_proc=data_args.preprocessing_num_workers,
-            load_from_cache_file=(not data_args.overwrite_cache),
+            load_from_cache_file=(not data_args.overwrite_cache) or (training_args.local_process_index != 0),
             desc="Converting format of dataset",
         )
 
     return dataset.map(
         convert_func,
-        batched=True,
+        batched=False,
         remove_columns=column_names,
-        features=features,
         **kwargs,
     )

src/llamafactory/data/collator.py

@@ -1,17 +1,135 @@
+# Copyright 2024 OpenAccess AI Collective and the LlamaFactory team.
+#
+# This code is inspired by the OpenAccess AI Collective's axolotl library.
+# https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/monkeypatch/utils.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from dataclasses import dataclass
-from typing import Any, Dict, Sequence
+from typing import TYPE_CHECKING, Any, Dict, Literal, Optional, Sequence
 
 import torch
 from transformers import DataCollatorForSeq2Seq
 
 
+if TYPE_CHECKING:
+    from transformers import ProcessorMixin
+
+    from .template import Template
+
+
+def prepare_4d_attention_mask(attention_mask_with_indices: "torch.Tensor", dtype: "torch.dtype") -> "torch.Tensor":
+    r"""
+    Expands the attention mask with indices from (batch_size, seq_len) to (batch_size, 1, seq_len, seq_len),
+    while handles packed sequences and transforms the mask to lower triangular form to prevent future peeking.
+
+    e.g.
+    ```python
+    # input
+    [[1, 1, 2, 2, 2, 0]]
+    # output
+    [
+        [
+            [
+                [o, x, x, x, x, x],
+                [o, o, x, x, x, x],
+                [x, x, o, x, x, x],
+                [x, x, o, o, x, x],
+                [x, x, o, o, o, x],
+                [x, x, x, x, x, x],
+            ]
+        ]
+    ]
+    ```
+    where `o` equals to `0.0`, `x` equals to `min_dtype`.
+    """
+    bsz, seq_len = attention_mask_with_indices.size()
+    min_dtype = torch.finfo(dtype).min
+    expanded_mask = attention_mask_with_indices[:, None, None, :].expand(bsz, 1, seq_len, seq_len)
+    # Create a binary mask from the original mask where zeros remain zeros and all other values are set to one
+    padding_mask = torch.where(expanded_mask != 0, 1, 0)
+    # Create a block-diagonal mask.
+    attention_mask_4d = torch.eq(expanded_mask, expanded_mask.transpose(-1, -2)).int() * padding_mask
+    # Use the lower triangular mask to zero out the upper triangular part
+    attention_mask_4d *= torch.tril(torch.ones((seq_len, seq_len), dtype=torch.long))
+    # Invert the attention mask.
+    attention_mask_4d = torch.where(attention_mask_4d != 0, torch.tensor(0, dtype=dtype), min_dtype)
+    return attention_mask_4d
+
+
 @dataclass
-class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
+class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
+    r"""
+    Data collator that supports VLMs.
+
+    Features should contain input_ids, attention_mask, labels and images.
+    """
+
+    template: Optional["Template"] = None
+    processor: Optional["ProcessorMixin"] = None
+
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
+        batch_images, batch_videos, batch_imglens, batch_vidlens, batch_input_ids = [], [], [], [], []
+        for feature in features:
+            images = feature.pop("images", None) or []
+            videos = feature.pop("videos", None) or []
+            batch_images.extend(images)
+            batch_videos.extend(videos)
+            batch_imglens.append(len(images))
+            batch_vidlens.append(len(videos))
+            batch_input_ids.append(feature["input_ids"])
+
+        mm_inputs = self.template.mm_plugin.get_mm_inputs(
+            batch_images, batch_videos, batch_imglens, batch_vidlens, batch_input_ids, self.processor
+        )
+        if "token_type_ids" in mm_inputs:
+            token_type_ids = mm_inputs.pop("token_type_ids")
+            for i, feature in enumerate(features):
+                feature["token_type_ids"] = token_type_ids[i]
+
+        features: Dict[str, "torch.Tensor"] = super().__call__(features)
+        features.update(mm_inputs)
+        if isinstance(features.get("pixel_values"), list):  # for pixtral inputs
+            features = features.data  # use default_collate() instead of BatchEncoding.to()
+
+        return features
+
+
+@dataclass
+class SFTDataCollatorWith4DAttentionMask(MultiModalDataCollatorForSeq2Seq):
+    r"""
+    Data collator for 4d attention mask.
+    """
+
+    block_diag_attn: bool = False
+    attn_implementation: Literal["eager", "sdpa", "flash_attention_2"] = "eager"
+    compute_dtype: "torch.dtype" = torch.float32
+
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
+        features = super().__call__(features)
+        if self.block_diag_attn and self.attn_implementation != "flash_attention_2":
+            features["attention_mask"] = prepare_4d_attention_mask(features["attention_mask"], self.compute_dtype)
+
+        return features
+
+
+@dataclass
+class PairwiseDataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
     r"""
     Data collator for pairwise data.
     """
 
-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
         r"""
         Pads batched data to the longest sequence in the batch.
 
@@ -22,28 +140,24 @@ class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
         for key in ("chosen", "rejected"):
             for feature in features:
                 target_feature = {
-                    "input_ids": feature["{}_input_ids".format(key)],
-                    "attention_mask": feature["{}_attention_mask".format(key)],
-                    "labels": feature["{}_labels".format(key)],
+                    "input_ids": feature[f"{key}_input_ids"],
+                    "attention_mask": feature[f"{key}_attention_mask"],
+                    "labels": feature[f"{key}_labels"],
+                    "images": feature["images"],
+                    "videos": feature["videos"],
                 }
-                if "pixel_values" in feature:
-                    target_feature["pixel_values"] = feature["pixel_values"]
-
-                if "{}_token_type_ids".format(key) in feature:
-                    target_feature["token_type_ids"] = feature["{}_token_type_ids".format(key)]
-
                 concatenated_features.append(target_feature)
 
         return super().__call__(concatenated_features)
 
 
 @dataclass
-class KTODataCollatorWithPadding(DataCollatorForSeq2Seq):
+class KTODataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
     r"""
     Data collator for KTO data.
     """
 
-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
         target_features = []
         kl_features = []
         kto_tags = []
@@ -52,19 +166,16 @@ class KTODataCollatorWithPadding(DataCollatorForSeq2Seq):
                 "input_ids": feature["input_ids"],
                 "attention_mask": feature["attention_mask"],
                 "labels": feature["labels"],
+                "images": feature["images"],
+                "videos": feature["videos"],
             }
             kl_feature = {
                 "input_ids": feature["kl_input_ids"],
                 "attention_mask": feature["kl_attention_mask"],
                 "labels": feature["kl_labels"],
+                "images": feature["images"],
+                "videos": feature["videos"],
             }
-            if "pixel_values" in feature:
-                target_feature["pixel_values"] = feature["pixel_values"]
-
-            if "token_type_ids" in feature:
-                target_feature["token_type_ids"] = feature["token_type_ids"]
-                kl_feature["token_type_ids"] = feature["kl_token_type_ids"]
-
             target_features.append(target_feature)
             kl_features.append(kl_feature)
             kto_tags.append(feature["kto_tags"])
@@ -74,7 +185,7 @@ class KTODataCollatorWithPadding(DataCollatorForSeq2Seq):
         batch["kl_input_ids"] = kl_batch["input_ids"]
         batch["kl_attention_mask"] = kl_batch["attention_mask"]
         batch["kl_labels"] = kl_batch["labels"]
-        if "token_type_ids" in batch:
+        if "token_type_ids" in kl_batch:
             batch["kl_token_type_ids"] = kl_batch["token_type_ids"]
 
         batch["kto_tags"] = torch.tensor(kto_tags)