Merge pull request #6124 from hiyouga/hiyouga/release

[release] release v0.9.1

Former-commit-id: f61cdd99fd282612884c92d36e111ad46b4e0d00

.dockerignore

@@ -7,6 +7,8 @@ data
 docker
 saves
 hf_cache
+ms_cache
+om_cache
 output
 .dockerignore
 .gitattributes

.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/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"
@@ -18,13 +18,27 @@ 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:
       - name: Checkout
@@ -33,7 +47,7 @@ jobs:
       - 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"
 

.gitignore

@@ -159,7 +159,13 @@ cython_debug/
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 .idea/
 
+# vscode
+.vscode/
+
 # custom .gitignore
+ms_cache/
+hf_cache/
+om_cache/
 cache/
 config/
 saves/

.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

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:
-	CUDA_VISIBLE_DEVICES= 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-72-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)
@@ -12,6 +12,7 @@
 [![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)
 
@@ -21,13 +22,22 @@
 
 **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**: 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, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA and Agent tuning.
+- **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, Unsloth, RoPE scaling, NEFTune and rsLoRA.
+- **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
 
+[24/10/09] 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.
+
+[24/09/19] We support fine-tuning the **[Qwen2.5](https://qwenlm.github.io/blog/qwen2.5/)** models.
+
+[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>
+
+[24/07/04] 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.
 
-<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 `paligemma` template for chat completion.
-
-[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.
 
 [24/05/18] 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.
 
@@ -152,7 +174,7 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 ## Supported Models
 
 | 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         |
@@ -161,20 +183,28 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 | [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             |
-| [InternLM2](https://huggingface.co/internlm)                 | 7B/20B                           | intern2   |
+| [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](https://huggingface.co/meta-llama)                 | 8B/70B                           | llama3    |
+| [Llama 3-3.2](https://huggingface.co/meta-llama)                  | 1B/3B/8B/70B                     | llama3           |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)                 | 7B/13B                           | vicuna    |
+| [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                               | gemma     |
+| [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       |
+| [Phi-3](https://huggingface.co/microsoft)                         | 4B/14B                           | phi              |
-| [Qwen/Qwen1.5/Qwen2 (Code/MoE)](https://huggingface.co/Qwen) | 0.5B/1.5B/4B/7B/14B/32B/72B/110B | qwen      |
+| [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](https://huggingface.co/01-ai)                    | 6B/9B/34B                        | yi        |
+| [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             |
 
@@ -200,6 +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 +292,9 @@ You also can add a custom chat template to [template.py](src/llamafactory/data/t
 - [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     |
+| torch        | 1.13.1  | 2.4.0     |
-| transformers | 4.41.2  | 4.41.2    |
+| transformers | 4.41.2  | 4.43.4    |
-| datasets     | 2.16.0  | 2.19.2    |
+| datasets     | 2.16.0  | 2.20.0    |
-| accelerate   | 0.30.1  | 0.30.1    |
+| accelerate   | 0.30.1  | 0.32.0    |
-| peft         | 0.11.1  | 0.11.1    |
+| peft         | 0.11.1  | 0.12.0    |
-| trl          | 0.8.6   | 0.9.4     |
+| 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     |
+| vllm         | 0.4.3   | 0.5.0     |
-| flash-attn   | 2.3.0   | 2.5.9     |
+| flash-attn   | 2.3.0   | 2.6.3     |
 
 ### Hardware Requirement
 
@@ -338,7 +375,7 @@ cd LLaMA-Factory
 pip install -e ".[torch,metrics]"
 ```
 
-Extra dependencies available: torch, torch-npu, metrics, deepspeed, bitsandbytes, hqq, eetq, gptq, awq, aqlm, vllm, galore, badam, 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.
@@ -390,7 +427,7 @@ Download the pre-built Docker images: [32GB](http://mirrors.cn-central-221.ovaij
 
 ### 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.
@@ -422,16 +459,24 @@ For CUDA users:
 
 ```bash
 cd docker/docker-cuda/
-docker-compose up -d
+docker compose up -d
-docker-compose exec llamafactory bash
+docker compose exec llamafactory bash
 ```
 
 For Ascend NPU users:
 
 ```bash
 cd docker/docker-npu/
-docker-compose up -d
+docker compose up -d
-docker-compose exec llamafactory bash
+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>
@@ -450,6 +495,7 @@ docker build -f ./docker/docker-cuda/Dockerfile \
 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 \
@@ -474,6 +520,7 @@ docker build -f ./docker/docker-npu/Dockerfile \
 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 \
@@ -493,13 +540,44 @@ docker run -dit \
 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>
 
 <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.
+- `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.
+- `ms_cache`: Similar to Hugging Face cache but for ModelScope users.
-- output: Set export dir to this location so that the merged result can be accessed directly on the host machine.
+- `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>
 
@@ -510,7 +588,9 @@ 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
 
@@ -522,6 +602,16 @@ 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 yaml files.
@@ -600,17 +690,39 @@ If you have a project that should be incorporated, please contact via email or c
 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. 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>
 
@@ -618,7 +730,7 @@ 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: [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) / [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/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)
+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
 

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-72-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)
@@ -12,6 +12,7 @@
 [![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)
 
@@ -21,13 +22,23 @@
 
 **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**：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、BAdam、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ、PiSSA 和 Agent 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、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。
+- **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
 
+[24/10/09] 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)。
+
+[24/09/19] 我们支持了 **[Qwen2.5](https://qwenlm.github.io/blog/qwen2.5/)** 模型的微调。
+
+[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>
+
+[24/07/04] 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)。
 
-<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 `paligemma` 模板进行微调使其获得对话能力。
-
-[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。
 
 [24/05/18] 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。
 
@@ -152,7 +175,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 ## 模型
 
 | 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         |
@@ -161,20 +184,27 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 | [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             |
-| [InternLM2](https://huggingface.co/internlm)                 | 7B/20B                           | intern2   |
+| [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](https://huggingface.co/meta-llama)                 | 8B/70B                           | llama3    |
+| [Llama 3-3.2](https://huggingface.co/meta-llama)                  | 1B/3B/8B/70B                     | llama3           |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)                 | 7B/13B                           | vicuna    |
+| [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                               | gemma     |
+| [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              |
-| [Qwen/Qwen1.5/Qwen2 (Code/MoE)](https://huggingface.co/Qwen) | 0.5B/1.5B/4B/7B/14B/32B/72B/110B | qwen      |
+| [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](https://huggingface.co/01-ai)                    | 6B/9B/34B                        | yi        |
+| [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             |
 
@@ -200,6 +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 +292,9 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 - [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     |
+| torch        | 1.13.1  | 2.4.0     |
-| transformers | 4.41.2  | 4.41.2    |
+| transformers | 4.41.2  | 4.43.4    |
-| datasets     | 2.16.0  | 2.19.2    |
+| datasets     | 2.16.0  | 2.20.0    |
-| accelerate   | 0.30.1  | 0.30.1    |
+| accelerate   | 0.30.1  | 0.32.0    |
-| peft         | 0.11.1  | 0.11.1    |
+| peft         | 0.11.1  | 0.12.0    |
-| trl          | 0.8.6   | 0.9.4     |
+| 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     |
+| vllm         | 0.4.3   | 0.5.0     |
-| flash-attn   | 2.3.0   | 2.5.9     |
+| flash-attn   | 2.3.0   | 2.6.3     |
 
 ### 硬件依赖
 
@@ -338,7 +375,7 @@ cd LLaMA-Factory
 pip install -e ".[torch,metrics]"
 ```
 
-可选的额外依赖项：torch、torch-npu、metrics、deepspeed、bitsandbytes、hqq、eetq、gptq、awq、aqlm、vllm、galore、badam、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。
@@ -390,7 +427,7 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 
 ### 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。
@@ -422,16 +459,24 @@ CUDA 用户：
 
 ```bash
 cd docker/docker-cuda/
-docker-compose up -d
+docker compose up -d
-docker-compose exec llamafactory bash
+docker compose exec llamafactory bash
 ```
 
 昇腾 NPU 用户：
 
 ```bash
 cd docker/docker-npu/
-docker-compose up -d
+docker compose up -d
-docker-compose exec llamafactory bash
+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>
@@ -450,6 +495,7 @@ docker build -f ./docker/docker-cuda/Dockerfile \
 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 \
@@ -474,6 +520,7 @@ docker build -f ./docker/docker-npu/Dockerfile \
 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 \
@@ -493,13 +540,44 @@ docker run -dit \
 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。
+- `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。
+- `ms_cache`：类似 Hugging Face 缓存文件夹，为 ModelScope 用户提供。
-- 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。
+- `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>
 
@@ -510,7 +588,9 @@ 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
 
@@ -522,6 +602,16 @@ 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，请在 yaml 文件中添加下面的参数。
@@ -600,17 +690,38 @@ run_name: test_run # 可选
 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. 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>
 
@@ -618,7 +729,7 @@ 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：[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) / [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/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)
+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
 

data/README.md

@@ -23,6 +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)"
@@ -107,7 +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.
 
@@ -139,67 +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
+An additional column `images` is required. Please refer to the [sharegpt](#sharegpt-format) format for details.
-[
-  {
-    "instruction": "human instruction (required)",
-    "input": "human input (optional)",
-    "output": "model response (required)",
-    "kto_tag": "human feedback [true/false] (required)"
-  }
-]
-```
 
-Regarding the above dataset, the *dataset description* in `dataset_info.json` should be:
+### Multimodal Video Dataset
 
-```json
+An additional column `videos` is required. Please refer to the [sharegpt](#sharegpt-format) format for details.
-"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"
-  }
-}
-```
 
 ## Sharegpt Format
 
@@ -252,6 +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)
@@ -302,6 +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.
@@ -345,7 +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

@@ -23,6 +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）"
@@ -107,7 +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。
 
@@ -139,67 +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
+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)。
-[
-  {
-    "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）"
-  }
-]
-```
 
-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
+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)。
-"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"
-  }
-}
-```
 
 ## Sharegpt 格式
 
@@ -252,6 +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)
@@ -302,6 +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。
@@ -345,7 +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/docker-cuda/Dockerfile

@@ -1,6 +1,7 @@
-# Use the NVIDIA official image with PyTorch 2.3.0
+# Default use the NVIDIA official image with PyTorch 2.3.0
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-24-02.html
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/index.html
-FROM nvcr.io/nvidia/pytorch:24.02-py3
+ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:24.02-py3
+FROM ${BASE_IMAGE}
 
 # Define environments
 ENV MAX_JOBS=4
@@ -12,6 +13,9 @@ 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
@@ -38,6 +42,15 @@ RUN EXTRA_PACKAGES="metrics"; \
     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

docker/docker-cuda/docker-compose.yml

@@ -8,11 +8,15 @@ services:
         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:
@@ -20,6 +24,7 @@ services:
       - "8000:8000"
     ipc: host
     tty: true
+    shm_size: '16gb'
     stdin_open: true
     command: bash
     deploy:

docker/docker-npu/Dockerfile

@@ -1,9 +1,9 @@
 # Use the Ubuntu 22.04 image with CANN 8.0.rc1
-# More versions can be found at https://hub.docker.com/r/cosdt/cann/tags
+# More versions can be found at https://hub.docker.com/r/ascendai/cann/tags
-# FROM cosdt/cann:8.0.rc1-910-ubuntu22.04
+# FROM ascendai/cann:8.0.rc1-910-ubuntu22.04-py3.8
-FROM cosdt/cann:8.0.rc1-910b-ubuntu22.04
+FROM ascendai/cann:8.0.rc1-910b-ubuntu22.04-py3.8
-# FROM cosdt/cann:8.0.rc1-910-openeuler22.03
+# FROM ascendai/cann:8.0.rc1-910-openeuler22.03-py3.8
-# FROM cosdt/cann:8.0.rc1-910b-openeuler22.03
+# FROM ascendai/cann:8.0.rc1-910b-openeuler22.03-py3.8
 
 # Define environments
 ENV DEBIAN_FRONTEND=noninteractive

docker/docker-npu/docker-compose.yml

@@ -10,6 +10,7 @@ services:
     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
@@ -21,6 +22,7 @@ services:
       - "8000:8000"
     ipc: host
     tty: true
+    shm_size: '16gb'
     stdin_open: true
     command: bash
     devices:

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/mmlu/mmlu.py

@@ -158,5 +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 from enumerate(df.to_dict(orient="records"))
-            yield i, instance

examples/README.md

@@ -33,6 +33,19 @@ llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 
 ```bash
 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
+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
@@ -47,12 +60,6 @@ llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
 llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
 ```
 
-#### DPO/ORPO/SimPO Training
-
-```bash
-llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
-```
-
 #### KTO Training
 
 ```bash
@@ -82,8 +89,8 @@ llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
 #### Supervised Fine-Tuning on Multiple Nodes
 
 ```bash
-FORCE_TORCHRUN=1 NNODES=2 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=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 RANK=1 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)
@@ -133,6 +140,12 @@ FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llama
 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
@@ -189,6 +202,12 @@ llamafactory-cli train examples/extras/galore/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

examples/README_zh.md

@@ -33,6 +33,19 @@ llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 
 ```bash
 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
+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
@@ -47,12 +60,6 @@ llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
 llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
 ```
 
-#### DPO/ORPO/SimPO 训练
-
-```bash
-llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
-```
-
 #### KTO 训练
 
 ```bash
@@ -82,8 +89,8 @@ llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
 #### 多机指令监督微调
 
 ```bash
-FORCE_TORCHRUN=1 NNODES=2 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=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 RANK=1 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
@@ -133,6 +140,12 @@ FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llama
 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
@@ -189,6 +202,12 @@ llamafactory-cli train examples/extras/galore/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

examples/extras/adam_mini/qwen2_full_sft.yaml

@@ -1,27 +1,22 @@
 ### model
-model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+model_name_or_path: Qwen/Qwen2-1.5B-Instruct
 
 ### method
 stage: sft
 do_train: true
 finetuning_type: full
-use_badam: true
+use_adam_mini: 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
+template: qwen
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/full/sft
+output_dir: saves/qwen2-1_5b/full/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -30,10 +25,12 @@ 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
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/badam/llama3_full_sft.yaml

@@ -10,11 +10,12 @@ 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
@@ -29,7 +30,7 @@ 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

examples/extras/fsdp_qlora/llama3_lora_sft.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

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

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

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
@@ -26,7 +26,7 @@ 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

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

@@ -13,7 +13,7 @@ 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

examples/inference/llava1_5.yaml

@@ -1,3 +1,2 @@
 model_name_or_path: llava-hf/llava-1.5-7b-hf
-template: vicuna
+template: llava
-visual_inputs: true

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

@@ -10,7 +10,7 @@ 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
@@ -25,7 +25,7 @@ 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

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

@@ -12,7 +12,7 @@ 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

examples/train_lora/llama3_lora_kto.yaml

@@ -11,7 +11,7 @@ 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

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

examples/train_lora/llama3_lora_predict.yaml

@@ -10,7 +10,7 @@ finetuning_type: lora
 ### dataset
 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_lora/llama3_lora_pretrain.yaml

@@ -9,7 +9,7 @@ lora_target: all
 
 ### dataset
 dataset: c4_demo
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16

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

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

examples/train_lora/llama3_lora_sft_ds0.yaml

@@ -11,7 +11,7 @@ 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

examples/train_lora/llama3_lora_sft_ds3.yaml

@@ -11,7 +11,7 @@ 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

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
+template: llava
-cutoff_len: 1024
+cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16

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

@@ -0,0 +1,39 @@
+### model
+model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+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/qwen2_vl-7b/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

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

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

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

examples/train_qlora/llama3_lora_sft_otfq.yaml

@@ -12,7 +12,7 @@ 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

requirements.txt

@@ -1,9 +1,9 @@
-transformers>=4.41.2
+transformers>=4.41.2,<=4.46.1
-datasets>=2.16.0
+datasets>=2.16.0,<=3.1.0
-accelerate>=0.30.1
+accelerate>=0.34.0,<=1.0.1
-peft>=0.11.1
+peft>=0.11.1,<=0.12.0
-trl>=0.8.6
+trl>=0.8.6,<=0.9.6
-gradio>=4.0.0
+gradio>=4.0.0,<5.0.0
 pandas>=2.0.0
 scipy
 einops
@@ -19,3 +19,5 @@ fire
 packaging
 pyyaml
 numpy<2.0.0
+av
+tyro<0.9.0

scripts/cal_flops.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 Microsoft Corporation and the LlamaFactory team.
 #
 # This code is inspired by the Microsoft's DeepSpeed library.
@@ -27,7 +26,7 @@ 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"""
@@ -36,9 +35,11 @@ def calculate_flops(
     """
     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

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 imoneoi and the LlamaFactory team.
 #
 # This code is inspired by the imoneoi's OpenChat library.
@@ -25,7 +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
@@ -39,16 +38,17 @@ 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 --cutoff_len 1024 --batch_size 16
+    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(
@@ -66,13 +66,14 @@ def calculate_lr(
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
+    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("Stage does not supported: {}.".format(stage))
+        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
@@ -84,7 +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

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -23,7 +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
@@ -55,12 +54,12 @@ 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,
@@ -69,7 +68,7 @@ def cal_ppl(
 ):
     r"""
     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
+    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(
@@ -88,7 +87,8 @@ def cal_ppl(
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
+    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)
@@ -99,7 +99,7 @@ def cal_ppl(
             tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX, train_on_prompt=train_on_prompt
         )
     else:
-        raise NotImplementedError("Stage does not supported: {}.".format(stage))
+        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")
@@ -124,9 +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(f"Average perplexity is {total_ppl / len(perplexities):.2f}")
-    print("Perplexities have been saved at {}.".format(save_name))
+    print(f"Perplexities have been saved at {save_name}.")
 
 
 if __name__ == "__main__":
-    fire.Fire(cal_ppl)
+    fire.Fire(calculate_ppl)

scripts/length_cdf.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -18,21 +17,21 @@ 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 --template default
+    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(
@@ -48,7 +47,8 @@ def length_cdf(
         )
     )
     tokenizer_module = load_tokenizer(model_args)
-    trainset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)["train_dataset"]
+    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"]):
@@ -60,7 +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

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 Tencent Inc. and the LlamaFactory team.
 #
 # This code is inspired by the Tencent's LLaMA-Pro library.
@@ -19,7 +18,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING
 
 import fire
 import torch
@@ -40,15 +39,15 @@ 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",
+    shard_size: str = "2GB",
-    save_safetensors: Optional[bool] = False,
+    save_safetensors: bool = True,
 ):
     r"""
     Performs block expansion for LLaMA, Mistral, Qwen1.5 or Yi models.
@@ -76,27 +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():
@@ -113,17 +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(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

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -16,7 +15,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import Any, Dict, Optional
+from typing import Any, Dict
 
 import fire
 import torch
@@ -63,16 +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"]
@@ -82,11 +81,14 @@ 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.

scripts/llamafy_qwen.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -16,7 +15,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import Any, Dict, Optional
+from typing import Any, Dict
 
 import fire
 import torch
@@ -86,7 +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)
@@ -98,18 +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()
@@ -135,11 +134,14 @@ 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.

scripts/loftq_init.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
 #
 # This code is based on the HuggingFace's PEFT library.
@@ -67,22 +66,22 @@ def quantize_loftq(
     loftq_dir = os.path.join(output_dir, "loftq_init")
 
     # Save LoftQ model
-    setattr(peft_model.peft_config["default"], "base_model_name_or_path", output_dir)
+    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("Adapter weights saved in {}".format(loftq_dir))
+    print(f"Adapter weights saved in {loftq_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("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(f"model_name_or_path: {output_dir}")
-    print("adapter_name_or_path: {}".format(loftq_dir))
+    print(f"adapter_name_or_path: {loftq_dir}")
     print("finetuning_type: lora")
-    print("quantization_bit: {}".format(loftq_bits))
+    print(f"quantization_bit: {loftq_bits}")
 
 
 if __name__ == "__main__":

scripts/pissa_init.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
 #
 # This code is based on the HuggingFace's PEFT library.
@@ -31,7 +30,7 @@ if TYPE_CHECKING:
 def quantize_pissa(
     model_name_or_path: str,
     output_dir: str,
-    pissa_iter: int = 4,
+    pissa_iter: int = 16,
     lora_alpha: int = None,
     lora_rank: int = 16,
     lora_dropout: float = 0,
@@ -54,7 +53,7 @@ def quantize_pissa(
         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 "pissa_niter_{}".format(pissa_iter),
+        init_lora_weights="pissa" if pissa_iter == -1 else f"pissa_niter_{pissa_iter}",
     )
 
     # Init PiSSA model
@@ -62,19 +61,20 @@ def quantize_pissa(
     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("Adapter weights saved in {}".format(pissa_dir))
+    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("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(f"model_name_or_path: {output_dir}")
-    print("adapter_name_or_path: {}".format(pissa_dir))
+    print(f"adapter_name_or_path: {pissa_dir}")
     print("finetuning_type: lora")
     print("pissa_init: false")
     print("pissa_convert: true")

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

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");

setup.py

@@ -14,42 +14,54 @@
 
 import os
 import re
+from typing import List
 
 from setuptools import find_packages, setup
 
 
-def get_version():
+def get_version() -> str:
-    with open(os.path.join("src", "llamafactory", "extras", "env.py"), "r", encoding="utf-8") as f:
+    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():
+def get_requires() -> List[str]:
-    with open("requirements.txt", "r", encoding="utf-8") as f:
+    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"],
+    "deepspeed": ["deepspeed>=0.10.0,<=0.14.4"],
+    "liger-kernel": ["liger-kernel"],
     "bitsandbytes": ["bitsandbytes>=0.39.0"],
     "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"],
+    "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"],
 }
 
 
@@ -60,7 +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",
@@ -70,7 +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

@@ -23,9 +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_host = os.getenv("API_HOST", "0.0.0.0")
-    api_port = int(os.environ.get("API_PORT", "8000"))
+    api_port = int(os.getenv("API_PORT", "8000"))
-    print("Visit http://localhost:{}/docs for API document.".format(api_port))
+    print(f"Visit http://localhost:{api_port}/docs for API document.")
     uvicorn.run(app, host=api_host, port=api_port)
 
 

src/llamafactory/__init__.py

@@ -20,22 +20,28 @@ Level:
 
 Dependency graph:
   main:
-    transformers>=4.41.2
+    transformers>=4.41.2,<=4.46.1
-    datasets>=2.16.0
+    datasets>=2.16.0,<=3.1.0
-    accelerate>=0.30.1
+    accelerate>=0.34.0,<=1.0.1
-    peft>=0.11.1
+    peft>=0.11.1,<=0.12.0
-    trl>=0.8.6
+    trl>=0.8.6,<=0.9.6
   attention:
     transformers>=4.42.4 (gemma+fa2)
   longlora:
-    transformers>=4.41.2,<=4.42.4
+    transformers>=4.41.2,<=4.46.1
   packing:
-    transformers>=4.41.2,<=4.42.4
+    transformers>=4.41.2,<=4.46.1
-  patcher:
+
-    transformers==4.41.2 (chatglm)
+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 .cli import VERSION
+from .extras.env import VERSION
 
 
 __version__ = VERSION

src/llamafactory/api/app.py

@@ -12,8 +12,10 @@
 # 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
@@ -50,14 +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=["*"],
@@ -65,7 +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)]):
@@ -79,7 +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(
@@ -116,7 +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_host = os.getenv("API_HOST", "0.0.0.0")
-    api_port = int(os.environ.get("API_PORT", "8000"))
+    api_port = int(os.getenv("API_PORT", "8000"))
-    print("Visit http://localhost:{}/docs for API document.".format(api_port))
+    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

@@ -16,11 +16,12 @@ 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 (
@@ -51,13 +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,
@@ -69,8 +69,8 @@ ROLE_MAPPING = {
 
 def _process_request(
     request: "ChatCompletionRequest",
-) -> Tuple[List[Dict[str, str]], Optional[str], Optional[str], Optional["NDArray"]]:
+) -> Tuple[List[Dict[str, str]], Optional[str], Optional[str], Optional[List["ImageInput"]]]:
-    logger.info("==== request ====\n{}".format(json.dumps(dictify(request), indent=2, ensure_ascii=False)))
+    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")
@@ -84,7 +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")
@@ -104,15 +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
+                    if re.match(r"^data:image\/(png|jpg|jpeg|gif|bmp);base64,(.+)$", image_url):  # base64 image
-                        image_data = base64.b64decode(image_url.split(",", maxsplit=1)[1])
+                        image_stream = io.BytesIO(base64.b64decode(image_url.split(",", maxsplit=1)[1]))
-                        image_path = io.BytesIO(image_data)
                     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})
 
@@ -125,7 +124,7 @@ def _process_request(
     else:
         tools = None
 
-    return input_messages, system, tools, image
+    return input_messages, system, tools, images or None
 
 
 def _create_stream_chat_completion_chunk(
@@ -143,13 +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)
+    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
-    input_messages, system, tools, image = _process_request(request)
+    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,
@@ -170,7 +169,7 @@ async def create_chat_completion_response(
             tool_calls = []
             for tool in result:
                 function = Function(name=tool[0], arguments=tool[1])
-                tool_calls.append(FunctionCall(id="call_{}".format(uuid.uuid4().hex), function=function))
+                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
@@ -194,8 +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)
+    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
-    input_messages, system, tools, image = _process_request(request)
+    input_messages, system, tools, images = _process_request(request)
     if tools:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Cannot stream function calls.")
 
@@ -209,7 +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,
@@ -230,8 +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/chat/base_engine.py

@@ -18,11 +18,11 @@ from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, List, Literal, Opti
 
 
 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
 
 
@@ -35,6 +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
@@ -48,7 +54,11 @@ class BaseEngine(ABC):
         data_args: "DataArguments",
         finetuning_args: "FinetuningArguments",
         generating_args: "GeneratingArguments",
-    ) -> None: ...
+    ) -> None:
+        r"""
+        Initializes an inference engine.
+        """
+        ...
 
     @abstractmethod
     async def chat(
@@ -56,9 +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(
@@ -66,13 +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

@@ -27,8 +27,7 @@ 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
 
 
@@ -38,14 +37,23 @@ def _start_background_loop(loop: "asyncio.AbstractEventLoop") -> None:
 
 
 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)
@@ -56,10 +64,16 @@ 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"]:
-        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(
@@ -67,20 +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)
@@ -93,10 +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(
@@ -104,6 +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()
 
@@ -112,6 +141,9 @@ 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)
 
 

src/llamafactory/chat/hf_engine.py

@@ -20,25 +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):
@@ -54,7 +55,7 @@ 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, data_args.tool_format)
+        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
@@ -62,11 +63,11 @@ class HuggingfaceEngine(BaseEngine):
         try:
             asyncio.get_event_loop()
         except RuntimeError:
-            logger.warning("There is no current event loop, creating a new one.")
+            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.environ.get("MAX_CONCURRENT", "1")))
+        self.semaphore = asyncio.Semaphore(int(os.getenv("MAX_CONCURRENT", "1")))
 
     @staticmethod
     def _process_args(
@@ -78,31 +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 (
+        mm_input_dict = {"images": [], "videos": [], "imglens": [0], "vidlens": [0]}
-            processor is not None
+        if images is not None:
-            and image is not None
+            mm_input_dict.update({"images": images, "imglens": [len(images)]})
-            and not hasattr(processor, "image_seq_length")
+            if not any(IMAGE_PLACEHOLDER in message["content"] for message in messages):
-            and template.image_token not in messages[0]["content"]
+                messages[0]["content"] = IMAGE_PLACEHOLDER * len(images) + messages[0]["content"]
-        ):  # llava-like models
-            messages[0]["content"] = template.image_token + 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, paired_messages, system, tools)
-        prompt_ids, _ = template.encode_oneturn(
+        prompt_ids, _ = template.mm_plugin.process_token_ids(
-            tokenizer=tokenizer, messages=paired_messages, system=system, tools=tools
+            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)
@@ -119,7 +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 by the huggingface engine yet.")
+            logger.warning_rank0("Stop parameter is not supported by the huggingface engine yet.")
 
         generating_args = generating_args.copy()
         generating_args.update(
@@ -164,8 +164,14 @@ class HuggingfaceEngine(BaseEngine):
             logits_processor=get_logits_processor(),
         )
 
-        if pixel_values is not None:
+        mm_inputs = template.mm_plugin.get_mm_inputs(**mm_input_dict, batch_ids=[prompt_ids], processor=processor)
-            gen_kwargs["pixel_values"] = pixel_values
+        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
 
@@ -180,11 +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:]
@@ -215,11 +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
@@ -242,37 +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)
-
+        values: "torch.Tensor" = model(**inputs, return_dict=True, use_cache=False)[-1]
-        input_ids: torch.Tensor = inputs["input_ids"]
+        scores = values.gather(dim=-1, index=(inputs["attention_mask"].sum(dim=-1, keepdim=True) - 1))
-        _, _, values = model(**inputs, output_hidden_states=True, return_dict=True)
-
-        if getattr(model.config, "model_type", None) == "chatglm":
-            values = torch.transpose(values, 0, 1)
-
-        scores = []
-        for i in range(input_ids.size(0)):
-            end_indexes = (input_ids[i] != tokenizer.pad_token_id).nonzero()
-            end_index = end_indexes[-1].item() if len(end_indexes) else 0
-            scores.append(values[i, end_index].nan_to_num().item())
-
         return scores
 
+    @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:
@@ -288,19 +307,22 @@ class HuggingfaceEngine(BaseEngine):
             messages,
             system,
             tools,
-            image,
+            images,
+            videos,
             input_kwargs,
         )
         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:
@@ -316,7 +338,8 @@ class HuggingfaceEngine(BaseEngine):
             messages,
             system,
             tools,
-            image,
+            images,
+            videos,
             input_kwargs,
         )
         async with self.semaphore:
@@ -328,6 +351,7 @@ class HuggingfaceEngine(BaseEngine):
                     except StopAsyncIteration:
                         break
 
+    @override
     async def get_scores(
         self,
         batch_input: List[str],

src/llamafactory/chat/vllm_engine.py

@@ -15,36 +15,35 @@
 import uuid
 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, is_vllm_version_greater_than_0_5, is_vllm_version_greater_than_0_5_1
+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
 
-    if is_vllm_version_greater_than_0_5_1():
-        pass
-    elif is_vllm_version_greater_than_0_5():
-        from vllm.multimodal.image import ImagePixelData
-    else:
-        from vllm.sequence import MultiModalData
-
 
 if TYPE_CHECKING:
-    from numpy.typing import NDArray
+    from ..data.mm_plugin import ImageInput, VideoInput
-    from transformers.image_processing_utils import BaseImageProcessor
-
     from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 class VllmEngine(BaseEngine):
@@ -67,7 +66,7 @@ class VllmEngine(BaseEngine):
         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, data_args.tool_format)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args)
         self.generating_args = generating_args.to_dict()
 
         engine_args = {
@@ -84,19 +83,13 @@ 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
 
-                logger.info("Detected Yi-VL model, applying projector patch.")
+            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))
@@ -110,40 +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 (
+        if self.template.mm_plugin.__class__.__name__ == "Qwen2vlPlugin":  # temporary solution
-            self.processor is not None
+            image_str = f"<|vision_start|>{self.template.mm_plugin.image_token}<|vision_end|>"
-            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"]
-            if is_vllm_version_greater_than_0_5_1():
-                multi_modal_data = {"image": pixel_values}
-            elif is_vllm_version_greater_than_0_5():
-                multi_modal_data = ImagePixelData(image=pixel_values)
-            else:  # TODO: remove vllm 0.4.3 support
-                multi_modal_data = MultiModalData(type=MultiModalData.Type.IMAGE, data=pixel_values)
         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)
@@ -154,6 +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:
@@ -177,32 +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
 
+    @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
 
@@ -219,21 +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

@@ -22,8 +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
@@ -47,7 +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
@@ -56,7 +56,7 @@ WELCOME = (
     + "-" * 58
 )
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 @unique
@@ -86,25 +86,26 @@ 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_addr = os.getenv("MASTER_ADDR", "127.0.0.1")
-            master_port = os.environ.get("MASTER_PORT", str(random.randint(20001, 29999)))
+            master_port = os.getenv("MASTER_PORT", str(random.randint(20001, 29999)))
-            logger.info("Initializing distributed tasks at: {}:{}".format(master_addr, master_port))
+            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"),
+                .format(
-                    node_rank=os.environ.get("RANK", "0"),
+                    nnodes=os.getenv("NNODES", "1"),
-                    nproc_per_node=os.environ.get("NPROC_PER_NODE", str(get_device_count())),
+                    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:
@@ -118,4 +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

@@ -12,7 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .collator import KTODataCollatorWithPadding, PairwiseDataCollatorWithPadding, SFTDataCollatorWith4DAttentionMask
+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
@@ -20,6 +25,7 @@ from .template import TEMPLATES, Template, get_template_and_fix_tokenizer
 
 __all__ = [
     "KTODataCollatorWithPadding",
+    "MultiModalDataCollatorForSeq2Seq",
     "PairwiseDataCollatorWithPadding",
     "SFTDataCollatorWith4DAttentionMask",
     "Role",

src/llamafactory/data/aligner.py

@@ -14,11 +14,9 @@
 
 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 import logging
-
-from ..extras.logging import get_logger
 from .data_utils import Role
 
 
@@ -27,88 +25,123 @@ if TYPE_CHECKING:
     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 not isinstance(images, list):
-    if dataset_attr.load_from in ["script", "file"]:
+        images = [images]
-        for image in images:
+    elif len(images) == 0:
-            if isinstance(image, str) and os.path.isfile(os.path.join(data_args.dataset_dir, image)):
+        return None
-                outputs.append(os.path.join(data_args.dataset_dir, image))
     else:
-                outputs.append(image)
+        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"
+    example: Dict[str, Any],
-) -> Dict[str, List[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": []}
-    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):
+    if dataset_attr.history and isinstance(example[dataset_attr.history], list):
-            for old_prompt, old_response in examples[dataset_attr.history][i]:
+        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})
 
-        content = []
+    query = []
-        if dataset_attr.prompt and examples[dataset_attr.prompt][i]:
+    if dataset_attr.prompt and example[dataset_attr.prompt]:
-            content.append(examples[dataset_attr.prompt][i])
+        query.append(example[dataset_attr.prompt])
 
-        if dataset_attr.query and examples[dataset_attr.query][i]:
+    if dataset_attr.query and example[dataset_attr.query]:
-            content.append(examples[dataset_attr.query][i])
+        query.append(example[dataset_attr.query])
 
-        prompt.append({"role": Role.USER.value, "content": "\n".join(content)})  # "prompt\nquery"
+    prompt.append({"role": Role.USER.value, "content": "\n".join(query)})  # "prompt\nquery"
 
-        if dataset_attr.kto_tag and isinstance(examples[dataset_attr.kto_tag][i], bool):  # kto example
+    if dataset_attr.kto_tag and isinstance(example[dataset_attr.kto_tag], bool):  # kto example
-            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
+        response = [{"role": Role.ASSISTANT.value, "content": example[dataset_attr.response]}]
-            if examples[dataset_attr.kto_tag][i]:
+        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(examples[dataset_attr.chosen][i], str)
+        and isinstance(example[dataset_attr.chosen], str)
-            and isinstance(examples[dataset_attr.rejected][i], str)
+        and isinstance(example[dataset_attr.rejected], str)
     ):  # pairwise example
         response = [
-                {"role": Role.ASSISTANT.value, "content": examples[dataset_attr.chosen][i]},
+            {"role": Role.ASSISTANT.value, "content": example[dataset_attr.chosen]},
-                {"role": Role.ASSISTANT.value, "content": examples[dataset_attr.rejected][i]},
+            {"role": Role.ASSISTANT.value, "content": example[dataset_attr.rejected]},
         ]
-        elif dataset_attr.response and isinstance(examples[dataset_attr.response][i], str):  # normal example
+    elif dataset_attr.response and isinstance(example[dataset_attr.response], str):  # normal example
-            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
+        response = [{"role": Role.ASSISTANT.value, "content": example[dataset_attr.response]}]
     else:  # unsupervised
         response = []
 
-        outputs["prompt"].append(prompt)
+    convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
-        outputs["response"].append(response)
+    convert_videos = partial(_convert_videos, dataset_attr=dataset_attr, data_args=data_args)
-        outputs["system"].append(examples[dataset_attr.system][i] if dataset_attr.system else "")
+    output = {
-        outputs["tools"].append(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
+        "_prompt": prompt,
-        outputs["images"].append(convert_images(examples[dataset_attr.images][i]) if dataset_attr.images else [])
+        "_response": response,
-
+        "_system": example[dataset_attr.system] if dataset_attr.system else "",
-    return outputs
+        "_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"
+    example: Dict[str, Any],
-) -> Dict[str, List[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,
@@ -119,21 +152,22 @@ 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]):
+    messages = example[dataset_attr.messages]
-        if dataset_attr.system_tag and messages[0][dataset_attr.role_tag] == dataset_attr.system_tag:
+    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 = examples[dataset_attr.system][i] if dataset_attr.system 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))
+            logger.warning_rank0(f"Invalid role tag in {messages}.")
             broken_data = True
 
         aligned_messages.append(
@@ -143,28 +177,28 @@ def convert_sharegpt(
     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))
+        logger.warning_rank0(f"Invalid message count in {messages}.")
         broken_data = True
 
-        if dataset_attr.kto_tag and isinstance(examples[dataset_attr.kto_tag][i], bool):  # kto example
+    if dataset_attr.kto_tag and isinstance(example[dataset_attr.kto_tag], bool):  # kto example
         prompt = aligned_messages[:-1]
         response = aligned_messages[-1:]
-            if examples[dataset_attr.kto_tag][i]:
+        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(examples[dataset_attr.chosen][i], dict)
+        and isinstance(example[dataset_attr.chosen], dict)
-            and isinstance(examples[dataset_attr.rejected][i], dict)
+        and isinstance(example[dataset_attr.rejected], dict)
     ):  # pairwise example
-            chosen = examples[dataset_attr.chosen][i]
+        chosen = example[dataset_attr.chosen]
-            rejected = examples[dataset_attr.rejected][i]
+        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("Invalid role tag in {}.".format([chosen, rejected]))
+            logger.warning_rank0(f"Invalid role tag in {[chosen, rejected]}.")
             broken_data = True
 
         prompt = aligned_messages
@@ -177,16 +211,20 @@ def convert_sharegpt(
         response = aligned_messages[-1:]
 
     if broken_data:
-            logger.warning("Skipping this abnormal example.")
+        logger.warning_rank0("Skipping this abnormal example.")
-            continue
+        prompt, response = [], []
 
-        outputs["prompt"].append(prompt)
+    convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
-        outputs["response"].append(response)
+    convert_videos = partial(_convert_videos, dataset_attr=dataset_attr, data_args=data_args)
-        outputs["system"].append(system)
+    output = {
-        outputs["tools"].append(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
+        "_prompt": prompt,
-        outputs["images"].append(convert_images(examples[dataset_attr.images][i]) if dataset_attr.images else [])
+        "_response": response,
-
+        "_system": system,
-    return outputs
+        "_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(
@@ -197,11 +235,12 @@ def align_dataset(
 ) -> Union["Dataset", "IterableDataset"]:
     r"""
     Aligned dataset:
-        prompt: [{"role": "user", "content": "..."}] * (2T - 1)
+        _prompt: [{"role": "user", "content": "..."}] * (2T - 1)
-        response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
+        _response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
-        system: "..."
+        _system: "..."
-        tools: "...",
+        _tools: "...",
-        images: [],
+        _images: [],
+        _videos: [],
     """
     if dataset_attr.formatting == "alpaca":
         convert_func = partial(convert_alpaca, dataset_attr=dataset_attr, data_args=data_args)
@@ -209,19 +248,6 @@ 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(
@@ -232,8 +258,7 @@ def align_dataset(
 
     return dataset.map(
         convert_func,
-        batched=True,
+        batched=False,
         remove_columns=column_names,
-        features=features,
         **kwargs,
     )

src/llamafactory/data/collator.py

@@ -16,12 +16,18 @@
 # limitations under the License.
 
 from dataclasses import dataclass
-from typing import Any, Dict, Literal, 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),
@@ -62,7 +68,45 @@ def prepare_4d_attention_mask(attention_mask_with_indices: "torch.Tensor", dtype
 
 
 @dataclass
-class SFTDataCollatorWith4DAttentionMask(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.
     """
@@ -80,7 +124,7 @@ class SFTDataCollatorWith4DAttentionMask(DataCollatorForSeq2Seq):
 
 
 @dataclass
-class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
+class PairwiseDataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
     r"""
     Data collator for pairwise data.
     """
@@ -96,23 +140,19 @@ class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
         for key in ("chosen", "rejected"):
             for feature in features:
                 target_feature = {
-                    "input_ids": feature["{}_input_ids".format(key)],
+                    "input_ids": feature[f"{key}_input_ids"],
-                    "attention_mask": feature["{}_attention_mask".format(key)],
+                    "attention_mask": feature[f"{key}_attention_mask"],
-                    "labels": feature["{}_labels".format(key)],
+                    "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.
     """
@@ -126,19 +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"])
@@ -148,7 +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)

src/llamafactory/data/data_utils.py

@@ -17,7 +17,7 @@ from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Set, TypedDict
 
 from datasets import DatasetDict, concatenate_datasets, interleave_datasets
 
-from ..extras.logging import get_logger
+from ..extras import logging
 
 
 if TYPE_CHECKING:
@@ -26,7 +26,7 @@ if TYPE_CHECKING:
     from ..hparams import DataArguments
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
@@ -49,16 +49,19 @@ class DatasetModule(TypedDict):
 def merge_dataset(
     all_datasets: List[Union["Dataset", "IterableDataset"]], data_args: "DataArguments", seed: int
 ) -> Union["Dataset", "IterableDataset"]:
+    r"""
+    Merges multiple datasets to a unified dataset.
+    """
     if len(all_datasets) == 1:
         return all_datasets[0]
     elif data_args.mix_strategy == "concat":
         if data_args.streaming:
-            logger.warning("The samples between different datasets will not be mixed in streaming mode.")
+            logger.warning_once("The samples between different datasets will not be mixed in streaming mode.")
 
         return concatenate_datasets(all_datasets)
     elif data_args.mix_strategy.startswith("interleave"):
         if not data_args.streaming:
-            logger.warning("We recommend using `mix_strategy=concat` in non-streaming mode.")
+            logger.warning_once("We recommend using `mix_strategy=concat` in non-streaming mode.")
 
         return interleave_datasets(
             datasets=all_datasets,
@@ -67,14 +70,16 @@ def merge_dataset(
             stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
         )
     else:
-        raise ValueError("Unknown mixing strategy.")
+        raise ValueError(f"Unknown mixing strategy: {data_args.mix_strategy}.")
 
 
 def split_dataset(
     dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", seed: int
 ) -> "DatasetDict":
     r"""
-    Splits the dataset and returns a dataset dict containing train set (required) and validation set (optional).
+    Splits the dataset and returns a dataset dict containing train set and validation set.
+
+    Supports both map dataset and iterable dataset.
     """
     if data_args.streaming:
         dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=seed)

src/llamafactory/data/formatter.py

@@ -16,21 +16,36 @@ import json
 import re
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
-from typing import List, Literal, Optional, Tuple, Union
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+from typing_extensions import override
 
 from .data_utils import SLOTS
-from .tool_utils import DefaultToolUtils, GLM4ToolUtils
+from .tool_utils import get_tool_utils
+
+
+if TYPE_CHECKING:
+    from .tool_utils import FunctionCall
 
 
 @dataclass
 class Formatter(ABC):
     slots: SLOTS = field(default_factory=list)
-    tool_format: Optional[Literal["default", "glm4"]] = None
+    tool_format: Optional[str] = None
 
     @abstractmethod
-    def apply(self, **kwargs) -> SLOTS: ...
+    def apply(self, **kwargs) -> SLOTS:
+        r"""
+        Forms a list of slots according to the inputs to encode.
+        """
+        ...
 
-    def extract(self, content: str) -> Union[str, List[Tuple[str, str]]]:
+    def extract(self, content: str) -> Union[str, List["FunctionCall"]]:
+        r"""
+        Extract a list of tuples from the response message if using tools.
+
+        Each tuple consists of function name and function arguments.
+        """
         raise NotImplementedError
 
 
@@ -45,6 +60,7 @@ class EmptyFormatter(Formatter):
         if has_placeholder:
             raise ValueError("Empty formatter should not contain any placeholder.")
 
+    @override
     def apply(self, **kwargs) -> SLOTS:
         return self.slots
 
@@ -60,20 +76,21 @@ class StringFormatter(Formatter):
         if not has_placeholder:
             raise ValueError("A placeholder is required in the string formatter.")
 
+    @override
     def apply(self, **kwargs) -> SLOTS:
         elements = []
         for slot in self.slots:
             if isinstance(slot, str):
                 for name, value in kwargs.items():
                     if not isinstance(value, str):
-                        raise RuntimeError("Expected a string, got {}".format(value))
+                        raise RuntimeError(f"Expected a string, got {value}")
 
                     slot = slot.replace("{{" + name + "}}", value, 1)
                 elements.append(slot)
             elif isinstance(slot, (dict, set)):
                 elements.append(slot)
             else:
-                raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+                raise RuntimeError(f"Input must be string, set[str] or dict[str, str], got {type(slot)}")
 
         return elements
 
@@ -81,13 +98,9 @@ class StringFormatter(Formatter):
 @dataclass
 class FunctionFormatter(Formatter):
     def __post_init__(self):
-        if self.tool_format == "default":
+        self.slots = get_tool_utils(self.tool_format).get_function_slots() + self.slots
-            self.slots = DefaultToolUtils.get_function_slots() + self.slots
-        elif self.tool_format == "glm4":
-            self.slots = GLM4ToolUtils.get_function_slots() + self.slots
-        else:
-            raise NotImplementedError("Tool format {} was not found.".format(self.tool_format))
 
+    @override
     def apply(self, **kwargs) -> SLOTS:
         content = kwargs.pop("content")
         functions: List[Tuple[str, str]] = []
@@ -100,7 +113,7 @@ class FunctionFormatter(Formatter):
                 functions.append((tool_call["name"], json.dumps(tool_call["arguments"], ensure_ascii=False)))
 
         except json.JSONDecodeError:
-            functions = []
+            raise RuntimeError(f"Invalid JSON format in function message: {str([content])}")  # flat string
 
         elements = []
         for name, arguments in functions:
@@ -111,7 +124,7 @@ class FunctionFormatter(Formatter):
                 elif isinstance(slot, (dict, set)):
                     elements.append(slot)
                 else:
-                    raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+                    raise RuntimeError(f"Input must be string, set[str] or dict[str, str], got {type(slot)}")
 
         return elements
 
@@ -119,22 +132,17 @@ class FunctionFormatter(Formatter):
 @dataclass
 class ToolFormatter(Formatter):
     def __post_init__(self):
-        if self.tool_format == "default":
+        self.tool_utils = get_tool_utils(self.tool_format)
-            self._tool_formatter = DefaultToolUtils.tool_formatter
-            self._tool_extractor = DefaultToolUtils.tool_extractor
-        elif self.tool_format == "glm4":
-            self._tool_formatter = GLM4ToolUtils.tool_formatter
-            self._tool_extractor = GLM4ToolUtils.tool_extractor
-        else:
-            raise NotImplementedError("Tool format {} was not found.".format(self.tool_format))
 
+    @override
     def apply(self, **kwargs) -> SLOTS:
         content = kwargs.pop("content")
         try:
             tools = json.loads(content)
-            return [self._tool_formatter(tools) if len(tools) != 0 else ""]
+            return [self.tool_utils.tool_formatter(tools) if len(tools) != 0 else ""]
         except json.JSONDecodeError:
-            return [""]
+            raise RuntimeError(f"Invalid JSON format in tool description: {str([content])}")  # flat string
 
-    def extract(self, content: str) -> Union[str, List[Tuple[str, str]]]:
+    @override
-        return self._tool_extractor(content)
+    def extract(self, content: str) -> Union[str, List["FunctionCall"]]:
+        return self.tool_utils.tool_extractor(content)

src/llamafactory/data/loader.py

@@ -20,14 +20,13 @@ import numpy as np
 from datasets import DatasetDict, load_dataset, load_from_disk
 from transformers.utils.versions import require_version
 
+from ..extras import logging
 from ..extras.constants import FILEEXT2TYPE
-from ..extras.logging import get_logger
 from ..extras.misc import has_tokenized_data
 from .aligner import align_dataset
 from .data_utils import merge_dataset, split_dataset
 from .parser import get_dataset_list
 from .preprocess import get_preprocess_and_print_func
-from .template import get_template_and_fix_tokenizer
 
 
 if TYPE_CHECKING:
@@ -40,7 +39,7 @@ if TYPE_CHECKING:
     from .template import Template
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 def _load_single_dataset(
@@ -49,9 +48,12 @@ def _load_single_dataset(
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
 ) -> Union["Dataset", "IterableDataset"]:
-    logger.info("Loading dataset {}...".format(dataset_attr))
+    r"""
+    Loads a single dataset and aligns it to the standard format.
+    """
+    logger.info_rank0(f"Loading dataset {dataset_attr}...")
     data_path, data_name, data_dir, data_files = None, None, None, None
-    if dataset_attr.load_from in ["hf_hub", "ms_hub"]:
+    if dataset_attr.load_from in ["hf_hub", "ms_hub", "om_hub"]:
         data_path = dataset_attr.dataset_name
         data_name = dataset_attr.subset
         data_dir = dataset_attr.folder
@@ -67,25 +69,24 @@ def _load_single_dataset(
         if os.path.isdir(local_path):  # is directory
             for file_name in os.listdir(local_path):
                 data_files.append(os.path.join(local_path, file_name))
-                if data_path is None:
-                    data_path = FILEEXT2TYPE.get(file_name.split(".")[-1], None)
-                elif data_path != FILEEXT2TYPE.get(file_name.split(".")[-1], None):
-                    raise ValueError("File types should be identical.")
         elif os.path.isfile(local_path):  # is file
             data_files.append(local_path)
-            data_path = FILEEXT2TYPE.get(local_path.split(".")[-1], None)
         else:
-            raise ValueError("File {} not found.".format(local_path))
+            raise ValueError(f"File {local_path} not found.")
 
+        data_path = FILEEXT2TYPE.get(os.path.splitext(data_files[0])[-1][1:], None)
         if data_path is None:
             raise ValueError("Allowed file types: {}.".format(",".join(FILEEXT2TYPE.keys())))
+
+        if any(data_path != FILEEXT2TYPE.get(os.path.splitext(data_file)[-1][1:], None) for data_file in data_files):
+            raise ValueError("File types should be identical.")
     else:
-        raise NotImplementedError("Unknown load type: {}.".format(dataset_attr.load_from))
+        raise NotImplementedError(f"Unknown load type: {dataset_attr.load_from}.")
 
     if dataset_attr.load_from == "ms_hub":
         require_version("modelscope>=1.11.0", "To fix: pip install modelscope>=1.11.0")
-        from modelscope import MsDataset
+        from modelscope import MsDataset  # type: ignore
-        from modelscope.utils.config_ds import MS_DATASETS_CACHE
+        from modelscope.utils.config_ds import MS_DATASETS_CACHE  # type: ignore
 
         cache_dir = model_args.cache_dir or MS_DATASETS_CACHE
         dataset = MsDataset.load(
@@ -96,10 +97,27 @@ def _load_single_dataset(
             split=dataset_attr.split,
             cache_dir=cache_dir,
             token=model_args.ms_hub_token,
-            use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
+            use_streaming=data_args.streaming,
         )
         if isinstance(dataset, MsDataset):
             dataset = dataset.to_hf_dataset()
+
+    elif dataset_attr.load_from == "om_hub":
+        require_version("openmind>=0.8.0", "To fix: pip install openmind>=0.8.0")
+        from openmind import OmDataset  # type: ignore
+        from openmind.utils.hub import OM_DATASETS_CACHE  # type: ignore
+
+        cache_dir = model_args.cache_dir or OM_DATASETS_CACHE
+        dataset = OmDataset.load_dataset(
+            path=data_path,
+            name=data_name,
+            data_dir=data_dir,
+            data_files=data_files,
+            split=dataset_attr.split,
+            cache_dir=cache_dir,
+            token=model_args.om_hub_token,
+            streaming=data_args.streaming,
+        )
     else:
         dataset = load_dataset(
             path=data_path,
@@ -109,16 +127,13 @@ def _load_single_dataset(
             split=dataset_attr.split,
             cache_dir=model_args.cache_dir,
             token=model_args.hf_hub_token,
-            streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
+            streaming=data_args.streaming,
             trust_remote_code=True,
         )
 
-    if data_args.streaming and (dataset_attr.load_from == "file"):  # faster than specifying streaming=True
-        dataset = dataset.to_iterable_dataset()  # TODO: add num shards parameter
-
     if dataset_attr.num_samples is not None and not data_args.streaming:
         target_num = dataset_attr.num_samples
-        indexes = np.random.permutation(len(dataset))[:target_num]
+        indexes = np.random.permutation(len(dataset))[:target_num]  # all samples should be included
         target_num -= len(indexes)
         if target_num > 0:
             expand_indexes = np.random.choice(len(dataset), target_num)
@@ -126,7 +141,7 @@ def _load_single_dataset(
 
         assert len(indexes) == dataset_attr.num_samples, "Sample num mismatched."
         dataset = dataset.select(indexes)
-        logger.info("Sampled {} examples from dataset {}.".format(dataset_attr.num_samples, dataset_attr))
+        logger.info_rank0(f"Sampled {dataset_attr.num_samples} examples from dataset {dataset_attr}.")
 
     if data_args.max_samples is not None:  # truncate dataset
         max_samples = min(data_args.max_samples, len(dataset))
@@ -142,6 +157,9 @@ def _get_merged_dataset(
     training_args: "Seq2SeqTrainingArguments",
     stage: Literal["pt", "sft", "rm", "ppo", "kto"],
 ) -> Optional[Union["Dataset", "IterableDataset"]]:
+    r"""
+    Gets the merged datasets in the standard format.
+    """
     if dataset_names is None:
         return None
 
@@ -165,6 +183,9 @@ def _get_preprocessed_dataset(
     processor: Optional["ProcessorMixin"] = None,
     is_eval: bool = False,
 ) -> Optional[Union["Dataset", "IterableDataset"]]:
+    r"""
+    Preprocesses the dataset, including format checking and tokenization.
+    """
     if dataset is None:
         return None
 
@@ -180,7 +201,13 @@ def _get_preprocessed_dataset(
             desc="Running tokenizer on dataset",
         )
 
-    dataset = dataset.map(preprocess_func, batched=True, remove_columns=column_names, **kwargs)
+    dataset = dataset.map(
+        preprocess_func,
+        batched=True,
+        batch_size=data_args.preprocessing_batch_size,
+        remove_columns=column_names,
+        **kwargs,
+    )
 
     if training_args.should_log:
         try:
@@ -196,6 +223,7 @@ def _get_preprocessed_dataset(
 
 
 def get_dataset(
+    template: "Template",
     model_args: "ModelArguments",
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
@@ -203,20 +231,20 @@ def get_dataset(
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"] = None,
 ) -> "DatasetModule":
-    template = get_template_and_fix_tokenizer(tokenizer, data_args.template, data_args.tool_format)
+    r"""
-    if data_args.train_on_prompt and template.efficient_eos:
+    Gets the train dataset and optionally gets the evaluation dataset.
-        raise ValueError("Current template does not support `train_on_prompt`.")
+    """
-
     # Load tokenized dataset
     if data_args.tokenized_path is not None:
         if has_tokenized_data(data_args.tokenized_path):
-            logger.warning("Loading dataset from disk will ignore other data arguments.")
+            logger.warning_rank0("Loading dataset from disk will ignore other data arguments.")
             dataset_dict: "DatasetDict" = load_from_disk(data_args.tokenized_path)
-            logger.info("Loaded tokenized dataset from {}.".format(data_args.tokenized_path))
+            logger.info_rank0(f"Loaded tokenized dataset from {data_args.tokenized_path}.")
 
             dataset_module: Dict[str, "Dataset"] = {}
             if "train" in dataset_dict:
                 dataset_module["train_dataset"] = dataset_dict["train"]
+
             if "validation" in dataset_dict:
                 dataset_module["eval_dataset"] = dataset_dict["validation"]
 
@@ -262,14 +290,15 @@ def get_dataset(
         if data_args.tokenized_path is not None:
             if training_args.should_save:
                 dataset_dict.save_to_disk(data_args.tokenized_path)
-                logger.info("Tokenized dataset saved at {}.".format(data_args.tokenized_path))
+                logger.info_rank0(f"Tokenized dataset saved at {data_args.tokenized_path}.")
-                logger.info("Please restart the training with `tokenized_path: {}`.".format(data_args.tokenized_path))
+                logger.info_rank0(f"Please restart the training with `tokenized_path: {data_args.tokenized_path}`.")
 
             sys.exit(0)
 
         dataset_module = {}
         if "train" in dataset_dict:
             dataset_module["train_dataset"] = dataset_dict["train"]
+
         if "validation" in dataset_dict:
             dataset_module["eval_dataset"] = dataset_dict["validation"]
 

src/llamafactory/data/mm_plugin.py

@@ -0,0 +1,787 @@
+import math
+from copy import deepcopy
+from io import BytesIO
+from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Tuple, TypedDict, Union
+
+import numpy as np
+import torch
+from transformers.image_utils import get_image_size, to_numpy_array
+from typing_extensions import override
+
+from ..extras.constants import IGNORE_INDEX, IMAGE_PLACEHOLDER, VIDEO_PLACEHOLDER
+from ..extras.packages import is_pillow_available, is_pyav_available, is_transformers_version_greater_than
+
+
+if is_pillow_available():
+    from PIL import Image
+    from PIL.Image import Image as ImageObject
+
+
+if is_pyav_available():
+    import av
+
+
+if is_transformers_version_greater_than("4.45.0"):
+    from transformers.models.mllama.processing_mllama import (
+        convert_sparse_cross_attention_mask_to_dense,
+        get_cross_attention_token_mask,
+    )
+
+
+if TYPE_CHECKING:
+    from av.stream import Stream
+    from transformers import PreTrainedTokenizer, ProcessorMixin
+    from transformers.image_processing_utils import BaseImageProcessor
+
+    class EncodedImage(TypedDict):
+        path: Optional[str]
+        bytes: Optional[bytes]
+
+    ImageInput = Union[str, bytes, EncodedImage, ImageObject]
+    VideoInput = str
+
+
+def _get_paligemma_token_type_ids(
+    imglens: Sequence[int], seqlens: Sequence[int], processor: "ProcessorMixin"
+) -> List[List[int]]:
+    r"""
+    Gets paligemma token type ids for computing loss.
+
+    Returns:
+        batch_token_type_ids: shape (batch_size, sequence_length)
+    """
+    batch_token_type_ids = []
+    for imglen, seqlen in zip(imglens, seqlens):
+        image_seqlen = imglen * getattr(processor, "image_seqlen")
+        batch_token_type_ids.append([0] * image_seqlen + [1] * (seqlen - image_seqlen))
+
+    return batch_token_type_ids
+
+
+class BasePlugin:
+    def __init__(self, image_token: Optional[str], video_token: Optional[str]) -> None:
+        self.image_token = image_token
+        self.video_token = video_token
+
+    def _validate_input(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+    ) -> None:
+        r"""
+        Validates if this model accepts the input modalities.
+        """
+        if len(images) != 0 and self.image_token is None:
+            raise ValueError("This model does not support image input.")
+
+        if len(videos) != 0 and self.video_token is None:
+            raise ValueError("This model does not support video input.")
+
+    def _preprocess_image(self, image: "ImageObject", **kwargs) -> "ImageObject":
+        r"""
+        Pre-processes a single image.
+        """
+        image_resolution: int = kwargs.get("image_resolution")
+        if (image.width * image.height) > image_resolution:
+            resize_factor = math.sqrt(image_resolution / (image.width * image.height))
+            width, height = int(image.width * resize_factor), int(image.height * resize_factor)
+            image = image.resize((width, height), resample=Image.NEAREST)
+
+        if image.mode != "RGB":
+            image = image.convert("RGB")
+
+        return image
+
+    def _get_video_sample_frames(self, video_stream: "Stream", **kwargs) -> int:
+        r"""
+        Computes video sample frames according to fps.
+        """
+        video_fps: float = kwargs.get("video_fps")
+        video_maxlen: int = kwargs.get("video_maxlen")
+        total_frames = video_stream.frames
+        sample_frames = float(video_stream.duration * video_stream.time_base) * video_fps
+        sample_frames = min(total_frames, video_maxlen, sample_frames)
+        return math.floor(sample_frames)
+
+    def _regularize_images(self, images: Sequence["ImageInput"], **kwargs) -> List["ImageObject"]:
+        r"""
+        Regularizes images to avoid error. Including reading and pre-processing.
+        """
+        results = []
+        for image in images:
+            if isinstance(image, str):
+                image = Image.open(image)
+            elif isinstance(image, bytes):
+                image = Image.open(BytesIO(image))
+            elif isinstance(image, dict):
+                if image["bytes"] is not None:
+                    image = Image.open(BytesIO(image["bytes"]))
+                else:
+                    image = Image.open(image["path"])
+
+            if not isinstance(image, ImageObject):
+                raise ValueError(f"Expect input is a list of Images, but got {type(image)}.")
+
+            results.append(self._preprocess_image(image, **kwargs))
+
+        return results
+
+    def _regularize_videos(self, videos: Sequence["VideoInput"], **kwargs) -> List[List["ImageObject"]]:
+        r"""
+        Regularizes videos to avoid error. Including reading, resizing and converting.
+        """
+        results = []
+        for video in videos:
+            container = av.open(video, "r")
+            video_stream = next(stream for stream in container.streams if stream.type == "video")
+            total_frames = video_stream.frames
+            sample_frames = self._get_video_sample_frames(video_stream, **kwargs)
+            sample_indices = np.linspace(0, total_frames - 1, sample_frames).astype(np.int32)
+            frames: List["ImageObject"] = []
+            container.seek(0)
+            for frame_idx, frame in enumerate(container.decode(video_stream)):
+                if frame_idx in sample_indices:
+                    frames.append(frame.to_image())
+
+            frames = self._regularize_images(frames, **kwargs)
+            results.append(frames)
+
+        return results
+
+    def _get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: "ProcessorMixin",
+    ) -> Dict[str, "torch.Tensor"]:
+        r"""
+        Processes visual inputs.
+
+        Returns: (llava and paligemma)
+            pixel_values: tensor with shape (B, C, H, W)
+
+        Returns: (qwen2-vl)
+            pixel_values: tensor with shape (num_patches, patch_dim)
+            image_grid_thw: tensor with shape (num_images, 3), where the three numbers are time, width, height
+
+        It holds num_patches == torch.prod(image_grid_thw)
+        """
+        image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
+        video_processor: "BaseImageProcessor" = getattr(processor, "video_processor", image_processor)
+        input_dict = {"images": None}  # default key
+        if len(images) != 0:
+            images = self._regularize_images(
+                images,
+                image_resolution=getattr(processor, "image_resolution", 512 * 512),
+            )
+            input_dict["images"] = images
+
+        if len(videos) != 0:
+            videos = self._regularize_videos(
+                videos,
+                image_resolution=getattr(processor, "video_resolution", 128 * 128),
+                video_fps=getattr(processor, "video_fps", 2.0),
+                video_maxlen=getattr(processor, "video_maxlen", 64),
+            )
+            input_dict["videos"] = videos
+
+        mm_inputs = {}
+        if image_processor != video_processor:
+            if input_dict.get("images") is not None:
+                mm_inputs.update(image_processor(input_dict["images"], return_tensors="pt"))
+            if input_dict.get("videos") is not None:
+                mm_inputs.update(video_processor(input_dict["videos"], return_tensors="pt"))
+        elif input_dict.get("images") is not None or input_dict.get("videos") is not None:  # same processor (qwen2-vl)
+            mm_inputs.update(image_processor(**input_dict, return_tensors="pt"))
+
+        return mm_inputs
+
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        r"""
+        Pre-processes input messages before tokenization for VLMs.
+        """
+        self._validate_input(images, videos)
+        return messages
+
+    def process_token_ids(
+        self,
+        input_ids: List[int],
+        labels: Optional[List[int]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        tokenizer: "PreTrainedTokenizer",
+        processor: Optional["ProcessorMixin"],
+    ) -> Tuple[List[int], Optional[List[int]]]:
+        r"""
+        Pre-processes token ids after tokenization for VLMs.
+        """
+        self._validate_input(images, videos)
+        return input_ids, labels
+
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        r"""
+        Builds batched multimodal inputs for VLMs.
+
+        Arguments:
+            images: a list of image inputs, shape (num_images,)
+            videos: a list of video inputs, shape (num_videos,)
+            imglens: number of images in each sample, shape (batch_size,)
+            vidlens: number of videos in each sample, shape (batch_size,)
+            batch_ids: input ids of samples, shape (batch_size, seq_len)
+            processor: a processor for pre-processing images and videos
+        """
+        self._validate_input(images, videos)
+        return {}
+
+
+class LlavaPlugin(BasePlugin):
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        num_image_tokens = 0
+        image_seqlen = getattr(processor, "image_seqlen")
+        messages = deepcopy(messages)
+        for message in messages:
+            content = message["content"]
+            while IMAGE_PLACEHOLDER in content:
+                num_image_tokens += 1
+                content = content.replace(IMAGE_PLACEHOLDER, "{{image}}" * image_seqlen, 1)
+
+            message["content"] = content.replace("{{image}}", self.image_token)
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        return self._get_mm_inputs(images, videos, processor)
+
+
+class LlavaNextPlugin(BasePlugin):
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        num_image_tokens = 0
+        messages = deepcopy(messages)
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        if "image_sizes" in mm_inputs:
+            image_sizes = iter(mm_inputs["image_sizes"])
+
+        if "pixel_values" in mm_inputs:
+            height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))
+
+        for message in messages:
+            content = message["content"]
+            while IMAGE_PLACEHOLDER in content:
+                image_size = next(image_sizes)
+                orig_height, orig_width = image_size
+                image_seqlen = processor._get_number_of_features(orig_height, orig_width, height, width)
+                if getattr(processor, "vision_feature_select_strategy") == "default":
+                    image_seqlen -= 1
+
+                num_image_tokens += 1
+                content = content.replace(IMAGE_PLACEHOLDER, "{{image}}" * image_seqlen, 1)
+
+            message["content"] = content.replace("{{image}}", self.image_token)
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        return self._get_mm_inputs(images, videos, processor)
+
+
+class LlavaNextVideoPlugin(BasePlugin):
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        num_image_tokens, num_video_tokens = 0, 0
+        messages = deepcopy(messages)
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        if "pixel_values" in mm_inputs:
+            image_sizes = iter(mm_inputs["image_sizes"])
+            height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))
+            for message in messages:
+                content = message["content"]
+                while IMAGE_PLACEHOLDER in content:
+                    image_size = next(image_sizes)
+                    orig_height, orig_width = image_size
+                    image_seqlen = processor._get_number_of_features(orig_height, orig_width, height, width)
+                    if getattr(processor, "vision_feature_select_strategy") == "default":
+                        image_seqlen -= 1
+
+                    num_image_tokens += 1
+                    content = content.replace(IMAGE_PLACEHOLDER, "{{image}}" * image_seqlen, 1)
+
+                message["content"] = content.replace("{{image}}", self.image_token)
+
+        if "pixel_values_videos" in mm_inputs:
+            pixel_values_video = to_numpy_array(mm_inputs.get("pixel_values_videos")[0])
+            height, width = get_image_size(pixel_values_video[0])
+            num_frames = pixel_values_video.shape[0]  # frame dim is always after batch dim
+            image_seqlen = (height // processor.patch_size) * (width // processor.patch_size)
+            video_seqlen = image_seqlen // 4 * num_frames  # divide by 4 needed for avg pooling layer
+            for message in messages:
+                content = message["content"]
+                while VIDEO_PLACEHOLDER in content:
+                    num_video_tokens += 1
+                    content = content.replace(VIDEO_PLACEHOLDER, "{{video}}" * video_seqlen, 1)
+
+                message["content"] = content.replace("{{video}}", self.video_token)
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        if len(videos) != num_video_tokens:
+            raise ValueError(f"The number of videos does not match the number of {VIDEO_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        return self._get_mm_inputs(images, videos, processor)
+
+
+class PaliGemmaPlugin(BasePlugin):
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        num_image_tokens = 0
+        messages = deepcopy(messages)
+        for message in messages:
+            content = message["content"]
+            while IMAGE_PLACEHOLDER in content:
+                num_image_tokens += 1
+                content = content.replace(IMAGE_PLACEHOLDER, "{{image}}", 1)
+
+            message["content"] = content.replace("{{image}}", "")
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def process_token_ids(
+        self,
+        input_ids: List[int],
+        labels: Optional[List[int]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        tokenizer: "PreTrainedTokenizer",
+        processor: Optional["ProcessorMixin"],
+    ) -> Tuple[List[int], Optional[List[int]]]:
+        self._validate_input(images, videos)
+        num_images = len(images)
+        image_seqlen = num_images * getattr(processor, "image_seqlen")
+        image_token_id = tokenizer.convert_tokens_to_ids(self.image_token)
+        input_ids = [image_token_id] * image_seqlen + input_ids
+        if labels is not None:
+            labels = [IGNORE_INDEX] * image_seqlen + labels
+
+        return input_ids, labels
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        seqlens = [len(input_ids) for input_ids in batch_ids]
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        mm_inputs["token_type_ids"] = _get_paligemma_token_type_ids(imglens, seqlens, processor)
+        return mm_inputs
+
+
+class PixtralPlugin(BasePlugin):
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        patch_size = getattr(processor, "patch_size")
+        image_token = getattr(processor, "image_token")
+        image_break_token = getattr(processor, "image_break_token")
+        image_end_token = getattr(processor, "image_end_token")
+
+        num_image_tokens = 0
+        messages = deepcopy(messages)
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        image_input_sizes = mm_inputs.get("image_sizes", None)
+        for message in messages:
+            content = message["content"]
+            while IMAGE_PLACEHOLDER in content:
+                if image_input_sizes is None:
+                    raise ValueError("Cannot get image input sizes.")
+
+                image_size = image_input_sizes[0][num_image_tokens]
+                height, width = image_size
+                num_height_tokens = height // patch_size
+                num_width_tokens = width // patch_size
+                replace_tokens = [[image_token] * num_width_tokens + [image_break_token]] * num_height_tokens
+                replace_tokens = [item for sublist in replace_tokens for item in sublist]  # flatten list
+                replace_tokens[-1] = image_end_token
+                replace_str = "".join(replace_tokens)
+                content = content.replace(IMAGE_PLACEHOLDER, replace_str, 1)
+                num_image_tokens += 1
+
+            message["content"] = content
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        if mm_inputs.get("pixel_values"):
+            mm_inputs["pixel_values"] = mm_inputs["pixel_values"][0]
+
+        mm_inputs.pop("image_sizes", None)
+        return mm_inputs
+
+
+class Qwen2vlPlugin(BasePlugin):
+    @override
+    def _preprocess_image(self, image: "ImageObject", **kwargs) -> "ImageObject":
+        image = super()._preprocess_image(image, **kwargs)
+        if min(image.width, image.height) < 28:
+            width, height = max(image.width, 28), max(image.height, 28)
+            image = image.resize((width, height), resample=Image.NEAREST)
+
+        if image.width / image.height > 200:
+            width, height = image.height * 180, image.height
+            image = image.resize((width, height), resample=Image.NEAREST)
+
+        if image.height / image.width > 200:
+            width, height = image.width, image.width * 180
+            image = image.resize((width, height), resample=Image.NEAREST)
+
+        return image
+
+    @override
+    def _get_video_sample_frames(self, video_stream: "Stream", **kwargs) -> int:
+        sample_frames = super()._get_video_sample_frames(video_stream, **kwargs)
+        sample_frames = sample_frames // 2 * 2
+        return sample_frames
+
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
+        merge_length: int = getattr(image_processor, "merge_size") ** 2
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        image_grid_thw = mm_inputs.get("image_grid_thw", [])
+        video_grid_thw = mm_inputs.get("video_grid_thw", [])
+
+        num_image_tokens, num_video_tokens = 0, 0
+        messages = deepcopy(messages)
+        for message in messages:
+            content = message["content"]
+            while IMAGE_PLACEHOLDER in content:
+                if num_image_tokens >= len(image_grid_thw):
+                    raise ValueError(f"`len(images)` is less than the number of {IMAGE_PLACEHOLDER} tokens.")
+
+                content = content.replace(
+                    IMAGE_PLACEHOLDER,
+                    "<|vision_start|>{}<|vision_end|>".format(
+                        self.image_token * (image_grid_thw[num_image_tokens].prod() // merge_length)
+                    ),
+                    1,
+                )
+                num_image_tokens += 1
+
+            while VIDEO_PLACEHOLDER in content:
+                if num_video_tokens >= len(video_grid_thw):
+                    raise ValueError(f"`len(videos)` is less than the number of {VIDEO_PLACEHOLDER} tokens.")
+
+                content = content.replace(
+                    VIDEO_PLACEHOLDER,
+                    "<|vision_start|>{}<|vision_end|>".format(
+                        self.video_token * (video_grid_thw[num_video_tokens].prod() // merge_length)
+                    ),
+                    1,
+                )
+                num_video_tokens += 1
+
+            message["content"] = content
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        if len(videos) != num_video_tokens:
+            raise ValueError(f"The number of videos does not match the number of {VIDEO_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        return self._get_mm_inputs(images, videos, processor)
+
+
+class VideoLlavaPlugin(BasePlugin):
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        num_image_tokens, num_video_tokens = 0, 0
+        messages = deepcopy(messages)
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        num_frames = 0
+        has_images = "pixel_values_images" in mm_inputs
+        has_videos = "pixel_values_videos" in mm_inputs
+        if has_images or has_videos:
+            if has_images:
+                height, width = get_image_size(to_numpy_array(mm_inputs.get("pixel_values_images")[0]))
+                num_frames = 1
+
+            if has_videos:
+                pixel_values_video = to_numpy_array(mm_inputs.get("pixel_values_videos")[0])
+                height, width = get_image_size(pixel_values_video[0])
+                num_frames = pixel_values_video.shape[0]  # frame dim is always after batch dim
+
+            image_seqlen = (height // processor.patch_size) * (width // processor.patch_size) + 1
+            video_seqlen = image_seqlen * num_frames
+            if getattr(processor, "vision_feature_select_strategy") == "default":
+                image_seqlen -= 1
+
+            for message in messages:
+                content = message["content"]
+                while IMAGE_PLACEHOLDER in content:
+                    num_image_tokens += 1
+                    content = content.replace(IMAGE_PLACEHOLDER, "{{image}}" * image_seqlen, 1)
+
+                while VIDEO_PLACEHOLDER in content:
+                    num_video_tokens += 1
+                    content = content.replace(VIDEO_PLACEHOLDER, "{{video}}" * video_seqlen, 1)
+
+                content = content.replace("{{image}}", self.image_token)
+                message["content"] = content.replace("{{video}}", self.video_token)
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        if len(videos) != num_video_tokens:
+            raise ValueError(f"The number of videos does not match the number of {VIDEO_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        return self._get_mm_inputs(images, videos, processor)
+
+
+class MllamaPlugin(BasePlugin):
+    @override
+    def process_messages(
+        self,
+        messages: Sequence[Dict[str, str]],
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: Optional["ProcessorMixin"],
+    ) -> List[Dict[str, str]]:
+        self._validate_input(images, videos)
+        num_image_tokens = 0
+        messages = deepcopy(messages)
+        for message in messages:
+            content = message["content"]
+            num_image_tokens += content.count(IMAGE_PLACEHOLDER)
+            message["content"] = content.replace(IMAGE_PLACEHOLDER, self.image_token)
+
+        if len(images) != num_image_tokens:
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens.")
+
+        return messages
+
+    @override
+    def _get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        processor: "ProcessorMixin",
+    ) -> Dict[str, "torch.Tensor"]:
+        r"""
+        Processes visual inputs for mllama because its image processor only accepts List[List[ImageInput]].
+
+        Returns:
+            pixel_values: tensor with shape
+                          (batch_size, max_num_images, max_image_tiles, channels, tile_height, tile_width)
+                          For example, (2, 1, 4, 3, 560, 560).
+            aspect_ratio_ids: tensor with shape (batch_size, max_num_images). For example, (2, 1).
+            aspect_ratio_mask: tensor with shape (batch_size, max_num_images, max_image_tiles). For example, (2, 1, 4).
+            num_tiles: List[List[int]] with shape (batch_size, num_images_in_batch). For example, (2, 1).
+        """
+        image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
+        images = self._regularize_images(images, image_resolution=getattr(processor, "image_resolution", 512 * 512))
+        return image_processor([[image] for image in images], return_tensors="pt")
+
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        batch_ids: Sequence[List[int]],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        if len(images) != len(batch_ids):
+            raise ValueError("Mllama only supports one image per sample.")
+
+        mm_inputs = self._get_mm_inputs(images, videos, processor)
+        num_tiles = mm_inputs.pop("num_tiles")
+        image_token_id = getattr(processor, "image_token_id")
+        max_image_tiles = getattr(processor.image_processor, "max_image_tiles")
+        cross_attention_token_mask = [
+            get_cross_attention_token_mask(input_ids, image_token_id) for input_ids in batch_ids
+        ]
+        mm_inputs["cross_attention_mask"] = convert_sparse_cross_attention_mask_to_dense(
+            cross_attention_token_mask,
+            num_tiles=num_tiles,
+            max_num_tiles=max_image_tiles,
+            length=max(len(input_ids) for input_ids in batch_ids),
+        )
+        return mm_inputs
+
+
+PLUGINS = {
+    "base": BasePlugin,
+    "llava": LlavaPlugin,
+    "llava_next": LlavaNextPlugin,
+    "llava_next_video": LlavaNextVideoPlugin,
+    "paligemma": PaliGemmaPlugin,
+    "pixtral": PixtralPlugin,
+    "qwen2_vl": Qwen2vlPlugin,
+    "video_llava": VideoLlavaPlugin,
+    "mllama": MllamaPlugin,
+}
+
+
+def get_mm_plugin(
+    name: str,
+    image_token: Optional[str] = None,
+    video_token: Optional[str] = None,
+) -> "BasePlugin":
+    plugin_class = PLUGINS.get(name, None)
+    if plugin_class is None:
+        raise ValueError(f"Multimodal plugin `{name}` not found.")
+
+    return plugin_class(image_token, video_token)

src/llamafactory/data/parser.py

@@ -20,7 +20,7 @@ from typing import Any, Dict, List, Literal, Optional, Sequence
 from transformers.utils import cached_file
 
 from ..extras.constants import DATA_CONFIG
-from ..extras.misc import use_modelscope
+from ..extras.misc import use_modelscope, use_openmind
 
 
 @dataclass
@@ -30,7 +30,7 @@ class DatasetAttr:
     """
 
     # basic configs
-    load_from: Literal["hf_hub", "ms_hub", "script", "file"]
+    load_from: Literal["hf_hub", "ms_hub", "om_hub", "script", "file"]
     dataset_name: str
     formatting: Literal["alpaca", "sharegpt"] = "alpaca"
     ranking: bool = False
@@ -43,6 +43,7 @@ class DatasetAttr:
     system: Optional[str] = None
     tools: Optional[str] = None
     images: Optional[str] = None
+    videos: Optional[str] = None
     # rlhf columns
     chosen: Optional[str] = None
     rejected: Optional[str] = None
@@ -86,31 +87,39 @@ def get_dataset_list(dataset_names: Optional[Sequence[str]], dataset_dir: str) -
             config_path = os.path.join(dataset_dir, DATA_CONFIG)
 
         try:
-            with open(config_path, "r") as f:
+            with open(config_path) as f:
                 dataset_info = json.load(f)
         except Exception as err:
             if len(dataset_names) != 0:
-                raise ValueError("Cannot open {} due to {}.".format(config_path, str(err)))
+                raise ValueError(f"Cannot open {config_path} due to {str(err)}.")
 
             dataset_info = None
 
     dataset_list: List["DatasetAttr"] = []
     for name in dataset_names:
         if dataset_info is None:  # dataset_dir is ONLINE
-            load_from = "ms_hub" if use_modelscope() else "hf_hub"
+            if use_modelscope():
+                load_from = "ms_hub"
+            elif use_openmind():
+                load_from = "om_hub"
+            else:
+                load_from = "hf_hub"
             dataset_attr = DatasetAttr(load_from, dataset_name=name)
             dataset_list.append(dataset_attr)
             continue
 
         if name not in dataset_info:
-            raise ValueError("Undefined dataset {} in {}.".format(name, DATA_CONFIG))
+            raise ValueError(f"Undefined dataset {name} in {DATA_CONFIG}.")
 
         has_hf_url = "hf_hub_url" in dataset_info[name]
         has_ms_url = "ms_hub_url" in dataset_info[name]
+        has_om_url = "om_hub_url" in dataset_info[name]
 
-        if has_hf_url or has_ms_url:
+        if has_hf_url or has_ms_url or has_om_url:
-            if (use_modelscope() and has_ms_url) or (not has_hf_url):
+            if has_ms_url and (use_modelscope() or not has_hf_url):
                 dataset_attr = DatasetAttr("ms_hub", dataset_name=dataset_info[name]["ms_hub_url"])
+            elif has_om_url and (use_openmind() or not has_hf_url):
+                dataset_attr = DatasetAttr("om_hub", dataset_name=dataset_info[name]["om_hub_url"])
             else:
                 dataset_attr = DatasetAttr("hf_hub", dataset_name=dataset_info[name]["hf_hub_url"])
         elif "script_url" in dataset_info[name]:
@@ -126,7 +135,7 @@ def get_dataset_list(dataset_names: Optional[Sequence[str]], dataset_dir: str) -
         dataset_attr.set_attr("num_samples", dataset_info[name])
 
         if "columns" in dataset_info[name]:
-            column_names = ["system", "tools", "images", "chosen", "rejected", "kto_tag"]
+            column_names = ["system", "tools", "images", "videos", "chosen", "rejected", "kto_tag"]
             if dataset_attr.formatting == "alpaca":
                 column_names.extend(["prompt", "query", "response", "history"])
             else:

src/llamafactory/data/preprocess.py

@@ -50,7 +50,7 @@ def get_preprocess_and_print_func(
         print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
     elif stage == "sft" and not do_generate:
         if data_args.packing:
-            if data_args.neat_packing:
+            if data_args.neat_packing:  # hack datasets to have int32 attention mask
                 from datasets.arrow_writer import OptimizedTypedSequence, TypedSequence
 
                 def __init__(self, data, **kwargs):
@@ -67,6 +67,7 @@ def get_preprocess_and_print_func(
                 preprocess_packed_supervised_dataset,
                 template=template,
                 tokenizer=tokenizer,
+                processor=processor,
                 data_args=data_args,
             )
         else:

src/llamafactory/data/processors/feedback.py

@@ -12,21 +12,23 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from collections import defaultdict
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
+from ...extras import logging
 from ...extras.constants import IGNORE_INDEX
-from ...extras.logging import get_logger
+from .processor_utils import infer_seqlen
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, infer_seqlen
 
 
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
+    from ..mm_plugin import ImageInput, VideoInput
     from ..template import Template
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 def _encode_feedback_example(
@@ -35,14 +37,13 @@ def _encode_feedback_example(
     kl_response: Sequence[Dict[str, str]],
     system: Optional[str],
     tools: Optional[str],
+    images: Sequence["ImageInput"],
+    videos: Sequence["VideoInput"],
     template: "Template",
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
-    data_args: "DataArguments",
+    cutoff_len: int,
 ) -> Tuple[List[int], List[int], List[int], List[int], bool]:
-    if processor is not None and not hasattr(processor, "image_seq_length"):  # llava-like models
-        prompt[0]["content"] = template.image_token + prompt[0]["content"]
-
     if response[0]["content"]:  # desired example
         kto_tag = True
         messages = prompt + [response[0]]
@@ -55,6 +56,8 @@ def _encode_feedback_example(
     else:
         kl_messages = prompt + [kl_response[1]]
 
+    messages = template.mm_plugin.process_messages(messages, images, videos, processor)
+    kl_messages = template.mm_plugin.process_messages(kl_messages, images, videos, processor)
     prompt_ids, response_ids = template.encode_oneturn(tokenizer, messages, system, tools)
     kl_prompt_ids, kl_response_ids = template.encode_oneturn(tokenizer, kl_messages, system, tools)
 
@@ -62,15 +65,13 @@ def _encode_feedback_example(
         response_ids += [tokenizer.eos_token_id]
         kl_response_ids += [tokenizer.eos_token_id]
 
-    if processor is not None and hasattr(processor, "image_seq_length"):  # paligemma models
+    prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, videos, tokenizer, processor)
-        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
+    kl_prompt_ids, _ = template.mm_plugin.process_token_ids(kl_prompt_ids, None, images, videos, tokenizer, processor)
-        prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
-        kl_prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + kl_prompt_ids
 
-    source_len, target_len = infer_seqlen(len(prompt_ids), len(response_ids), data_args.cutoff_len)
+    source_len, target_len = infer_seqlen(len(prompt_ids), len(response_ids), cutoff_len)
     prompt_ids = prompt_ids[:source_len]
     response_ids = response_ids[:target_len]
-    kl_source_len, kl_target_len = infer_seqlen(len(kl_prompt_ids), len(kl_response_ids), data_args.cutoff_len)
+    kl_source_len, kl_target_len = infer_seqlen(len(kl_prompt_ids), len(kl_response_ids), cutoff_len)
     kl_prompt_ids = kl_prompt_ids[:kl_source_len]
     kl_response_ids = kl_response_ids[:kl_target_len]
 
@@ -78,7 +79,6 @@ def _encode_feedback_example(
     labels = [IGNORE_INDEX] * source_len + response_ids
     kl_input_ids = kl_prompt_ids + kl_response_ids
     kl_labels = [IGNORE_INDEX] * kl_source_len + kl_response_ids
-
     return input_ids, labels, kl_input_ids, kl_labels, kto_tag
 
 
@@ -88,39 +88,29 @@ def preprocess_feedback_dataset(
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
     data_args: "DataArguments",
-) -> Dict[str, List[List[int]]]:
+) -> Dict[str, List[Any]]:
     # create unrelated input-output pairs for estimating the KL term by flipping the matched pairs
-    kl_response = examples["response"][::-1]
+    kl_response = examples["_response"][::-1]
-    model_inputs = {
+    model_inputs = defaultdict(list)
-        "input_ids": [],
+    for i in range(len(examples["_prompt"])):
-        "attention_mask": [],
+        if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) < 2:
-        "labels": [],
+            logger.warning_rank0(
-        "kl_input_ids": [],
+                "Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i])
-        "kl_attention_mask": [],
+            )
-        "kl_labels": [],
-        "kto_tags": [],
-    }
-    if processor is not None:
-        model_inputs["pixel_values"] = []
-        if hasattr(processor, "image_seq_length"):  # paligemma models
-            model_inputs["token_type_ids"] = []
-            model_inputs["kl_token_type_ids"] = []
-
-    for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
-            logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
             continue
 
         input_ids, labels, kl_input_ids, kl_labels, kto_tag = _encode_feedback_example(
-            prompt=examples["prompt"][i],
+            prompt=examples["_prompt"][i],
-            response=examples["response"][i],
+            response=examples["_response"][i],
             kl_response=kl_response[i],
-            system=examples["system"][i],
+            system=examples["_system"][i],
-            tools=examples["tools"][i],
+            tools=examples["_tools"][i],
+            images=examples["_images"][i] or [],
+            videos=examples["_videos"][i] or [],
             template=template,
             tokenizer=tokenizer,
             processor=processor,
-            data_args=data_args,
+            cutoff_len=data_args.cutoff_len,
         )
         model_inputs["input_ids"].append(input_ids)
         model_inputs["attention_mask"].append([1] * len(input_ids))
@@ -129,15 +119,12 @@ def preprocess_feedback_dataset(
         model_inputs["kl_attention_mask"].append([1] * len(kl_input_ids))
         model_inputs["kl_labels"].append(kl_labels)
         model_inputs["kto_tags"].append(kto_tag)
-        if processor is not None:
+        model_inputs["images"].append(examples["_images"][i])
-            model_inputs["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
+        model_inputs["videos"].append(examples["_videos"][i])
-            if hasattr(processor, "image_seq_length"):  # paligemma models
-                model_inputs["token_type_ids"].append(get_paligemma_token_type_ids(len(input_ids), processor))
-                model_inputs["kl_token_type_ids"].append(get_paligemma_token_type_ids(len(kl_input_ids), processor))
 
     desirable_num = sum([1 for tag in model_inputs["kto_tags"] if tag])
     undesirable_num = len(model_inputs["kto_tags"]) - desirable_num
     if desirable_num == 0 or undesirable_num == 0:
-        logger.warning("Your dataset only has one preference type.")
+        logger.warning_rank0("Your dataset only has one preference type.")
 
     return model_inputs

src/llamafactory/data/processors/pairwise.py

@@ -12,21 +12,23 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from collections import defaultdict
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
+from ...extras import logging
 from ...extras.constants import IGNORE_INDEX
-from ...extras.logging import get_logger
+from .processor_utils import infer_seqlen
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, infer_seqlen
 
 
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
+    from ..mm_plugin import ImageInput, VideoInput
     from ..template import Template
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 def _encode_pairwise_example(
@@ -34,16 +36,15 @@ def _encode_pairwise_example(
     response: Sequence[Dict[str, str]],
     system: Optional[str],
     tools: Optional[str],
+    images: Sequence["ImageInput"],
+    videos: Sequence["VideoInput"],
     template: "Template",
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
-    data_args: "DataArguments",
+    cutoff_len: int,
 ) -> Tuple[List[int], List[int], List[int], List[int]]:
-    if processor is not None and not hasattr(processor, "image_seq_length"):  # llava-like models
+    chosen_messages = template.mm_plugin.process_messages(prompt + [response[0]], images, videos, processor)
-        prompt[0]["content"] = template.image_token + prompt[0]["content"]
+    rejected_messages = template.mm_plugin.process_messages(prompt + [response[1]], images, videos, processor)
-
-    chosen_messages = prompt + [response[0]]
-    rejected_messages = prompt + [response[1]]
     prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, chosen_messages, system, tools)
     _, rejected_ids = template.encode_oneturn(tokenizer, rejected_messages, system, tools)
 
@@ -51,13 +52,9 @@ def _encode_pairwise_example(
         chosen_ids += [tokenizer.eos_token_id]
         rejected_ids += [tokenizer.eos_token_id]
 
-    if processor is not None and hasattr(processor, "image_seq_length"):  # paligemma models
+    prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, videos, tokenizer, processor)
-        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
+    # consider the response is more important
-        prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
+    source_len, target_len = infer_seqlen(len(prompt_ids), max(len(chosen_ids), len(rejected_ids)), cutoff_len)
-
-    source_len, target_len = infer_seqlen(
-        len(prompt_ids), max(len(chosen_ids), len(rejected_ids)), data_args.cutoff_len
-    )  # consider the response is more important
     prompt_ids = prompt_ids[:source_len]
     chosen_ids = chosen_ids[:target_len]
     rejected_ids = rejected_ids[:target_len]
@@ -66,7 +63,6 @@ def _encode_pairwise_example(
     chosen_labels = [IGNORE_INDEX] * source_len + chosen_ids
     rejected_input_ids = prompt_ids + rejected_ids
     rejected_labels = [IGNORE_INDEX] * source_len + rejected_ids
-
     return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels
 
 
@@ -76,36 +72,27 @@ def preprocess_pairwise_dataset(
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
     data_args: "DataArguments",
-) -> Dict[str, List[List[int]]]:
+) -> Dict[str, List[Any]]:
     # build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
-    model_inputs = {
+    model_inputs = defaultdict(list)
-        "chosen_input_ids": [],
+    for i in range(len(examples["_prompt"])):
-        "chosen_attention_mask": [],
+        if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) < 2:
-        "chosen_labels": [],
+            logger.warning_rank0(
-        "rejected_input_ids": [],
+                "Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i])
-        "rejected_attention_mask": [],
+            )
-        "rejected_labels": [],
-    }
-    if processor is not None:
-        model_inputs["pixel_values"] = []
-        if hasattr(processor, "image_seq_length"):  # paligemma models
-            model_inputs["chosen_token_type_ids"] = []
-            model_inputs["rejected_token_type_ids"] = []
-
-    for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
-            logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
             continue
 
         chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels = _encode_pairwise_example(
-            prompt=examples["prompt"][i],
+            prompt=examples["_prompt"][i],
-            response=examples["response"][i],
+            response=examples["_response"][i],
-            system=examples["system"][i],
+            system=examples["_system"][i],
-            tools=examples["tools"][i],
+            tools=examples["_tools"][i],
+            images=examples["_images"][i] or [],
+            videos=examples["_videos"][i] or [],
             template=template,
             tokenizer=tokenizer,
             processor=processor,
-            data_args=data_args,
+            cutoff_len=data_args.cutoff_len,
         )
         model_inputs["chosen_input_ids"].append(chosen_input_ids)
         model_inputs["chosen_attention_mask"].append([1] * len(chosen_input_ids))
@@ -113,15 +100,8 @@ def preprocess_pairwise_dataset(
         model_inputs["rejected_input_ids"].append(rejected_input_ids)
         model_inputs["rejected_attention_mask"].append([1] * len(rejected_input_ids))
         model_inputs["rejected_labels"].append(rejected_labels)
-        if processor is not None:
+        model_inputs["images"].append(examples["_images"][i])
-            model_inputs["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
+        model_inputs["videos"].append(examples["_videos"][i])
-            if hasattr(processor, "image_seq_length"):  # paligemma models
-                model_inputs["chosen_token_type_ids"].append(
-                    get_paligemma_token_type_ids(len(chosen_input_ids), processor)
-                )
-                model_inputs["rejected_token_type_ids"].append(
-                    get_paligemma_token_type_ids(len(rejected_input_ids), processor)
-                )
 
     return model_inputs
 
@@ -132,8 +112,8 @@ def print_pairwise_dataset_example(example: Dict[str, List[int]], tokenizer: "Pr
     print("chosen_input_ids:\n{}".format(example["chosen_input_ids"]))
     print("chosen_inputs:\n{}".format(tokenizer.decode(example["chosen_input_ids"], skip_special_tokens=False)))
     print("chosen_label_ids:\n{}".format(example["chosen_labels"]))
-    print("chosen_labels:\n{}".format(tokenizer.decode(valid_chosen_labels, skip_special_tokens=False)))
+    print(f"chosen_labels:\n{tokenizer.decode(valid_chosen_labels, skip_special_tokens=False)}")
     print("rejected_input_ids:\n{}".format(example["rejected_input_ids"]))
     print("rejected_inputs:\n{}".format(tokenizer.decode(example["rejected_input_ids"], skip_special_tokens=False)))
     print("rejected_label_ids:\n{}".format(example["rejected_labels"]))
-    print("rejected_labels:\n{}".format(tokenizer.decode(valid_rejected_labels, skip_special_tokens=False)))
+    print(f"rejected_labels:\n{tokenizer.decode(valid_rejected_labels, skip_special_tokens=False)}")

src/llamafactory/data/processors/pretrain.py

@@ -27,16 +27,16 @@ if TYPE_CHECKING:
 
 def preprocess_pretrain_dataset(
     examples: Dict[str, List[Any]], tokenizer: "PreTrainedTokenizer", data_args: "DataArguments"
-) -> Dict[str, List[List[int]]]:
+) -> Dict[str, List[Any]]:
     # build grouped texts with format `X1 X2 X3 ...` if packing is enabled
     eos_token = "<|end_of_text|>" if data_args.template == "llama3" else tokenizer.eos_token
-    text_examples = [messages[0]["content"] + eos_token for messages in examples["prompt"]]
+    text_examples = [messages[0]["content"] + eos_token for messages in examples["_prompt"]]
 
     if not data_args.packing:
         if data_args.template == "gemma":
             text_examples = [tokenizer.bos_token + example for example in text_examples]
 
-        result = tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len, truncation=True)
+        result = tokenizer(text_examples, add_special_tokens=False, truncation=True, max_length=data_args.cutoff_len)
     else:
         tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
         concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}

src/llamafactory/data/processors/processor_utils.py

@@ -13,20 +13,7 @@
 # limitations under the License.
 
 import bisect
-from typing import TYPE_CHECKING, List, Sequence, Tuple
+from typing import List, Sequence, Tuple
-
-from ...extras.packages import is_pillow_available
-
-
-if is_pillow_available():
-    from PIL import Image
-
-
-if TYPE_CHECKING:
-    from numpy.typing import NDArray
-    from PIL.Image import Image as ImageObject
-    from transformers import ProcessorMixin
-    from transformers.image_processing_utils import BaseImageProcessor
 
 
 def search_for_fit(numbers: Sequence[int], capacity: int) -> int:
@@ -61,23 +48,6 @@ def greedy_knapsack(numbers: List[int], capacity: int) -> List[List[int]]:
     return knapsacks
 
 
-def get_pixel_values(images: Sequence["ImageObject"], processor: "ProcessorMixin") -> "NDArray":
-    r"""
-    Processes visual inputs. (currently only supports a single image)
-    """
-    image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
-    image = images[0] if len(images) != 0 else Image.new("RGB", (100, 100), (255, 255, 255))
-    return image_processor(image, return_tensors="pt")["pixel_values"][0]  # shape (C, H, W)
-
-
-def get_paligemma_token_type_ids(input_len: int, processor: "ProcessorMixin") -> List[int]:
-    r"""
-    Gets paligemma token type ids for computing loss.
-    """
-    image_seq_length = getattr(processor, "image_seq_length")
-    return [0] * image_seq_length + [1] * (input_len - image_seq_length)
-
-
 def infer_seqlen(source_len: int, target_len: int, cutoff_len: int) -> Tuple[int, int]:
     r"""
     Computes the real sequence length after truncation by the cutoff_len.

src/llamafactory/data/processors/supervised.py

@@ -15,19 +15,20 @@
 from collections import defaultdict
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
+from ...extras import logging
 from ...extras.constants import IGNORE_INDEX
-from ...extras.logging import get_logger
+from .processor_utils import greedy_knapsack, infer_seqlen
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, greedy_knapsack, infer_seqlen
 
 
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
+    from ..mm_plugin import ImageInput, VideoInput
     from ..template import Template
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 def _encode_supervised_example(
@@ -35,45 +36,47 @@ def _encode_supervised_example(
     response: Sequence[Dict[str, str]],
     system: Optional[str],
     tools: Optional[str],
+    images: Sequence["ImageInput"],
+    videos: Sequence["VideoInput"],
     template: "Template",
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
-    data_args: "DataArguments",
+    cutoff_len: int,
+    train_on_prompt: bool,
+    mask_history: bool,
 ) -> Tuple[List[int], List[int]]:
-    if processor is not None and not hasattr(processor, "image_seq_length"):  # llava-like models
+    messages = template.mm_plugin.process_messages(prompt + response, images, videos, processor)
-        prompt[0]["content"] = template.image_token + prompt[0]["content"]
+    input_ids, labels = template.mm_plugin.process_token_ids([], [], images, videos, tokenizer, processor)
-
-    messages = prompt + response
-    input_ids, labels = [], []
-
-    if processor is not None and hasattr(processor, "image_seq_length"):  # paligemma models
-        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
-        input_ids += [image_token_id] * getattr(processor, "image_seq_length")
-        labels += [IGNORE_INDEX] * getattr(processor, "image_seq_length")
-
     encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools)
-    total_length = 1 if template.efficient_eos else 0
+    total_length = len(input_ids) + (1 if template.efficient_eos else 0)
+    if mask_history:
+        encoded_pairs = encoded_pairs[::-1]  # high priority for last turns
+
     for turn_idx, (source_ids, target_ids) in enumerate(encoded_pairs):
-        if total_length >= data_args.cutoff_len:
+        if total_length >= cutoff_len:
             break
 
-        source_len, target_len = infer_seqlen(len(source_ids), len(target_ids), data_args.cutoff_len - total_length)
+        source_len, target_len = infer_seqlen(len(source_ids), len(target_ids), cutoff_len - total_length)
         source_ids = source_ids[:source_len]
         target_ids = target_ids[:target_len]
         total_length += source_len + target_len
 
-        if data_args.train_on_prompt:
+        if train_on_prompt:
             source_label = source_ids
-        elif turn_idx != 0 and template.efficient_eos:
+        elif template.efficient_eos:
             source_label = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (source_len - 1)
         else:
             source_label = [IGNORE_INDEX] * source_len
 
-        if data_args.mask_history and turn_idx != len(encoded_pairs) - 1:
+        if mask_history and turn_idx != 0:  # train on the last turn only
             target_label = [IGNORE_INDEX] * target_len
         else:
             target_label = target_ids
 
+        if mask_history:  # reversed sequences
+            input_ids = source_ids + target_ids + input_ids
+            labels = source_label + target_label + labels
+        else:
             input_ids += source_ids + target_ids
             labels += source_label + target_label
 
@@ -90,37 +93,36 @@ def preprocess_supervised_dataset(
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
     data_args: "DataArguments",
-) -> Dict[str, List[List[int]]]:
+) -> Dict[str, List[Any]]:
     # build inputs with format `<bos> X Y <eos>` and labels with format `<ignore> ... <ignore> Y <eos>`
     # for multiturn examples, we only mask the prompt part in each prompt-response pair.
-    model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
+    model_inputs = defaultdict(list)
-    if processor is not None:
+    for i in range(len(examples["_prompt"])):
-        model_inputs["pixel_values"] = []
+        if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) != 1:
-        if hasattr(processor, "image_seq_length"):  # paligemma models
+            logger.warning_rank0(
-            model_inputs["token_type_ids"] = []
+                "Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i])
-
+            )
-    for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
-            logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
             continue
 
         input_ids, labels = _encode_supervised_example(
-            prompt=examples["prompt"][i],
+            prompt=examples["_prompt"][i],
-            response=examples["response"][i],
+            response=examples["_response"][i],
-            system=examples["system"][i],
+            system=examples["_system"][i],
-            tools=examples["tools"][i],
+            tools=examples["_tools"][i],
+            images=examples["_images"][i] or [],
+            videos=examples["_videos"][i] or [],
             template=template,
             tokenizer=tokenizer,
             processor=processor,
-            data_args=data_args,
+            cutoff_len=data_args.cutoff_len,
+            train_on_prompt=data_args.train_on_prompt,
+            mask_history=data_args.mask_history,
         )
         model_inputs["input_ids"].append(input_ids)
         model_inputs["attention_mask"].append([1] * len(input_ids))
         model_inputs["labels"].append(labels)
-        if processor is not None:
+        model_inputs["images"].append(examples["_images"][i])
-            model_inputs["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
+        model_inputs["videos"].append(examples["_videos"][i])
-            if hasattr(processor, "image_seq_length"):  # paligemma models
-                model_inputs["token_type_ids"].append(get_paligemma_token_type_ids(len(input_ids), processor))
 
     return model_inputs
 
@@ -129,47 +131,60 @@ def preprocess_packed_supervised_dataset(
     examples: Dict[str, List[Any]],
     template: "Template",
     tokenizer: "PreTrainedTokenizer",
+    processor: Optional["ProcessorMixin"],
     data_args: "DataArguments",
-) -> Dict[str, List[List[int]]]:
+) -> Dict[str, List[Any]]:
+    # TODO: use `position_ids` to achieve packing
     # build inputs with format `<bos> X1 Y1 <eos> <bos> X2 Y2 <eos>`
     # and labels with format `<ignore> ... <ignore> Y1 <eos> <ignore> ... <ignore> Y2 <eos>`
     valid_num = 0
-    batch_input_ids, batch_labels = [], []
+    batch_input_ids, batch_labels, batch_images, batch_videos = [], [], [], []
     lengths = []
     length2indexes = defaultdict(list)
-    for i in range(len(examples["prompt"])):
+    for i in range(len(examples["_prompt"])):
-        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
+        if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) != 1:
-            logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
+            logger.warning_rank0(
+                "Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i])
+            )
             continue
 
         input_ids, labels = _encode_supervised_example(
-            prompt=examples["prompt"][i],
+            prompt=examples["_prompt"][i],
-            response=examples["response"][i],
+            response=examples["_response"][i],
-            system=examples["system"][i],
+            system=examples["_system"][i],
-            tools=examples["tools"][i],
+            tools=examples["_tools"][i],
+            images=examples["_images"][i] or [],
+            videos=examples["_videos"][i] or [],
             template=template,
             tokenizer=tokenizer,
-            processor=None,
+            processor=processor,
-            data_args=data_args,
+            cutoff_len=data_args.cutoff_len - 1,  # reserved for the padding token
+            train_on_prompt=data_args.train_on_prompt,
+            mask_history=data_args.mask_history,
         )
         length = len(input_ids)
         if length > data_args.cutoff_len:
-            logger.warning("Dropped lengthy example with length {} > {}.".format(length, data_args.cutoff_len))
+            logger.warning_rank0(f"Dropped lengthy example with length {length} > {data_args.cutoff_len}.")
         else:
             lengths.append(length)
             length2indexes[length].append(valid_num)
             batch_input_ids.append(input_ids)
             batch_labels.append(labels)
+            batch_images.append(examples["_images"][i] or [])
+            batch_videos.append(examples["_videos"][i] or [])
             valid_num += 1
 
-    model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
+    model_inputs = defaultdict(list)
-    knapsacks = greedy_knapsack(lengths, data_args.cutoff_len)
+    knapsacks = greedy_knapsack(lengths, data_args.cutoff_len - 1)  # reserved for the padding token
     for knapsack in knapsacks:
         packed_input_ids, packed_attention_masks, packed_labels = [], [], []
+        packed_images, packed_videos = [], []
         for i, length in enumerate(knapsack):
             index = length2indexes[length].pop()
             packed_input_ids += batch_input_ids[index]
             packed_labels += batch_labels[index]
+            packed_images += batch_images[index]
+            packed_videos += batch_videos[index]
             if data_args.neat_packing:
                 packed_attention_masks += [i + 1] * len(batch_input_ids[index])  # start from 1
             else:
@@ -190,6 +205,8 @@ def preprocess_packed_supervised_dataset(
         model_inputs["input_ids"].append(packed_input_ids)
         model_inputs["attention_mask"].append(packed_attention_masks)
         model_inputs["labels"].append(packed_labels)
+        model_inputs["images"].append(packed_images or None)
+        model_inputs["videos"].append(packed_videos or None)
 
     return model_inputs
 
@@ -199,4 +216,4 @@ def print_supervised_dataset_example(example: Dict[str, List[int]], tokenizer: "
     print("input_ids:\n{}".format(example["input_ids"]))
     print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
     print("label_ids:\n{}".format(example["labels"]))
-    print("labels:\n{}".format(tokenizer.decode(valid_labels, skip_special_tokens=False)))
+    print(f"labels:\n{tokenizer.decode(valid_labels, skip_special_tokens=False)}")

src/llamafactory/data/processors/unsupervised.py

@@ -12,21 +12,23 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from collections import defaultdict
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
-from ...extras.logging import get_logger
+from ...extras import logging
 from ..data_utils import Role
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, infer_seqlen
+from .processor_utils import infer_seqlen
 
 
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
+    from ..mm_plugin import ImageInput, VideoInput
     from ..template import Template
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 def _encode_unsupervised_example(
@@ -34,28 +36,25 @@ def _encode_unsupervised_example(
     response: Sequence[Dict[str, str]],
     system: Optional[str],
     tools: Optional[str],
+    images: Sequence["ImageInput"],
+    videos: Sequence["VideoInput"],
     template: "Template",
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
-    data_args: "DataArguments",
+    cutoff_len: int,
 ) -> Tuple[List[int], List[int]]:
-    if processor is not None and not hasattr(processor, "image_seq_length"):  # llava-like models
-        prompt[0]["content"] = template.image_token + prompt[0]["content"]
-
     if len(response) == 1:
         messages = prompt + response
     else:
         messages = prompt + [{"role": Role.ASSISTANT.value, "content": ""}]
 
+    messages = template.mm_plugin.process_messages(messages, images, videos, processor)
     input_ids, labels = template.encode_oneturn(tokenizer, messages, system, tools)
     if template.efficient_eos:
         labels += [tokenizer.eos_token_id]
 
-    if processor is not None and hasattr(processor, "image_seq_length"):  # paligemma models
+    input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, images, videos, tokenizer, processor)
-        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
+    source_len, target_len = infer_seqlen(len(input_ids), len(labels), cutoff_len)
-        input_ids = [image_token_id] * getattr(processor, "image_seq_length") + input_ids
-
-    source_len, target_len = infer_seqlen(len(input_ids), len(labels), data_args.cutoff_len)
     input_ids = input_ids[:source_len]
     labels = labels[:target_len]
     return input_ids, labels
@@ -67,36 +66,33 @@ def preprocess_unsupervised_dataset(
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
     data_args: "DataArguments",
-) -> Dict[str, List[List[int]]]:
+) -> Dict[str, List[Any]]:
     # build inputs with format `<bos> X` and labels with format `Y <eos>`
-    model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
+    model_inputs = defaultdict(list)
-    if processor is not None:
+    for i in range(len(examples["_prompt"])):
-        model_inputs["pixel_values"] = []
+        if len(examples["_prompt"][i]) % 2 != 1:
-        if hasattr(processor, "image_seq_length"):  # paligemma models
+            logger.warning_rank0(
-            model_inputs["token_type_ids"] = []
+                "Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i])
-
+            )
-    for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) % 2 != 1:
-            logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
             continue
 
         input_ids, labels = _encode_unsupervised_example(
-            prompt=examples["prompt"][i],
+            prompt=examples["_prompt"][i],
-            response=examples["response"][i],
+            response=examples["_response"][i],
-            system=examples["system"][i],
+            system=examples["_system"][i],
-            tools=examples["tools"][i],
+            tools=examples["_tools"][i],
+            images=examples["_images"][i] or [],
+            videos=examples["_videos"][i] or [],
             template=template,
             tokenizer=tokenizer,
             processor=processor,
-            data_args=data_args,
+            cutoff_len=data_args.cutoff_len,
         )
         model_inputs["input_ids"].append(input_ids)
         model_inputs["attention_mask"].append([1] * len(input_ids))
         model_inputs["labels"].append(labels)
-        if processor is not None:
+        model_inputs["images"].append(examples["_images"][i])
-            model_inputs["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
+        model_inputs["videos"].append(examples["_videos"][i])
-            if hasattr(processor, "image_seq_length"):  # paligemma models
-                model_inputs["token_type_ids"].append(get_paligemma_token_type_ids(len(input_ids), processor))
 
     return model_inputs
 

src/llamafactory/data/template.py

@@ -15,18 +15,24 @@
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Tuple, Union
 
-from ..extras.logging import get_logger
+from transformers.utils.versions import require_version
+from typing_extensions import override
+
+from ..extras import logging
 from .data_utils import Role
 from .formatter import EmptyFormatter, FunctionFormatter, StringFormatter, ToolFormatter
+from .mm_plugin import get_mm_plugin
 
 
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer
 
+    from ..hparams import DataArguments
     from .formatter import SLOTS, Formatter
+    from .mm_plugin import BasePlugin
 
 
-logger = get_logger(__name__)
+logger = logging.get_logger(__name__)
 
 
 @dataclass
@@ -41,9 +47,10 @@ class Template:
     format_prefix: "Formatter"
     default_system: str
     stop_words: List[str]
-    image_token: str
     efficient_eos: bool
     replace_eos: bool
+    replace_jinja_template: bool
+    mm_plugin: "BasePlugin"
 
     def encode_oneturn(
         self,
@@ -140,13 +147,14 @@ class Template:
                 elif "eos_token" in elem and tokenizer.eos_token_id is not None:
                     token_ids += [tokenizer.eos_token_id]
             else:
-                raise ValueError("Input must be string, set[str] or dict[str, str], got {}".format(type(elem)))
+                raise ValueError(f"Input must be string, set[str] or dict[str, str], got {type(elem)}")
 
         return token_ids
 
 
 @dataclass
 class Llama2Template(Template):
+    @override
     def _encode(
         self,
         tokenizer: "PreTrainedTokenizer",
@@ -190,7 +198,7 @@ class Llama2Template(Template):
         return encoded_messages
 
 
-TEMPLATES: Dict[str, Template] = {}
+TEMPLATES: Dict[str, "Template"] = {}
 
 
 def _register_template(
@@ -205,9 +213,10 @@ def _register_template(
     format_prefix: Optional["Formatter"] = None,
     default_system: str = "",
     stop_words: Sequence[str] = [],
-    image_token: str = "<image>",
     efficient_eos: bool = False,
     replace_eos: bool = False,
+    replace_jinja_template: bool = True,
+    mm_plugin: "BasePlugin" = get_mm_plugin(name="base"),
 ) -> None:
     r"""
     Registers a chat template.
@@ -254,9 +263,10 @@ def _register_template(
         format_prefix=format_prefix or default_prefix_formatter,
         default_system=default_system,
         stop_words=stop_words,
-        image_token=image_token,
         efficient_eos=efficient_eos,
         replace_eos=replace_eos,
+        replace_jinja_template=replace_jinja_template,
+        mm_plugin=mm_plugin,
     )
 
 
@@ -265,12 +275,12 @@ def _add_or_replace_eos_token(tokenizer: "PreTrainedTokenizer", eos_token: str)
     num_added_tokens = tokenizer.add_special_tokens({"eos_token": eos_token})
 
     if is_added:
-        logger.info("Add eos token: {}".format(tokenizer.eos_token))
+        logger.info_rank0(f"Add eos token: {tokenizer.eos_token}")
     else:
-        logger.info("Replace eos token: {}".format(tokenizer.eos_token))
+        logger.info_rank0(f"Replace eos token: {tokenizer.eos_token}")
 
     if num_added_tokens > 0:
-        logger.warning("New tokens have been added, make sure `resize_vocab` is True.")
+        logger.warning_rank0("New tokens have been added, make sure `resize_vocab` is True.")
 
 
 def _jinja_escape(content: str) -> str:
@@ -300,6 +310,9 @@ def _convert_slots_to_jinja(slots: "SLOTS", tokenizer: "PreTrainedTokenizer", pl
 
 
 def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer") -> str:
+    r"""
+    Returns the jinja template.
+    """
     jinja_template = ""
 
     prefix = _convert_slots_to_jinja(template.format_prefix.apply(), tokenizer)
@@ -310,14 +323,15 @@ def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer")
         jinja_template += "{% set system_message = '" + _jinja_escape(template.default_system) + "' %}"
 
     jinja_template += (
-        "{% if messages[0]['role'] == 'system' %}{% set system_message = messages[0]['content'] %}{% endif %}"
+        "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}"
+        "{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% endif %}"
     )
 
     system_message = _convert_slots_to_jinja(template.format_system.apply(), tokenizer, placeholder="system_message")
     if not isinstance(template, Llama2Template):
         jinja_template += "{% if system_message is defined %}{{ " + system_message + " }}{% endif %}"
 
-    jinja_template += "{% for message in messages %}"
+    jinja_template += "{% for message in loop_messages %}"
     jinja_template += "{% set content = message['content'] %}"
     if isinstance(template, Llama2Template):
         jinja_template += "{% if loop.index0 == 0 and system_message is defined %}"
@@ -338,23 +352,28 @@ def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer")
     return jinja_template
 
 
-def get_template_and_fix_tokenizer(
+def get_template_and_fix_tokenizer(tokenizer: "PreTrainedTokenizer", data_args: "DataArguments") -> "Template":
-    tokenizer: "PreTrainedTokenizer",
+    r"""
-    name: Optional[str] = None,
+    Gets chat template and fixes the tokenizer.
-    tool_format: Optional[str] = None,
+    """
-) -> Template:
+    if data_args.template is None:
-    if name is None:
         template = TEMPLATES["empty"]  # placeholder
     else:
-        template = TEMPLATES.get(name, None)
+        template = TEMPLATES.get(data_args.template, None)
         if template is None:
-            raise ValueError("Template {} does not exist.".format(name))
+            raise ValueError(f"Template {data_args.template} does not exist.")
 
-    if tool_format is not None:
+    if template.mm_plugin.__class__.__name__ != "BasePlugin":
-        logger.info("Using tool format: {}.".format(tool_format))
+        require_version("transformers>=4.45.0", "To fix: pip install transformers>=4.45.0")
+
+    if data_args.train_on_prompt and template.efficient_eos:
+        raise ValueError("Current template does not support `train_on_prompt`.")
+
+    if data_args.tool_format is not None:
+        logger.info_rank0(f"Using tool format: {data_args.tool_format}.")
         eos_slots = [] if template.efficient_eos else [{"eos_token"}]
-        template.format_tools = ToolFormatter(tool_format=tool_format)
+        template.format_function = FunctionFormatter(slots=eos_slots, tool_format=data_args.tool_format)
-        template.format_function = FunctionFormatter(slots=eos_slots, tool_format=tool_format)
+        template.format_tools = ToolFormatter(tool_format=data_args.tool_format)
 
     stop_words = template.stop_words
     if template.replace_eos:
@@ -369,20 +388,21 @@ def get_template_and_fix_tokenizer(
 
     if tokenizer.pad_token_id is None:
         tokenizer.pad_token = tokenizer.eos_token
-        logger.info("Add pad token: {}".format(tokenizer.pad_token))
+        logger.info_rank0(f"Add pad token: {tokenizer.pad_token}")
 
     if stop_words:
         num_added_tokens = tokenizer.add_special_tokens(
             dict(additional_special_tokens=stop_words), replace_additional_special_tokens=False
         )
-        logger.info("Add {} to stop words.".format(",".join(stop_words)))
+        logger.info_rank0("Add {} to stop words.".format(",".join(stop_words)))
         if num_added_tokens > 0:
-            logger.warning("New tokens have been added, make sure `resize_vocab` is True.")
+            logger.warning_rank0("New tokens have been added, make sure `resize_vocab` is True.")
 
+    if tokenizer.chat_template is None or template.replace_jinja_template:
         try:
             tokenizer.chat_template = _get_jinja_template(template, tokenizer)
-    except ValueError:
+        except ValueError as e:
-        logger.info("Cannot add this chat template to tokenizer.")
+            logger.info_rank0(f"Cannot add this chat template to tokenizer: {e}.")
 
     return template
 
@@ -549,6 +569,15 @@ _register_template(
 )
 
 
+_register_template(
+    name="cpm3",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+    stop_words=["<|im_end|>"],
+)
+
+
 _register_template(
     name="dbrx",
     format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
@@ -578,6 +607,7 @@ _register_template(
 _register_template(
     name="deepseek",
     format_user=StringFormatter(slots=["User: {{content}}\n\nAssistant:"]),
+    format_system=StringFormatter(slots=["{{content}}\n\n"]),
     format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
@@ -585,14 +615,14 @@ _register_template(
 _register_template(
     name="deepseekcoder",
     format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n### Response:"]),
-    format_assistant=StringFormatter(slots=["\n{{content}}\n"]),
+    format_assistant=StringFormatter(slots=["\n{{content}}\n<|EOT|>"]),
     format_separator=EmptyFormatter(slots=["\n"]),
     format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     default_system=(
-        "You are an AI programming assistant, utilizing the Deepseek Coder model, "
+        "You are an AI programming assistant, utilizing the DeepSeek Coder model, "
-        "developed by Deepseek Company, and you only answer questions related to computer science. "
+        "developed by DeepSeek Company, and you only answer questions related to computer science. "
         "For politically sensitive questions, security and privacy issues, "
-        "and other non-computer science questions, you will refuse to answer\n"
+        "and other non-computer science questions, you will refuse to answer.\n"
     ),
 )
 
@@ -611,6 +641,14 @@ _register_template(
 )
 
 
+_register_template(
+    name="exaone",
+    format_user=StringFormatter(slots=["[|user|]{{content}}\n[|assistant|]"]),
+    format_system=StringFormatter(slots=["[|system|]{{content}}[|endofturn|]\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+)
+
+
 _register_template(
     name="falcon",
     format_user=StringFormatter(slots=["User: {{content}}\nFalcon:"]),
@@ -635,6 +673,7 @@ _register_template(
     format_separator=EmptyFormatter(slots=["<end_of_turn>\n"]),
     format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     efficient_eos=True,
+    replace_jinja_template=False,
 )
 
 
@@ -652,6 +691,14 @@ _register_template(
 )
 
 
+_register_template(
+    name="index",
+    format_user=StringFormatter(slots=["reserved_0{{content}}reserved_1"]),
+    format_system=StringFormatter(slots=["<unk>{{content}}"]),
+    efficient_eos=True,
+)
+
+
 _register_template(
     name="intern",
     format_user=StringFormatter(slots=["<|User|>:{{content}}\n<|Bot|>:"]),
@@ -711,6 +758,146 @@ _register_template(
     format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     stop_words=["<|eot_id|>"],
     replace_eos=True,
+    replace_jinja_template=False,
+)
+
+
+_register_template(
+    name="mllama",
+    format_user=StringFormatter(
+        slots=[
+            (
+                "<|start_header_id|>user<|end_header_id|>\n\n{{content}}<|eot_id|>"
+                "<|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
+        ]
+    ),
+    format_system=StringFormatter(slots=["<|start_header_id|>system<|end_header_id|>\n\n{{content}}<|eot_id|>"]),
+    format_observation=StringFormatter(
+        slots=[
+            (
+                "<|start_header_id|>tool<|end_header_id|>\n\n{{content}}<|eot_id|>"
+                "<|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
+        ]
+    ),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+    stop_words=["<|eot_id|>"],
+    replace_eos=True,
+    replace_jinja_template=False,
+    mm_plugin=get_mm_plugin(name="mllama", image_token="<|image|>"),
+)
+
+
+_register_template(
+    name="llava",
+    format_user=StringFormatter(slots=["USER: {{content}} ASSISTANT:"]),
+    default_system=(
+        "A chat between a curious user and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the user's questions."
+    ),
+    mm_plugin=get_mm_plugin(name="llava", image_token="<image>"),
+)
+
+
+_register_template(
+    name="llava_next",
+    format_user=StringFormatter(slots=["USER: {{content}} ASSISTANT:"]),
+    default_system=(
+        "A chat between a curious user and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the user's questions."
+    ),
+    mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
+)
+
+
+_register_template(
+    name="llava_next_llama3",
+    format_user=StringFormatter(
+        slots=[
+            (
+                "<|start_header_id|>user<|end_header_id|>\n\n{{content}}<|eot_id|>"
+                "<|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
+        ]
+    ),
+    format_system=StringFormatter(slots=["<|start_header_id|>system<|end_header_id|>\n\n{{content}}<|eot_id|>"]),
+    format_observation=StringFormatter(
+        slots=[
+            (
+                "<|start_header_id|>tool<|end_header_id|>\n\n{{content}}<|eot_id|>"
+                "<|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
+        ]
+    ),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+    stop_words=["<|eot_id|>"],
+    replace_eos=True,
+    replace_jinja_template=False,
+    mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
+)
+
+
+_register_template(
+    name="llava_next_mistral",
+    format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+    mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
+)
+
+
+_register_template(
+    name="llava_next_qwen",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_observation=StringFormatter(slots=["<|im_start|>tool\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system="You are a helpful assistant.",
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+    replace_jinja_template=False,
+    mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
+)
+
+
+_register_template(
+    name="llava_next_yi",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+    mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
+)
+
+
+_register_template(
+    name="llava_next_video",
+    format_user=StringFormatter(slots=["USER: {{content}} ASSISTANT:"]),
+    default_system=(
+        "A chat between a curious user and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the user's questions."
+    ),
+    mm_plugin=get_mm_plugin(name="llava_next_video", image_token="<image>", video_token="<video>"),
+)
+
+
+_register_template(
+    name="llava_next_video_mistral",
+    format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+    mm_plugin=get_mm_plugin(name="llava_next_video", image_token="<image>", video_token="<video>"),
+)
+
+
+_register_template(
+    name="llava_next_video_yi",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+    mm_plugin=get_mm_plugin(name="llava_next_video", image_token="<image>", video_token="<video>"),
 )
 
 
@@ -751,6 +938,19 @@ _register_template(
 )
 
 
+_register_template(
+    name="opencoder",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_observation=StringFormatter(slots=["<|im_start|>tool\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system="You are OpenCoder, created by OpenCoder Team.",
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+    replace_jinja_template=False,
+)
+
+
 _register_template(
     name="orion",
     format_user=StringFormatter(slots=["Human: {{content}}\n\nAssistant: ", {"eos_token"}]),
@@ -758,6 +958,19 @@ _register_template(
 )
 
 
+_register_template(
+    name="paligemma",
+    format_user=StringFormatter(slots=["<start_of_turn>user\n{{content}}<end_of_turn>\n<start_of_turn>model\n"]),
+    format_observation=StringFormatter(
+        slots=["<start_of_turn>tool\n{{content}}<end_of_turn>\n<start_of_turn>model\n"]
+    ),
+    format_separator=EmptyFormatter(slots=["<end_of_turn>\n"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+    efficient_eos=True,
+    mm_plugin=get_mm_plugin(name="paligemma", image_token="<image>"),
+)
+
+
 _register_template(
     name="phi",
     format_user=StringFormatter(slots=["<|user|>\n{{content}}<|end|>\n<|assistant|>\n"]),
@@ -769,6 +982,25 @@ _register_template(
 )
 
 
+_register_template(
+    name="phi_small",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|end|>\n<|assistant|>\n"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}<|end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    format_prefix=EmptyFormatter(slots=[{"<|endoftext|>"}]),
+    stop_words=["<|end|>"],
+    replace_eos=True,
+)
+
+
+_register_template(
+    name="pixtral",
+    format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+    mm_plugin=get_mm_plugin(name="pixtral", image_token="[IMG]"),
+)
+
+
 _register_template(
     name="qwen",
     format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
@@ -778,6 +1010,35 @@ _register_template(
     default_system="You are a helpful assistant.",
     stop_words=["<|im_end|>"],
     replace_eos=True,
+    replace_jinja_template=False,
+)
+
+
+_register_template(
+    name="qwen2_vl",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_observation=StringFormatter(slots=["<|im_start|>tool\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system="You are a helpful assistant.",
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+    replace_jinja_template=False,
+    mm_plugin=get_mm_plugin(name="qwen2_vl", image_token="<|image_pad|>", video_token="<|video_pad|>"),
+)
+
+
+_register_template(
+    name="sailor",
+    format_user=StringFormatter(slots=["<|im_start|>question\n{{content}}<|im_end|>\n<|im_start|>answer\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system=(
+        "You are an AI assistant named Sailor created by Sea AI Lab. "
+        "Your answer should be friendly, unbiased, faithful, informative and detailed."
+    ),
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
 )
 
 
@@ -818,6 +1079,17 @@ _register_template(
 )
 
 
+_register_template(
+    name="video_llava",
+    format_user=StringFormatter(slots=["USER: {{content}} ASSISTANT:"]),
+    default_system=(
+        "A chat between a curious user and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the user's questions."
+    ),
+    mm_plugin=get_mm_plugin(name="video_llava", image_token="<image>", video_token="<video>"),
+)
+
+
 _register_template(
     name="xuanyuan",
     format_user=StringFormatter(slots=["Human: {{content}} Assistant:"]),
@@ -878,6 +1150,7 @@ _register_template(
     ),
     stop_words=["###"],
     efficient_eos=True,
+    mm_plugin=get_mm_plugin(name="llava", image_token="<image>"),
 )
 
 

src/llamafactory/data/tool_utils.py

@@ -15,9 +15,12 @@
 import json
 import re
 from abc import ABC, abstractmethod
+from collections import namedtuple
 from dataclasses import dataclass
 from typing import Any, Dict, List, Tuple, Union
 
+from typing_extensions import override
+
 from .data_utils import SLOTS
 
 
@@ -38,26 +41,47 @@ GLM4_TOOL_PROMPT = (
 )
 
 
+FunctionCall = namedtuple("FunctionCall", ["name", "arguments"])
+
+
 @dataclass
 class ToolUtils(ABC):
-    @staticmethod
+    """
-    @abstractmethod
+    Base class for tool utilities.
-    def get_function_slots() -> SLOTS: ...
+    """
 
     @staticmethod
     @abstractmethod
-    def tool_formatter(tools: List[Dict[str, Any]]) -> str: ...
+    def get_function_slots() -> SLOTS:
+        r"""
+        Gets a list of slots corresponding to a single function call.
+        """
+        ...
 
     @staticmethod
     @abstractmethod
-    def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]: ...
+    def tool_formatter(tools: List[Dict[str, Any]]) -> str:
+        r"""
+        Generates the system message describing all the available tools.
+        """
+        ...
+
+    @staticmethod
+    @abstractmethod
+    def tool_extractor(content: str) -> Union[str, List["FunctionCall"]]:
+        r"""
+        Extracts all the function calls from the response message.
+        """
+        ...
 
 
 class DefaultToolUtils(ToolUtils):
+    @override
     @staticmethod
     def get_function_slots() -> SLOTS:
         return ["Action: {{name}}\nAction Input: {{arguments}}\n"]
 
+    @override
     @staticmethod
     def tool_formatter(tools: List[Dict[str, Any]]) -> str:
         tool_text = ""
@@ -91,8 +115,9 @@ class DefaultToolUtils(ToolUtils):
 
         return DEFAULT_TOOL_PROMPT.format(tool_text=tool_text, tool_names=", ".join(tool_names))
 
+    @override
     @staticmethod
-    def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
+    def tool_extractor(content: str) -> Union[str, List["FunctionCall"]]:
         regex = re.compile(r"Action:\s*([a-zA-Z0-9_]+)\s*Action Input:\s*(.+?)(?=\s*Action:|\s*$)", re.DOTALL)
         action_match: List[Tuple[str, str]] = re.findall(regex, content)
         if not action_match:
@@ -112,10 +137,12 @@ class DefaultToolUtils(ToolUtils):
 
 
 class GLM4ToolUtils(ToolUtils):
+    @override
     @staticmethod
     def get_function_slots() -> SLOTS:
         return ["{{name}}\n{{arguments}}"]
 
+    @override
     @staticmethod
     def tool_formatter(tools: List[Dict[str, Any]]) -> str:
         tool_text = ""
@@ -126,8 +153,9 @@ class GLM4ToolUtils(ToolUtils):
 
         return GLM4_TOOL_PROMPT.format(tool_text=tool_text)
 
+    @override
     @staticmethod
-    def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
+    def tool_extractor(content: str) -> Union[str, List["FunctionCall"]]:
         if "\n" not in content:
             return content
 
@@ -138,3 +166,17 @@ class GLM4ToolUtils(ToolUtils):
             return content
 
         return [(tool_name, json.dumps(arguments, ensure_ascii=False))]
+
+
+TOOLS = {
+    "default": DefaultToolUtils(),
+    "glm4": GLM4ToolUtils(),
+}
+
+
+def get_tool_utils(name: str) -> "ToolUtils":
+    tool_utils = TOOLS.get(name, None)
+    if tool_utils is None:
+        raise ValueError(f"Tool utils `{name}` not found.")
+
+    return tool_utils

src/llamafactory/eval/evaluator.py

@@ -39,7 +39,7 @@
 
 import json
 import os
-from typing import Any, Dict, List, Optional
+from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 import numpy as np
 import torch
@@ -54,18 +54,22 @@ from ..model import load_model, load_tokenizer
 from .template import get_eval_template
 
 
+if TYPE_CHECKING:
+    from numpy.typing import NDArray
+
+
 class Evaluator:
     def __init__(self, args: Optional[Dict[str, Any]] = None) -> None:
         self.model_args, self.data_args, self.eval_args, finetuning_args = get_eval_args(args)
         self.tokenizer = load_tokenizer(self.model_args)["tokenizer"]
         self.tokenizer.padding_side = "right"  # avoid overflow issue in batched inference for llama2
-        self.template = get_template_and_fix_tokenizer(self.tokenizer, self.data_args.template)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, self.data_args)
         self.model = load_model(self.tokenizer, self.model_args, finetuning_args)
         self.eval_template = get_eval_template(self.eval_args.lang)
         self.choice_inputs = [self.tokenizer.encode(ch, add_special_tokens=False)[-1] for ch in CHOICES]
 
     @torch.inference_mode()
-    def batch_inference(self, batch_input: Dict[str, torch.Tensor]) -> List[str]:
+    def batch_inference(self, batch_input: Dict[str, "torch.Tensor"]) -> List[str]:
         logits = self.model(**batch_input).logits
         lengths = torch.sum(batch_input["attention_mask"], dim=-1)
         word_probs = torch.stack([logits[i, lengths[i] - 1] for i in range(len(lengths))], dim=0)
@@ -83,7 +87,7 @@ class Evaluator:
             token=self.model_args.hf_hub_token,
         )
 
-        with open(mapping, "r", encoding="utf-8") as f:
+        with open(mapping, encoding="utf-8") as f:
             categorys: Dict[str, Dict[str, str]] = json.load(f)
 
         category_corrects = {subj: np.array([], dtype="bool") for subj in SUBJECTS}
@@ -132,10 +136,10 @@ class Evaluator:
         pbar.close()
         self._save_results(category_corrects, results)
 
-    def _save_results(self, category_corrects: Dict[str, np.ndarray], results: Dict[str, Dict[int, str]]) -> None:
+    def _save_results(self, category_corrects: Dict[str, "NDArray"], results: Dict[str, Dict[int, str]]) -> None:
         score_info = "\n".join(
             [
-                "{:>15}: {:.2f}".format(category_name, 100 * np.mean(category_correct))
+                f"{category_name:>15}: {100 * np.mean(category_correct):.2f}"
                 for category_name, category_correct in category_corrects.items()
                 if len(category_correct)
             ]