release v0.9.0 (real)

Former-commit-id: 8ff781c8ae5654680f738f69a6db9d7b95d76baf

.dockerignore

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

.env.local

@@ -0,0 +1,35 @@
+# Note: actually we do not support .env, just for reference
+# api
+API_HOST=0.0.0.0
+API_PORT=8000
+API_KEY=
+API_MODEL_NAME=gpt-3.5-turbo
+FASTAPI_ROOT_PATH=
+# general
+DISABLE_VERSION_CHECK=
+FORCE_CHECK_IMPORTS=
+FORCE_TORCHRUN=
+LLAMAFACTORY_VERBOSITY=
+USE_MODELSCOPE_HUB=
+RECORD_VRAM=
+# torchrun
+FORCE_TORCHRUN=
+MASTER_ADDR=
+MASTER_PORT=
+NNODES=
+RANK=
+NPROC_PER_NODE=
+# wandb
+WANDB_DISABLED=
+WANDB_PROJECT=huggingface
+WANDB_API_KEY=
+# gradio ui
+GRADIO_SHARE=False
+GRADIO_SERVER_NAME=0.0.0.0
+GRADIO_SERVER_PORT=
+GRADIO_ROOT_PATH=
+# setup
+ENABLE_SHORT_CONSOLE=1
+# reserved (do not use)
+LLAMABOARD_ENABLED=
+LLAMABOARD_WORKDIR=

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,13 +1,19 @@
 name: "\U0001F41B Bug / Help"
 description: Create a report to help us improve the LLaMA Factory
 body:
+  - type: markdown
+    attributes:
+      value: |
+        Issues included in **FAQs** or those with **insufficient** information may be closed without a response.
+        包含在**常见问题**内或提供信息**不完整**的 issues 可能不会被回复。
+
   - type: checkboxes
     id: reminder
     attributes:
       label: Reminder
       description: |
-        Please ensure you have read the README carefully and searched the existing issues.
-        请确保您已经认真阅读了 README 并且搜索过现有的 Issue。
+        Please ensure you have read the README carefully and searched the existing issues (including FAQs).
+        请确保您已经认真阅读了 README 并且搜索过现有的 issues（包括常见问题）。
 
       options:
         - label: I have read the README and searched the existing issues.

.github/workflows/label_issue.yml

@@ -9,9 +9,22 @@ jobs:
   label_issue:
     runs-on: ubuntu-latest
 
+    permissions:
+      issues: write
+
     steps:
       - env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           ISSUE_URL: ${{ github.event.issue.html_url }}
+          ISSUE_TITLE: ${{ github.event.issue.title }}
         run: |
-          gh issue edit $ISSUE_URL --add-label "pending"
+          LABEL=pending
+          NPU_KEYWORDS=(npu huawei ascend 华为 昇腾)
+          ISSUE_TITLE_LOWER=$(echo $ISSUE_TITLE | tr '[:upper:]' '[:lower:]')
+          for KEYWORD in ${NPU_KEYWORDS[@]}; do
+            if [[ $ISSUE_TITLE_LOWER == *$KEYWORD* ]] && [[ $ISSUE_TITLE_LOWER != *input* ]]; then
+              LABEL=pending,npu
+              break
+            fi
+          done
+          gh issue edit $ISSUE_URL --add-label $LABEL

.github/workflows/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"
@@ -18,7 +18,27 @@ on:
 
 jobs:
   tests:
-    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version:
+          - "3.8"
+          - "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
@@ -27,14 +47,15 @@ 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"
 
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install .[torch,dev]
+          python -m pip install git+https://github.com/huggingface/transformers.git
+          python -m pip install ".[torch,dev]"
 
       - name: Check quality
         run: |

.gitignore

@@ -160,6 +160,10 @@ cython_debug/
 .idea/
 
 # custom .gitignore
-user.config
-saves/
+ms_cache/
+hf_cache/
 cache/
+config/
+saves/
+output/
+wandb/

CITATION.cff

@@ -12,12 +12,16 @@ authors:
   given-names: "Yanhan"
 - family-names: "Luo"
   given-names: "Zheyan"
+- family-names: "Feng"
+  given-names: "Zhangchi"
 - family-names: "Ma"
   given-names: "Yongqiang"
 title: "LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models"
 url: "https://arxiv.org/abs/2403.13372"
 preferred-citation:
-  type: article
+  type: conference-paper
+  conference:
+    name: "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)"
   authors:
     - family-names: "Zheng"
       given-names: "Yaowei"
@@ -29,9 +33,12 @@ preferred-citation:
       given-names: "Yanhan"
     - family-names: "Luo"
       given-names: "Zheyan"
+    - family-names: "Feng"
+      given-names: "Zhangchi"
     - family-names: "Ma"
       given-names: "Yongqiang"
-  journal: "arXiv preprint arXiv:2403.13372"
   title: "LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models"
   url: "https://arxiv.org/abs/2403.13372"
   year: 2024
+  publisher: "Association for Computational Linguistics"
+  address: "Bangkok, Thailand"

Dockerfile

@@ -1,47 +0,0 @@
-# Use the NVIDIA official image with PyTorch 2.3.0
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-24-02.html
-FROM nvcr.io/nvidia/pytorch:24.02-py3
-
-# Define installation arguments
-ARG INSTALL_BNB=false
-ARG INSTALL_VLLM=false
-ARG INSTALL_DEEPSPEED=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 $PIP_INDEX
-RUN python -m pip install --upgrade pip
-RUN python -m pip install -r requirements.txt
-
-# Copy the rest of the application into the image
-COPY . /app/
-
-# Install the LLaMA Factory
-RUN EXTRA_PACKAGES="metrics"; \
-    if [ "$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="${EXTRA_PACKAGES},deepspeed"; \
-    fi; \
-    pip install -e .[$EXTRA_PACKAGES] && \
-    pip uninstall -y transformer-engine flash-attn
-
-# Set up volumes
-VOLUME [ "/root/.cache/huggingface/", "/app/data", "/app/output" ]
-
-# Expose port 7860 for the LLaMA Board
-EXPOSE 7860
-
-# Expose port 8000 for the API service
-EXPOSE 8000
-
-# Launch LLaMA Board
-CMD [ "llamafactory-cli", "webui" ]

Makefile

@@ -1,6 +1,6 @@
 .PHONY: quality style test
 
-check_dirs := scripts src tests
+check_dirs := scripts src tests setup.py
 
 quality:
 	ruff check $(check_dirs)

README.md

@@ -4,7 +4,7 @@
 [![GitHub Code License](https://img.shields.io/github/license/hiyouga/LLaMA-Factory)](LICENSE)
 [![GitHub last commit](https://img.shields.io/github/last-commit/hiyouga/LLaMA-Factory)](https://github.com/hiyouga/LLaMA-Factory/commits/main)
 [![PyPI](https://img.shields.io/pypi/v/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-44-green)](#projects-using-llama-factory)
+[![Citation](https://img.shields.io/badge/citation-91-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)
@@ -15,19 +15,20 @@
 
 [![GitHub Tread](https://trendshift.io/api/badge/repositories/4535)](https://trendshift.io/repositories/4535)
 
-👋 Join our [WeChat](assets/wechat.jpg).
+👋 Join our [WeChat](assets/wechat.jpg) or [NPU user group](assets/wechat_npu.jpg).
 
 \[ English | [中文](README_zh.md) \]
 
 **Fine-tuning a large language model can be easy as...**
 
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/9840a653-7e9c-41c8-ae89-7ace5698baf6
+https://github.com/user-attachments/assets/7c96b465-9df7-45f4-8053-bf03e58386d3
 
 Choose your path:
 
 - **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)
+- **Documentation (WIP)**: https://llamafactory.readthedocs.io/zh-cn/latest/
 
 ## Table of Contents
 
@@ -46,11 +47,11 @@ Choose your path:
 
 ## Features
 
-- **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
+- **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Qwen2-VL, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
 - **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc.
-- **Scalable resources**: 32-bit full-tuning, 16-bit freeze-tuning, 16-bit LoRA and 2/4/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8.
-- **Advanced algorithms**: GaLore, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA and Agent tuning.
-- **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA.
+- **Scalable resources**: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ.
+- **Advanced algorithms**: [GaLore](https://github.com/jiaweizzhao/GaLore), [BAdam](https://github.com/Ledzy/BAdam), [Adam-mini](https://github.com/zyushun/Adam-mini), DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA and Agent tuning.
+- **Practical tricks**: [FlashAttention-2](https://github.com/Dao-AILab/flash-attention), [Unsloth](https://github.com/unslothai/unsloth), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), RoPE scaling, NEFTune and rsLoRA.
 - **Experiment monitors**: LlamaBoard, TensorBoard, Wandb, MLflow, etc.
 - **Faster inference**: OpenAI-style API, Gradio UI and CLI with vLLM worker.
 
@@ -71,15 +72,23 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Changelog
 
+[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 `gemma` 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 `paligemma` template for chat completion.
 
 [24/05/18] We supported **[KTO](https://arxiv.org/abs/2402.01306)** algorithm for preference learning. See [examples](examples/README.md) for usage.
 
@@ -91,7 +100,7 @@ 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 +112,7 @@ 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.
 
@@ -151,35 +160,34 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Supported Models
 
-| Model                                                     | Model size                       | Template  |
-| --------------------------------------------------------- | -------------------------------- | --------- |
-| [Baichuan2](https://huggingface.co/baichuan-inc)          | 7B/13B                           | baichuan2 |
-| [BLOOM](https://huggingface.co/bigscience)                | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [BLOOMZ](https://huggingface.co/bigscience)               | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [ChatGLM3](https://huggingface.co/THUDM)                  | 6B                               | chatglm3  |
-| [Command-R](https://huggingface.co/CohereForAI)           | 35B/104B                         | cohere    |
-| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai) | 7B/16B/67B/236B                  | deepseek  |
-| [Falcon](https://huggingface.co/tiiuae)                   | 7B/11B/40B/180B                  | falcon    |
-| [Gemma/CodeGemma](https://huggingface.co/google)          | 2B/7B                            | gemma     |
-| [GLM4](https://huggingface.co/THUDM)                      | 9B                               | glm4      |
-| [InternLM2](https://huggingface.co/internlm)              | 7B/20B                           | intern2   |
-| [LLaMA](https://github.com/facebookresearch/llama)        | 7B/13B/33B/65B                   | -         |
-| [LLaMA-2](https://huggingface.co/meta-llama)              | 7B/13B/70B                       | llama2    |
-| [LLaMA-3](https://huggingface.co/meta-llama)              | 8B/70B                           | llama3    |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)              | 7B/13B                           | vicuna    |
-| [Mistral/Mixtral](https://huggingface.co/mistralai)       | 7B/8x7B/8x22B                    | mistral   |
-| [OLMo](https://huggingface.co/allenai)                    | 1B/7B                            | -         |
-| [PaliGemma](https://huggingface.co/google)                | 3B                               | gemma     |
-| [Phi-1.5/2](https://huggingface.co/microsoft)             | 1.3B/2.7B                        | -         |
-| [Phi-3](https://huggingface.co/microsoft)                 | 4B/7B/14B                        | phi       |
-| [Qwen](https://huggingface.co/Qwen)                       | 1.8B/7B/14B/72B                  | qwen      |
-| [Qwen1.5 (Code/MoE)](https://huggingface.co/Qwen)         | 0.5B/1.8B/4B/7B/14B/32B/72B/110B | qwen      |
-| [Qwen2 (MoE)](https://huggingface.co/Qwen)                | 0.5B/1.5B/7B/57B/72B             | qwen      |
-| [StarCoder2](https://huggingface.co/bigcode)              | 3B/7B/15B                        | -         |
-| [XVERSE](https://huggingface.co/xverse)                   | 7B/13B/65B                       | xverse    |
-| [Yi (1/1.5)](https://huggingface.co/01-ai)                | 6B/9B/34B                        | yi        |
-| [Yi-VL](https://huggingface.co/01-ai)                     | 6B/34B                           | yi_vl     |
-| [Yuan](https://huggingface.co/IEITYuan)                   | 2B/51B/102B                      | yuan      |
+| Model                                                             | Model size                       | Template  |
+| ----------------------------------------------------------------- | -------------------------------- | --------- |
+| [Baichuan 2](https://huggingface.co/baichuan-inc)                 | 7B/13B                           | baichuan2 |
+| [BLOOM/BLOOMZ](https://huggingface.co/bigscience)                 | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
+| [ChatGLM3](https://huggingface.co/THUDM)                          | 6B                               | chatglm3  |
+| [Command R](https://huggingface.co/CohereForAI)                   | 35B/104B                         | cohere    |
+| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)         | 7B/16B/67B/236B                  | deepseek  |
+| [Falcon](https://huggingface.co/tiiuae)                           | 7B/11B/40B/180B                  | falcon    |
+| [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google)          | 2B/7B/9B/27B                     | gemma     |
+| [GLM-4](https://huggingface.co/THUDM)                             | 9B                               | glm4      |
+| [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/Llama 3.1](https://huggingface.co/meta-llama)            | 8B/70B                           | llama3    |
+| [LLaVA-1.5](https://huggingface.co/llava-hf)                      | 7B/13B                           | llava     |
+| [MiniCPM](https://huggingface.co/openbmb)                         | 1B/2B/4B                         | cpm/cpm3  |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)               | 7B/8x7B/8x22B                    | mistral   |
+| [OLMo](https://huggingface.co/allenai)                            | 1B/7B                            | -         |
+| [PaliGemma](https://huggingface.co/google)                        | 3B                               | paligemma |
+| [Phi-1.5/Phi-2](https://huggingface.co/microsoft)                 | 1.3B/2.7B                        | -         |
+| [Phi-3](https://huggingface.co/microsoft)                         | 4B/7B/14B                        | phi       |
+| [Qwen/Qwen1.5/Qwen2 (Code/Math/MoE)](https://huggingface.co/Qwen) | 0.5B/1.5B/4B/7B/14B/32B/72B/110B | qwen      |
+| [Qwen2-VL](https://huggingface.co/Qwen)                           | 2B/7B                            | qwen2_vl  |
+| [StarCoder 2](https://huggingface.co/bigcode)                     | 3B/7B/15B                        | -         |
+| [XVERSE](https://huggingface.co/xverse)                           | 7B/13B/65B                       | xverse    |
+| [Yi/Yi-1.5 (Code)](https://huggingface.co/01-ai)                  | 1.5B/6B/9B/34B                   | yi        |
+| [Yi-VL](https://huggingface.co/01-ai)                             | 6B/34B                           | yi_vl     |
+| [Yuan 2](https://huggingface.co/IEITYuan)                         | 2B/51B/102B                      | yuan      |
 
 > [!NOTE]
 > For the "base" models, the `template` argument can be chosen from `default`, `alpaca`, `vicuna` etc. But make sure to use the **corresponding template** for the "instruct/chat" models.
@@ -203,6 +211,9 @@ 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>
@@ -262,7 +273,9 @@ 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)
@@ -279,6 +292,8 @@ 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)
@@ -299,20 +314,20 @@ huggingface-cli login
 | Mandatory    | Minimum | Recommend |
 | ------------ | ------- | --------- |
 | python       | 3.8     | 3.11      |
-| torch        | 1.13.1  | 2.3.0     |
-| transformers | 4.41.2  | 4.41.2    |
-| datasets     | 2.16.0  | 2.19.2    |
-| accelerate   | 0.30.1  | 0.30.1    |
-| peft         | 0.11.1  | 0.11.1    |
-| trl          | 0.8.6   | 0.9.4     |
+| torch        | 1.13.1  | 2.4.0     |
+| transformers | 4.41.2  | 4.43.4    |
+| datasets     | 2.16.0  | 2.20.0    |
+| accelerate   | 0.30.1  | 0.32.0    |
+| peft         | 0.11.1  | 0.12.0    |
+| trl          | 0.8.6   | 0.9.6     |
 
 | Optional     | Minimum | Recommend |
 | ------------ | ------- | --------- |
 | CUDA         | 11.6    | 12.2      |
 | deepspeed    | 0.10.0  | 0.14.0    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.4.3     |
-| flash-attn   | 2.3.0   | 2.5.9     |
+| vllm         | 0.4.3   | 0.5.0     |
+| flash-attn   | 2.3.0   | 2.6.3     |
 
 ### Hardware Requirement
 
@@ -341,7 +356,7 @@ cd LLaMA-Factory
 pip install -e ".[torch,metrics]"
 ```
 
-Extra dependencies available: torch, torch_npu, metrics, deepspeed, bitsandbytes, vllm, galore, badam, gptq, awq, aqlm, qwen, modelscope, quality
+Extra dependencies available: torch, torch-npu, metrics, deepspeed, liger-kernel, bitsandbytes, hqq, eetq, gptq, awq, aqlm, vllm, galore, badam, adam-mini, qwen, modelscope, quality
 
 > [!TIP]
 > Use `pip install --no-deps -e .` to resolve package conflicts.
@@ -360,9 +375,7 @@ To enable FlashAttention-2 on the Windows platform, you need to install the prec
 
 <details><summary>For Ascend NPU users</summary>
 
-Join [NPU user group](assets/wechat_npu.jpg).
-
-To install LLaMA Factory on Ascend NPU devices, please specify extra dependencies: `pip install -e '.[torch-npu,metrics]'`. Additionally, you need to install the **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**. Please follow the [installation tutorial](https://www.hiascend.com/document/detail/en/CANNCommunityEdition/600alphaX/softwareinstall/instg/atlasdeploy_03_0031.html) or use the following commands:
+To install LLaMA Factory on Ascend NPU devices, please specify extra dependencies: `pip install -e ".[torch-npu,metrics]"`. Additionally, you need to install the **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**. Please follow the [installation tutorial](https://www.hiascend.com/document/detail/en/CANNCommunityEdition/600alphaX/softwareinstall/instg/atlasdeploy_03_0031.html) or use the following commands:
 
 ```bash
 # replace the url according to your CANN version and devices
@@ -385,15 +398,12 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch-npu    | 2.1.0   | 2.1.0.post3 |
 | deepspeed    | 0.13.2  | 0.13.2      |
 
-Docker image:
-
-- 32GB: [Download page](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html)
-- 64GB: [Download page](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
-
 Remember to use `ASCEND_RT_VISIBLE_DEVICES` instead of `CUDA_VISIBLE_DEVICES` to specify the device to use.
 
 If you cannot infer model on NPU devices, try setting `do_sample: false` in the configurations.
 
+Download the pre-built Docker images: [32GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html) | [64GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
+
 </details>
 
 ### Data Preparation
@@ -426,18 +436,46 @@ llamafactory-cli webui
 
 ### Build Docker
 
-#### Use Docker
+For CUDA users:
 
 ```bash
-docker build -f ./Dockerfile \
+cd docker/docker-cuda/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+For Ascend NPU users:
+
+```bash
+cd docker/docker-npu/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+For AMD ROCm users:
+
+```bash
+cd docker/docker-rocm/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+<details><summary>Build without Docker Compose</summary>
+
+For CUDA users:
+
+```bash
+docker build -f ./docker/docker-cuda/Dockerfile \
     --build-arg INSTALL_BNB=false \
     --build-arg INSTALL_VLLM=false \
     --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
     --build-arg PIP_INDEX=https://pypi.org/simple \
     -t llamafactory:latest .
 
-docker run -it --gpus=all \
-    -v ./hf_cache:/root/.cache/huggingface/ \
+docker run -dit --gpus=all \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
     -v ./data:/app/data \
     -v ./output:/app/output \
     -p 7860:7860 \
@@ -445,20 +483,78 @@ docker run -it --gpus=all \
     --shm-size 16G \
     --name llamafactory \
     llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
-#### Use Docker Compose
+For Ascend NPU users:
 
 ```bash
-docker-compose up -d
-docker-compose exec llamafactory bash
+# Choose docker image upon your environment
+docker build -f ./docker/docker-npu/Dockerfile \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+# Change `device` upon your resources
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
+    -v /etc/ascend_install.info:/etc/ascend_install.info \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/davinci0 \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
+For AMD ROCm users:
+
+```bash
+docker build -f ./docker/docker-rocm/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./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.
-- data: Place datasets on this dir of the host machine so that they can be selected on LLaMA Board GUI.
-- output: Set export dir to this location so that the merged result can be accessed directly on the host machine.
+- `hf_cache`: Utilize Hugging Face cache on the host machine. Reassignable if a cache already exists in a different directory.
+- `ms_cache`: Similar to Hugging Face cache but for ModelScope users.
+- `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>
 
@@ -469,7 +565,7 @@ 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.
 
 ### Download from ModelScope Hub
 
@@ -503,38 +599,82 @@ If you have a project that should be incorporated, please contact via email or c
 1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
 1. Luceri et al. Leveraging Large Language Models to Detect Influence Campaigns in Social Media. 2023. [[arxiv]](https://arxiv.org/abs/2311.07816)
 1. Zhang et al. Alleviating Hallucinations of Large Language Models through Induced Hallucinations. 2023. [[arxiv]](https://arxiv.org/abs/2312.15710)
-1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
-1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
+1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. KDD 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
+1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
 1. Choi et al. FACT-GPT: Fact-Checking Augmentation via Claim Matching with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2402.05904)
 1. Zhang et al. AutoMathText: Autonomous Data Selection with Language Models for Mathematical Texts. 2024. [[arxiv]](https://arxiv.org/abs/2402.07625)
 1. Lyu et al. KnowTuning: Knowledge-aware Fine-tuning for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11176)
 1. Yang et al. LaCo: Large Language Model Pruning via Layer Collaps. 2024. [[arxiv]](https://arxiv.org/abs/2402.11187)
 1. Bhardwaj et al. Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic. 2024. [[arxiv]](https://arxiv.org/abs/2402.11746)
 1. Yang et al. Enhancing Empathetic Response Generation by Augmenting LLMs with Small-scale Empathetic Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11801)
-1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. 2024. [[arxiv]](https://arxiv.org/abs/2402.11809)
+1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2402.11809)
 1. Cao et al. Head-wise Shareable Attention for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11819)
 1. Zhang et al. Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languages. 2024. [[arxiv]](https://arxiv.org/abs/2402.12204)
 1. Kim et al. Efficient and Effective Vocabulary Expansion Towards Multilingual Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.14714)
-1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
+1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. ACL 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
 1. Huang et al. Key-Point-Driven Data Synthesis with its Enhancement on Mathematical Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2403.02333)
 1. Duan et al. Negating Negatives: Alignment without Human Positive Samples via Distributional Dispreference Optimization. 2024. [[arxiv]](https://arxiv.org/abs/2403.03419)
 1. Xie and Schwertfeger. Empowering Robotics with Large Language Models: osmAG Map Comprehension with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2403.08228)
 1. Wu et al. Large Language Models are Parallel Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2403.09073)
 1. Zhang et al. EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling. 2024. [[arxiv]](https://arxiv.org/abs/2403.14541)
 1. Weller et al. FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2403.15246)
-1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
+1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. COLING 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
 1. Zan et al. CodeS: Natural Language to Code Repository via Multi-Layer Sketch. 2024. [[arxiv]](https://arxiv.org/abs/2403.16443)
 1. Liu et al. Extensive Self-Contrast Enables Feedback-Free Language Model Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2404.00604)
 1. Luo et al. BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.02827)
 1. Du et al. Chinese Tiny LLM: Pretraining a Chinese-Centric Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2404.04167)
-1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
+1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. ICML 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
 1. Liu et al. Dynamic Generation of Personalities with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.07084)
 1. Shang et al. How Far Have We Gone in Stripped Binary Code Understanding Using Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.09836)
 1. Huang et al. LLMTune: Accelerate Database Knob Tuning with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.11581)
 1. Deng et al. Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction. 2024. [[arxiv]](https://arxiv.org/abs/2404.14215)
 1. Acikgoz et al. Hippocrates: An Open-Source Framework for Advancing Large Language Models in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2404.16621)
-1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2404.17140)
-1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2404.17140)
+1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. NAACL 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Xu et al. Large Language Models for Cyber Security: A Systematic Literature Review. 2024. [[arxiv]](https://arxiv.org/abs/2405.04760)
+1. Dammu et al. "They are uncultured": Unveiling Covert Harms and Social Threats in LLM Generated Conversations. 2024. [[arxiv]](https://arxiv.org/abs/2405.05378)
+1. Yi et al. A safety realignment framework via subspace-oriented model fusion for large language models. 2024. [[arxiv]](https://arxiv.org/abs/2405.09055)
+1. Lou et al. SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling. 2024. [[arxiv]](https://arxiv.org/abs/2405.12739)
+1. Zhang et al. Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2405.13816)
+1. Zhang et al. TS-Align: A Teacher-Student Collaborative Framework for Scalable Iterative Finetuning of Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2405.20215)
+1. Zihong Chen. Sentence Segmentation and Sentence Punctuation Based on XunziALLM. 2024. [[paper]](https://aclanthology.org/2024.lt4hala-1.30)
+1. Gao et al. The Best of Both Worlds: Toward an Honest and Helpful Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2406.00380)
+1. Wang and Song. MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset. 2024. [[arxiv]](https://arxiv.org/abs/2406.02106)
+1. Hu et al. Computational Limits of Low-Rank Adaptation (LoRA) for Transformer-Based Models. 2024. [[arxiv]](https://arxiv.org/abs/2406.03136)
+1. Ge et al. Time Sensitive Knowledge Editing through Efficient Finetuning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2406.04496)
+1. Tan et al. Peer Review as A Multi-Turn and Long-Context Dialogue with Role-Based Interactions. 2024. [[arxiv]](https://arxiv.org/abs/2406.05688)
+1. Song et al. Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters. 2024. [[arxiv]](https://arxiv.org/abs/2406.05955)
+1. Gu et al. RWKV-CLIP: A Robust Vision-Language Representation Learner. 2024. [[arxiv]](https://arxiv.org/abs/2406.06973)
+1. Chen et al. Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees. 2024. [[arxiv]](https://arxiv.org/abs/2406.07115)
+1. Zhu et al. Are Large Language Models Good Statisticians?. 2024. [[arxiv]](https://arxiv.org/abs/2406.07815)
+1. Li et al. Know the Unknown: An Uncertainty-Sensitive Method for LLM Instruction Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2406.10099)
+1. Ding et al. IntentionQA: A Benchmark for Evaluating Purchase Intention Comprehension Abilities of Language Models in E-commerce. 2024. [[arxiv]](https://arxiv.org/abs/2406.10173)
+1. He et al. COMMUNITY-CROSS-INSTRUCT: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities. 2024. [[arxiv]](https://arxiv.org/abs/2406.12074)
+1. Lin et al. FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving. 2024. [[arxiv]](https://arxiv.org/abs/2406.14408)
+1. Treutlein et al. Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data. 2024. [[arxiv]](https://arxiv.org/abs/2406.14546)
+1. Feng et al. SS-Bench: A Benchmark for Social Story Generation and Evaluation. 2024. [[arxiv]](https://arxiv.org/abs/2406.15695)
+1. Feng et al. Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement. 2024. [[arxiv]](https://arxiv.org/abs/2406.17233)
+1. Liu et al. Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals. 2024. [[arxiv]](https://arxiv.org/abs/2406.18069)
+1. Iyer et al. Exploring Very Low-Resource Translation with LLMs: The University of Edinburgh's Submission to AmericasNLP 2024 Translation Task. AmericasNLP 2024. [[paper]](https://aclanthology.org/2024.americasnlp-1.25)
+1. Li et al. Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring. 2024. [[arxiv]](https://arxiv.org/abs/2406.19949)
+1. Yang et al. Financial Knowledge Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2407.00365)
+1. Lin et al. DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging. 2024. [[arxiv]](https://arxiv.org/abs/2407.01470)
+1. Bako et al. Evaluating the Semantic Profiling Abilities of LLMs for Natural Language Utterances in Data Visualization. 2024. [[arxiv]](https://arxiv.org/abs/2407.06129)
+1. Huang et al. RoLoRA: Fine-tuning Rotated Outlier-free LLMs for Effective Weight-Activation Quantization. 2024. [[arxiv]](https://arxiv.org/abs/2407.08044)
+1. Jiang et al. LLM-Collaboration on Automatic Science Journalism for the General Audience. 2024. [[arxiv]](https://arxiv.org/abs/2407.09756)
+1. Inouye et al. Applied Auto-tuning on LoRA Hyperparameters. 2024. [[paper]](https://scholarcommons.scu.edu/cseng_senior/272/)
+1. Qi et al. Research on Tibetan Tourism Viewpoints information generation system based on LLM. 2024. [[arxiv]](https://arxiv.org/abs/2407.13561)
+1. Xu et al. Course-Correction: Safety Alignment Using Synthetic Preferences. 2024. [[arxiv]](https://arxiv.org/abs/2407.16637)
+1. Sun et al. LAMBDA: A Large Model Based Data Agent. 2024. [[arxiv]](https://arxiv.org/abs/2407.17535)
+1. Zhu et al. CollectiveSFT: Scaling Large Language Models for Chinese Medical Benchmark with Collective Instructions in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2407.19705)
+1. Yu et al. Correcting Negative Bias in Large Language Models through Negative Attention Score Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2408.00137)
+1. Xie et al. The Power of Personalized Datasets: Advancing Chinese Composition Writing for Elementary School through Targeted Model Fine-Tuning. IALP 2024. [[paper]](https://www.asianlp.sg/conferences/ialp2024/proceedings/papers/IALP2024_P055.pdf)
+1. Liu et al. Instruct-Code-Llama: Improving Capabilities of Language Model in Competition Level Code Generation by Online Judge Feedback. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_11)
+1. Wang et al. Cybernetic Sentinels: Unveiling the Impact of Safety Data Selection on Model Security in Supervised Fine-Tuning. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_23)
+1. Xia et al. Understanding the Performance and Estimating the Cost of LLM Fine-Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2408.04693)
+1. Zeng et al. Perceive, Reflect, and Plan: Designing LLM Agent for Goal-Directed City Navigation without Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2408.04168)
+1. Xia et al. Using Pre-trained Language Model for Accurate ESG Prediction. FinNLP 2024. [[paper]](https://aclanthology.org/2024.finnlp-2.1/)
+1. Liang et al. I-SHEEP: Self-Alignment of LLM from Scratch through an Iterative Self-Enhancement Paradigm. 2024. [[arxiv]](https://arxiv.org/abs/2408.08072)
 1. **[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.
@@ -542,6 +682,9 @@ If you have a project that should be incorporated, please contact via email or c
 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. **[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.
 
 </details>
 
@@ -549,17 +692,19 @@ If you have a project that should be incorporated, please contact via email or c
 
 This repository is licensed under the [Apache-2.0 License](LICENSE).
 
-Please follow the model licenses to use the corresponding model weights: [Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command-R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [LLaMA-3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yi-1.5](LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+Please follow the model licenses to use the corresponding model weights: [Baichuan 2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [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](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder 2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yi-1.5](LICENSE) / [Yuan 2](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## Citation
 
 If this work is helpful, please kindly cite as:
 
 ```bibtex
-@article{zheng2024llamafactory,
+@inproceedings{zheng2024llamafactory,
   title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models},
-  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Yongqiang Ma},
-  journal={arXiv preprint arXiv:2403.13372},
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Zhangchi Feng and Yongqiang Ma},
+  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
+  address={Bangkok, Thailand},
+  publisher={Association for Computational Linguistics},
   year={2024},
   url={http://arxiv.org/abs/2403.13372}
 }

README_zh.md

@@ -4,7 +4,7 @@
 [![GitHub Code License](https://img.shields.io/github/license/hiyouga/LLaMA-Factory)](LICENSE)
 [![GitHub last commit](https://img.shields.io/github/last-commit/hiyouga/LLaMA-Factory)](https://github.com/hiyouga/LLaMA-Factory/commits/main)
 [![PyPI](https://img.shields.io/pypi/v/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-44-green)](#使用了-llama-factory-的项目)
+[![Citation](https://img.shields.io/badge/citation-91-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)
@@ -15,19 +15,21 @@
 
 [![GitHub Tread](https://trendshift.io/api/badge/repositories/4535)](https://trendshift.io/repositories/4535)
 
-👋 加入我们的[微信群](assets/wechat.jpg)。
+👋 加入我们的[微信群](assets/wechat.jpg)或 [NPU 用户群](assets/wechat_npu.jpg)。
 
 \[ [English](README.md) | 中文 \]
 
 **微调大模型可以像这样轻松…**
 
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd-d76c6d0a6594
+https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 
 选择你的打开方式：
 
 - **Colab**：https://colab.research.google.com/drive/1d5KQtbemerlSDSxZIfAaWXhKr30QypiK?usp=sharing
-- **PAI-DSW**: https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory
+- **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)
+- **入门教程**：https://zhuanlan.zhihu.com/p/695287607
+- **框架文档**：https://llamafactory.readthedocs.io/zh-cn/latest/
 
 ## 目录
 
@@ -46,11 +48,11 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 项目特色
 
-- **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
+- **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Qwen2-VL、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
 - **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO 训练、DPO 训练、KTO 训练、ORPO 训练等等。
-- **多种精度**：32 比特全参数微调、16 比特冻结微调、16 比特 LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8 的 2/4/8 比特 QLoRA 微调。
-- **先进算法**：GaLore、BAdam、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ、PiSSA 和 Agent 微调。
-- **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。
+- **多种精度**：16 比特全参数微调、冻结微调、LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ 的 2/3/4/5/6/8 比特 QLoRA 微调。
+- **先进算法**：[GaLore](https://github.com/jiaweizzhao/GaLore)、[BAdam](https://github.com/Ledzy/BAdam)、[Adam-mini](https://github.com/zyushun/Adam-mini)、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ、PiSSA 和 Agent 微调。
+- **实用技巧**：[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)、[Unsloth](https://github.com/unslothai/unsloth)、[Liger Kernel](https://github.com/linkedin/Liger-Kernel)、RoPE scaling、NEFTune 和 rsLoRA。
 - **实验监控**：LlamaBoard、TensorBoard、Wandb、MLflow 等等。
 - **极速推理**：基于 vLLM 的 OpenAI 风格 API、浏览器界面和命令行接口。
 
@@ -71,15 +73,23 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 更新日志
 
+[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 `gemma` 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 `paligemma` 模板进行微调使其获得对话能力。
 
 [24/05/18] 我们支持了 **[KTO](https://arxiv.org/abs/2402.01306)** 偏好对齐算法。详细用法请参照 [examples](examples/README_zh.md)。
 
@@ -91,7 +101,7 @@ 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 +113,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 [24/03/13] 我们支持了 **[LoRA+](https://arxiv.org/abs/2402.12354)**。详细用法请参照 [examples](examples/README_zh.md)。
 
-[24/03/07] 我们支持了梯度低秩投影（**[GaLore](https://arxiv.org/abs/2403.03507)**）算法。详细用法请参照 [examples](examples/README_zh.md)。
+[24/03/07] 我们支持了 **[GaLore](https://arxiv.org/abs/2403.03507)** 优化器。详细用法请参照 [examples](examples/README_zh.md)。
 
 [24/03/07] 我们集成了 **[vLLM](https://github.com/vllm-project/vllm)** 以实现极速并发推理。请使用 `infer_backend: vllm` 来获得 **270%** 的推理速度。
 
@@ -151,35 +161,34 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 模型
 
-| 模型名                                                    | 模型大小                          | Template  |
-| --------------------------------------------------------- | -------------------------------- | --------- |
-| [Baichuan2](https://huggingface.co/baichuan-inc)          | 7B/13B                           | baichuan2 |
-| [BLOOM](https://huggingface.co/bigscience)                | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [BLOOMZ](https://huggingface.co/bigscience)               | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [ChatGLM3](https://huggingface.co/THUDM)                  | 6B                               | chatglm3  |
-| [Command-R](https://huggingface.co/CohereForAI)           | 35B/104B                         | cohere    |
-| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai) | 7B/16B/67B/236B                  | deepseek  |
-| [Falcon](https://huggingface.co/tiiuae)                   | 7B/11B/40B/180B                  | falcon    |
-| [Gemma/CodeGemma](https://huggingface.co/google)          | 2B/7B                            | gemma     |
-| [GLM4](https://huggingface.co/THUDM)                      | 9B                               | glm4      |
-| [InternLM2](https://huggingface.co/internlm)              | 7B/20B                           | intern2   |
-| [LLaMA](https://github.com/facebookresearch/llama)        | 7B/13B/33B/65B                   | -         |
-| [LLaMA-2](https://huggingface.co/meta-llama)              | 7B/13B/70B                       | llama2    |
-| [LLaMA-3](https://huggingface.co/meta-llama)              | 8B/70B                           | llama3    |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)              | 7B/13B                           | vicuna    |
-| [Mistral/Mixtral](https://huggingface.co/mistralai)       | 7B/8x7B/8x22B                    | mistral   |
-| [OLMo](https://huggingface.co/allenai)                    | 1B/7B                            | -         |
-| [PaliGemma](https://huggingface.co/google)                | 3B                               | gemma     |
-| [Phi-1.5/2](https://huggingface.co/microsoft)             | 1.3B/2.7B                        | -         |
-| [Phi-3](https://huggingface.co/microsoft)                 | 4B/7B/14B                        | phi       |
-| [Qwen](https://huggingface.co/Qwen)                       | 1.8B/7B/14B/72B                  | qwen      |
-| [Qwen1.5 (Code/MoE)](https://huggingface.co/Qwen)         | 0.5B/1.8B/4B/7B/14B/32B/72B/110B | qwen      |
-| [Qwen2 (MoE)](https://huggingface.co/Qwen)                | 0.5B/1.5B/7B/57B/72B             | qwen      |
-| [StarCoder2](https://huggingface.co/bigcode)              | 3B/7B/15B                        | -         |
-| [XVERSE](https://huggingface.co/xverse)                   | 7B/13B/65B                       | xverse    |
-| [Yi (1/1.5)](https://huggingface.co/01-ai)                | 6B/9B/34B                        | yi        |
-| [Yi-VL](https://huggingface.co/01-ai)                     | 6B/34B                           | yi_vl     |
-| [Yuan](https://huggingface.co/IEITYuan)                   | 2B/51B/102B                      | yuan      |
+| 模型名                                                            | 模型大小                          | Template  |
+| ----------------------------------------------------------------- | -------------------------------- | --------- |
+| [Baichuan 2](https://huggingface.co/baichuan-inc)                 | 7B/13B                           | baichuan2 |
+| [BLOOM/BLOOMZ](https://huggingface.co/bigscience)                 | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
+| [ChatGLM3](https://huggingface.co/THUDM)                          | 6B                               | chatglm3  |
+| [Command R](https://huggingface.co/CohereForAI)                   | 35B/104B                         | cohere    |
+| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)         | 7B/16B/67B/236B                  | deepseek  |
+| [Falcon](https://huggingface.co/tiiuae)                           | 7B/11B/40B/180B                  | falcon    |
+| [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google)          | 2B/7B/9B/27B                     | gemma     |
+| [GLM-4](https://huggingface.co/THUDM)                             | 9B                               | glm4      |
+| [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/Llama 3.1](https://huggingface.co/meta-llama)            | 8B/70B                           | llama3    |
+| [LLaVA-1.5](https://huggingface.co/llava-hf)                      | 7B/13B                           | llava     |
+| [MiniCPM](https://huggingface.co/openbmb)                         | 1B/2B/4B                         | cpm/cpm3  |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)               | 7B/8x7B/8x22B                    | mistral   |
+| [OLMo](https://huggingface.co/allenai)                            | 1B/7B                            | -         |
+| [PaliGemma](https://huggingface.co/google)                        | 3B                               | paligemma |
+| [Phi-1.5/Phi-2](https://huggingface.co/microsoft)                 | 1.3B/2.7B                        | -         |
+| [Phi-3](https://huggingface.co/microsoft)                         | 4B/7B/14B                        | phi       |
+| [Qwen/Qwen1.5/Qwen2 (Code/Math/MoE)](https://huggingface.co/Qwen) | 0.5B/1.5B/4B/7B/14B/32B/72B/110B | qwen      |
+| [Qwen2-VL](https://huggingface.co/Qwen)                           | 2B/7B                            | qwen2_vl  |
+| [StarCoder 2](https://huggingface.co/bigcode)                     | 3B/7B/15B                        | -         |
+| [XVERSE](https://huggingface.co/xverse)                           | 7B/13B/65B                       | xverse    |
+| [Yi/Yi-1.5 (Code)](https://huggingface.co/01-ai)                  | 1.5B/6B/9B/34B                   | yi        |
+| [Yi-VL](https://huggingface.co/01-ai)                             | 6B/34B                           | yi_vl     |
+| [Yuan 2](https://huggingface.co/IEITYuan)                         | 2B/51B/102B                      | yuan      |
 
 > [!NOTE]
 > 对于所有“基座”（Base）模型，`template` 参数可以是 `default`, `alpaca`, `vicuna` 等任意值。但“对话”（Instruct/Chat）模型请务必使用**对应的模板**。
@@ -203,6 +212,9 @@ 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>
@@ -262,7 +274,9 @@ 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)
@@ -279,6 +293,8 @@ 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)
@@ -299,20 +315,20 @@ huggingface-cli login
 | 必需项       | 至少     | 推荐      |
 | ------------ | ------- | --------- |
 | python       | 3.8     | 3.11      |
-| torch        | 1.13.1  | 2.3.0     |
-| transformers | 4.41.2  | 4.41.2    |
-| datasets     | 2.16.0  | 2.19.2    |
-| accelerate   | 0.30.1  | 0.30.1    |
-| peft         | 0.11.1  | 0.11.1    |
-| trl          | 0.8.6   | 0.9.4     |
+| torch        | 1.13.1  | 2.4.0     |
+| transformers | 4.41.2  | 4.43.4    |
+| datasets     | 2.16.0  | 2.20.0    |
+| accelerate   | 0.30.1  | 0.32.0    |
+| peft         | 0.11.1  | 0.12.0    |
+| trl          | 0.8.6   | 0.9.6     |
 
 | 可选项       | 至少     | 推荐      |
 | ------------ | ------- | --------- |
 | CUDA         | 11.6    | 12.2      |
 | deepspeed    | 0.10.0  | 0.14.0    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.4.3     |
-| flash-attn   | 2.3.0   | 2.5.9     |
+| vllm         | 0.4.3   | 0.5.0     |
+| flash-attn   | 2.3.0   | 2.6.3     |
 
 ### 硬件依赖
 
@@ -341,7 +357,7 @@ cd LLaMA-Factory
 pip install -e ".[torch,metrics]"
 ```
 
-可选的额外依赖项：torch、torch_npu、metrics、deepspeed、bitsandbytes、vllm、galore、badam、gptq、awq、aqlm、qwen、modelscope、quality
+可选的额外依赖项：torch、torch-npu、metrics、deepspeed、liger-kernel、bitsandbytes、hqq、eetq、gptq、awq、aqlm、vllm、galore、badam、adam-mini、qwen、modelscope、quality
 
 > [!TIP]
 > 遇到包冲突时，可使用 `pip install --no-deps -e .` 解决。
@@ -360,9 +376,7 @@ pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/downl
 
 <details><summary>昇腾 NPU 用户指南</summary>
 
-加入 [NPU 用户群](assets/wechat_npu.jpg)。
-
-在昇腾 NPU 设备上安装 LLaMA Factory 时，需要指定额外依赖项，使用 `pip install -e '.[torch-npu,metrics]'` 命令安装。此外，还需要安装 **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**，安装方法请参考[安装教程](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/quickstart/quickstart/quickstart_18_0004.html)或使用以下命令：
+在昇腾 NPU 设备上安装 LLaMA Factory 时，需要指定额外依赖项，使用 `pip install -e ".[torch-npu,metrics]"` 命令安装。此外，还需要安装 **[Ascend CANN Toolkit 与 Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**，安装方法请参考[安装教程](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/quickstart/quickstart/quickstart_18_0004.html)或使用以下命令：
 
 ```bash
 # 请替换 URL 为 CANN 版本和设备型号对应的 URL
@@ -385,15 +399,12 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch-npu    | 2.1.0   | 2.1.0.post3 |
 | deepspeed    | 0.13.2  | 0.13.2      |
 
-Docker 镜像：
-
-- 32GB：[下载地址](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html)
-- 64GB：[下载地址](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
-
 请使用 `ASCEND_RT_VISIBLE_DEVICES` 而非 `CUDA_VISIBLE_DEVICES` 来指定运算设备。
 
 如果遇到无法正常推理的情况，请尝试设置 `do_sample: false`。
 
+下载预构建 Docker 镜像：[32GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html) | [64GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
+
 </details>
 
 ### 数据准备
@@ -426,18 +437,46 @@ llamafactory-cli webui
 
 ### 构建 Docker
 
-#### 使用 Docker
+CUDA 用户：
 
 ```bash
-docker build -f ./Dockerfile \
+cd docker/docker-cuda/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+昇腾 NPU 用户：
+
+```bash
+cd docker/docker-npu/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+AMD ROCm 用户：
+
+```bash
+cd docker/docker-rocm/
+docker compose up -d
+docker compose exec llamafactory bash
+```
+
+<details><summary>不使用 Docker Compose 构建</summary>
+
+CUDA 用户：
+
+```bash
+docker build -f ./docker/docker-cuda/Dockerfile \
     --build-arg INSTALL_BNB=false \
     --build-arg INSTALL_VLLM=false \
     --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
     --build-arg PIP_INDEX=https://pypi.org/simple \
     -t llamafactory:latest .
 
-docker run -it --gpus=all \
-    -v ./hf_cache:/root/.cache/huggingface/ \
+docker run -dit --gpus=all \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
     -v ./data:/app/data \
     -v ./output:/app/output \
     -p 7860:7860 \
@@ -445,20 +484,78 @@ docker run -it --gpus=all \
     --shm-size 16G \
     --name llamafactory \
     llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
-#### 使用 Docker Compose
+昇腾 NPU 用户：
 
 ```bash
-docker-compose up -d
-docker-compose exec llamafactory bash
+# 根据您的环境选择镜像
+docker build -f ./docker/docker-npu/Dockerfile \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+# 根据您的资源更改 `device`
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
+    -v /etc/ascend_install.info:/etc/ascend_install.info \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/davinci0 \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
+AMD ROCm 用户：
+
+```bash
+docker build -f ./docker/docker-rocm/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v ./saves:/app/saves \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/kfd \
+    --device /dev/dri \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
+```
+
+</details>
+
 <details><summary>数据卷详情</summary>
 
-- hf_cache：使用宿主机的 Hugging Face 缓存文件夹，允许更改为新的目录。
-- data：宿主机中存放数据集的文件夹路径。
-- output：将导出目录设置为该路径后，即可在宿主机中访问导出后的模型。
+- `hf_cache`：使用宿主机的 Hugging Face 缓存文件夹，允许更改为新的目录。
+- `ms_cache`：类似 Hugging Face 缓存文件夹，为 ModelScope 用户提供。
+- `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>
 
@@ -469,7 +566,7 @@ 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)。
 
 ### 从魔搭社区下载
 
@@ -503,38 +600,82 @@ run_name: test_run # 可选
 1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
 1. Luceri et al. Leveraging Large Language Models to Detect Influence Campaigns in Social Media. 2023. [[arxiv]](https://arxiv.org/abs/2311.07816)
 1. Zhang et al. Alleviating Hallucinations of Large Language Models through Induced Hallucinations. 2023. [[arxiv]](https://arxiv.org/abs/2312.15710)
-1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
-1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
+1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. KDD 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
+1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
 1. Choi et al. FACT-GPT: Fact-Checking Augmentation via Claim Matching with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2402.05904)
 1. Zhang et al. AutoMathText: Autonomous Data Selection with Language Models for Mathematical Texts. 2024. [[arxiv]](https://arxiv.org/abs/2402.07625)
 1. Lyu et al. KnowTuning: Knowledge-aware Fine-tuning for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11176)
 1. Yang et al. LaCo: Large Language Model Pruning via Layer Collaps. 2024. [[arxiv]](https://arxiv.org/abs/2402.11187)
 1. Bhardwaj et al. Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic. 2024. [[arxiv]](https://arxiv.org/abs/2402.11746)
 1. Yang et al. Enhancing Empathetic Response Generation by Augmenting LLMs with Small-scale Empathetic Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11801)
-1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. 2024. [[arxiv]](https://arxiv.org/abs/2402.11809)
+1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2402.11809)
 1. Cao et al. Head-wise Shareable Attention for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11819)
 1. Zhang et al. Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languages. 2024. [[arxiv]](https://arxiv.org/abs/2402.12204)
 1. Kim et al. Efficient and Effective Vocabulary Expansion Towards Multilingual Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.14714)
-1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
+1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. ACL 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
 1. Huang et al. Key-Point-Driven Data Synthesis with its Enhancement on Mathematical Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2403.02333)
 1. Duan et al. Negating Negatives: Alignment without Human Positive Samples via Distributional Dispreference Optimization. 2024. [[arxiv]](https://arxiv.org/abs/2403.03419)
 1. Xie and Schwertfeger. Empowering Robotics with Large Language Models: osmAG Map Comprehension with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2403.08228)
 1. Wu et al. Large Language Models are Parallel Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2403.09073)
 1. Zhang et al. EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling. 2024. [[arxiv]](https://arxiv.org/abs/2403.14541)
 1. Weller et al. FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2403.15246)
-1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
+1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. COLING 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
 1. Zan et al. CodeS: Natural Language to Code Repository via Multi-Layer Sketch. 2024. [[arxiv]](https://arxiv.org/abs/2403.16443)
 1. Liu et al. Extensive Self-Contrast Enables Feedback-Free Language Model Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2404.00604)
 1. Luo et al. BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.02827)
 1. Du et al. Chinese Tiny LLM: Pretraining a Chinese-Centric Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2404.04167)
-1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
+1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. ICML 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
 1. Liu et al. Dynamic Generation of Personalities with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.07084)
 1. Shang et al. How Far Have We Gone in Stripped Binary Code Understanding Using Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.09836)
 1. Huang et al. LLMTune: Accelerate Database Knob Tuning with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.11581)
 1. Deng et al. Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction. 2024. [[arxiv]](https://arxiv.org/abs/2404.14215)
 1. Acikgoz et al. Hippocrates: An Open-Source Framework for Advancing Large Language Models in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2404.16621)
-1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2404.17140)
-1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2404.17140)
+1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. NAACL 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Xu et al. Large Language Models for Cyber Security: A Systematic Literature Review. 2024. [[arxiv]](https://arxiv.org/abs/2405.04760)
+1. Dammu et al. "They are uncultured": Unveiling Covert Harms and Social Threats in LLM Generated Conversations. 2024. [[arxiv]](https://arxiv.org/abs/2405.05378)
+1. Yi et al. A safety realignment framework via subspace-oriented model fusion for large language models. 2024. [[arxiv]](https://arxiv.org/abs/2405.09055)
+1. Lou et al. SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling. 2024. [[arxiv]](https://arxiv.org/abs/2405.12739)
+1. Zhang et al. Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2405.13816)
+1. Zhang et al. TS-Align: A Teacher-Student Collaborative Framework for Scalable Iterative Finetuning of Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2405.20215)
+1. Zihong Chen. Sentence Segmentation and Sentence Punctuation Based on XunziALLM. 2024. [[paper]](https://aclanthology.org/2024.lt4hala-1.30)
+1. Gao et al. The Best of Both Worlds: Toward an Honest and Helpful Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2406.00380)
+1. Wang and Song. MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset. 2024. [[arxiv]](https://arxiv.org/abs/2406.02106)
+1. Hu et al. Computational Limits of Low-Rank Adaptation (LoRA) for Transformer-Based Models. 2024. [[arxiv]](https://arxiv.org/abs/2406.03136)
+1. Ge et al. Time Sensitive Knowledge Editing through Efficient Finetuning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2406.04496)
+1. Tan et al. Peer Review as A Multi-Turn and Long-Context Dialogue with Role-Based Interactions. 2024. [[arxiv]](https://arxiv.org/abs/2406.05688)
+1. Song et al. Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters. 2024. [[arxiv]](https://arxiv.org/abs/2406.05955)
+1. Gu et al. RWKV-CLIP: A Robust Vision-Language Representation Learner. 2024. [[arxiv]](https://arxiv.org/abs/2406.06973)
+1. Chen et al. Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees. 2024. [[arxiv]](https://arxiv.org/abs/2406.07115)
+1. Zhu et al. Are Large Language Models Good Statisticians?. 2024. [[arxiv]](https://arxiv.org/abs/2406.07815)
+1. Li et al. Know the Unknown: An Uncertainty-Sensitive Method for LLM Instruction Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2406.10099)
+1. Ding et al. IntentionQA: A Benchmark for Evaluating Purchase Intention Comprehension Abilities of Language Models in E-commerce. 2024. [[arxiv]](https://arxiv.org/abs/2406.10173)
+1. He et al. COMMUNITY-CROSS-INSTRUCT: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities. 2024. [[arxiv]](https://arxiv.org/abs/2406.12074)
+1. Lin et al. FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving. 2024. [[arxiv]](https://arxiv.org/abs/2406.14408)
+1. Treutlein et al. Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data. 2024. [[arxiv]](https://arxiv.org/abs/2406.14546)
+1. Feng et al. SS-Bench: A Benchmark for Social Story Generation and Evaluation. 2024. [[arxiv]](https://arxiv.org/abs/2406.15695)
+1. Feng et al. Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement. 2024. [[arxiv]](https://arxiv.org/abs/2406.17233)
+1. Liu et al. Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals. 2024. [[arxiv]](https://arxiv.org/abs/2406.18069)
+1. Iyer et al. Exploring Very Low-Resource Translation with LLMs: The University of Edinburgh's Submission to AmericasNLP 2024 Translation Task. AmericasNLP 2024. [[paper]](https://aclanthology.org/2024.americasnlp-1.25)
+1. Li et al. Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring. 2024. [[arxiv]](https://arxiv.org/abs/2406.19949)
+1. Yang et al. Financial Knowledge Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2407.00365)
+1. Lin et al. DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging. 2024. [[arxiv]](https://arxiv.org/abs/2407.01470)
+1. Bako et al. Evaluating the Semantic Profiling Abilities of LLMs for Natural Language Utterances in Data Visualization. 2024. [[arxiv]](https://arxiv.org/abs/2407.06129)
+1. Huang et al. RoLoRA: Fine-tuning Rotated Outlier-free LLMs for Effective Weight-Activation Quantization. 2024. [[arxiv]](https://arxiv.org/abs/2407.08044)
+1. Jiang et al. LLM-Collaboration on Automatic Science Journalism for the General Audience. 2024. [[arxiv]](https://arxiv.org/abs/2407.09756)
+1. Inouye et al. Applied Auto-tuning on LoRA Hyperparameters. 2024. [[paper]](https://scholarcommons.scu.edu/cseng_senior/272/)
+1. Qi et al. Research on Tibetan Tourism Viewpoints information generation system based on LLM. 2024. [[arxiv]](https://arxiv.org/abs/2407.13561)
+1. Xu et al. Course-Correction: Safety Alignment Using Synthetic Preferences. 2024. [[arxiv]](https://arxiv.org/abs/2407.16637)
+1. Sun et al. LAMBDA: A Large Model Based Data Agent. 2024. [[arxiv]](https://arxiv.org/abs/2407.17535)
+1. Zhu et al. CollectiveSFT: Scaling Large Language Models for Chinese Medical Benchmark with Collective Instructions in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2407.19705)
+1. Yu et al. Correcting Negative Bias in Large Language Models through Negative Attention Score Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2408.00137)
+1. Xie et al. The Power of Personalized Datasets: Advancing Chinese Composition Writing for Elementary School through Targeted Model Fine-Tuning. IALP 2024. [[paper]](https://www.asianlp.sg/conferences/ialp2024/proceedings/papers/IALP2024_P055.pdf)
+1. Liu et al. Instruct-Code-Llama: Improving Capabilities of Language Model in Competition Level Code Generation by Online Judge Feedback. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_11)
+1. Wang et al. Cybernetic Sentinels: Unveiling the Impact of Safety Data Selection on Model Security in Supervised Fine-Tuning. ICIC 2024. [[paper]](https://link.springer.com/chapter/10.1007/978-981-97-5669-8_23)
+1. Xia et al. Understanding the Performance and Estimating the Cost of LLM Fine-Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2408.04693)
+1. Zeng et al. Perceive, Reflect, and Plan: Designing LLM Agent for Goal-Directed City Navigation without Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2408.04168)
+1. Xia et al. Using Pre-trained Language Model for Accurate ESG Prediction. FinNLP 2024. [[paper]](https://aclanthology.org/2024.finnlp-2.1/)
+1. Liang et al. I-SHEEP: Self-Alignment of LLM from Scratch through an Iterative Self-Enhancement Paradigm. 2024. [[arxiv]](https://arxiv.org/abs/2408.08072)
 1. **[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。
@@ -542,6 +683,9 @@ run_name: test_run # 可选
 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. **[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.
 
 </details>
 
@@ -549,17 +693,19 @@ run_name: test_run # 可选
 
 本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源。
 
-使用模型权重时，请遵循对应的模型协议：[Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command-R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [LLaMA-3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yi-1.5](LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+使用模型权重时，请遵循对应的模型协议：[Baichuan 2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [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](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder 2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yi-1.5](LICENSE) / [Yuan 2](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## 引用
 
 如果您觉得此项目有帮助，请考虑以下列格式引用
 
 ```bibtex
-@article{zheng2024llamafactory,
-  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models}, 
-  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Yongqiang Ma},
-  journal={arXiv preprint arXiv:2403.13372},
+@inproceedings{zheng2024llamafactory,
+  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models},
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Zhangchi Feng and Yongqiang Ma},
+  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
+  address={Bangkok, Thailand},
+  publisher={Association for Computational Linguistics},
   year={2024},
   url={http://arxiv.org/abs/2403.13372}
 }

data/README.md

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

data/README_zh.md

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

docker/docker-cuda/Dockerfile

@@ -0,0 +1,59 @@
+# Use the NVIDIA official image with PyTorch 2.3.0
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-24-02.html
+FROM nvcr.io/nvidia/pytorch:24.02-py3
+
+# 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 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; \
+    pip install -e ".[$EXTRA_PACKAGES]"
+
+# Rebuild flash attention
+RUN pip uninstall -y transformer-engine flash-attn && \
+    if [ "$INSTALL_FLASHATTN" == "true" ]; then \
+        pip uninstall -y ninja && pip install ninja && \
+        pip install --no-cache-dir flash-attn --no-build-isolation; \
+    fi
+
+# Set up volumes
+VOLUME [ "/root/.cache/huggingface", "/root/.cache/modelscope", "/app/data", "/app/output" ]
+
+# Expose port 7860 for the LLaMA Board
+ENV GRADIO_SERVER_PORT 7860
+EXPOSE 7860
+
+# Expose port 8000 for the API service
+ENV API_PORT 8000
+EXPOSE 8000

docker/docker-cuda/docker-compose.yml

@@ -1,18 +1,20 @@
 services:
   llamafactory:
     build:
-      dockerfile: Dockerfile
-      context: .
+      dockerfile: ./docker/docker-cuda/Dockerfile
+      context: ../..
       args:
         INSTALL_BNB: false
         INSTALL_VLLM: false
         INSTALL_DEEPSPEED: false
+        INSTALL_FLASHATTN: false
         PIP_INDEX: https://pypi.org/simple
     container_name: llamafactory
     volumes:
-      - ./hf_cache:/root/.cache/huggingface/
-      - ./data:/app/data
-      - ./output:/app/output
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../data:/app/data
+      - ../../output:/app/output
     ports:
       - "7860:7860"
       - "8000:8000"

docker/docker-npu/Dockerfile

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

docker/docker-npu/docker-compose.yml

@@ -0,0 +1,31 @@
+services:
+  llamafactory:
+    build:
+      dockerfile: ./docker/docker-npu/Dockerfile
+      context: ../..
+      args:
+        INSTALL_DEEPSPEED: false
+        PIP_INDEX: https://pypi.org/simple
+    container_name: llamafactory
+    volumes:
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../data:/app/data
+      - ../../output:/app/output
+      - /usr/local/dcmi:/usr/local/dcmi
+      - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
+      - /usr/local/Ascend/driver:/usr/local/Ascend/driver
+      - /etc/ascend_install.info:/etc/ascend_install.info
+    ports:
+      - "7860:7860"
+      - "8000:8000"
+    ipc: host
+    tty: true
+    stdin_open: true
+    command: bash
+    devices:
+      - /dev/davinci0
+      - /dev/davinci_manager
+      - /dev/devmm_svm
+      - /dev/hisi_hdc
+    restart: unless-stopped

docker/docker-rocm/Dockerfile

@@ -0,0 +1,57 @@
+FROM hardandheavy/transformers-rocm:2.1.0
+
+# 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 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; \
+    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 +1,29 @@
+services:
+  llamafactory:
+    build:
+      dockerfile: ./docker/docker-rocm/Dockerfile
+      context: ../..
+      args:
+        INSTALL_BNB: false
+        INSTALL_VLLM: false
+        INSTALL_DEEPSPEED: false
+        INSTALL_FLASHATTN: false
+        PIP_INDEX: https://pypi.org/simple
+    container_name: llamafactory
+    volumes:
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../data:/app/data
+      - ../../output:/app/output
+      - ../../saves:/app/saves
+    ports:
+      - "7860:7860"
+      - "8000:8000"
+    ipc: host
+    tty: true
+    stdin_open: true
+    command: bash
+    devices:
+      - /dev/kfd:/dev/kfd
+      - /dev/dri:/dev/dri
+    restart: unless-stopped

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
@@ -94,10 +101,10 @@ FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.
 
 ### QLoRA Fine-Tuning
 
-#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes Quantization (Recommended)
+#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes/HQQ/EETQ Quantization (Recommended)
 
 ```bash
-llamafactory-cli train examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_otfq.yaml
 ```
 
 #### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization
@@ -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
@@ -94,10 +101,10 @@ FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.
 
 ### QLoRA 微调
 
-#### 基于 4/8 比特 Bitsandbytes 量化进行指令监督微调（推荐）
+#### 基于 4/8 比特 Bitsandbytes/HQQ/EETQ 量化进行指令监督微调（推荐）
 
 ```bash
-llamafactory-cli train examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_otfq.yaml
 ```
 
 #### 基于 4/8 比特 GPTQ 量化进行指令监督微调
@@ -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

@@ -0,0 +1,39 @@
+### model
+model_name_or_path: Qwen/Qwen2-1.5B-Instruct
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+use_adam_mini: true
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: qwen
+cutoff_len: 1024
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/qwen2-1_5b/full/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-5
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

examples/extras/badam/llama3_full_sft.yaml

@@ -6,9 +6,11 @@ stage: sft
 do_train: true
 finetuning_type: full
 use_badam: true
+badam_mode: layer
 badam_switch_mode: ascending
 badam_switch_interval: 50
 badam_verbose: 2
+# deepspeed: examples/deepspeed/ds_z3_config.json
 
 ### dataset
 dataset: identity,alpaca_en_demo
@@ -28,11 +30,10 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 8
-learning_rate: 1.0e-4
+learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-pure_bf16: true
 
 ### eval
 val_size: 0.1

examples/extras/fsdp_qlora/llama3_lora_sft.yaml

@@ -30,7 +30,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/extras/galore/llama3_full_sft.yaml

@@ -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
@@ -18,7 +18,7 @@ overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b-instruct-pro/freeze/sft
+output_dir: saves/llama3-8b-pro/freeze/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -31,7 +31,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/extras/loraplus/llama3_lora_sft.yaml

@@ -30,7 +30,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/extras/mod/llama3_full_sft.yaml

@@ -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

@@ -7,7 +7,7 @@ do_train: true
 finetuning_type: lora
 lora_target: all
 pissa_init: true
-pissa_iter: 4
+pissa_iter: 16
 pissa_convert: true
 
 ### dataset
@@ -32,7 +32,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/inference/llava1_5.yaml

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

examples/inference/qwen2_vl.yaml

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

examples/merge_lora/qwen2vl_lora_sft.yaml

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

examples/train_full/llama3_full_predict.yaml

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

examples/train_full/llama3_full_sft_ds3.yaml

@@ -25,11 +25,11 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 2
-learning_rate: 1.0e-4
+learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

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: 1024
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/qwen2_vl-7b/full/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 2
+learning_rate: 1.0e-5
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

examples/train_lora/llama3_lora_dpo.yaml

@@ -7,7 +7,7 @@ do_train: true
 finetuning_type: lora
 lora_target: all
 pref_beta: 0.1
-pref_loss: sigmoid  # [sigmoid (dpo), orpo, simpo]
+pref_loss: sigmoid  # choices: [sigmoid (dpo), orpo, simpo]
 
 ### dataset
 dataset: dpo_en_demo
@@ -31,7 +31,7 @@ learning_rate: 5.0e-6
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_lora/llama3_lora_eval.yaml

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

examples/train_lora/llama3_lora_kto.yaml

@@ -30,7 +30,7 @@ learning_rate: 5.0e-6
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_lora/llama3_lora_ppo.yaml

@@ -30,7 +30,7 @@ learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### generate

examples/train_lora/llama3_lora_predict.yaml

@@ -8,7 +8,7 @@ do_predict: true
 finetuning_type: lora
 
 ### dataset
-dataset: identity,alpaca_en_demo
+eval_dataset: identity,alpaca_en_demo
 template: llama3
 cutoff_len: 1024
 max_samples: 50

examples/train_lora/llama3_lora_pretrain.yaml

@@ -15,7 +15,7 @@ overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/lora/sft
+output_dir: saves/llama3-8b/lora/pretrain
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -28,7 +28,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_lora/llama3_lora_reward.yaml

@@ -25,11 +25,11 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 8
-learning_rate: 1.0e-5
+learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_lora/llama3_lora_sft.yaml

@@ -29,7 +29,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_lora/llama3_lora_sft_ds0.yaml

@@ -30,7 +30,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_lora/llama3_lora_sft_ds3.yaml

@@ -30,7 +30,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

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
@@ -10,7 +9,7 @@ lora_target: all
 
 ### dataset
 dataset: mllm_demo
-template: vicuna
+template: llava
 cutoff_len: 1024
 max_samples: 1000
 overwrite_cache: true
@@ -30,7 +29,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

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: 1024
+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: 1024
+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

@@ -29,7 +29,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_qlora/llama3_lora_sft_awq.yaml

@@ -29,7 +29,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_qlora/llama3_lora_sft_gptq.yaml

@@ -29,7 +29,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

examples/train_qlora/llama3_lora_sft_otfq.yaml

@@ -1,6 +1,7 @@
 ### model
 model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 quantization_bit: 4
+quantization_method: bitsandbytes  # choices: [bitsandbytes (4/8), hqq (2/3/4/5/6/8), eetq (8)]
 
 ### method
 stage: sft
@@ -30,7 +31,7 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
 ddp_timeout: 180000000
 
 ### eval

requirements.txt

@@ -1,8 +1,8 @@
-transformers>=4.41.2
-datasets>=2.16.0
-accelerate>=0.30.1
-peft>=0.11.1
-trl>=0.8.6
+transformers>=4.41.2,<=4.45.0
+datasets>=2.16.0,<=2.21.0
+accelerate>=0.30.1,<=0.33.0
+peft>=0.11.1,<=0.12.0
+trl>=0.8.6,<=0.9.6
 gradio>=4.0.0
 pandas>=2.0.0
 scipy
@@ -18,3 +18,4 @@ matplotlib>=3.7.0
 fire
 packaging
 pyyaml
+numpy<2.0.0

scripts/cal_flops.py

@@ -27,7 +27,7 @@ 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 +36,11 @@ 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

@@ -25,7 +25,7 @@ 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
@@ -39,15 +39,17 @@ 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(
@@ -57,19 +59,22 @@ def calculate_lr(
             dataset_dir=dataset_dir,
             template=template,
             cutoff_len=cutoff_len,
+            packing=packing,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    trainset = get_dataset(template, model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
     if stage == "pt":
         data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
     elif stage == "sft":
         data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX)
     else:
-        raise NotImplementedError
+        raise NotImplementedError("Stage does not supported: {}.".format(stage))
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     valid_tokens, total_tokens = 0, 0
@@ -81,7 +86,7 @@ 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 +1,164 @@
+# coding=utf-8
+# 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("Device not supported: {}.".format(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"] = "examples/deepspeed/ds_z{}_config.json".format(deepspeed_stage)
+
+    run_exp(args)
+    with open(os.path.join("saves", "test_mfu", "all_results.json"), "r", 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("MFU: {:.2f}%".format(mfu_value * 100))
+
+
+if __name__ == "__main__":
+    fire.Fire(calculate_mfu)

scripts/cal_ppl.py

@@ -23,7 +23,7 @@ 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 +55,12 @@ 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 +69,7 @@ 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(
@@ -83,11 +83,13 @@ def cal_ppl(
             train_on_prompt=train_on_prompt,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    trainset = get_dataset(template, model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
     model = load_model(tokenizer, model_args, finetuning_args, is_trainable=False)
     if stage == "pt":
         data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
@@ -98,7 +100,7 @@ def cal_ppl(
             tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX, train_on_prompt=train_on_prompt
         )
     else:
-        raise NotImplementedError
+        raise NotImplementedError("Stage does not supported: {}.".format(stage))
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     criterion = torch.nn.CrossEntropyLoss(reduction="none")
@@ -128,4 +130,4 @@ def cal_ppl(
 
 
 if __name__ == "__main__":
-    fire.Fire(cal_ppl)
+    fire.Fire(calculate_ppl)

scripts/length_cdf.py

@@ -18,21 +18,21 @@ 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(
@@ -44,10 +44,12 @@ def length_cdf(
             cutoff_len=1_000_000,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
-    trainset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
+    template = get_template_and_fix_tokenizer(tokenizer_module["tokenizer"], data_args)
+    trainset = get_dataset(template, model_args, data_args, training_args, "sft", **tokenizer_module)["train_dataset"]
     total_num = len(trainset)
     length_dict = defaultdict(int)
     for sample in tqdm(trainset["input_ids"]):

scripts/llama_pro.py

@@ -19,7 +19,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING
 
 import fire
 import torch
@@ -47,8 +47,8 @@ def block_expansion(
     model_name_or_path: str,
     output_dir: str,
     num_expand: int,
-    shard_size: Optional[str] = "2GB",
-    save_safetensors: Optional[bool] = False,
+    shard_size: str = "2GB",
+    save_safetensors: bool = True,
 ):
     r"""
     Performs block expansion for LLaMA, Mistral, Qwen1.5 or Yi models.

scripts/llamafy_baichuan2.py

@@ -16,7 +16,7 @@
 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 +86,10 @@ def save_config(input_dir: str, output_dir: str):
 
 
 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

@@ -16,7 +16,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import Any, Dict, Optional
+from typing import Any, Dict
 
 import fire
 import torch
@@ -139,7 +139,10 @@ def save_config(input_dir: str, output_dir: str, torch_dtype: str):
 
 
 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

@@ -36,15 +36,19 @@ def quantize_loftq(
     lora_alpha: int = None,
     lora_rank: int = 16,
     lora_dropout: float = 0,
-    lora_target: str = "q_proj,v_proj",
+    lora_target: tuple = ("q_proj", "v_proj"),
     save_safetensors: bool = True,
 ):
     r"""
     Initializes LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ)
     Usage: python loftq_init.py --model_name_or_path path_to_model --output_dir output_dir
     """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
     tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
     model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
+
     loftq_config = LoftQConfig(loftq_bits=loftq_bits, loftq_iter=loftq_iter)
     lora_config = LoraConfig(
         task_type=TaskType.CAUSAL_LM,
@@ -52,7 +56,7 @@ def quantize_loftq(
         r=lora_rank,
         lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
         lora_dropout=lora_dropout,
-        target_modules=[name.strip() for name in lora_target.split(",")],
+        target_modules=lora_target,
         init_lora_weights="loftq",
         loftq_config=loftq_config,
     )
@@ -63,7 +67,7 @@ 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))

scripts/pissa_init.py

@@ -31,25 +31,29 @@ 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,
-    lora_target: str = "q_proj,v_proj",
+    lora_target: tuple = ("q_proj", "v_proj"),
     save_safetensors: bool = True,
 ):
     r"""
     Initializes LoRA weights with Principal Singular values and Singular vectors Adaptation (PiSSA)
     Usage: python pissa_init.py --model_name_or_path path_to_model --output_dir output_dir
     """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
     tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
     model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
+
     lora_config = LoraConfig(
         task_type=TaskType.CAUSAL_LM,
         r=lora_rank,
         lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
         lora_dropout=lora_dropout,
-        target_modules=[name.strip() for name in lora_target.split(",")],
+        target_modules=lora_target,
         init_lora_weights="pissa" if pissa_iter == -1 else "pissa_niter_{}".format(pissa_iter),
     )
 
@@ -58,6 +62,7 @@ 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))

setup.py

@@ -14,11 +14,12 @@
 
 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:
         file_content = f.read()
         pattern = r"{}\W*=\W*\"([^\"]+)\"".format("VERSION")
@@ -26,25 +27,37 @@ def get_version():
         return version
 
 
-def get_requires():
+def get_requires() -> List[str]:
     with open("requirements.txt", "r", 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"],
-    "vllm": ["vllm>=0.4.3"],
-    "galore": ["galore-torch"],
-    "badam": ["badam"],
-    "gptq": ["optimum>=1.16.0", "auto-gptq>=0.5.0"],
+    "hqq": ["hqq"],
+    "eetq": ["eetq"],
+    "gptq": ["optimum>=1.17.0", "auto-gptq>=0.5.0"],
     "awq": ["autoawq"],
     "aqlm": ["aqlm[gpu]>=1.1.0"],
+    "vllm": ["vllm>=0.4.3,<=0.6.0"],
+    "galore": ["galore-torch"],
+    "badam": ["badam>=1.2.1"],
+    "adam-mini": ["adam-mini"],
     "qwen": ["transformers_stream_generator"],
     "modelscope": ["modelscope"],
     "dev": ["ruff", "pytest"],
@@ -68,7 +81,7 @@ 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/llamafactory/__init__.py

@@ -12,9 +12,35 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Level: api, webui > chat, eval, train > data, model > hparams > extras
+r"""
+Efficient fine-tuning of large language models.
 
-from .cli import VERSION
+Level:
+  api, webui > chat, eval, train > data, model > hparams > extras
+
+Dependency graph:
+  main:
+    transformers>=4.41.2,<=4.45.0
+    datasets>=2.16.0,<=2.21.0
+    accelerate>=0.30.1,<=0.33.0
+    peft>=0.11.1,<=0.12.0
+    trl>=0.8.6,<=0.9.6
+  attention:
+    transformers>=4.42.4 (gemma+fa2)
+  longlora:
+    transformers>=4.41.2,<=4.45.0
+  packing:
+    transformers>=4.41.2,<=4.45.0
+
+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
+"""
+
+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.environ.get("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.environ.get("API_KEY", None)
     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.environ.get("API_MODEL_NAME", "gpt-3.5-turbo"))
         return ModelList(data=[model_card])
 
     @app.post(

src/llamafactory/api/chat.py

@@ -16,6 +16,7 @@ import base64
 import io
 import json
 import os
+import re
 import uuid
 from typing import TYPE_CHECKING, AsyncGenerator, Dict, List, Optional, Tuple
 
@@ -51,9 +52,8 @@ 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
 
 
@@ -69,7 +69,7 @@ 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["ImageInput"]]:
     logger.info("==== request ====\n{}".format(json.dumps(dictify(request), indent=2, ensure_ascii=False)))
 
     if len(request.messages) == 0:
@@ -93,7 +93,7 @@ def _process_request(
 
         if message.role == Role.ASSISTANT and isinstance(message.tool_calls, list) and len(message.tool_calls):
             tool_calls = [
-                {"name": tool_call.function.name, "argument": tool_call.function.arguments}
+                {"name": tool_call.function.name, "arguments": tool_call.function.arguments}
                 for tool_call in message.tool_calls
             ]
             content = json.dumps(tool_calls, ensure_ascii=False)
@@ -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
-                        image_data = base64.b64decode(image_url.split(",", maxsplit=1)[1])
-                        image_path = io.BytesIO(image_data)
+                    if re.match(r"^data:image\/(png|jpg|jpeg|gif|bmp);base64,(.+)$", image_url):  # base64 image
+                        image_stream = io.BytesIO(base64.b64decode(image_url.split(",", maxsplit=1)[1]))
                     elif os.path.isfile(image_url):  # local file
-                        image_path = open(image_url, "rb")
+                        image_stream = open(image_url, "rb")
                     else:  # web uri
-                        image_path = requests.get(image_url, stream=True).raw
+                        image_stream = requests.get(image_url, stream=True).raw
 
-                    image = Image.open(image_path).convert("RGB")
+                    image = Image.open(image_stream).convert("RGB")
         else:
             input_messages.append({"role": ROLE_MAPPING[message.role], "content": message.content})
 

src/llamafactory/api/protocol.py

@@ -96,7 +96,7 @@ class ChatCompletionRequest(BaseModel):
     model: str
     messages: List[ChatMessage]
     tools: Optional[List[FunctionAvailable]] = None
-    do_sample: bool = True
+    do_sample: Optional[bool] = None
     temperature: Optional[float] = None
     top_p: Optional[float] = None
     n: int = 1

src/llamafactory/chat/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,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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

@@ -16,6 +16,7 @@
 # limitations under the License.
 
 import asyncio
+import os
 from threading import Thread
 from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, Generator, List, Optional, Sequence
 
@@ -26,8 +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
 
 
@@ -37,8 +37,17 @@ 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":
@@ -55,10 +64,16 @@ class ChatModel:
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, **input_kwargs), self._loop
+        )
         return task.result()
 
     async def achat(
@@ -66,20 +81,28 @@ class ChatModel:
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, **input_kwargs)
 
     def stream_chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, **input_kwargs)
         while True:
             try:
                 task = asyncio.run_coroutine_threadsafe(generator.__anext__(), self._loop)
@@ -92,10 +115,14 @@ class ChatModel:
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, **input_kwargs):
             yield new_token
 
     def get_scores(
@@ -103,6 +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()
 
@@ -111,17 +141,18 @@ class ChatModel:
         batch_input: List[str],
         **input_kwargs,
     ) -> List[float]:
+        r"""
+        Asynchronously gets a list of scores of the reward model.
+        """
         return await self.engine.get_scores(batch_input, **input_kwargs)
 
 
 def run_chat() -> None:
-    try:
-        import platform
-
-        if platform.system() != "Windows":
+    if os.name != "nt":
+        try:
             import readline  # noqa: F401
-    except ImportError:
-        print("Install `readline` for a better experience.")
+        except ImportError:
+            print("Install `readline` for a better experience.")
 
     chat_model = ChatModel()
     messages = []

src/llamafactory/chat/hf_engine.py

@@ -20,8 +20,10 @@ 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.constants import IMAGE_PLACEHOLDER, VIDEO_PLACEHOLDER
 from ..extras.logging import get_logger
 from ..extras.misc import get_logits_processor
 from ..model import load_model, load_tokenizer
@@ -29,12 +31,11 @@ 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
 
 
@@ -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)
+        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
@@ -78,31 +79,30 @@ class HuggingfaceEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["VideoInput"] = None,
         input_kwargs: Optional[Dict[str, Any]] = {},
     ) -> Tuple[Dict[str, Any], int]:
-        if (
-            processor is not None
-            and image is not None
-            and not hasattr(processor, "image_seq_length")
-            and template.image_token not in messages[0]["content"]
-        ):  # llava-like models
-            messages[0]["content"] = template.image_token + messages[0]["content"]
+        mm_input_dict = {"images": [], "videos": [], "imglens": [0], "vidlens": [0]}
+        if image is not None:
+            mm_input_dict.update({"images": [image], "imglens": [1]})
+            if IMAGE_PLACEHOLDER not in messages[0]["content"]:
+                messages[0]["content"] = IMAGE_PLACEHOLDER + messages[0]["content"]
 
+        if video is not None:
+            mm_input_dict.update({"videos": [video], "vidlens": [1]})
+            if VIDEO_PLACEHOLDER not in messages[0]["content"]:
+                messages[0]["content"] = VIDEO_PLACEHOLDER + messages[0]["content"]
+
+        messages = template.mm_plugin.process_messages(
+            messages, mm_input_dict["images"], mm_input_dict["videos"], processor
+        )
         paired_messages = messages + [{"role": "assistant", "content": ""}]
         system = system or generating_args["default_system"]
-        pixel_values = None
-        prompt_ids, _ = template.encode_oneturn(
-            tokenizer=tokenizer, messages=paired_messages, system=system, tools=tools
+        prompt_ids, _ = template.encode_oneturn(tokenizer, paired_messages, system, tools)
+        prompt_ids, _ = template.mm_plugin.process_token_ids(
+            prompt_ids, None, mm_input_dict["images"], mm_input_dict["videos"], tokenizer, processor
         )
-        if processor is not None and image is not None:  # add image features
-            image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
-            batch_feature = image_processor(image, return_tensors="pt")
-            pixel_values = batch_feature.to(model.device)["pixel_values"]  # shape (B, C, H, W)
-            if hasattr(processor, "image_seq_length"):  # paligemma models
-                image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
-                prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
-
         prompt_length = len(prompt_ids)
         inputs = torch.tensor([prompt_ids], device=model.device)
         attention_mask = torch.ones_like(inputs, dtype=torch.bool)
@@ -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 in Huggingface engine yet.")
+            logger.warning("Stop parameter is not supported by the huggingface engine yet.")
 
         generating_args = generating_args.copy()
         generating_args.update(
@@ -164,8 +164,10 @@ class HuggingfaceEngine(BaseEngine):
             logits_processor=get_logits_processor(),
         )
 
-        if pixel_values is not None:
-            gen_kwargs["pixel_values"] = pixel_values
+        mm_inputs = template.mm_plugin.get_mm_inputs(**mm_input_dict, seqlens=[prompt_length], processor=processor)
+        for key, value in mm_inputs.items():
+            value = value if isinstance(value, torch.Tensor) else torch.tensor(value)
+            gen_kwargs[key] = value.to(model.device)
 
         return gen_kwargs, prompt_length
 
@@ -180,11 +182,12 @@ class HuggingfaceEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, input_kwargs
         )
         generate_output = model.generate(**gen_kwargs)
         response_ids = generate_output[:, prompt_length:]
@@ -215,11 +218,12 @@ class HuggingfaceEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, input_kwargs
         )
         streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
         gen_kwargs["streamer"] = streamer
@@ -267,12 +271,14 @@ class HuggingfaceEngine(BaseEngine):
 
         return scores
 
+    @override
     async def chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["VideoInput"] = None,
         **input_kwargs,
     ) -> List["Response"]:
         if not self.can_generate:
@@ -289,18 +295,21 @@ class HuggingfaceEngine(BaseEngine):
             system,
             tools,
             image,
+            video,
             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,
+        image: Optional["ImageInput"] = None,
+        video: Optional["VideoInput"] = None,
         **input_kwargs,
     ) -> AsyncGenerator[str, None]:
         if not self.can_generate:
@@ -317,6 +326,7 @@ class HuggingfaceEngine(BaseEngine):
             system,
             tools,
             image,
+            video,
             input_kwargs,
         )
         async with self.semaphore:
@@ -328,6 +338,7 @@ class HuggingfaceEngine(BaseEngine):
                     except StopAsyncIteration:
                         break
 
+    @override
     async def get_scores(
         self,
         batch_input: List[str],

src/llamafactory/chat/vllm_engine.py

@@ -13,31 +13,33 @@
 # limitations under the License.
 
 import uuid
-from typing import TYPE_CHECKING, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence, Union
+from typing import TYPE_CHECKING, Any, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence, Union
+
+from typing_extensions import override
 
 from ..data import get_template_and_fix_tokenizer
+from ..extras.constants import IMAGE_PLACEHOLDER
 from ..extras.logging import get_logger
 from ..extras.misc import get_device_count
-from ..extras.packages import is_vllm_available, is_vllm_version_greater_than_0_5
+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():
-        from vllm.multimodal.image import ImagePixelData
-    else:
-        from vllm.sequence import MultiModalData
-
 
 if TYPE_CHECKING:
-    from numpy.typing import NDArray
-    from transformers.image_processing_utils import BaseImageProcessor
-
+    from ..data.mm_plugin import ImageInput, VideoInput
     from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
 
 
@@ -53,13 +55,18 @@ class VllmEngine(BaseEngine):
         generating_args: "GeneratingArguments",
     ) -> None:
         config = load_config(model_args)  # may download model from ms hub
+        if getattr(config, "quantization_config", None):  # gptq models should use float16
+            quantization_config: Dict[str, Any] = getattr(config, "quantization_config", None)
+            quant_method = quantization_config.get("quant_method", "")
+            if quant_method == QuantizationMethod.GPTQ and model_args.infer_dtype == "auto":
+                model_args.infer_dtype = "float16"
 
         self.can_generate = finetuning_args.stage == "sft"
         tokenizer_module = load_tokenizer(model_args)
         self.tokenizer = tokenizer_module["tokenizer"]
         self.processor = tokenizer_module["processor"]
         self.tokenizer.padding_side = "left"
-        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args)
         self.generating_args = generating_args.to_dict()
 
         engine_args = {
@@ -77,19 +84,11 @@ class VllmEngine(BaseEngine):
             "max_lora_rank": model_args.vllm_max_lora_rank,
         }
 
-        if model_args.visual_inputs:
-            image_size = config.vision_config.image_size
-            patch_size = config.vision_config.patch_size
-            self.image_feature_size = (image_size // patch_size) ** 2
-            engine_args["image_input_type"] = "pixel_values"
-            engine_args["image_token_id"] = self.tokenizer.convert_tokens_to_ids(self.template.image_token)
-            engine_args["image_input_shape"] = "1,3,{},{}".format(image_size, image_size)
-            engine_args["image_feature_size"] = self.image_feature_size
-            if getattr(config, "is_yi_vl_derived_model", None):
-                import vllm.model_executor.models.llava
+        if getattr(config, "is_yi_vl_derived_model", None):
+            import vllm.model_executor.models.llava
 
-                logger.info("Detected Yi-VL model, applying projector patch.")
-                vllm.model_executor.models.llava.LlavaMultiModalProjector = LlavaMultiModalProjectorForYiVLForVLLM
+            logger.info("Detected Yi-VL model, applying projector patch.")
+            vllm.model_executor.models.llava.LlavaMultiModalProjector = LlavaMultiModalProjectorForYiVLForVLLM
 
         self.model = AsyncLLMEngine.from_engine_args(AsyncEngineArgs(**engine_args))
         if model_args.adapter_name_or_path is not None:
@@ -102,35 +101,18 @@ class VllmEngine(BaseEngine):
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        image: Optional["NDArray"] = None,
+        image: Optional["ImageInput"] = None,
+        video: Optional["VideoInput"] = None,
         **input_kwargs,
     ) -> AsyncIterator["RequestOutput"]:
         request_id = "chatcmpl-{}".format(uuid.uuid4().hex)
-
-        if (
-            self.processor is not None
-            and image is not None
-            and not hasattr(self.processor, "image_seq_length")
-            and self.template.image_token not in messages[0]["content"]
-        ):  # llava-like models (TODO: paligemma models)
-            messages[0]["content"] = self.template.image_token * self.image_feature_size + messages[0]["content"]
+        if image is not None:
+            if IMAGE_PLACEHOLDER not in messages[0]["content"]:
+                messages[0]["content"] = IMAGE_PLACEHOLDER + 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():
-                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
-
+        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
@@ -175,6 +157,17 @@ class VllmEngine(BaseEngine):
             skip_special_tokens=True,
         )
 
+        if image is not None:  # add image features
+            if not isinstance(image, (str, ImageObject)):
+                raise ValueError("Expected image input is a path or PIL.Image, but got {}.".format(type(image)))
+
+            if isinstance(image, str):
+                image = Image.open(image).convert("RGB")
+
+            multi_modal_data = {"image": image}
+        else:
+            multi_modal_data = None
+
         result_generator = self.model.generate(
             inputs={"prompt_token_ids": prompt_ids, "multi_modal_data": multi_modal_data},
             sampling_params=sampling_params,
@@ -183,16 +176,18 @@ class VllmEngine(BaseEngine):
         )
         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,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, **input_kwargs)
         async for request_output in generator:
             final_output = request_output
 
@@ -209,21 +204,24 @@ 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,
+        image: Optional["ImageInput"] = None,
+        video: Optional["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, image, video, **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

@@ -74,7 +74,7 @@ class Command(str, Enum):
 
 
 def main():
-    command = sys.argv.pop(1)
+    command = sys.argv.pop(1) if len(sys.argv) != 1 else Command.HELP
     if command == Command.API:
         run_api()
     elif command == Command.CHAT:
@@ -91,7 +91,7 @@ def main():
             master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1")
             master_port = os.environ.get("MASTER_PORT", str(random.randint(20001, 29999)))
             logger.info("Initializing distributed tasks at: {}:{}".format(master_addr, master_port))
-            subprocess.run(
+            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}"
@@ -106,6 +106,7 @@ def main():
                 ),
                 shell=True,
             )
+            sys.exit(process.returncode)
         else:
             run_exp()
     elif command == Command.WEBDEMO:
@@ -117,4 +118,4 @@ def main():
     elif command == Command.HELP:
         print(USAGE)
     else:
-        raise NotImplementedError("Unknown command: {}".format(command))
+        raise NotImplementedError("Unknown command: {}.".format(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
+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
@@ -20,7 +25,9 @@ from .template import TEMPLATES, Template, get_template_and_fix_tokenizer
 
 __all__ = [
     "KTODataCollatorWithPadding",
+    "MultiModalDataCollatorForSeq2Seq",
     "PairwiseDataCollatorWithPadding",
+    "SFTDataCollatorWith4DAttentionMask",
     "Role",
     "split_dataset",
     "get_dataset",

src/llamafactory/data/aligner.py

@@ -14,9 +14,7 @@
 
 import os
 from functools import partial
-from typing import TYPE_CHECKING, Any, Dict, List, Union
-
-from datasets import Features
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Union
 
 from ..extras.logging import get_logger
 from .data_utils import Role
@@ -27,88 +25,117 @@ 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__)
 
 
-def _convert_images(images: List[Any], dataset_attr: "DatasetAttr", data_args: "DataArguments") -> List[Any]:
+def _convert_images(
+    images: Sequence["ImageInput"],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+) -> Optional[List["ImageInput"]]:
     r"""
     Optionally concatenates image path to dataset dir when loading from local disk.
     """
-    outputs = []
-    if dataset_attr.load_from in ["script", "file"]:
-        for image in images:
-            if isinstance(image, str) and os.path.isfile(os.path.join(data_args.dataset_dir, image)):
-                outputs.append(os.path.join(data_args.dataset_dir, image))
-            else:
-                outputs.append(image)
+    if len(images) == 0:
+        return None
 
-    return outputs
+    images = images[:]
+    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.dataset_dir, images[i])):
+                images[i] = os.path.join(data_args.dataset_dir, images[i])
+
+    return images
+
+
+def _convert_videos(
+    videos: 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 len(videos) == 0:
+        return None
+
+    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.dataset_dir, videos[i])):
+                videos[i] = os.path.join(data_args.dataset_dir, videos[i])
+
+    return videos
 
 
 def convert_alpaca(
-    examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr", data_args: "DataArguments"
-) -> Dict[str, List[Any]]:
+    example: Dict[str, Any],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+) -> Dict[str, Any]:
     r"""
     Converts alpaca format dataset to the standard format.
     """
-    outputs = {"prompt": [], "response": [], "system": [], "tools": [], "images": []}
+    prompt = []
+    if dataset_attr.history and isinstance(example[dataset_attr.history], list):
+        for old_prompt, old_response in example[dataset_attr.history]:
+            prompt.append({"role": Role.USER.value, "content": old_prompt})
+            prompt.append({"role": Role.ASSISTANT.value, "content": old_response})
+
+    query = []
+    if dataset_attr.prompt and example[dataset_attr.prompt]:
+        query.append(example[dataset_attr.prompt])
+
+    if dataset_attr.query and example[dataset_attr.query]:
+        query.append(example[dataset_attr.query])
+
+    prompt.append({"role": Role.USER.value, "content": "\n".join(query)})  # "prompt\nquery"
+
+    if dataset_attr.kto_tag and isinstance(example[dataset_attr.kto_tag], bool):  # kto example
+        response = [{"role": Role.ASSISTANT.value, "content": example[dataset_attr.response]}]
+        if example[dataset_attr.kto_tag]:
+            response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
+        else:
+            response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
+    elif (
+        dataset_attr.ranking
+        and isinstance(example[dataset_attr.chosen], str)
+        and isinstance(example[dataset_attr.rejected], str)
+    ):  # pairwise example
+        response = [
+            {"role": Role.ASSISTANT.value, "content": example[dataset_attr.chosen]},
+            {"role": Role.ASSISTANT.value, "content": example[dataset_attr.rejected]},
+        ]
+    elif dataset_attr.response and isinstance(example[dataset_attr.response], str):  # normal example
+        response = [{"role": Role.ASSISTANT.value, "content": example[dataset_attr.response]}]
+    else:  # unsupervised
+        response = []
+
     convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
-    for i in range(len(examples[dataset_attr.prompt])):
-        prompt = []
-        if dataset_attr.history and isinstance(examples[dataset_attr.history][i], list):
-            for old_prompt, old_response in examples[dataset_attr.history][i]:
-                prompt.append({"role": Role.USER.value, "content": old_prompt})
-                prompt.append({"role": Role.ASSISTANT.value, "content": old_response})
-
-        content = []
-        if dataset_attr.prompt and examples[dataset_attr.prompt][i]:
-            content.append(examples[dataset_attr.prompt][i])
-
-        if dataset_attr.query and examples[dataset_attr.query][i]:
-            content.append(examples[dataset_attr.query][i])
-
-        prompt.append({"role": Role.USER.value, "content": "\n".join(content)})  # "prompt\nquery"
-
-        if dataset_attr.kto_tag and isinstance(examples[dataset_attr.kto_tag][i], bool):  # kto example
-            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
-            if examples[dataset_attr.kto_tag][i]:
-                response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
-            else:
-                response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
-        elif (
-            dataset_attr.ranking
-            and isinstance(examples[dataset_attr.chosen][i], str)
-            and isinstance(examples[dataset_attr.rejected][i], str)
-        ):  # pairwise example
-            response = [
-                {"role": Role.ASSISTANT.value, "content": examples[dataset_attr.chosen][i]},
-                {"role": Role.ASSISTANT.value, "content": examples[dataset_attr.rejected][i]},
-            ]
-        elif dataset_attr.response and isinstance(examples[dataset_attr.response][i], str):  # normal example
-            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
-        else:  # unsupervised
-            response = []
-
-        outputs["prompt"].append(prompt)
-        outputs["response"].append(response)
-        outputs["system"].append(examples[dataset_attr.system][i] if dataset_attr.system else "")
-        outputs["tools"].append(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
-        outputs["images"].append(convert_images(examples[dataset_attr.images][i]) if dataset_attr.images else [])
-
-    return outputs
+    convert_videos = partial(_convert_videos, dataset_attr=dataset_attr, data_args=data_args)
+    output = {
+        "_prompt": prompt,
+        "_response": response,
+        "_system": example[dataset_attr.system] if dataset_attr.system else "",
+        "_tools": example[dataset_attr.tools] if dataset_attr.tools else "",
+        "_images": convert_images(example[dataset_attr.images]) if dataset_attr.images else None,
+        "_videos": convert_videos(example[dataset_attr.videos]) if dataset_attr.videos else None,
+    }
+    return output
 
 
 def convert_sharegpt(
-    examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr", data_args: "DataArguments"
-) -> Dict[str, List[Any]]:
+    example: Dict[str, Any],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+) -> Dict[str, Any]:
     r"""
     Converts sharegpt format dataset to the standard format.
     """
-    outputs = {"prompt": [], "response": [], "system": [], "tools": [], "images": []}
-    convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
     tag_mapping = {
         dataset_attr.user_tag: Role.USER.value,
         dataset_attr.assistant_tag: Role.ASSISTANT.value,
@@ -119,74 +146,79 @@ def convert_sharegpt(
     odd_tags = (dataset_attr.user_tag, dataset_attr.observation_tag)
     even_tags = (dataset_attr.assistant_tag, dataset_attr.function_tag)
     accept_tags = (odd_tags, even_tags)
-    for i, messages in enumerate(examples[dataset_attr.messages]):
-        if dataset_attr.system_tag and messages[0][dataset_attr.role_tag] == dataset_attr.system_tag:
-            system = messages[0][dataset_attr.content_tag]
-            messages = messages[1:]
-        else:
-            system = examples[dataset_attr.system][i] if dataset_attr.system else ""
+    messages = example[dataset_attr.messages]
+    if (
+        dataset_attr.system_tag
+        and len(messages) != 0
+        and messages[0][dataset_attr.role_tag] == dataset_attr.system_tag
+    ):
+        system = messages[0][dataset_attr.content_tag]
+        messages = messages[1:]
+    else:
+        system = example[dataset_attr.system] if dataset_attr.system else ""
 
-        if len(messages) == 0:
-            continue
-
-        aligned_messages = []
-        broken_data = False
-        for turn_idx, message in enumerate(messages):
-            if message[dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
-                logger.warning("Invalid role tag in {}.".format(messages))
-                broken_data = True
-
-            aligned_messages.append(
-                {"role": tag_mapping[message[dataset_attr.role_tag]], "content": message[dataset_attr.content_tag]}
-            )
-
-        if (not dataset_attr.ranking and len(aligned_messages) % 2 != 0) or (
-            dataset_attr.ranking and len(aligned_messages) % 2 == 0
-        ):
-            logger.warning("Invalid message count in {}.".format(messages))
+    aligned_messages = []
+    broken_data = False
+    for turn_idx, message in enumerate(messages):
+        if message[dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
+            logger.warning("Invalid role tag in {}.".format(messages))
             broken_data = True
 
-        if dataset_attr.kto_tag and isinstance(examples[dataset_attr.kto_tag][i], bool):  # kto example
-            prompt = aligned_messages[:-1]
-            response = aligned_messages[-1:]
-            if examples[dataset_attr.kto_tag][i]:
-                response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
-            else:
-                response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
-        elif (
-            dataset_attr.ranking
-            and isinstance(examples[dataset_attr.chosen][i], dict)
-            and isinstance(examples[dataset_attr.rejected][i], dict)
-        ):  # pairwise example
-            chosen = examples[dataset_attr.chosen][i]
-            rejected = examples[dataset_attr.rejected][i]
-            if (
-                chosen[dataset_attr.role_tag] not in accept_tags[-1]
-                or rejected[dataset_attr.role_tag] not in accept_tags[-1]
-            ):
-                logger.warning("Invalid role tag in {}.".format([chosen, rejected]))
-                broken_data = True
+        aligned_messages.append(
+            {"role": tag_mapping[message[dataset_attr.role_tag]], "content": message[dataset_attr.content_tag]}
+        )
 
-            prompt = aligned_messages
-            response = [
-                {"role": tag_mapping[chosen[dataset_attr.role_tag]], "content": chosen[dataset_attr.content_tag]},
-                {"role": tag_mapping[rejected[dataset_attr.role_tag]], "content": rejected[dataset_attr.content_tag]},
-            ]
-        else:  # normal example
-            prompt = aligned_messages[:-1]
-            response = aligned_messages[-1:]
+    if (not dataset_attr.ranking and len(aligned_messages) % 2 != 0) or (
+        dataset_attr.ranking and len(aligned_messages) % 2 == 0
+    ):
+        logger.warning("Invalid message count in {}.".format(messages))
+        broken_data = True
 
-        if broken_data:
-            logger.warning("Skipping this abnormal example.")
-            continue
+    if dataset_attr.kto_tag and isinstance(example[dataset_attr.kto_tag], bool):  # kto example
+        prompt = aligned_messages[:-1]
+        response = aligned_messages[-1:]
+        if example[dataset_attr.kto_tag]:
+            response = response + [{"role": Role.ASSISTANT.value, "content": ""}]
+        else:
+            response = [{"role": Role.ASSISTANT.value, "content": ""}] + response
+    elif (
+        dataset_attr.ranking
+        and isinstance(example[dataset_attr.chosen], dict)
+        and isinstance(example[dataset_attr.rejected], dict)
+    ):  # pairwise example
+        chosen = example[dataset_attr.chosen]
+        rejected = example[dataset_attr.rejected]
+        if (
+            chosen[dataset_attr.role_tag] not in accept_tags[-1]
+            or rejected[dataset_attr.role_tag] not in accept_tags[-1]
+        ):
+            logger.warning("Invalid role tag in {}.".format([chosen, rejected]))
+            broken_data = True
 
-        outputs["prompt"].append(prompt)
-        outputs["response"].append(response)
-        outputs["system"].append(system)
-        outputs["tools"].append(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
-        outputs["images"].append(convert_images(examples[dataset_attr.images][i]) if dataset_attr.images else [])
+        prompt = aligned_messages
+        response = [
+            {"role": tag_mapping[chosen[dataset_attr.role_tag]], "content": chosen[dataset_attr.content_tag]},
+            {"role": tag_mapping[rejected[dataset_attr.role_tag]], "content": rejected[dataset_attr.content_tag]},
+        ]
+    else:  # normal example
+        prompt = aligned_messages[:-1]
+        response = aligned_messages[-1:]
 
-    return outputs
+    if broken_data:
+        logger.warning("Skipping this abnormal example.")
+        prompt, response = [], []
+
+    convert_images = partial(_convert_images, dataset_attr=dataset_attr, data_args=data_args)
+    convert_videos = partial(_convert_videos, dataset_attr=dataset_attr, data_args=data_args)
+    output = {
+        "_prompt": prompt,
+        "_response": response,
+        "_system": system,
+        "_tools": example[dataset_attr.tools] if dataset_attr.tools else "",
+        "_images": convert_images(example[dataset_attr.images]) if dataset_attr.images else None,
+        "_videos": convert_videos(example[dataset_attr.videos]) if dataset_attr.videos else None,
+    }
+    return output
 
 
 def align_dataset(
@@ -197,11 +229,12 @@ def align_dataset(
 ) -> Union["Dataset", "IterableDataset"]:
     r"""
     Aligned dataset:
-        prompt: [{"role": "user", "content": "..."}] * (2T - 1)
-        response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
-        system: "..."
-        tools: "...",
-        images: [],
+        _prompt: [{"role": "user", "content": "..."}] * (2T - 1)
+        _response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
+        _system: "..."
+        _tools: "...",
+        _images: [],
+        _videos: [],
     """
     if dataset_attr.formatting == "alpaca":
         convert_func = partial(convert_alpaca, dataset_attr=dataset_attr, data_args=data_args)
@@ -209,19 +242,6 @@ 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 +252,7 @@ def align_dataset(
 
     return dataset.map(
         convert_func,
-        batched=True,
+        batched=False,
         remove_columns=column_names,
-        features=features,
         **kwargs,
     )

src/llamafactory/data/collator.py

@@ -1,4 +1,7 @@
-# Copyright 2024 the LlamaFactory team.
+# Copyright 2024 OpenAccess AI Collective and the LlamaFactory team.
+#
+# This code is inspired by the OpenAccess AI Collective's axolotl library.
+# https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/monkeypatch/utils.py
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -13,19 +16,117 @@
 # limitations under the License.
 
 from dataclasses import dataclass
-from typing import Any, Dict, Sequence
+from typing import TYPE_CHECKING, Any, Dict, Literal, Optional, Sequence
 
 import torch
 from transformers import DataCollatorForSeq2Seq
 
 
+if TYPE_CHECKING:
+    from transformers import ProcessorMixin
+
+    from .template import Template
+
+
+def prepare_4d_attention_mask(attention_mask_with_indices: "torch.Tensor", dtype: "torch.dtype") -> "torch.Tensor":
+    r"""
+    Expands the attention mask with indices from (batch_size, seq_len) to (batch_size, 1, seq_len, seq_len),
+    while handles packed sequences and transforms the mask to lower triangular form to prevent future peeking.
+
+    e.g.
+    ```python
+    # input
+    [[1, 1, 2, 2, 2, 0]]
+    # output
+    [
+        [
+            [
+                [o, x, x, x, x, x],
+                [o, o, x, x, x, x],
+                [x, x, o, x, x, x],
+                [x, x, o, o, x, x],
+                [x, x, o, o, o, x],
+                [x, x, x, x, x, x],
+            ]
+        ]
+    ]
+    ```
+    where `o` equals to `0.0`, `x` equals to `min_dtype`.
+    """
+    bsz, seq_len = attention_mask_with_indices.size()
+    min_dtype = torch.finfo(dtype).min
+    expanded_mask = attention_mask_with_indices[:, None, None, :].expand(bsz, 1, seq_len, seq_len)
+    # Create a binary mask from the original mask where zeros remain zeros and all other values are set to one
+    padding_mask = torch.where(expanded_mask != 0, 1, 0)
+    # Create a block-diagonal mask.
+    attention_mask_4d = torch.eq(expanded_mask, expanded_mask.transpose(-1, -2)).int() * padding_mask
+    # Use the lower triangular mask to zero out the upper triangular part
+    attention_mask_4d *= torch.tril(torch.ones((seq_len, seq_len), dtype=torch.long))
+    # Invert the attention mask.
+    attention_mask_4d = torch.where(attention_mask_4d != 0, torch.tensor(0, dtype=dtype), min_dtype)
+    return attention_mask_4d
+
+
 @dataclass
-class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
+class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
+    r"""
+    Data collator that supports VLMs.
+
+    Features should contain input_ids, attention_mask, labels and images.
+    """
+
+    template: Optional["Template"] = None
+    processor: Optional["ProcessorMixin"] = None
+
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
+        batch_images, batch_videos, batch_imglens, batch_vidlens, batch_seqlens = [], [], [], [], []
+        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_seqlens.append(len(feature["input_ids"]))
+
+        mm_inputs = self.template.mm_plugin.get_mm_inputs(
+            batch_images, batch_videos, batch_imglens, batch_vidlens, batch_seqlens, 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)
+        return features
+
+
+@dataclass
+class SFTDataCollatorWith4DAttentionMask(MultiModalDataCollatorForSeq2Seq):
+    r"""
+    Data collator for 4d attention mask.
+    """
+
+    block_diag_attn: bool = False
+    attn_implementation: Literal["eager", "sdpa", "flash_attention_2"] = "eager"
+    compute_dtype: "torch.dtype" = torch.float32
+
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
+        features = super().__call__(features)
+        if self.block_diag_attn and self.attn_implementation != "flash_attention_2":
+            features["attention_mask"] = prepare_4d_attention_mask(features["attention_mask"], self.compute_dtype)
+
+        return features
+
+
+@dataclass
+class PairwiseDataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
     r"""
     Data collator for pairwise data.
     """
 
-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
         r"""
         Pads batched data to the longest sequence in the batch.
 
@@ -39,25 +140,21 @@ class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
                     "input_ids": feature["{}_input_ids".format(key)],
                     "attention_mask": feature["{}_attention_mask".format(key)],
                     "labels": feature["{}_labels".format(key)],
+                    "images": feature["images"],
+                    "videos": feature["videos"],
                 }
-                if "pixel_values" in feature:
-                    target_feature["pixel_values"] = feature["pixel_values"]
-
-                if "{}_token_type_ids".format(key) in feature:
-                    target_feature["token_type_ids"] = feature["{}_token_type_ids".format(key)]
-
                 concatenated_features.append(target_feature)
 
         return super().__call__(concatenated_features)
 
 
 @dataclass
-class KTODataCollatorWithPadding(DataCollatorForSeq2Seq):
+class KTODataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
     r"""
     Data collator for KTO data.
     """
 
-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
         target_features = []
         kl_features = []
         kto_tags = []
@@ -66,19 +163,16 @@ 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"])
@@ -88,7 +182,7 @@ 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

@@ -13,16 +13,15 @@
 # limitations under the License.
 
 from enum import Enum, unique
-from typing import TYPE_CHECKING, Dict, List, Tuple, Union
+from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Set, TypedDict, Union
 
-from datasets import concatenate_datasets, interleave_datasets
+from datasets import DatasetDict, concatenate_datasets, interleave_datasets
 
 from ..extras.logging import get_logger
 
 
 if TYPE_CHECKING:
     from datasets import Dataset, IterableDataset
-    from transformers import Seq2SeqTrainingArguments
 
     from ..hparams import DataArguments
 
@@ -30,6 +29,9 @@ if TYPE_CHECKING:
 logger = get_logger(__name__)
 
 
+SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
+
+
 @unique
 class Role(str, Enum):
     USER = "user"
@@ -39,54 +41,52 @@ class Role(str, Enum):
     OBSERVATION = "observation"
 
 
-def infer_max_len(source_len: int, target_len: int, max_len: int, reserved_label_len: int) -> Tuple[int, int]:
-    max_target_len = int(max_len * (target_len / (source_len + target_len)))
-    max_target_len = max(max_target_len, reserved_label_len)
-    max_source_len = max_len - min(max_target_len, target_len)
-    return max_source_len, max_target_len
+class DatasetModule(TypedDict):
+    train_dataset: Optional[Union["Dataset", "IterableDataset"]]
+    eval_dataset: Optional[Union["Dataset", "IterableDataset"]]
 
 
 def merge_dataset(
-    all_datasets: List[Union["Dataset", "IterableDataset"]],
-    data_args: "DataArguments",
-    training_args: "Seq2SeqTrainingArguments",
+    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.")
+
         return concatenate_datasets(all_datasets)
     elif data_args.mix_strategy.startswith("interleave"):
         if not data_args.streaming:
             logger.warning("We recommend using `mix_strategy=concat` in non-streaming mode.")
+
         return interleave_datasets(
             datasets=all_datasets,
             probabilities=data_args.interleave_probs,
-            seed=training_args.seed,
+            seed=seed,
             stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
         )
     else:
-        raise ValueError("Unknown mixing strategy.")
+        raise ValueError("Unknown mixing strategy: {}.".format(data_args.mix_strategy))
 
 
 def split_dataset(
-    dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", training_args: "Seq2SeqTrainingArguments"
-) -> Dict[str, "Dataset"]:
-    if training_args.do_train:
-        if data_args.val_size > 1e-6:  # Split the dataset
-            if data_args.streaming:
-                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
-                val_set = dataset.take(int(data_args.val_size))
-                train_set = dataset.skip(int(data_args.val_size))
-                return {"train_dataset": train_set, "eval_dataset": val_set}
-            else:
-                val_size = int(data_args.val_size) if data_args.val_size > 1 else data_args.val_size
-                dataset = dataset.train_test_split(test_size=val_size, seed=training_args.seed)
-                return {"train_dataset": dataset["train"], "eval_dataset": dataset["test"]}
-        else:
-            if data_args.streaming:
-                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
-            return {"train_dataset": dataset}
-    else:  # do_eval or do_predict
-        return {"eval_dataset": dataset}
+    dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", seed: int
+) -> "DatasetDict":
+    r"""
+    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)
+        val_set = dataset.take(int(data_args.val_size))
+        train_set = dataset.skip(int(data_args.val_size))
+        return DatasetDict({"train": train_set, "validation": val_set})
+    else:
+        val_size = int(data_args.val_size) if data_args.val_size > 1 else data_args.val_size
+        dataset = dataset.train_test_split(test_size=val_size, seed=seed)
+        return DatasetDict({"train": dataset["train"], "validation": dataset["test"]})

src/llamafactory/data/formatter.py

@@ -16,108 +16,36 @@ import json
 import re
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Literal, Optional, Sequence, Set, Tuple, Union
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+from typing_extensions import override
+
+from .data_utils import SLOTS
+from .tool_utils import get_tool_utils
 
 
-SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
-
-
-DEFAULT_TOOL_PROMPT = (
-    "You have access to the following tools:\n{tool_text}"
-    "Use the following format if using a tool:\n"
-    "```\n"
-    "Action: tool name (one of [{tool_names}]).\n"
-    "Action Input: the input to the tool, in a JSON format representing the kwargs "
-    """(e.g. ```{{"input": "hello world", "num_beams": 5}}```).\n"""
-    "```\n"
-)
-
-
-GLM4_TOOL_PROMPT = (
-    "你是一个名为 GLM-4 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，"
-    "你的任务是针对用户的问题和要求提供适当的答复和支持。{tool_text}"
-)
-
-
-def default_tool_formatter(tools: List[Dict[str, Any]]) -> str:
-    tool_text = ""
-    tool_names = []
-    for tool in tools:
-        param_text = ""
-        for name, param in tool["parameters"]["properties"].items():
-            required = ", required" if name in tool["parameters"].get("required", []) else ""
-            enum = ", should be one of [{}]".format(", ".join(param["enum"])) if param.get("enum", None) else ""
-            items = (
-                ", where each item should be {}".format(param["items"].get("type", "")) if param.get("items") else ""
-            )
-            param_text += "  - {name} ({type}{required}): {desc}{enum}{items}\n".format(
-                name=name,
-                type=param.get("type", ""),
-                required=required,
-                desc=param.get("description", ""),
-                enum=enum,
-                items=items,
-            )
-
-        tool_text += "> Tool Name: {name}\nTool Description: {desc}\nTool Args:\n{args}\n".format(
-            name=tool["name"], desc=tool.get("description", ""), args=param_text
-        )
-        tool_names.append(tool["name"])
-
-    return DEFAULT_TOOL_PROMPT.format(tool_text=tool_text, tool_names=", ".join(tool_names))
-
-
-def default_tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
-    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:
-        return content
-
-    results = []
-    for match in action_match:
-        tool_name = match[0].strip()
-        tool_input = match[1].strip().strip('"').strip("```")
-        try:
-            arguments = json.loads(tool_input)
-            results.append((tool_name, json.dumps(arguments, ensure_ascii=False)))
-        except json.JSONDecodeError:
-            return content
-
-    return results
-
-
-def glm4_tool_formatter(tools: List[Dict[str, Any]]) -> str:
-    tool_text = ""
-    for tool in tools:
-        tool_text += "\n\n## {name}\n\n{body}\n在调用上述函数时，请使用 Json 格式表示调用的参数。".format(
-            name=tool["name"], body=json.dumps(tool, indent=4, ensure_ascii=False)
-        )
-
-    return GLM4_TOOL_PROMPT.format(tool_text=tool_text)
-
-
-def glm4_tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
-    if "\n" not in content:
-        return content
-
-    tool_name, tool_input = content.split("\n", maxsplit=1)
-    try:
-        arguments = json.loads(tool_input)
-    except json.JSONDecodeError:
-        return content
-
-    return [(tool_name, json.dumps(arguments, ensure_ascii=False))]
+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
 
 
@@ -132,6 +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
 
@@ -147,6 +76,7 @@ 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:
@@ -168,16 +98,9 @@ class StringFormatter(Formatter):
 @dataclass
 class FunctionFormatter(Formatter):
     def __post_init__(self):
-        has_name, has_args = False, False
-        for slot in filter(lambda s: isinstance(s, str), self.slots):
-            if "{{name}}" in slot:
-                has_name = True
-            if "{{arguments}}" in slot:
-                has_args = True
-
-        if not has_name or not has_args:
-            raise ValueError("Name and arguments placeholders are required in the function formatter.")
+        self.slots = get_tool_utils(self.tool_format).get_function_slots() + self.slots
 
+    @override
     def apply(self, **kwargs) -> SLOTS:
         content = kwargs.pop("content")
         functions: List[Tuple[str, str]] = []
@@ -209,22 +132,17 @@ class FunctionFormatter(Formatter):
 @dataclass
 class ToolFormatter(Formatter):
     def __post_init__(self):
-        if self.tool_format == "default":
-            self._tool_formatter = default_tool_formatter
-            self._tool_extractor = default_tool_extractor
-        elif self.tool_format == "glm4":
-            self._tool_formatter = glm4_tool_formatter
-            self._tool_extractor = glm4_tool_extractor
-        else:
-            raise ValueError("Tool format was not found.")
+        self.tool_utils = get_tool_utils(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 [""]
 
-    def extract(self, content: str) -> Union[str, List[Tuple[str, str]]]:
-        return self._tool_extractor(content)
+    @override
+    def extract(self, content: str) -> Union[str, List["FunctionCall"]]:
+        return self.tool_utils.tool_extractor(content)

src/llamafactory/data/loader.py

@@ -12,22 +12,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import inspect
 import os
 import sys
-from typing import TYPE_CHECKING, Literal, Optional, Union
+from typing import TYPE_CHECKING, Dict, Literal, Optional, Sequence, Union
 
 import numpy as np
-from datasets import load_dataset, load_from_disk
+from datasets import DatasetDict, load_dataset, load_from_disk
+from transformers.utils.versions import require_version
 
 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
+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:
@@ -35,18 +34,23 @@ if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer, ProcessorMixin, Seq2SeqTrainingArguments
 
     from ..hparams import DataArguments, ModelArguments
+    from .data_utils import DatasetModule
     from .parser import DatasetAttr
+    from .template import Template
 
 
 logger = get_logger(__name__)
 
 
-def load_single_dataset(
+def _load_single_dataset(
     dataset_attr: "DatasetAttr",
     model_args: "ModelArguments",
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
 ) -> Union["Dataset", "IterableDataset"]:
+    r"""
+    Loads a single dataset and aligns it to the standard format.
+    """
     logger.info("Loading dataset {}...".format(dataset_attr))
     data_path, data_name, data_dir, data_files = None, None, None, None
     if dataset_attr.load_from in ["hf_hub", "ms_hub"]:
@@ -81,41 +85,34 @@ def load_single_dataset(
         raise NotImplementedError("Unknown load type: {}.".format(dataset_attr.load_from))
 
     if dataset_attr.load_from == "ms_hub":
-        try:
-            from modelscope import MsDataset
-            from modelscope.utils.config_ds import MS_DATASETS_CACHE
+        require_version("modelscope>=1.11.0", "To fix: pip install modelscope>=1.11.0")
+        from modelscope import MsDataset
+        from modelscope.utils.config_ds import MS_DATASETS_CACHE
 
-            cache_dir = model_args.cache_dir or MS_DATASETS_CACHE
-            dataset = MsDataset.load(
-                dataset_name=data_path,
-                subset_name=data_name,
-                data_dir=data_dir,
-                data_files=data_files,
-                split=data_args.split,
-                cache_dir=cache_dir,
-                token=model_args.ms_hub_token,
-                use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
-            )
-            if isinstance(dataset, MsDataset):
-                dataset = dataset.to_hf_dataset()
-        except ImportError:
-            raise ImportError("Please install modelscope via `pip install modelscope -U`")
+        cache_dir = model_args.cache_dir or MS_DATASETS_CACHE
+        dataset = MsDataset.load(
+            dataset_name=data_path,
+            subset_name=data_name,
+            data_dir=data_dir,
+            data_files=data_files,
+            split=dataset_attr.split,
+            cache_dir=cache_dir,
+            token=model_args.ms_hub_token,
+            use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
+        )
+        if isinstance(dataset, MsDataset):
+            dataset = dataset.to_hf_dataset()
     else:
-        if "trust_remote_code" in inspect.signature(load_dataset).parameters:  # for datasets==2.16.0
-            kwargs = {"trust_remote_code": True}
-        else:
-            kwargs = {}
-
         dataset = load_dataset(
             path=data_path,
             name=data_name,
             data_dir=data_dir,
             data_files=data_files,
-            split=data_args.split,
+            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")),
-            **kwargs,
+            trust_remote_code=True,
         )
 
     if data_args.streaming and (dataset_attr.load_from == "file"):  # faster than specifying streaming=True
@@ -123,7 +120,7 @@ def load_single_dataset(
 
     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)
@@ -140,71 +137,156 @@ def load_single_dataset(
     return align_dataset(dataset, dataset_attr, data_args, training_args)
 
 
+def _get_merged_dataset(
+    dataset_names: Optional[Sequence[str]],
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    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
+
+    datasets = []
+    for dataset_attr in get_dataset_list(dataset_names, data_args.dataset_dir):
+        if (stage == "rm" and dataset_attr.ranking is False) or (stage != "rm" and dataset_attr.ranking is True):
+            raise ValueError("The dataset is not applicable in the current training stage.")
+
+        datasets.append(_load_single_dataset(dataset_attr, model_args, data_args, training_args))
+
+    return merge_dataset(datasets, data_args, seed=training_args.seed)
+
+
+def _get_preprocessed_dataset(
+    dataset: Optional[Union["Dataset", "IterableDataset"]],
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    stage: Literal["pt", "sft", "rm", "ppo", "kto"],
+    template: "Template",
+    tokenizer: "PreTrainedTokenizer",
+    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
+
+    preprocess_func, print_function = get_preprocess_and_print_func(
+        data_args, stage, template, tokenizer, processor, do_generate=(training_args.predict_with_generate and is_eval)
+    )
+    column_names = list(next(iter(dataset)).keys())
+    kwargs = {}
+    if not data_args.streaming:
+        kwargs = dict(
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=(not data_args.overwrite_cache) or (training_args.local_process_index != 0),
+            desc="Running tokenizer on dataset",
+        )
+
+    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:
+            print("eval example:" if is_eval else "training example:")
+            print_function(next(iter(dataset)))
+        except StopIteration:
+            if stage == "pt":
+                raise RuntimeError("Cannot find sufficient samples, consider increasing dataset size.")
+            else:
+                raise RuntimeError("Cannot find valid samples, check `data/README.md` for the data format.")
+
+    return dataset
+
+
 def get_dataset(
+    template: "Template",
     model_args: "ModelArguments",
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
     stage: Literal["pt", "sft", "rm", "ppo", "kto"],
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"] = None,
-) -> Union["Dataset", "IterableDataset"]:
-    template = get_template_and_fix_tokenizer(tokenizer, data_args.template)
-    if data_args.train_on_prompt and template.efficient_eos:
-        raise ValueError("Current template does not support `train_on_prompt`.")
-
+) -> "DatasetModule":
+    r"""
+    Gets the train dataset and optionally gets the evaluation dataset.
+    """
     # 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.")
-            dataset = load_from_disk(data_args.tokenized_path)
+            dataset_dict: "DatasetDict" = load_from_disk(data_args.tokenized_path)
             logger.info("Loaded tokenized dataset from {}.".format(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"]
+
             if data_args.streaming:
-                dataset = dataset.to_iterable_dataset()
-            return dataset
+                dataset_module = {k: v.to_iterable_dataset() for k, v in dataset_module.items()}
+
+            return dataset_module
 
         if data_args.streaming:
             raise ValueError("Turn off `streaming` when saving dataset to disk.")
 
+    # Load and preprocess dataset
     with training_args.main_process_first(desc="load dataset"):
-        all_datasets = []
-        for dataset_attr in get_dataset_list(data_args):
-            if (stage == "rm" and dataset_attr.ranking is False) or (stage != "rm" and dataset_attr.ranking is True):
-                raise ValueError("The dataset is not applicable in the current training stage.")
-
-            all_datasets.append(load_single_dataset(dataset_attr, model_args, data_args, training_args))
-
-        dataset = merge_dataset(all_datasets, data_args, training_args)
+        dataset = _get_merged_dataset(data_args.dataset, model_args, data_args, training_args, stage)
+        eval_dataset = _get_merged_dataset(data_args.eval_dataset, model_args, data_args, training_args, stage)
 
     with training_args.main_process_first(desc="pre-process dataset"):
-        preprocess_func, print_function = get_preprocess_and_print_func(
-            data_args, training_args, stage, template, tokenizer, processor
+        dataset = _get_preprocessed_dataset(
+            dataset, data_args, training_args, stage, template, tokenizer, processor, is_eval=False
+        )
+        eval_dataset = _get_preprocessed_dataset(
+            eval_dataset, data_args, training_args, stage, template, tokenizer, processor, is_eval=True
         )
-        column_names = list(next(iter(dataset)).keys())
-        kwargs = {}
-        if not data_args.streaming:
-            kwargs = dict(
-                num_proc=data_args.preprocessing_num_workers,
-                load_from_cache_file=(not data_args.overwrite_cache) or (training_args.local_process_index != 0),
-                desc="Running tokenizer on dataset",
-            )
 
-        dataset = dataset.map(preprocess_func, batched=True, remove_columns=column_names, **kwargs)
+        if data_args.val_size > 1e-6:
+            dataset_dict = split_dataset(dataset, data_args, seed=training_args.seed)
+        else:
+            dataset_dict = {}
+            if dataset is not None:
+                if data_args.streaming:
+                    dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
+
+                dataset_dict["train"] = dataset
+
+            if eval_dataset is not None:
+                if data_args.streaming:
+                    eval_dataset = eval_dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
+
+                dataset_dict["validation"] = eval_dataset
+
+            dataset_dict = DatasetDict(dataset_dict)
 
         if data_args.tokenized_path is not None:
             if training_args.should_save:
-                dataset.save_to_disk(data_args.tokenized_path)
+                dataset_dict.save_to_disk(data_args.tokenized_path)
                 logger.info("Tokenized dataset saved at {}.".format(data_args.tokenized_path))
                 logger.info("Please restart the training with `tokenized_path: {}`.".format(data_args.tokenized_path))
 
             sys.exit(0)
 
-        if training_args.should_log:
-            try:
-                print_function(next(iter(dataset)))
-            except StopIteration:
-                if stage == "pt":
-                    raise RuntimeError("Cannot find sufficient samples, consider increasing dataset size.")
-                else:
-                    raise RuntimeError("Cannot find valid samples, check `data/README.md` for the data format.")
+        dataset_module = {}
+        if "train" in dataset_dict:
+            dataset_module["train_dataset"] = dataset_dict["train"]
 
-        return dataset
+        if "validation" in dataset_dict:
+            dataset_module["eval_dataset"] = dataset_dict["validation"]
+
+        return dataset_module

src/llamafactory/data/mm_plugin.py

@@ -0,0 +1,400 @@
+from copy import deepcopy
+from io import BytesIO
+from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Tuple, TypedDict, Union
+
+import numpy as np
+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
+
+
+if is_pillow_available():
+    from PIL import Image
+    from PIL.Image import Image as ImageObject
+
+
+if is_pyav_available():
+    import av
+
+
+if TYPE_CHECKING:
+    import torch
+    from transformers import PreTrainedTokenizer, ProcessorMixin
+    from transformers.image_processing_utils import BaseImageProcessor
+
+    class EncodedImage(TypedDict):
+        path: Optional[str]
+        bytes: Optional[bytes]
+
+    ImageInput = Union[str, EncodedImage, ImageObject]
+    VideoInput = str
+
+
+def _regularize_images(
+    images: Sequence["ImageInput"],
+    processor: "ProcessorMixin",
+    max_resolution: Optional[int] = None,
+) -> List["ImageObject"]:
+    r"""
+    Regularizes images to avoid error. Including reading, resizing and converting.
+    """
+    if max_resolution is None:
+        max_resolution: int = getattr(processor, "image_resolution", 512)
+
+    results = []
+    for image in images:
+        if isinstance(image, str):
+            image = Image.open(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("Expect input is a list of Images, but got {}.".format(type(image)))
+
+        if max(image.width, image.height) > max_resolution:
+            factor = max_resolution / max(image.width, image.height)
+            image = image.resize((int(image.width * factor), int(image.height * factor)), resample=Image.NEAREST)
+
+        if image.mode != "RGB":
+            image = image.convert("RGB")
+
+        results.append(image)
+
+    return results
+
+
+def _regularize_videos(
+    videos: Sequence["VideoInput"],
+    processor: "ProcessorMixin",
+) -> List[List["ImageObject"]]:
+    r"""
+    Regularizes videos to avoid error. Including reading, resizing and converting.
+    """
+    video_resolution: int = getattr(processor, "video_resolution", 128)
+    video_fps: float = getattr(processor, "video_fps", 1.0)
+    video_maxlen: int = getattr(processor, "video_maxlen", 64)
+    video_factor: int = getattr(processor, "video_factor", 1)
+    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 = float(video_stream.duration * video_stream.time_base) * video_fps
+        sample_frames = min(video_maxlen, sample_frames)  # reduce length <= maxlen
+        sample_frames = round(sample_frames / video_factor) * video_factor  # for qwen2_vl
+        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 = _regularize_images(frames, processor, video_resolution)
+        results.append(frames)
+
+    return results
+
+
+def _get_mm_inputs(
+    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")
+    input_dict = {"images": None}  # default key
+    if len(images) != 0:
+        images = _regularize_images(images, processor)
+        input_dict["images"] = images
+
+    if len(videos) != 0:
+        videos = _regularize_videos(videos, processor)
+        input_dict["videos"] = videos
+
+    if input_dict.get("images", None) is not None or input_dict.get("videos", None) is not None:
+        return image_processor(**input_dict, return_tensors="pt")
+    else:
+        return {}
+
+
+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:
+        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 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],
+        seqlens: Sequence[int],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        r"""
+        Builds batched multimodal inputs for VLMs.
+        """
+        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}}", 1)
+
+            message["content"] = content.replace("{{image}}", self.image_token * image_seqlen)
+
+        if len(images) != num_image_tokens:
+            raise ValueError("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        seqlens: Sequence[int],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        return _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("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+
+        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],
+        seqlens: Sequence[int],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        mm_inputs = _get_mm_inputs(images, videos, processor)
+        mm_inputs["token_type_ids"] = _get_paligemma_token_type_ids(imglens, seqlens, processor)
+        return mm_inputs
+
+
+class Qwen2vlPlugin(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)
+        image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
+        merge_length: int = getattr(image_processor, "merge_size") ** 2
+        mm_inputs = _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("`len(images)` is less than the number of {} tokens.".format(IMAGE_PLACEHOLDER))
+
+                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("`len(videos)` is less than the number of {} tokens.".format(VIDEO_PLACEHOLDER))
+
+                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("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+
+        if len(videos) != num_video_tokens:
+            raise ValueError("The number of videos does not match the number of {} tokens".format(VIDEO_PLACEHOLDER))
+
+        return messages
+
+    @override
+    def get_mm_inputs(
+        self,
+        images: Sequence["ImageInput"],
+        videos: Sequence["VideoInput"],
+        imglens: Sequence[int],
+        vidlens: Sequence[int],
+        seqlens: Sequence[int],
+        processor: Optional["ProcessorMixin"],
+    ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
+        self._validate_input(images, videos)
+        return _get_mm_inputs(images, videos, processor)
+
+
+PLUGINS = {
+    "base": BasePlugin,
+    "llava": LlavaPlugin,
+    "paligemma": PaliGemmaPlugin,
+    "qwen2_vl": Qwen2vlPlugin,
+}
+
+
+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("Multimodal plugin `{}` not found.".format(name))
+
+    return plugin_class(image_token, video_token)

src/llamafactory/data/parser.py

@@ -15,47 +15,47 @@
 import json
 import os
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional
+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
 
 
-if TYPE_CHECKING:
-    from ..hparams import DataArguments
-
-
 @dataclass
 class DatasetAttr:
     r"""
     Dataset attributes.
     """
 
-    """ basic configs """
+    # basic configs
     load_from: Literal["hf_hub", "ms_hub", "script", "file"]
     dataset_name: str
     formatting: Literal["alpaca", "sharegpt"] = "alpaca"
     ranking: bool = False
-    """ extra configs """
+    # extra configs
     subset: Optional[str] = None
+    split: str = "train"
     folder: Optional[str] = None
     num_samples: Optional[int] = None
-    """ common columns """
+    # common columns
     system: Optional[str] = None
     tools: Optional[str] = None
     images: Optional[str] = None
-    """ rlhf columns """
+    videos: Optional[str] = None
+    # rlhf columns
     chosen: Optional[str] = None
     rejected: Optional[str] = None
     kto_tag: Optional[str] = None
-    """ alpaca columns """
+    # alpaca columns
     prompt: Optional[str] = "instruction"
     query: Optional[str] = "input"
     response: Optional[str] = "output"
     history: Optional[str] = None
-    """ sharegpt columns """
+    # sharegpt columns
     messages: Optional[str] = "conversations"
-    """ sharegpt tags """
+    # sharegpt tags
     role_tag: Optional[str] = "from"
     content_tag: Optional[str] = "value"
     user_tag: Optional[str] = "human"
@@ -71,31 +71,33 @@ class DatasetAttr:
         setattr(self, key, obj.get(key, default))
 
 
-def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
-    if data_args.dataset is not None:
-        dataset_names = [ds.strip() for ds in data_args.dataset.split(",")]
-    else:
+def get_dataset_list(dataset_names: Optional[Sequence[str]], dataset_dir: str) -> List["DatasetAttr"]:
+    r"""
+    Gets the attributes of the datasets.
+    """
+    if dataset_names is None:
         dataset_names = []
 
-    if data_args.dataset_dir == "ONLINE":
+    if dataset_dir == "ONLINE":
         dataset_info = None
     else:
+        if dataset_dir.startswith("REMOTE:"):
+            config_path = cached_file(path_or_repo_id=dataset_dir[7:], filename=DATA_CONFIG, repo_type="dataset")
+        else:
+            config_path = os.path.join(dataset_dir, DATA_CONFIG)
+
         try:
-            with open(os.path.join(data_args.dataset_dir, DATA_CONFIG), "r") as f:
+            with open(config_path, "r") as f:
                 dataset_info = json.load(f)
         except Exception as err:
             if len(dataset_names) != 0:
-                raise ValueError(
-                    "Cannot open {} due to {}.".format(os.path.join(data_args.dataset_dir, DATA_CONFIG), str(err))
-                )
+                raise ValueError("Cannot open {} due to {}.".format(config_path, str(err)))
+
             dataset_info = None
 
-    if data_args.interleave_probs is not None:
-        data_args.interleave_probs = [float(prob.strip()) for prob in data_args.interleave_probs.split(",")]
-
-    dataset_list: List[DatasetAttr] = []
+    dataset_list: List["DatasetAttr"] = []
     for name in dataset_names:
-        if dataset_info is None:
+        if dataset_info is None:  # dataset_dir is ONLINE
             load_from = "ms_hub" if use_modelscope() else "hf_hub"
             dataset_attr = DatasetAttr(load_from, dataset_name=name)
             dataset_list.append(dataset_attr)
@@ -120,11 +122,12 @@ def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
         dataset_attr.set_attr("formatting", dataset_info[name], default="alpaca")
         dataset_attr.set_attr("ranking", dataset_info[name], default=False)
         dataset_attr.set_attr("subset", dataset_info[name])
+        dataset_attr.set_attr("split", dataset_info[name], default="train")
         dataset_attr.set_attr("folder", dataset_info[name])
         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

@@ -27,7 +27,7 @@ from .processors.unsupervised import preprocess_unsupervised_dataset, print_unsu
 
 
 if TYPE_CHECKING:
-    from transformers import PreTrainedTokenizer, ProcessorMixin, Seq2SeqTrainingArguments
+    from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ..hparams import DataArguments
     from .template import Template
@@ -35,11 +35,11 @@ if TYPE_CHECKING:
 
 def get_preprocess_and_print_func(
     data_args: "DataArguments",
-    training_args: "Seq2SeqTrainingArguments",
     stage: Literal["pt", "sft", "rm", "ppo", "kto"],
     template: "Template",
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
+    do_generate: bool = False,
 ) -> Tuple[Callable, Callable]:
     if stage == "pt":
         preprocess_func = partial(
@@ -48,12 +48,26 @@ def get_preprocess_and_print_func(
             data_args=data_args,
         )
         print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
-    elif stage == "sft" and not training_args.predict_with_generate:
+    elif stage == "sft" and not do_generate:
         if data_args.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):
+                    return TypedSequence.__init__(
+                        self,
+                        data,
+                        type=kwargs.pop("type", None),
+                        try_type=kwargs.pop("try_type", None),
+                        optimized_int_type=kwargs.pop("optimized_int_type", None),
+                    )
+
+                OptimizedTypedSequence.__init__ = __init__
             preprocess_func = partial(
                 preprocess_packed_supervised_dataset,
                 template=template,
                 tokenizer=tokenizer,
+                processor=processor,
                 data_args=data_args,
             )
         else:

src/llamafactory/data/processors/feedback.py

@@ -12,17 +12,19 @@
 # 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.constants import IGNORE_INDEX
 from ...extras.logging import get_logger
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values
+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
 
 
@@ -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]]
@@ -55,26 +56,29 @@ def _encode_feedback_example(
     else:
         kl_messages = prompt + [kl_response[1]]
 
-    prompt_ids, response_ids = template.encode_oneturn(
-        tokenizer, messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
-    _, kl_response_ids = template.encode_oneturn(
-        tokenizer, kl_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
+    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)
 
     if template.efficient_eos:
         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
-        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
-        prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
+    prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, videos, tokenizer, processor)
+    kl_prompt_ids, _ = template.mm_plugin.process_token_ids(kl_prompt_ids, None, images, videos, tokenizer, processor)
+
+    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), cutoff_len)
+    kl_prompt_ids = kl_prompt_ids[:kl_source_len]
+    kl_response_ids = kl_response_ids[:kl_target_len]
 
     input_ids = prompt_ids + response_ids
-    labels = [IGNORE_INDEX] * len(prompt_ids) + response_ids
-    kl_input_ids = prompt_ids + kl_response_ids
-    kl_labels = [IGNORE_INDEX] * len(prompt_ids) + kl_response_ids
-
+    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
 
 
@@ -84,39 +88,27 @@ 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]
-    model_inputs = {
-        "input_ids": [],
-        "attention_mask": [],
-        "labels": [],
-        "kl_input_ids": [],
-        "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]))
+    kl_response = examples["_response"][::-1]
+    model_inputs = defaultdict(list)
+    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],
-            response=examples["response"][i],
+            prompt=examples["_prompt"][i],
+            response=examples["_response"][i],
             kl_response=kl_response[i],
-            system=examples["system"][i],
-            tools=examples["tools"][i],
+            system=examples["_system"][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))
@@ -125,11 +117,8 @@ 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["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
-            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))
+        model_inputs["images"].append(examples["_images"][i])
+        model_inputs["videos"].append(examples["_videos"][i])
 
     desirable_num = sum([1 for tag in model_inputs["kto_tags"] if tag])
     undesirable_num = len(model_inputs["kto_tags"]) - desirable_num

src/llamafactory/data/processors/pairwise.py

@@ -12,17 +12,19 @@
 # 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.constants import IGNORE_INDEX
 from ...extras.logging import get_logger
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values
+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
 
 
@@ -34,36 +36,33 @@ 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
-        prompt[0]["content"] = template.image_token + prompt[0]["content"]
-
-    chosen_messages = prompt + [response[0]]
-    rejected_messages = prompt + [response[1]]
-    prompt_ids, chosen_ids = template.encode_oneturn(
-        tokenizer, chosen_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
-    _, rejected_ids = template.encode_oneturn(
-        tokenizer, rejected_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
+    chosen_messages = template.mm_plugin.process_messages(prompt + [response[0]], images, videos, processor)
+    rejected_messages = template.mm_plugin.process_messages(prompt + [response[1]], images, videos, processor)
+    prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, chosen_messages, system, tools)
+    _, rejected_ids = template.encode_oneturn(tokenizer, rejected_messages, system, tools)
 
     if template.efficient_eos:
         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
-        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
-        prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
+    prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, images, videos, tokenizer, processor)
+    # consider the response is more important
+    source_len, target_len = infer_seqlen(len(prompt_ids), max(len(chosen_ids), len(rejected_ids)), cutoff_len)
+    prompt_ids = prompt_ids[:source_len]
+    chosen_ids = chosen_ids[:target_len]
+    rejected_ids = rejected_ids[:target_len]
 
     chosen_input_ids = prompt_ids + chosen_ids
-    chosen_labels = [IGNORE_INDEX] * len(prompt_ids) + chosen_ids
+    chosen_labels = [IGNORE_INDEX] * source_len + chosen_ids
     rejected_input_ids = prompt_ids + rejected_ids
-    rejected_labels = [IGNORE_INDEX] * len(prompt_ids) + rejected_ids
-
+    rejected_labels = [IGNORE_INDEX] * source_len + rejected_ids
     return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels
 
 
@@ -73,36 +72,25 @@ 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 = {
-        "chosen_input_ids": [],
-        "chosen_attention_mask": [],
-        "chosen_labels": [],
-        "rejected_input_ids": [],
-        "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]))
+    model_inputs = defaultdict(list)
+    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],
-            response=examples["response"][i],
-            system=examples["system"][i],
-            tools=examples["tools"][i],
+            prompt=examples["_prompt"][i],
+            response=examples["_response"][i],
+            system=examples["_system"][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))
@@ -110,15 +98,8 @@ 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["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
-            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)
-                )
+        model_inputs["images"].append(examples["_images"][i])
+        model_inputs["videos"].append(examples["_videos"][i])
 
     return model_inputs
 

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
-
-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
+from typing import List, Sequence, Tuple
 
 
 def search_for_fit(numbers: Sequence[int], capacity: int) -> int:
@@ -61,18 +48,18 @@ def greedy_knapsack(numbers: List[int], capacity: int) -> List[List[int]]:
     return knapsacks
 
 
-def get_pixel_values(images: Sequence["ImageObject"], processor: "ProcessorMixin") -> "NDArray":
+def infer_seqlen(source_len: int, target_len: int, cutoff_len: int) -> Tuple[int, int]:
     r"""
-    Processes visual inputs. (currently only supports a single image)
+    Computes the real sequence length after truncation by the cutoff_len.
     """
-    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)
+    if target_len * 2 < cutoff_len:  # truncate source
+        max_target_len = cutoff_len
+    elif source_len * 2 < cutoff_len:  # truncate target
+        max_target_len = cutoff_len - source_len
+    else:  # truncate both
+        max_target_len = int(cutoff_len * (target_len / (source_len + target_len)))
 
-
-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)
+    new_target_len = min(max_target_len, target_len)
+    max_source_len = max(cutoff_len - new_target_len, 0)
+    new_source_len = min(max_source_len, source_len)
+    return new_source_len, new_target_len

src/llamafactory/data/processors/supervised.py

@@ -17,13 +17,14 @@ from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
 from ...extras.constants import IGNORE_INDEX
 from ...extras.logging import get_logger
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, greedy_knapsack
+from .processor_utils import 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
 
 
@@ -35,35 +36,49 @@ 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
-        prompt[0]["content"] = template.image_token + prompt[0]["content"]
+    messages = template.mm_plugin.process_messages(prompt + response, images, videos, processor)
+    input_ids, labels = template.mm_plugin.process_token_ids([], [], images, videos, tokenizer, processor)
+    encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools)
+    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
 
-    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, data_args.cutoff_len, data_args.reserved_label_len
-    )
     for turn_idx, (source_ids, target_ids) in enumerate(encoded_pairs):
-        if data_args.train_on_prompt:
-            source_mask = source_ids
-        elif turn_idx != 0 and template.efficient_eos:
-            source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
-        else:
-            source_mask = [IGNORE_INDEX] * len(source_ids)
+        if total_length >= cutoff_len:
+            break
 
-        input_ids += source_ids + target_ids
-        labels += source_mask + target_ids
+        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 train_on_prompt:
+            source_label = source_ids
+        elif template.efficient_eos:
+            source_label = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (source_len - 1)
+        else:
+            source_label = [IGNORE_INDEX] * source_len
+
+        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
 
     if template.efficient_eos:
         input_ids += [tokenizer.eos_token_id]
@@ -78,37 +93,34 @@ 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": []}
-    if processor is not None:
-        model_inputs["pixel_values"] = []
-        if hasattr(processor, "image_seq_length"):  # paligemma models
-            model_inputs["token_type_ids"] = []
-
-    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]))
+    model_inputs = defaultdict(list)
+    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],
-            response=examples["response"][i],
-            system=examples["system"][i],
-            tools=examples["tools"][i],
+            prompt=examples["_prompt"][i],
+            response=examples["_response"][i],
+            system=examples["_system"][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["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
-            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["images"].append(examples["_images"][i])
+        model_inputs["videos"].append(examples["_videos"][i])
 
     return model_inputs
 
@@ -117,28 +129,34 @@ 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"])):
-        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]))
+    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],
-            response=examples["response"][i],
-            system=examples["system"][i],
-            tools=examples["tools"][i],
+            prompt=examples["_prompt"][i],
+            response=examples["_response"][i],
+            system=examples["_system"][i],
+            tools=examples["_tools"][i],
+            images=examples["_images"][i] or [],
+            videos=examples["_videos"][i] or [],
             template=template,
             tokenizer=tokenizer,
-            processor=None,
-            data_args=data_args,
+            processor=processor,
+            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:
@@ -148,28 +166,43 @@ def preprocess_packed_supervised_dataset(
             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": []}
-    knapsacks = greedy_knapsack(lengths, data_args.cutoff_len)
+    model_inputs = defaultdict(list)
+    knapsacks = greedy_knapsack(lengths, data_args.cutoff_len - 1)  # reserved for the padding token
     for knapsack in knapsacks:
-        packed_input_ids, packed_labels = [], []
-        for length in knapsack:
+        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:
+                packed_attention_masks += [1] * len(batch_input_ids[index])
 
         if len(packed_input_ids) < data_args.cutoff_len:
             pad_length = data_args.cutoff_len - len(packed_input_ids)
             packed_input_ids += [tokenizer.pad_token_id] * pad_length
             packed_labels += [IGNORE_INDEX] * pad_length
+            if data_args.neat_packing:
+                packed_attention_masks += [0] * pad_length
+            else:
+                packed_attention_masks += [1] * pad_length  # more efficient flash_attn
 
         if len(packed_input_ids) != data_args.cutoff_len:
             raise ValueError("The length of packed example should be identical to the cutoff length.")
 
         model_inputs["input_ids"].append(packed_input_ids)
-        model_inputs["attention_mask"].append([1] * data_args.cutoff_len)
+        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
 

src/llamafactory/data/processors/unsupervised.py

@@ -12,17 +12,19 @@
 # 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 ..data_utils import Role
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values
+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
 
 
@@ -34,29 +36,27 @@ 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": ""}]
 
-    input_ids, labels = template.encode_oneturn(
-        tokenizer, messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
+    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
-        image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
-        input_ids = [image_token_id] * getattr(processor, "image_seq_length") + input_ids
-
+    input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, images, videos, tokenizer, processor)
+    source_len, target_len = infer_seqlen(len(input_ids), len(labels), cutoff_len)
+    input_ids = input_ids[:source_len]
+    labels = labels[:target_len]
     return input_ids, labels
 
 
@@ -66,36 +66,31 @@ 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": []}
-    if processor is not None:
-        model_inputs["pixel_values"] = []
-        if hasattr(processor, "image_seq_length"):  # paligemma models
-            model_inputs["token_type_ids"] = []
-
-    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]))
+    model_inputs = defaultdict(list)
+    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],
-            response=examples["response"][i],
-            system=examples["system"][i],
-            tools=examples["tools"][i],
+            prompt=examples["_prompt"][i],
+            response=examples["_response"][i],
+            system=examples["_system"][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["pixel_values"].append(get_pixel_values(examples["images"][i], processor))
-            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["images"].append(examples["_images"][i])
+        model_inputs["videos"].append(examples["_videos"][i])
 
     return model_inputs
 

src/llamafactory/data/template.py

@@ -15,15 +15,21 @@
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Tuple, Union
 
+from transformers.utils.versions import require_version
+from typing_extensions import override
+
 from ..extras.logging import get_logger
-from .data_utils import Role, infer_max_len
+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__)
@@ -41,43 +47,40 @@ class Template:
     format_prefix: "Formatter"
     default_system: str
     stop_words: List[str]
-    image_token: str
     efficient_eos: bool
     replace_eos: bool
+    mm_plugin: "BasePlugin"
 
     def encode_oneturn(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        cutoff_len: int = 1_000_000,
-        reserved_label_len: int = 1,
     ) -> Tuple[List[int], List[int]]:
         r"""
         Returns a single pair of token ids representing prompt and response respectively.
         """
-        encoded_pairs = self._encode(tokenizer, messages, system, tools, cutoff_len, reserved_label_len)
+        encoded_messages = self._encode(tokenizer, messages, system, tools)
         prompt_ids = []
-        for query_ids, resp_ids in encoded_pairs[:-1]:
-            prompt_ids += query_ids + resp_ids
-        prompt_ids = prompt_ids + encoded_pairs[-1][0]
-        answer_ids = encoded_pairs[-1][1]
+        for encoded_ids in encoded_messages[:-1]:
+            prompt_ids += encoded_ids
+
+        answer_ids = encoded_messages[-1]
         return prompt_ids, answer_ids
 
     def encode_multiturn(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        cutoff_len: int = 1_000_000,
-        reserved_label_len: int = 1,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
+    ) -> List[Tuple[List[int], List[int]]]:
         r"""
         Returns multiple pairs of token ids representing prompts and responses respectively.
         """
-        return self._encode(tokenizer, messages, system, tools, cutoff_len, reserved_label_len)
+        encoded_messages = self._encode(tokenizer, messages, system, tools)
+        return [(encoded_messages[i], encoded_messages[i + 1]) for i in range(0, len(encoded_messages), 2)]
 
     def extract_tool(self, content: str) -> Union[str, List[Tuple[str, str]]]:
         r"""
@@ -88,16 +91,14 @@ class Template:
     def _encode(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: Optional[str],
         tools: Optional[str],
-        cutoff_len: int,
-        reserved_label_len: int,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
+    ) -> List[List[int]]:
         r"""
         Encodes formatted inputs to pairs of token ids.
-        Turn 0: system + query        resp
-        Turn t: sep + query           resp
+        Turn 0: prefix + system + query        resp
+        Turn t: sep + query                    resp
         """
         system = system or self.default_system
         encoded_messages = []
@@ -106,10 +107,9 @@ class Template:
 
             if i == 0:
                 elements += self.format_prefix.apply()
-
-            if i == 0 and (system or tools):
-                tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
-                elements += self.format_system.apply(content=(system + tool_text))
+                if system or tools:
+                    tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
+                    elements += self.format_system.apply(content=(system + tool_text))
 
             if i > 0 and i % 2 == 0:
                 elements += self.format_separator.apply()
@@ -127,11 +127,9 @@ class Template:
 
             encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
 
-        return self._make_pairs(encoded_messages, cutoff_len, reserved_label_len)
+        return encoded_messages
 
-    def _convert_elements_to_ids(
-        self, tokenizer: "PreTrainedTokenizer", elements: List[Union[str, Dict[str, str]]]
-    ) -> List[int]:
+    def _convert_elements_to_ids(self, tokenizer: "PreTrainedTokenizer", elements: "SLOTS") -> List[int]:
         r"""
         Converts elements to token ids.
         """
@@ -152,60 +150,33 @@ class Template:
 
         return token_ids
 
-    def _make_pairs(
-        self,
-        encoded_messages: Sequence[List[int]],
-        cutoff_len: int,
-        reserved_label_len: int,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
-        encoded_pairs = []
-        total_length = 0
-        for i in range(0, len(encoded_messages), 2):
-            if total_length >= cutoff_len:
-                break
-
-            max_source_len, max_target_len = infer_max_len(
-                source_len=len(encoded_messages[i]),
-                target_len=len(encoded_messages[i + 1]),
-                max_len=(cutoff_len - total_length),
-                reserved_label_len=reserved_label_len,
-            )
-            source_ids = encoded_messages[i][:max_source_len]
-            target_ids = encoded_messages[i + 1][:max_target_len]
-            total_length += len(source_ids) + len(target_ids)
-            encoded_pairs.append((source_ids, target_ids))
-
-        return encoded_pairs
-
 
 @dataclass
 class Llama2Template(Template):
+    @override
     def _encode(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: str,
         tools: str,
-        cutoff_len: int,
-        reserved_label_len: int,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
+    ) -> List[List[int]]:
         r"""
         Encodes formatted inputs to pairs of token ids.
-        Turn 0: system + query        resp
-        Turn t: sep + query           resp
+        Turn 0: prefix + system + query        resp
+        Turn t: sep + query                    resp
         """
         system = system or self.default_system
         encoded_messages = []
         for i, message in enumerate(messages):
             elements = []
 
+            system_text = ""
             if i == 0:
                 elements += self.format_prefix.apply()
-
-            system_text = ""
-            if i == 0 and (system or tools):
-                tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
-                system_text = self.format_system.apply(content=(system + tool_text))[0]
+                if system or tools:
+                    tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
+                    system_text = self.format_system.apply(content=(system + tool_text))[0]
 
             if i > 0 and i % 2 == 0:
                 elements += self.format_separator.apply()
@@ -223,10 +194,10 @@ class Llama2Template(Template):
 
             encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
 
-        return self._make_pairs(encoded_messages, cutoff_len, reserved_label_len)
+        return encoded_messages
 
 
-TEMPLATES: Dict[str, Template] = {}
+TEMPLATES: Dict[str, "Template"] = {}
 
 
 def _register_template(
@@ -240,10 +211,10 @@ def _register_template(
     format_separator: Optional["Formatter"] = None,
     format_prefix: Optional["Formatter"] = None,
     default_system: str = "",
-    stop_words: List[str] = [],
-    image_token: str = "<image>",
+    stop_words: Sequence[str] = [],
     efficient_eos: bool = False,
     replace_eos: bool = False,
+    mm_plugin: "BasePlugin" = get_mm_plugin(name="base"),
 ) -> None:
     r"""
     Registers a chat template.
@@ -275,9 +246,7 @@ def _register_template(
     template_class = Llama2Template if name.startswith("llama2") else Template
     default_user_formatter = StringFormatter(slots=["{{content}}"])
     default_assistant_formatter = StringFormatter(slots=["{{content}}"] + eos_slots)
-    default_function_formatter = FunctionFormatter(
-        slots=["Action: {{name}}\nAction Input: {{arguments}}\n"] + eos_slots
-    )
+    default_function_formatter = FunctionFormatter(slots=eos_slots, tool_format="default")
     default_tool_formatter = ToolFormatter(tool_format="default")
     default_separator_formatter = EmptyFormatter()
     default_prefix_formatter = EmptyFormatter()
@@ -292,9 +261,9 @@ 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,
+        mm_plugin=mm_plugin,
     )
 
 
@@ -338,6 +307,9 @@ 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)
@@ -348,14 +320,15 @@ 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 %}"
@@ -376,16 +349,30 @@ def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer")
     return jinja_template
 
 
-def get_template_and_fix_tokenizer(
-    tokenizer: "PreTrainedTokenizer",
-    name: Optional[str] = None,
-) -> Template:
-    if name is None:
+def get_template_and_fix_tokenizer(tokenizer: "PreTrainedTokenizer", data_args: "DataArguments") -> "Template":
+    r"""
+    Gets chat template and fixes the tokenizer.
+    """
+    if data_args.template in ["llava", "paligemma", "qwen2_vl"]:
+        require_version(
+            "transformers>=4.45.0.dev0", "To fix: pip install git+https://github.com/huggingface/transformers.git"
+        )
+
+    if data_args.template 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("Template {} does not exist.".format(data_args.template))
+
+    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("Using tool format: {}.".format(data_args.tool_format))
+        eos_slots = [] if template.efficient_eos else [{"eos_token"}]
+        template.format_function = FunctionFormatter(slots=eos_slots, tool_format=data_args.tool_format)
+        template.format_tools = ToolFormatter(tool_format=data_args.tool_format)
 
     stop_words = template.stop_words
     if template.replace_eos:
@@ -501,35 +488,17 @@ _register_template(
     format_user=StringFormatter(slots=[{"token": "<|user|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]),
     format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
     format_system=StringFormatter(slots=[{"token": "<|system|>"}, "\n", "{{content}}"]),
-    format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
+    format_function=FunctionFormatter(slots=[], tool_format="glm4"),
     format_observation=StringFormatter(
         slots=[{"token": "<|observation|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]
     ),
+    format_tools=ToolFormatter(tool_format="glm4"),
     format_prefix=EmptyFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}]),
     stop_words=["<|user|>", "<|observation|>"],
     efficient_eos=True,
 )
 
 
-_register_template(
-    name="chatglm3_system",
-    format_user=StringFormatter(slots=[{"token": "<|user|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]),
-    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
-    format_system=StringFormatter(slots=[{"token": "<|system|>"}, "\n", "{{content}}"]),
-    format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
-    format_observation=StringFormatter(
-        slots=[{"token": "<|observation|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]
-    ),
-    format_prefix=EmptyFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}]),
-    default_system=(
-        "You are ChatGLM3, a large language model trained by Zhipu.AI. "
-        "Follow the user's instructions carefully. Respond using markdown."
-    ),
-    stop_words=["<|user|>", "<|observation|>"],
-    efficient_eos=True,
-)
-
-
 _register_template(
     name="chatml",
     format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
@@ -559,6 +528,23 @@ _register_template(
 )
 
 
+_register_template(
+    name="codegeex4",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>\n"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}"]),
+    format_function=FunctionFormatter(slots=[], tool_format="glm4"),
+    format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>\n"]),
+    format_tools=ToolFormatter(tool_format="glm4"),
+    format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
+    default_system=(
+        "你是一位智能编程助手，你叫CodeGeeX。你会为用户回答关于编程、代码、计算机方面的任何问题，"
+        "并提供格式规范、可以执行、准确安全的代码，并在必要时提供详细的解释。"
+    ),
+    stop_words=["<|user|>", "<|observation|>"],
+    efficient_eos=True,
+)
+
+
 _register_template(
     name="cohere",
     format_user=StringFormatter(
@@ -581,6 +567,15 @@ _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"]),
@@ -610,6 +605,7 @@ _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"}]),
 )
 
@@ -617,22 +613,21 @@ _register_template(
 _register_template(
     name="deepseekcoder",
     format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n### Response:"]),
-    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
-    format_separator=EmptyFormatter(slots=["\n<|EOT|>\n"]),
+    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, "
-        "developed by Deepseek Company, and you only answer questions related to computer science. "
+        "You are an AI programming assistant, utilizing the DeepSeek Coder model, "
+        "developed by DeepSeek Company, and you only answer questions related to computer science. "
         "For politically sensitive questions, security and privacy issues, "
-        "and other non-computer science questions, you will refuse to answer\n"
+        "and other non-computer science questions, you will refuse to answer.\n"
     ),
-    stop_words=["<|EOT|>"],
-    efficient_eos=True,
 )
 
 
 _register_template(
     name="default",
-    format_user=StringFormatter(slots=["Human: {{content}}\nAssistant: "]),
+    format_user=StringFormatter(slots=["Human: {{content}}\nAssistant:"]),
     format_system=StringFormatter(slots=["{{content}}\n"]),
     format_separator=EmptyFormatter(slots=["\n"]),
 )
@@ -640,7 +635,6 @@ _register_template(
 
 _register_template(
     name="empty",
-    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     efficient_eos=True,
 )
 
@@ -677,7 +671,7 @@ _register_template(
     format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>"]),
     format_assistant=StringFormatter(slots=["\n{{content}}"]),
     format_system=StringFormatter(slots=["<|system|>\n{{content}}"]),
-    format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
+    format_function=FunctionFormatter(slots=[], tool_format="glm4"),
     format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>"]),
     format_tools=ToolFormatter(tool_format="glm4"),
     format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
@@ -748,6 +742,17 @@ _register_template(
 )
 
 
+_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="mistral",
     format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
@@ -792,6 +797,19 @@ _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"]),
@@ -815,6 +833,33 @@ _register_template(
 )
 
 
+_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,
+    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,
+)
+
+
 _register_template(
     name="solar",
     format_user=StringFormatter(slots=["### User:\n{{content}}\n\n### Assistant:\n"]),
@@ -912,6 +957,7 @@ _register_template(
     ),
     stop_words=["###"],
     efficient_eos=True,
+    mm_plugin=get_mm_plugin(name="llava", image_token="<image>"),
 )
 
 
@@ -926,8 +972,7 @@ _register_template(
 
 _register_template(
     name="zephyr",
-    format_user=StringFormatter(slots=["<|user|>\n{{content}}", {"eos_token"}, "<|assistant|>"]),
-    format_assistant=StringFormatter(slots=["\n{{content}}", {"eos_token"}]),
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}", {"eos_token"}, "<|assistant|>\n"]),
     format_system=StringFormatter(slots=["<|system|>\n{{content}}", {"eos_token"}]),
     default_system="You are Zephyr, a helpful assistant.",
 )

src/llamafactory/data/tool_utils.py

@@ -0,0 +1,182 @@
+# 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 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
+
+
+DEFAULT_TOOL_PROMPT = (
+    "You have access to the following tools:\n{tool_text}"
+    "Use the following format if using a tool:\n"
+    "```\n"
+    "Action: tool name (one of [{tool_names}])\n"
+    "Action Input: the input to the tool, in a JSON format representing the kwargs "
+    """(e.g. ```{{"input": "hello world", "num_beams": 5}}```)\n"""
+    "```\n"
+)
+
+
+GLM4_TOOL_PROMPT = (
+    "你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，"
+    "你的任务是针对用户的问题和要求提供适当的答复和支持。# 可用工具{tool_text}"
+)
+
+
+FunctionCall = namedtuple("FunctionCall", ["name", "arguments"])
+
+
+@dataclass
+class ToolUtils(ABC):
+    """
+    Base class for tool utilities.
+    """
+
+    @staticmethod
+    @abstractmethod
+    def get_function_slots() -> SLOTS:
+        r"""
+        Gets a list of slots corresponding to a single function call.
+        """
+        ...
+
+    @staticmethod
+    @abstractmethod
+    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 = ""
+        tool_names = []
+        for tool in tools:
+            param_text = ""
+            for name, param in tool["parameters"]["properties"].items():
+                required, enum, items = "", "", ""
+                if name in tool["parameters"].get("required", []):
+                    required = ", required"
+
+                if param.get("enum", None):
+                    enum = ", should be one of [{}]".format(", ".join(param["enum"]))
+
+                if param.get("items", None):
+                    items = ", where each item should be {}".format(param["items"].get("type", ""))
+
+                param_text += "  - {name} ({type}{required}): {desc}{enum}{items}\n".format(
+                    name=name,
+                    type=param.get("type", ""),
+                    required=required,
+                    desc=param.get("description", ""),
+                    enum=enum,
+                    items=items,
+                )
+
+            tool_text += "> Tool Name: {name}\nTool Description: {desc}\nTool Args:\n{args}\n".format(
+                name=tool["name"], desc=tool.get("description", ""), args=param_text
+            )
+            tool_names.append(tool["name"])
+
+        return DEFAULT_TOOL_PROMPT.format(tool_text=tool_text, tool_names=", ".join(tool_names))
+
+    @override
+    @staticmethod
+    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:
+            return content
+
+        results = []
+        for match in action_match:
+            tool_name = match[0].strip()
+            tool_input = match[1].strip().strip('"').strip("```")
+            try:
+                arguments = json.loads(tool_input)
+                results.append((tool_name, json.dumps(arguments, ensure_ascii=False)))
+            except json.JSONDecodeError:
+                return content
+
+        return results
+
+
+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 = ""
+        for tool in tools:
+            tool_text += "\n\n## {name}\n\n{body}\n在调用上述函数时，请使用 Json 格式表示调用的参数。".format(
+                name=tool["name"], body=json.dumps(tool, indent=4, ensure_ascii=False)
+            )
+
+        return GLM4_TOOL_PROMPT.format(tool_text=tool_text)
+
+    @override
+    @staticmethod
+    def tool_extractor(content: str) -> Union[str, List["FunctionCall"]]:
+        if "\n" not in content:
+            return content
+
+        tool_name, tool_input = content.split("\n", maxsplit=1)
+        try:
+            arguments = json.loads(tool_input)
+        except json.JSONDecodeError:
+            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("Tool utils `{}` not found.".format(name))
+
+    return tool_utils

src/llamafactory/eval/evaluator.py

@@ -37,10 +37,9 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
-import inspect
 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
@@ -55,18 +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)
@@ -74,8 +77,11 @@ class Evaluator:
         return [chr(ord("A") + offset.item()) for offset in torch.argmax(choice_probs, dim=-1)]
 
     def eval(self) -> None:
+        eval_task = self.eval_args.task.split("_")[0]
+        eval_split = self.eval_args.task.split("_")[1]
+
         mapping = cached_file(
-            path_or_repo_id=os.path.join(self.eval_args.task_dir, self.eval_args.task),
+            path_or_repo_id=os.path.join(self.eval_args.task_dir, eval_task),
             filename="mapping.json",
             cache_dir=self.model_args.cache_dir,
             token=self.model_args.hf_hub_token,
@@ -88,27 +94,22 @@ class Evaluator:
         pbar = tqdm(categorys.keys(), desc="Processing subjects", position=0)
         results = {}
         for subject in pbar:
-            if "trust_remote_code" in inspect.signature(load_dataset).parameters:  # for datasets==2.16.0
-                kwargs = {"trust_remote_code": True}
-            else:
-                kwargs = {}
-
             dataset = load_dataset(
-                path=os.path.join(self.eval_args.task_dir, self.eval_args.task),
+                path=os.path.join(self.eval_args.task_dir, eval_task),
                 name=subject,
                 cache_dir=self.model_args.cache_dir,
                 download_mode=self.eval_args.download_mode,
                 token=self.model_args.hf_hub_token,
-                **kwargs,
+                trust_remote_code=True,
             )
             pbar.set_postfix_str(categorys[subject]["name"])
             inputs, outputs, labels = [], [], []
-            for i in trange(len(dataset[self.data_args.split]), desc="Formatting batches", position=1, leave=False):
+            for i in trange(len(dataset[eval_split]), desc="Formatting batches", position=1, leave=False):
                 support_set = (
                     dataset["train"].shuffle().select(range(min(self.eval_args.n_shot, len(dataset["train"]))))
                 )
                 messages = self.eval_template.format_example(
-                    target_data=dataset[self.data_args.split][i],
+                    target_data=dataset[eval_split][i],
                     support_set=support_set,
                     subject_name=categorys[subject]["name"],
                 )
@@ -135,7 +136,7 @@ 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))

src/llamafactory/extras/constants.py

@@ -47,6 +47,8 @@ FILEEXT2TYPE = {
 
 IGNORE_INDEX = -100
 
+IMAGE_PLACEHOLDER = "<image>"
+
 LAYERNORM_NAMES = {"norm", "ln"}
 
 LLAMABOARD_CONFIG = "llamaboard_config.yaml"
@@ -78,8 +80,23 @@ TRAINING_STAGES = {
 
 STAGES_USE_PAIR_DATA = {"rm", "dpo"}
 
+SUPPORTED_CLASS_FOR_BLOCK_DIAG_ATTN = {
+    "cohere",
+    "falcon",
+    "gemma",
+    "gemma2",
+    "llama",
+    "mistral",
+    "phi",
+    "phi3",
+    "qwen2",
+    "starcoder2",
+}
+
 SUPPORTED_CLASS_FOR_S2ATTN = {"llama"}
 
+VIDEO_PLACEHOLDER = "<video>"
+
 V_HEAD_WEIGHTS_NAME = "value_head.bin"
 
 V_HEAD_SAFE_WEIGHTS_NAME = "value_head.safetensors"
@@ -286,6 +303,17 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "CodeGeeX4-9B-Chat": {
+            DownloadSource.DEFAULT: "THUDM/codegeex4-all-9b",
+            DownloadSource.MODELSCOPE: "ZhipuAI/codegeex4-all-9b",
+        },
+    },
+    template="codegeex4",
+)
+
+
 register_model_group(
     models={
         "CodeGemma-7B": {
@@ -507,6 +535,30 @@ register_model_group(
         "Gemma-1.1-7B-Chat": {
             DownloadSource.DEFAULT: "google/gemma-1.1-7b-it",
         },
+        "Gemma-2-2B": {
+            DownloadSource.DEFAULT: "google/gemma-2-2b",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-2b",
+        },
+        "Gemma-2-9B": {
+            DownloadSource.DEFAULT: "google/gemma-2-9b",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-9b",
+        },
+        "Gemma-2-27B": {
+            DownloadSource.DEFAULT: "google/gemma-2-27b",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-27b",
+        },
+        "Gemma-2-2B-Chat": {
+            DownloadSource.DEFAULT: "google/gemma-2-2b-it",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-2b-it",
+        },
+        "Gemma-2-9B-Chat": {
+            DownloadSource.DEFAULT: "google/gemma-2-9b-it",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-9b-it",
+        },
+        "Gemma-2-27B-Chat": {
+            DownloadSource.DEFAULT: "google/gemma-2-27b-it",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-27b-it",
+        },
     },
     template="gemma",
 )
@@ -579,7 +631,42 @@ register_model_group(
 
 register_model_group(
     models={
-        "Jambda-v0.1": {
+        "InternLM2.5-1.8B": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-1_8b",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-1_8b",
+        },
+        "InternLM2.5-7B": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-7b",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-7b",
+        },
+        "InternLM2.5-20B": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-20b",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-20b",
+        },
+        "InternLM2.5-1.8B-Chat": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-1_8b-chat",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-1_8b-chat",
+        },
+        "InternLM2.5-7B-Chat": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-7b-chat",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-7b-chat",
+        },
+        "InternLM2.5-7B-1M-Chat": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-7b-chat-1m",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-7b-chat-1m",
+        },
+        "InternLM2.5-20B-Chat": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-20b-chat",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-20b-chat",
+        },
+    },
+    template="intern2",
+)
+
+
+register_model_group(
+    models={
+        "Jamba-v0.1": {
             DownloadSource.DEFAULT: "ai21labs/Jamba-v0.1",
             DownloadSource.MODELSCOPE: "AI-ModelScope/Jamba-v0.1",
         }
@@ -680,6 +767,37 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "LLaMA3.1-8B": {
+            DownloadSource.DEFAULT: "meta-llama/Meta-Llama-3.1-8B",
+            DownloadSource.MODELSCOPE: "LLM-Research/Meta-Llama-3.1-8B",
+        },
+        "LLaMA3.1-70B": {
+            DownloadSource.DEFAULT: "meta-llama/Meta-Llama-3.1-70B",
+            DownloadSource.MODELSCOPE: "LLM-Research/Meta-Llama-3.1-70B",
+        },
+        "LLaMA3.1-405B": {
+            DownloadSource.DEFAULT: "meta-llama/Meta-Llama-3.1-405B",
+            DownloadSource.MODELSCOPE: "LLM-Research/Meta-Llama-3.1-405B",
+        },
+        "LLaMA3.1-8B-Chat": {
+            DownloadSource.DEFAULT: "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            DownloadSource.MODELSCOPE: "LLM-Research/Meta-Llama-3.1-8B-Instruct",
+        },
+        "LLaMA3.1-70B-Chat": {
+            DownloadSource.DEFAULT: "meta-llama/Meta-Llama-3.1-70B-Instruct",
+            DownloadSource.MODELSCOPE: "LLM-Research/Meta-Llama-3.1-70B-Instruct",
+        },
+        "LLaMA3.1-405B-Chat": {
+            DownloadSource.DEFAULT: "meta-llama/Meta-Llama-3.1-405B-Instruct",
+            DownloadSource.MODELSCOPE: "LLM-Research/Meta-Llama-3.1-405B-Instruct",
+        },
+    },
+    template="llama3",
+)
+
+
 register_model_group(
     models={
         "LLaVA1.5-7B-Chat": {
@@ -689,7 +807,7 @@ register_model_group(
             DownloadSource.DEFAULT: "llava-hf/llava-1.5-13b-hf",
         },
     },
-    template="vicuna",
+    template="llava",
     vision=True,
 )
 
@@ -709,6 +827,17 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "MiniCPM3-4B-Chat": {
+            DownloadSource.DEFAULT: "openbmb/MiniCPM3-4B",
+            DownloadSource.MODELSCOPE: "OpenBMB/MiniCPM3-4B",
+        },
+    },
+    template="cpm3",
+)
+
+
 register_model_group(
     models={
         "Mistral-7B-v0.1": {
@@ -732,6 +861,11 @@ register_model_group(
         },
         "Mistral-7B-v0.3-Chat": {
             DownloadSource.DEFAULT: "mistralai/Mistral-7B-Instruct-v0.3",
+            DownloadSource.MODELSCOPE: "LLM-Research/Mistral-7B-Instruct-v0.3",
+        },
+        "Mistral-Nemo-Chat": {
+            DownloadSource.DEFAULT: "mistralai/Mistral-Nemo-Instruct-2407",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/Mistral-Nemo-Instruct-2407",
         },
     },
     template="mistral",
@@ -829,27 +963,28 @@ register_model_group(
 
 register_model_group(
     models={
-        "PaliGemma-3B-pt-224": {
+        "PaliGemma-3B-pt-224-Chat": {
             DownloadSource.DEFAULT: "google/paligemma-3b-pt-224",
             DownloadSource.MODELSCOPE: "AI-ModelScope/paligemma-3b-pt-224",
         },
-        "PaliGemma-3B-pt-448": {
+        "PaliGemma-3B-pt-448-Chat": {
             DownloadSource.DEFAULT: "google/paligemma-3b-pt-448",
             DownloadSource.MODELSCOPE: "AI-ModelScope/paligemma-3b-pt-448",
         },
-        "PaliGemma-3B-pt-896": {
+        "PaliGemma-3B-pt-896-Chat": {
             DownloadSource.DEFAULT: "google/paligemma-3b-pt-896",
             DownloadSource.MODELSCOPE: "AI-ModelScope/paligemma-3b-pt-896",
         },
-        "PaliGemma-3B-mix-224": {
+        "PaliGemma-3B-mix-224-Chat": {
             DownloadSource.DEFAULT: "google/paligemma-3b-mix-224",
             DownloadSource.MODELSCOPE: "AI-ModelScope/paligemma-3b-mix-224",
         },
-        "PaliGemma-3B-mix-448": {
+        "PaliGemma-3B-mix-448-Chat": {
             DownloadSource.DEFAULT: "google/paligemma-3b-mix-448",
             DownloadSource.MODELSCOPE: "AI-ModelScope/paligemma-3b-mix-448",
         },
     },
+    template="paligemma",
     vision=True,
 )
 
@@ -1143,6 +1278,18 @@ register_model_group(
             DownloadSource.DEFAULT: "Qwen/Qwen2-57B-A14B",
             DownloadSource.MODELSCOPE: "qwen/Qwen2-57B-A14B",
         },
+        "Qwen2-Math-1.5B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-Math-1.5B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-Math-1.5B",
+        },
+        "Qwen2-Math-7B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-Math-7B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-Math-7B",
+        },
+        "Qwen2-Math-72B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-Math-72B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-Math-72B",
+        },
         "Qwen2-0.5B-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen2-0.5B-Instruct",
             DownloadSource.MODELSCOPE: "qwen/Qwen2-0.5B-Instruct",
@@ -1163,6 +1310,18 @@ register_model_group(
             DownloadSource.DEFAULT: "Qwen/Qwen2-57B-A14B-Instruct",
             DownloadSource.MODELSCOPE: "qwen/Qwen2-57B-A14B-Instruct",
         },
+        "Qwen2-Math-1.5B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-Math-1.5B-Instruct",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-Math-1.5B-Instruct",
+        },
+        "Qwen2-Math-7B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-Math-7B-Instruct",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-Math-7B-Instruct",
+        },
+        "Qwen2-Math-72B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-Math-72B-Instruct",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-Math-72B-Instruct",
+        },
         "Qwen2-0.5B-int8-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen2-0.5B-Instruct-GPTQ-Int8",
             DownloadSource.MODELSCOPE: "qwen/Qwen2-0.5B-Instruct-GPTQ-Int8",
@@ -1204,6 +1363,38 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "Qwen2VL-2B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-VL-2B-Instruct",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-VL-2B-Instruct",
+        },
+        "Qwen2VL-7B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-VL-7B-Instruct",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-VL-7B-Instruct",
+        },
+        "Qwen2VL-2B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-VL-2B-Instruct-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-VL-2B-Instruct-GPTQ-Int8",
+        },
+        "Qwen2VL-2B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-VL-2B-Instruct-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-VL-2B-Instruct-AWQ",
+        },
+        "Qwen2VL-7B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-VL-7B-Instruct-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-VL-7B-Instruct-GPTQ-Int8",
+        },
+        "Qwen2VL-7B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen2-VL-7B-Instruct-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen2-VL-7B-Instruct-AWQ",
+        },
+    },
+    template="qwen2_vl",
+    vision=True,
+)
+
+
 register_model_group(
     models={
         "SOLAR-10.7B": {
@@ -1248,6 +1439,10 @@ register_model_group(
 
 register_model_group(
     models={
+        "TeleChat-1B-Chat": {
+            DownloadSource.DEFAULT: "Tele-AI/TeleChat-1B",
+            DownloadSource.MODELSCOPE: "TeleAI/TeleChat-1B",
+        },
         "TeleChat-7B-Chat": {
             DownloadSource.DEFAULT: "Tele-AI/telechat-7B",
             DownloadSource.MODELSCOPE: "TeleAI/telechat-7B",
@@ -1471,6 +1666,22 @@ register_model_group(
             DownloadSource.DEFAULT: "01-ai/Yi-1.5-34B-Chat",
             DownloadSource.MODELSCOPE: "01ai/Yi-1.5-34B-Chat",
         },
+        "Yi-Coder-1.5B": {
+            DownloadSource.DEFAULT: "01-ai/Yi-Coder-1.5B",
+            DownloadSource.MODELSCOPE: "01ai/Yi-Coder-1.5B",
+        },
+        "Yi-Coder-9B": {
+            DownloadSource.DEFAULT: "01-ai/Yi-Coder-9B",
+            DownloadSource.MODELSCOPE: "01ai/Yi-Coder-9B",
+        },
+        "Yi-Coder-1.5B-Chat": {
+            DownloadSource.DEFAULT: "01-ai/Yi-Coder-1.5B-Chat",
+            DownloadSource.MODELSCOPE: "01ai/Yi-Coder-1.5B-Chat",
+        },
+        "Yi-Coder-9B-Chat": {
+            DownloadSource.DEFAULT: "01-ai/Yi-Coder-9B-Chat",
+            DownloadSource.MODELSCOPE: "01ai/Yi-Coder-9B-Chat",
+        },
     },
     template="yi",
 )

src/llamafactory/extras/env.py

@@ -1,4 +1,7 @@
-# Copyright 2024 the LlamaFactory team.
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/commands/env.py
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -23,7 +26,7 @@ import trl
 from transformers.utils import is_torch_cuda_available, is_torch_npu_available
 
 
-VERSION = "0.8.2"
+VERSION = "0.9.0"
 
 
 def print_env() -> None:

src/llamafactory/extras/logging.py

@@ -1,4 +1,7 @@
-# Copyright 2024 the LlamaFactory team.
+# Copyright 2024 Optuna, HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/utils/logging.py
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -15,14 +18,21 @@
 import logging
 import os
 import sys
+import threading
 from concurrent.futures import ThreadPoolExecutor
+from typing import Optional
 
 from .constants import RUNNING_LOG
 
 
+_thread_lock = threading.RLock()
+_default_handler: Optional["logging.Handler"] = None
+_default_log_level: "logging._Level" = logging.INFO
+
+
 class LoggerHandler(logging.Handler):
     r"""
-    Logger handler used in Web UI.
+    Redirects the logging output to the logging file for LLaMA Board.
     """
 
     def __init__(self, output_dir: str) -> None:
@@ -56,27 +66,56 @@ class LoggerHandler(logging.Handler):
         return super().close()
 
 
-def get_logger(name: str) -> logging.Logger:
+def _get_default_logging_level() -> "logging._Level":
     r"""
-    Gets a standard logger with a stream hander to stdout.
+    Returns the default logging level.
     """
-    formatter = logging.Formatter(
-        fmt="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S"
-    )
-    handler = logging.StreamHandler(sys.stdout)
-    handler.setFormatter(formatter)
+    env_level_str = os.environ.get("LLAMAFACTORY_VERBOSITY", None)
+    if env_level_str:
+        if env_level_str.upper() in logging._nameToLevel:
+            return logging._nameToLevel[env_level_str.upper()]
+        else:
+            raise ValueError("Unknown logging level: {}.".format(env_level_str))
 
-    logger = logging.getLogger(name)
-    logger.setLevel(logging.INFO)
-    logger.addHandler(handler)
-
-    return logger
+    return _default_log_level
 
 
-def reset_logging() -> None:
+def _get_library_name() -> str:
+    return __name__.split(".")[0]
+
+
+def _get_library_root_logger() -> "logging.Logger":
+    return logging.getLogger(_get_library_name())
+
+
+def _configure_library_root_logger() -> None:
     r"""
-    Removes basic config of root logger. (unused in script)
+    Configures root logger using a stdout stream handler with an explicit format.
     """
-    root = logging.getLogger()
-    list(map(root.removeHandler, root.handlers))
-    list(map(root.removeFilter, root.filters))
+    global _default_handler
+
+    with _thread_lock:
+        if _default_handler:
+            return
+
+        formatter = logging.Formatter(
+            fmt="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+            datefmt="%m/%d/%Y %H:%M:%S",
+        )
+        _default_handler = logging.StreamHandler(sys.stdout)
+        _default_handler.setFormatter(formatter)
+        library_root_logger = _get_library_root_logger()
+        library_root_logger.addHandler(_default_handler)
+        library_root_logger.setLevel(_get_default_logging_level())
+        library_root_logger.propagate = False
+
+
+def get_logger(name: Optional[str] = None) -> "logging.Logger":
+    r"""
+    Returns a logger with the specified name. It it not supposed to be accessed externally.
+    """
+    if name is None:
+        name = _get_library_name()
+
+    _configure_library_root_logger()
+    return logging.getLogger(name)

src/llamafactory/extras/misc.py

@@ -1,4 +1,7 @@
-# Copyright 2024 the LlamaFactory team.
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.10.0/src/peft/peft_model.py
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,15 +17,13 @@
 
 import gc
 import os
-from typing import TYPE_CHECKING, Dict, Tuple
+from typing import TYPE_CHECKING, Tuple, Union
 
 import torch
-from peft import PeftModel
-from transformers import InfNanRemoveLogitsProcessor, LogitsProcessorList, PreTrainedModel
+import transformers.dynamic_module_utils
+from transformers import InfNanRemoveLogitsProcessor, LogitsProcessorList
+from transformers.dynamic_module_utils import get_relative_imports
 from transformers.utils import (
-    SAFE_WEIGHTS_NAME,
-    WEIGHTS_NAME,
-    is_safetensors_available,
     is_torch_bf16_gpu_available,
     is_torch_cuda_available,
     is_torch_mps_available,
@@ -31,24 +32,18 @@ from transformers.utils import (
 )
 from transformers.utils.versions import require_version
 
-from .constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from .logging import get_logger
 
 
-if is_safetensors_available():
-    from safetensors import safe_open
-    from safetensors.torch import save_file
-
-
 _is_fp16_available = is_torch_npu_available() or is_torch_cuda_available()
 try:
-    _is_bf16_available = is_torch_bf16_gpu_available()
+    _is_bf16_available = is_torch_bf16_gpu_available() or (is_torch_npu_available() and torch.npu.is_bf16_supported())
 except Exception:
     _is_bf16_available = False
 
 
 if TYPE_CHECKING:
-    from trl import AutoModelForCausalLMWithValueHead
+    from numpy.typing import NDArray
 
     from ..hparams import ModelArguments
 
@@ -78,17 +73,20 @@ class AverageMeter:
 
 
 def check_dependencies() -> None:
+    r"""
+    Checks the version of the required packages.
+    """
     if os.environ.get("DISABLE_VERSION_CHECK", "0").lower() in ["true", "1"]:
         logger.warning("Version checking has been disabled, may lead to unexpected behaviors.")
     else:
-        require_version("transformers>=4.41.2", "To fix: pip install transformers>=4.41.2")
-        require_version("datasets>=2.16.0", "To fix: pip install datasets>=2.16.0")
-        require_version("accelerate>=0.30.1", "To fix: pip install accelerate>=0.30.1")
-        require_version("peft>=0.11.1", "To fix: pip install peft>=0.11.1")
-        require_version("trl>=0.8.6", "To fix: pip install trl>=0.8.6")
+        require_version("transformers>=4.41.2,<=4.45.0", "To fix: pip install transformers>=4.41.2,<=4.45.0")
+        require_version("datasets>=2.16.0,<=2.21.0", "To fix: pip install datasets>=2.16.0,<=2.21.0")
+        require_version("accelerate>=0.30.1,<=0.33.0", "To fix: pip install accelerate>=0.30.1,<=0.33.0")
+        require_version("peft>=0.11.1,<=0.12.0", "To fix: pip install peft>=0.11.1,<=0.12.0")
+        require_version("trl>=0.8.6,<=0.9.6", "To fix: pip install trl>=0.8.6,<=0.9.6")
 
 
-def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
+def count_parameters(model: "torch.nn.Module") -> Tuple[int, int]:
     r"""
     Returns the number of trainable parameters and number of all parameters in the model.
     """
@@ -99,7 +97,7 @@ def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
         if num_params == 0 and hasattr(param, "ds_numel"):
             num_params = param.ds_numel
 
-        # Due to the design of 4bit linear layers from bitsandbytes, multiply the number of parameters by 2
+        # Due to the design of 4bit linear layers from bitsandbytes, multiply the number of parameters by itemsize
         if param.__class__.__name__ == "Params4bit":
             if hasattr(param, "quant_storage") and hasattr(param.quant_storage, "itemsize"):
                 num_bytes = param.quant_storage.itemsize
@@ -117,52 +115,7 @@ def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
     return trainable_params, all_param
 
 
-def fix_valuehead_checkpoint(
-    model: "AutoModelForCausalLMWithValueHead", output_dir: str, safe_serialization: bool
-) -> None:
-    r"""
-    The model is already unwrapped.
-
-    There are three cases:
-    1. full tuning without ds_zero3: state_dict = {"model.layers.*": ..., "v_head.summary.*": ...}
-    2. lora tuning without ds_zero3: state_dict = {"v_head.summary.*": ...}
-    3. under deepspeed zero3: state_dict = {"pretrained_model.model.layers.*": ..., "v_head.summary.*": ...}
-
-    We assume `stage3_gather_16bit_weights_on_model_save=true`.
-    """
-    if not isinstance(model.pretrained_model, (PreTrainedModel, PeftModel)):
-        return
-
-    if safe_serialization:
-        path_to_checkpoint = os.path.join(output_dir, SAFE_WEIGHTS_NAME)
-        with safe_open(path_to_checkpoint, framework="pt", device="cpu") as f:
-            state_dict: Dict[str, torch.Tensor] = {key: f.get_tensor(key) for key in f.keys()}
-    else:
-        path_to_checkpoint = os.path.join(output_dir, WEIGHTS_NAME)
-        state_dict: Dict[str, torch.Tensor] = torch.load(path_to_checkpoint, map_location="cpu")
-
-    decoder_state_dict = {}
-    v_head_state_dict = {}
-    for name, param in state_dict.items():
-        if name.startswith("v_head."):
-            v_head_state_dict[name] = param
-        else:
-            decoder_state_dict[name.replace("pretrained_model.", "")] = param
-
-    os.remove(path_to_checkpoint)
-    model.pretrained_model.save_pretrained(
-        output_dir, state_dict=decoder_state_dict or None, safe_serialization=safe_serialization
-    )
-
-    if safe_serialization:
-        save_file(v_head_state_dict, os.path.join(output_dir, V_HEAD_SAFE_WEIGHTS_NAME), metadata={"format": "pt"})
-    else:
-        torch.save(v_head_state_dict, os.path.join(output_dir, V_HEAD_WEIGHTS_NAME))
-
-    logger.info("Value head model saved at: {}".format(output_dir))
-
-
-def get_current_device() -> torch.device:
+def get_current_device() -> "torch.device":
     r"""
     Gets the current available device.
     """
@@ -184,7 +137,9 @@ def get_device_count() -> int:
     r"""
     Gets the number of available GPU or NPU devices.
     """
-    if is_torch_npu_available():
+    if is_torch_xpu_available():
+        return torch.xpu.device_count()
+    elif is_torch_npu_available():
         return torch.npu.device_count()
     elif is_torch_cuda_available():
         return torch.cuda.device_count()
@@ -201,7 +156,26 @@ def get_logits_processor() -> "LogitsProcessorList":
     return logits_processor
 
 
-def infer_optim_dtype(model_dtype: torch.dtype) -> torch.dtype:
+def get_peak_memory() -> Tuple[int, int]:
+    r"""
+    Gets the peak memory usage for the current device (in Bytes).
+    """
+    if is_torch_npu_available():
+        return torch.npu.max_memory_allocated(), torch.npu.max_memory_reserved()
+    elif is_torch_cuda_available():
+        return torch.cuda.max_memory_allocated(), torch.cuda.max_memory_reserved()
+    else:
+        return 0, 0
+
+
+def has_tokenized_data(path: "os.PathLike") -> bool:
+    r"""
+    Checks if the path has a tokenized dataset.
+    """
+    return os.path.isdir(path) and len(os.listdir(path)) > 0
+
+
+def infer_optim_dtype(model_dtype: "torch.dtype") -> "torch.dtype":
     r"""
     Infers the optimal dtype according to the model_dtype and device compatibility.
     """
@@ -220,11 +194,26 @@ def is_gpu_or_npu_available() -> bool:
     return is_torch_npu_available() or is_torch_cuda_available()
 
 
-def has_tokenized_data(path: os.PathLike) -> bool:
+def numpify(inputs: Union["NDArray", "torch.Tensor"]) -> "NDArray":
     r"""
-    Checks if the path has a tokenized dataset.
+    Casts a torch tensor or a numpy array to a numpy array.
     """
-    return os.path.isdir(path) and len(os.listdir(path)) > 0
+    if isinstance(inputs, torch.Tensor):
+        inputs = inputs.cpu()
+        if inputs.dtype == torch.bfloat16:  # numpy does not support bfloat16 until 1.21.4
+            inputs = inputs.to(torch.float32)
+
+        inputs = inputs.numpy()
+
+    return inputs
+
+
+def skip_check_imports() -> None:
+    r"""
+    Avoids flash attention import error in custom model files.
+    """
+    if os.environ.get("FORCE_CHECK_IMPORTS", "0").lower() not in ["true", "1"]:
+        transformers.dynamic_module_utils.check_imports = get_relative_imports
 
 
 def torch_gc() -> None: