[model] fix vlm utils (#8388 )

[data] fix qwen2vl pos ids (#8387 )
[version] release v0.9.3 (#8386 )
2025-06-17 01:08:49 +08:00 · 2025-06-17 00:48:54 +08:00 · 2025-06-16 19:21:32 +08:00 · 2025-06-16 18:23:22 +08:00 · 2025-06-12 16:10:38 +08:00 · 2025-06-10 15:30:58 +08:00
218 changed files with 7334 additions and 4564 deletions
--- a/.dockerignore
+++ b/.dockerignore
@@ -3,12 +3,12 @@
 .github
 .venv
 cache
-data
 docker
 saves
 hf_cache
 ms_cache
 om_cache
+shared_data
 output
 .dockerignore
 .gitattributes
--- a/.env.local
+++ b/.env.local
@@ -16,6 +16,8 @@ USE_MODELSCOPE_HUB=
 USE_OPENMIND_HUB=
 USE_RAY=
 RECORD_VRAM=
+OPTIM_TORCH=
+NPU_JIT_COMPILE=
 # torchrun
 FORCE_TORCHRUN=
 MASTER_ADDR=
--- a/.github/ISSUE_TEMPLATE/1-bug-report.yml
+++ b/.github/ISSUE_TEMPLATE/1-bug-report.yml
@@ -12,7 +12,7 @@ body:
    attributes:
      value: |
        Please do not create issues that are not related to framework bugs under this category, use **[Discussions](https://github.com/hiyouga/LLaMA-Factory/discussions/categories/q-a)** instead.
-        请勿在此分类下创建和框架 bug 无关的 issues，请使用 **[讨论区](https://github.com/hiyouga/LLaMA-Factory/discussions/categories/q-a)**。
+        请勿在此分类下创建和框架 bug 无关的 issues，训练问题求助请使用 **[讨论区](https://github.com/hiyouga/LLaMA-Factory/discussions/categories/q-a)**。

  - type: checkboxes
    id: reminder
@@ -47,8 +47,6 @@ body:
      description: |
        Please provide entry arguments, error messages and stack traces that reproduces the problem.
        请提供入口参数，错误日志以及异常堆栈以便于我们复现问题。
-        Remember to wrap your log messages with \`\`\`.
-        请务必使用 Markdown 标签 \`\`\` 来包裹您的日志信息。

      value: |
        ```text
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -0,0 +1,66 @@
+name: docker
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - "main"
+    paths:
+      - "**/*.py"
+      - "requirements.txt"
+      - "docker/**"
+      - ".github/workflows/*.yml"
+  pull_request:
+    branches:
+      - "main"
+    paths:
+      - "**/*.py"
+      - "requirements.txt"
+      - "docker/**"
+      - ".github/workflows/*.yml"
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+
+    concurrency:
+      group: ${{ github.workflow }}-${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+
+    environment:
+      name: docker
+      url: https://hub.docker.com/r/hiyouga/llamafactory
+
+    steps:
+      - name: Free up disk space
+        run: |
+          df -h
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /opt/hostedtoolcache
+          df -h
+
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Login to Docker Hub
+        if: github.event_name != 'pull_request'
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Build and push Docker image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: ./docker/docker-cuda/Dockerfile
+          build-args: |
+            EXTRAS=metrics,deepspeed,liger-kernel
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: docker.io/hiyouga/llamafactory:latest
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -1,6 +1,7 @@
 name: publish

 on:
+  workflow_dispatch:
  release:
    types:
      - published
@@ -27,14 +28,9 @@ jobs:
        with:
          python-version: "3.9"

-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          python -m pip install build
-
      - name: Build package
        run: |
-          python -m build
+          make build

      - name: Publish package
        uses: pypa/gh-action-pypi-publish@release/v1
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -1,6 +1,7 @@
 name: tests

 on:
+  workflow_dispatch:
  push:
    branches:
      - "main"
@@ -21,7 +22,7 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        python-version:
+        python:
          - "3.9"
          - "3.10"
          - "3.11"
@@ -30,9 +31,25 @@ jobs:
          - "ubuntu-latest"
          - "windows-latest"
          - "macos-13"
+        transformers:
+          - null
+        include:  # test backward compatibility
+          - python: "3.9"
+            os: "ubuntu-latest"
+            transformers: "4.45.0"
+          - python: "3.9"
+            os: "ubuntu-latest"
+            transformers: "4.49.0"
+          - python: "3.9"
+            os: "ubuntu-latest"
+            transformers: "4.51.0"

    runs-on: ${{ matrix.os }}

+    concurrency:
+      group: ${{ github.workflow }}-${{ github.ref }}-${{ matrix.os }}-${{ matrix.python }}-${{ matrix.transformers }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+
    env:
      HF_TOKEN: ${{ secrets.HF_TOKEN }}
      OS_NAME: ${{ matrix.os }}
@@ -44,19 +61,42 @@ jobs:
      - name: Set up Python
        uses: actions/setup-python@v5
        with:
-          python-version: ${{ matrix.python-version }}
+          python-version: ${{ matrix.python }}
          cache: "pip"
-          cache-dependency-path: "setup.py"
+          cache-dependency-path: "**/requirements*.txt"

      - name: Install dependencies
        run: |
          python -m pip install --upgrade pip
          python -m pip install ".[torch,dev]"

+      - name: Install transformers
+        if: ${{ matrix.transformers }}
+        run: |
+          python -m pip install "transformers==${{ matrix.transformers }}"
+
+      - name: Cache files
+        id: hf-hub-cache
+        uses: actions/cache@v4
+        with:
+          path: ${{ runner.temp }}/huggingface
+          key: huggingface-${{ matrix.os }}-${{ matrix.python }}-${{ matrix.transformers }}-${{ hashFiles('tests/version.txt') }}
+
      - name: Check quality
        run: |
          make style && make quality

+      - name: Check license
+        run: |
+          make license
+
+      - name: Check build
+        run: |
+          make build
+
      - name: Test with pytest
        run: |
          make test
+        env:
+          HF_HOME: ${{ runner.temp }}/huggingface
+          HF_HUB_OFFLINE: "${{ steps.hf-hub-cache.outputs.cache-hit == 'true' && '1' || '0' }}"
--- a/.gitignore
+++ b/.gitignore
@@ -166,8 +166,8 @@ cython_debug/
 uv.lock

 # custom .gitignore
-ms_cache/
 hf_cache/
+ms_cache/
 om_cache/
 cache/
 config/
@@ -176,3 +176,4 @@ output/
 wandb/
 swanlog/
 generated_predictions.jsonl
+predictions_score.json
--- a/7
+++ b/7
@@ -1,14 +1,17 @@
-.PHONY: build commit quality style test
+.PHONY: build commit license quality style test

 check_dirs := scripts src tests setup.py

 build:
-	pip install build && python -m build
+	pip3 install build && python3 -m build

 commit:
 	pre-commit install
 	pre-commit run --all-files

+license:
+	python3 tests/check_license.py $(check_dirs)
+
 quality:
 	ruff check $(check_dirs)
 	ruff format --check $(check_dirs)
--- a/README.md
+++ b/README.md
@@ -5,8 +5,8 @@
 [![GitHub contributors](https://img.shields.io/github/contributors/hiyouga/LLaMA-Factory?color=orange)](https://github.com/hiyouga/LLaMA-Factory/graphs/contributors)
 [![GitHub workflow](https://github.com/hiyouga/LLaMA-Factory/actions/workflows/tests.yml/badge.svg)](https://github.com/hiyouga/LLaMA-Factory/actions/workflows/tests.yml)
 [![PyPI](https://img.shields.io/pypi/v/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-349-green)](https://scholar.google.com/scholar?cites=12620864006390196564)
-[![GitHub pull request](https://img.shields.io/badge/PRs-welcome-blue)](https://github.com/hiyouga/LLaMA-Factory/pulls)
+[![Citation](https://img.shields.io/badge/citation-614-green)](https://scholar.google.com/scholar?cites=12620864006390196564)
+[![Docker Pulls](https://img.shields.io/docker/pulls/hiyouga/llamafactory)](https://hub.docker.com/r/hiyouga/llamafactory/tags)

 [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
 [![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
@@ -14,34 +14,48 @@

 [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing)
 [![Open in DSW](https://gallery.pai-ml.com/assets/open-in-dsw.svg)](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory)
-[![Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
-[![Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
-[![SageMaker](https://img.shields.io/badge/SageMaker-Open%20in%20AWS-blue)](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
+[![Open in Alaya](assets/alaya_new.svg)](https://docs.alayanew.com/docs/documents/newActivities/llamafactory/?utm_source=LLaMA-Factory)
+[![Open in Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
+[![Open in Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Open in Novita](https://img.shields.io/badge/Novita-Deploy%20Template-blue)](https://novita.ai/templates-library/105981?sharer=88115474-394e-4bda-968e-b88e123d0c47)

-<h3 align="center">
-    Easily fine-tune 100+ large language models with zero-code <a href="#quickstart">CLI</a> and <a href="#fine-tuning-with-llama-board-gui-powered-by-gradio">Web UI</a>
-</h3>
-<p align="center">
-    <picture>
-        <img alt="Github trend" src="https://trendshift.io/api/badge/repositories/4535">
-    </picture>
-</p>
+### Used by [Amazon](https://aws.amazon.com/cn/blogs/machine-learning/how-apoidea-group-enhances-visual-information-extraction-from-banking-documents-with-multimodal-models-using-llama-factory-on-amazon-sagemaker-hyperpod/), [NVIDIA](https://developer.nvidia.com/rtx/ai-toolkit), [Aliyun](https://help.aliyun.com/zh/pai/use-cases/fine-tune-a-llama-3-model-with-llama-factory), etc.

-👋 Join our [WeChat](assets/wechat.jpg) or [NPU user group](assets/wechat_npu.jpg).
+<div align="center" markdown="1">
+
+### Supporters ❤️
+
+<a href="https://warp.dev/llama-factory">
+    <img alt="Warp sponsorship" width="400" src="https://github.com/user-attachments/assets/ab8dd143-b0fd-4904-bdc5-dd7ecac94eae">
+</a>
+
+#### [Warp, the agentic terminal for developers](https://warp.dev/llama-factory)
+
+[Available for MacOS, Linux, & Windows](https://warp.dev/llama-factory)
+
+----
+
+### Easily fine-tune 100+ large language models with zero-code [CLI](#quickstart) and [Web UI](#fine-tuning-with-llama-board-gui-powered-by-gradio)
+
+![GitHub Trend](https://trendshift.io/api/badge/repositories/4535)
+
+</div>
+
+👋 Join our [WeChat group](assets/wechat.jpg), [NPU user group](assets/wechat_npu.jpg) or [Alaya NeW user group](assets/wechat_alaya.png).

 \[ English | [中文](README_zh.md) \]

 **Fine-tuning a large language model can be easy as...**

-https://github.com/user-attachments/assets/7c96b465-9df7-45f4-8053-bf03e58386d3
+https://github.com/user-attachments/assets/3991a3a8-4276-4d30-9cab-4cb0c4b9b99e

 Choose your path:

 - **Documentation**: https://llamafactory.readthedocs.io/en/latest/
 - **Colab (free)**: https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing
 - **Local machine**: Please refer to [usage](#getting-started)
- **PAI-DSW (free trial)**: [Llama3 Example](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory) | [Qwen2-VL Example](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_qwen2vl) | [DeepSeek-R1-Distill Example](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_deepseek_r1_distill_7b)
- **Amazon SageMaker**: [Blog](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
+- **PAI-DSW (free trial)**: https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory
+- **Alaya NeW (cloud GPU deal)**: https://docs.alayanew.com/docs/documents/useGuide/LLaMAFactory/mutiple/?utm_source=LLaMA-Factory

 > [!NOTE]
 > Except for the above links, all other websites are unauthorized third-party websites. Please carefully use them.
@@ -49,7 +63,7 @@ Choose your path:
 ## Table of Contents

 - [Features](#features)
- [Benchmark](#benchmark)
+- [Blogs](#blogs)
 - [Changelog](#changelog)
 - [Supported Models](#supported-models)
 - [Supported Training Approaches](#supported-training-approaches)
@@ -76,51 +90,67 @@ Choose your path:
 - **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Qwen2-VL, DeepSeek, Yi, Gemma, ChatGLM, Phi, etc.
 - **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc.
 - **Scalable resources**: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ.
- **Advanced algorithms**: [GaLore](https://github.com/jiaweizzhao/GaLore), [BAdam](https://github.com/Ledzy/BAdam), [APOLLO](https://github.com/zhuhanqing/APOLLO), [Adam-mini](https://github.com/zyushun/Adam-mini), DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ and PiSSA.
+- **Advanced algorithms**: [GaLore](https://github.com/jiaweizzhao/GaLore), [BAdam](https://github.com/Ledzy/BAdam), [APOLLO](https://github.com/zhuhanqing/APOLLO), [Adam-mini](https://github.com/zyushun/Adam-mini), [Muon](https://github.com/KellerJordan/Muon), DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ and PiSSA.
 - **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.
 - **Wide tasks**: Multi-turn dialogue, tool using, image understanding, visual grounding, video recognition, audio understanding, etc.
- **Experiment monitors**: LlamaBoard, TensorBoard, Wandb, MLflow, SwanLab, etc.
- **Faster inference**: OpenAI-style API, Gradio UI and CLI with vLLM worker.
+- **Experiment monitors**: LlamaBoard, TensorBoard, Wandb, MLflow, [SwanLab](https://github.com/SwanHubX/SwanLab), etc.
+- **Faster inference**: OpenAI-style API, Gradio UI and CLI with [vLLM worker](https://github.com/vllm-project/vllm) or [SGLang worker](https://github.com/sgl-project/sglang).

 ### Day-N Support for Fine-Tuning Cutting-Edge Models

-| Support Date | Model Name                                                 |
-| ------------ | ---------------------------------------------------------- |
-| Day 0        | Qwen2.5 / Qwen2-VL / QwQ / QvQ / InternLM3 / MiniCPM-o-2.6 |
-| Day 1        | Llama 3 / GLM-4 / Mistral Small / PaliGemma2               |
+| Support Date | Model Name                                                   |
+| ------------ | ------------------------------------------------------------ |
+| Day 0        | Qwen3 / Qwen2.5-VL / Gemma 3 / InternLM 3 / MiniCPM-o-2.6    |
+| Day 1        | Llama 3 / GLM-4 / Mistral Small / PaliGemma2 / Llama 4       |

-## Benchmark
+## Blogs

-Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning), LLaMA Factory's LoRA tuning offers up to **3.7 times faster** training speed with a better Rouge score on the advertising text generation task. By leveraging 4-bit quantization technique, LLaMA Factory's QLoRA further improves the efficiency regarding the GPU memory.
+- [Fine-tune Qwen2.5-VL for Autonomous Driving using LLaMA-Factory](https://docs.alayanew.com/docs/documents/useGuide/LLaMAFactory/mutiple/?utm_source=LLaMA-Factory) (Chinese)
+- [How Apoidea Group enhances visual information extraction from banking documents with multimodal models using LLaMA-Factory on Amazon SageMaker HyperPod](https://aws.amazon.com/cn/blogs/machine-learning/how-apoidea-group-enhances-visual-information-extraction-from-banking-documents-with-multimodal-models-using-llama-factory-on-amazon-sagemaker-hyperpod/) (English)
+- [Easy Dataset × LLaMA Factory: Enabling LLMs to Efficiently Learn Domain Knowledge](https://buaa-act.feishu.cn/wiki/GVzlwYcRFiR8OLkHbL6cQpYin7g) (English)

-![benchmark](assets/benchmark.svg)
+<details><summary>All Blogs</summary>

-<details><summary>Definitions</summary>
-
- **Training Speed**: the number of training samples processed per second during the training. (bs=4, cutoff_len=1024)
- **Rouge Score**: Rouge-2 score on the development set of the [advertising text generation](https://aclanthology.org/D19-1321.pdf) task. (bs=4, cutoff_len=1024)
- **GPU Memory**: Peak GPU memory usage in 4-bit quantized training. (bs=1, cutoff_len=1024)
- We adopt `pre_seq_len=128` for ChatGLM's P-Tuning and `lora_rank=32` for LLaMA Factory's LoRA tuning.
+- [LLaMA Factory: Fine-tuning the DeepSeek-R1-Distill-Qwen-7B Model for News Classifier](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_deepseek_r1_distill_7b) (Chinese)
+- [A One-Stop Code-Free Model Fine-Tuning \& Deployment Platform based on SageMaker and LLaMA-Factory](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/) (Chinese)
+- [LLaMA Factory Multi-Modal Fine-Tuning Practice: Fine-Tuning Qwen2-VL for Personal Tourist Guide](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_qwen2vl) (Chinese)
+- [LLaMA Factory: Fine-tuning the LLaMA3 Model for Role-Playing](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory) (Chinese)

 </details>

 ## Changelog

+[25/04/28] We supported fine-tuning the **[Qwen3](https://qwenlm.github.io/blog/qwen3/)** model family.
+
+[25/04/21] We supported the **[Muon](https://github.com/KellerJordan/Muon)** optimizer. See [examples](examples/README.md) for usage. Thank [@tianshijing](https://github.com/tianshijing)'s PR.
+
+[25/04/16] We supported fine-tuning the **[InternVL3](https://huggingface.co/OpenGVLab/InternVL3-8B)** model. See [PR #7258](https://github.com/hiyouga/LLaMA-Factory/pull/7258) to get started.
+
+[25/04/14] We supported fine-tuning the **[GLM-Z1](https://huggingface.co/THUDM/GLM-Z1-9B-0414)** and **[Kimi-VL](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct)** models.
+
+[25/04/06] We supported fine-tuning the **[Llama 4](https://ai.meta.com/blog/llama-4-multimodal-intelligence/)** model. See [PR #7611](https://github.com/hiyouga/LLaMA-Factory/pull/7611) to get started.
+
+<details><summary>Full Changelog</summary>
+
+[25/03/31] We supported fine-tuning the **[Qwen2.5 Omni](https://qwenlm.github.io/blog/qwen2.5-omni/)** model. See [PR #7537](https://github.com/hiyouga/LLaMA-Factory/pull/7537) to get started.
+
+[25/03/15] We supported **[SGLang](https://github.com/sgl-project/sglang)** as inference backend. Try `infer_backend: sglang` to accelerate inference.
+
+[25/03/12] We supported fine-tuning the **[Gemma 3](https://huggingface.co/blog/gemma3)** model.
+
 [25/02/24] Announcing **[EasyR1](https://github.com/hiyouga/EasyR1)**, an efficient, scalable and multi-modality RL training framework for efficient GRPO training.

 [25/02/11] We supported saving the **[Ollama](https://github.com/ollama/ollama)** modelfile when exporting the model checkpoints. See [examples](examples/README.md) for usage.

 [25/02/05] We supported fine-tuning the **[Qwen2-Audio](Qwen/Qwen2-Audio-7B-Instruct)** and **[MiniCPM-o-2.6](https://huggingface.co/openbmb/MiniCPM-o-2_6)** on audio understanding tasks.

-[25/01/31] We supported fine-tuning the **[DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)** and **[Qwen2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)** model.
-
-<details><summary>Full Changelog</summary>
+[25/01/31] We supported fine-tuning the **[DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)** and **[Qwen2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)** models.

 [25/01/15] We supported **[APOLLO](https://arxiv.org/abs/2412.05270)** optimizer. See [examples](examples/README.md) for usage.

 [25/01/14] We supported fine-tuning the **[MiniCPM-o-2.6](https://huggingface.co/openbmb/MiniCPM-o-2_6)** and **[MiniCPM-V-2.6](https://huggingface.co/openbmb/MiniCPM-V-2_6)** models. Thank [@BUAADreamer](https://github.com/BUAADreamer)'s PR.

-[25/01/14] We supported fine-tuning the **[InternLM3](https://huggingface.co/collections/internlm/)** models. Thank [@hhaAndroid](https://github.com/hhaAndroid)'s PR.
+[25/01/14] We supported fine-tuning the **[InternLM 3](https://huggingface.co/collections/internlm/)** models. Thank [@hhaAndroid](https://github.com/hhaAndroid)'s PR.

 [25/01/10] We supported fine-tuning the **[Phi-4](https://huggingface.co/microsoft/phi-4)** model.

@@ -216,6 +246,9 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/

 </details>

+> [!TIP]
+> If you cannot use the latest feature, please pull the latest code and install LLaMA-Factory again.
+
 ## Supported Models

 | Model                                                             | Model size                       | Template            |
@@ -226,22 +259,28 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 | [Command R](https://huggingface.co/CohereForAI)                   | 35B/104B                         | cohere              |
 | [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)         | 7B/16B/67B/236B                  | deepseek            |
 | [DeepSeek 2.5/3](https://huggingface.co/deepseek-ai)              | 236B/671B                        | deepseek3           |
-| [DeepSeek R1 (Distill)](https://huggingface.co/deepseek-ai)       | 1.5B/7B/8B/14B/32B/70B/671B      | deepseek3           |
+| [DeepSeek R1 (Distill)](https://huggingface.co/deepseek-ai)       | 1.5B/7B/8B/14B/32B/70B/671B      | deepseekr1          |
 | [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                |
+| [Gemma 3](https://huggingface.co/google)                          | 1B/4B/12B/27B                    | gemma3/gemma (1B)   |
+| [GLM-4/GLM-4-0414/GLM-Z1](https://huggingface.co/THUDM)           | 9B/32B                           | glm4/glmz1          |
 | [GPT-2](https://huggingface.co/openai-community)                  | 0.1B/0.4B/0.8B/1.5B              | -                   |
-| [Granite 3.0-3.1](https://huggingface.co/ibm-granite)             | 1B/2B/3B/8B                      | granite3            |
+| [Granite 3.0-3.3](https://huggingface.co/ibm-granite)             | 1B/2B/3B/8B                      | granite3            |
+| [Hunyuan](https://huggingface.co/tencent/)                        | 7B                               | hunyuan             |
 | [Index](https://huggingface.co/IndexTeam)                         | 1.9B                             | index               |
 | [InternLM 2-3](https://huggingface.co/internlm)                   | 7B/8B/20B                        | intern2             |
+| [InternVL 2.5-3](https://huggingface.co/OpenGVLab)                | 1B/2B/8B/14B/38B/78B             | intern_vl           |
+| [Kimi-VL](https://huggingface.co/moonshotai)                      | 16B                              | kimi_vl             |
 | [Llama](https://github.com/facebookresearch/llama)                | 7B/13B/33B/65B                   | -                   |
 | [Llama 2](https://huggingface.co/meta-llama)                      | 7B/13B/70B                       | llama2              |
 | [Llama 3-3.3](https://huggingface.co/meta-llama)                  | 1B/3B/8B/70B                     | llama3              |
+| [Llama 4](https://huggingface.co/meta-llama)                      | 109B/402B                        | llama4              |
 | [Llama 3.2 Vision](https://huggingface.co/meta-llama)             | 11B/90B                          | mllama              |
 | [LLaVA-1.5](https://huggingface.co/llava-hf)                      | 7B/13B                           | llava               |
 | [LLaVA-NeXT](https://huggingface.co/llava-hf)                     | 7B/8B/13B/34B/72B/110B           | llava_next          |
 | [LLaVA-NeXT-Video](https://huggingface.co/llava-hf)               | 7B/34B                           | llava_next_video    |
-| [MiniCPM](https://huggingface.co/openbmb)                         | 1B/2B/4B                         | cpm/cpm3            |
+| [MiMo](https://huggingface.co/XiaomiMiMo)                         | 7B                               | mimo                |
+| [MiniCPM](https://huggingface.co/openbmb)                         | 0.5B/1B/2B/4B/8B                 | cpm/cpm3/cpm4       |
 | [MiniCPM-o-2.6/MiniCPM-V-2.6](https://huggingface.co/openbmb)     | 8B                               | minicpm_o/minicpm_v |
 | [Ministral/Mistral-Nemo](https://huggingface.co/mistralai)        | 8B/12B                           | ministral           |
 | [Mistral/Mixtral](https://huggingface.co/mistralai)               | 7B/8x7B/8x22B                    | mistral             |
@@ -253,9 +292,12 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 | [Phi-3-small](https://huggingface.co/microsoft)                   | 7B                               | phi_small           |
 | [Phi-4](https://huggingface.co/microsoft)                         | 14B                              | phi4                |
 | [Pixtral](https://huggingface.co/mistralai)                       | 12B                              | pixtral             |
-| [Qwen/QwQ (1-2.5) (Code/Math/MoE)](https://huggingface.co/Qwen)   | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen                |
+| [Qwen (1-2.5) (Code/Math/MoE/QwQ)](https://huggingface.co/Qwen)   | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen                |
+| [Qwen3 (MoE)](https://huggingface.co/Qwen)                        | 0.6B/1.7B/4B/8B/14B/32B/235B     | qwen3               |
 | [Qwen2-Audio](https://huggingface.co/Qwen)                        | 7B                               | qwen2_audio         |
-| [Qwen2-VL/Qwen2.5-VL/QVQ](https://huggingface.co/Qwen)            | 2B/3B/7B/72B                     | qwen2_vl            |
+| [Qwen2.5-Omni](https://huggingface.co/Qwen)                       | 3B/7B                            | qwen2_omni          |
+| [Qwen2-VL/Qwen2.5-VL/QVQ](https://huggingface.co/Qwen)            | 2B/3B/7B/32B/72B                 | qwen2_vl            |
+| [Seed Coder](https://huggingface.co/ByteDance-Seed)               | 8B                               | seed_coder          |
 | [Skywork o1](https://huggingface.co/Skywork)                      | 8B                               | skywork_o1          |
 | [StarCoder 2](https://huggingface.co/bigcode)                     | 3B/7B/15B                        | -                   |
 | [TeleChat2](https://huggingface.co/Tele-AI)                       | 3B/7B/35B/115B                   | telechat2           |
@@ -268,6 +310,10 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 > 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.
 >
 > Remember to use the **SAME** template in training and inference.
+>
+> \*: You should install the `transformers` from main branch and use `DISABLE_VERSION_CHECK=1` to skip version check.
+>
+> \*\*: You need to install a specific version of `transformers` to use the corresponding model.

 Please refer to [constants.py](src/llamafactory/extras/constants.py) for a full list of models we supported.

@@ -371,8 +417,10 @@ 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)
+- [COIG-P (zh)](https://huggingface.co/datasets/m-a-p/COIG-P)
 - [RLHF-V (en)](https://huggingface.co/datasets/openbmb/RLHF-V-Dataset)
 - [VLFeedback (en)](https://huggingface.co/datasets/Zhihui/VLFeedback)
+- [RLAIF-V (en)](https://huggingface.co/datasets/openbmb/RLAIF-V-Dataset)
 - [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)
@@ -393,11 +441,12 @@ huggingface-cli login
 | Mandatory    | Minimum | Recommend |
 | ------------ | ------- | --------- |
 | python       | 3.9     | 3.10      |
-| torch        | 1.13.1  | 2.5.1     |
-| transformers | 4.41.2  | 4.49.0    |
+| torch        | 2.0.0   | 2.6.0     |
+| torchvision  | 0.15.0  | 0.21.0    |
+| transformers | 4.45.0  | 4.50.0    |
 | datasets     | 2.16.0  | 3.2.0     |
 | accelerate   | 0.34.0  | 1.2.1     |
-| peft         | 0.11.1  | 0.12.0    |
+| peft         | 0.14.0  | 0.15.1    |
 | trl          | 0.8.6   | 0.9.6     |

 | Optional     | Minimum | Recommend |
@@ -405,8 +454,8 @@ huggingface-cli login
 | CUDA         | 11.6    | 12.2      |
 | deepspeed    | 0.10.0  | 0.16.4    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.7.3     |
-| flash-attn   | 2.3.0   | 2.7.2     |
+| vllm         | 0.4.3   | 0.8.2     |
+| flash-attn   | 2.5.6   | 2.7.2     |

 ### Hardware Requirement

@@ -428,16 +477,27 @@ huggingface-cli login
 > [!IMPORTANT]
 > Installation is mandatory.

+#### Install from Source
+
 ```bash
 git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
 cd LLaMA-Factory
-pip install -e ".[torch,metrics]"
+pip install -e ".[torch,metrics]" --no-build-isolation
 ```

-Extra dependencies available: torch, torch-npu, metrics, deepspeed, liger-kernel, bitsandbytes, hqq, eetq, gptq, awq, aqlm, vllm, galore, apollo, badam, adam-mini, qwen, minicpm_v, modelscope, openmind, swanlab, quality
+Extra dependencies available: torch, torch-npu, metrics, deepspeed, liger-kernel, bitsandbytes, hqq, eetq, gptq, aqlm, vllm, sglang, galore, apollo, badam, adam-mini, qwen, minicpm_v, modelscope, openmind, swanlab, dev

-> [!TIP]
-> Use `pip install --no-deps -e .` to resolve package conflicts.
+#### Install from Docker Image
+
+```bash
+docker run -it --rm --gpus=all --ipc=host hiyouga/llamafactory:latest
+```
+
+This image is built on Ubuntu 22.04 (x86\_64), CUDA 12.4, Python 3.11, PyTorch 2.6.0, and Flash-attn 2.7.4.
+
+Find the pre-built images: https://hub.docker.com/r/hiyouga/llamafactory/tags
+
+Please refer to [build docker](#build-docker) to build the image yourself.

 <details><summary>Setting up a virtual environment with <b>uv</b></summary>

@@ -457,6 +517,20 @@ uv run --prerelease=allow llamafactory-cli train examples/train_lora/llama3_lora

 <details><summary>For Windows users</summary>

+#### Install PyTorch
+
+You need to manually install the GPU version of PyTorch on the Windows platform. Please refer to the [official website](https://pytorch.org/get-started/locally/) and the following command to install PyTorch with CUDA support:
+
+```bash
+pip uninstall torch torchvision torchaudio
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
+python -c "import torch; print(torch.cuda.is_available())"
+```
+
+If you see `True` then you have successfully installed PyTorch with CUDA support.
+
+Try `dataloader_num_workers: 0` if you encounter `Can't pickle local object` error.
+
 #### Install BitsAndBytes

 If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you need to install a pre-built version of `bitsandbytes` library, which supports CUDA 11.1 to 12.2, please select the appropriate [release version](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels) based on your CUDA version.
@@ -495,6 +569,7 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch        | 2.1.0   | 2.4.0          |
 | torch-npu    | 2.1.0   | 2.4.0.post2    |
 | deepspeed    | 0.13.2  | 0.13.2         |
+| vllm-ascend  | -       | 0.7.3          |

 Remember to use `ASCEND_RT_VISIBLE_DEVICES` instead of `CUDA_VISIBLE_DEVICES` to specify the device to use.

@@ -540,11 +615,13 @@ pip install .

 ### Data Preparation

-Please refer to [data/README.md](data/README.md) for checking the details about the format of dataset files. You can either use datasets on HuggingFace / ModelScope / Modelers hub or load the dataset in local disk.
+Please refer to [data/README.md](data/README.md) for checking the details about the format of dataset files. You can use datasets on HuggingFace / ModelScope / Modelers hub, load the dataset in local disk, or specify a path to s3/gcs cloud storage.

 > [!NOTE]
 > Please update `data/dataset_info.json` to use your custom dataset.

+You can also use **[Easy Dataset](https://github.com/ConardLi/easy-dataset)** or **[GraphGen](https://github.com/open-sciencelab/GraphGen)** to create synthetic data for fine-tuning.
+
 ### Quickstart

 Use the following 3 commands to run LoRA **fine-tuning**, **inference** and **merging** of the Llama3-8B-Instruct model, respectively.
@@ -600,22 +677,13 @@ 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 \
+    --build-arg EXTRAS=metrics \
    -t llamafactory:latest .

-docker run -dit --gpus=all \
-    -v ./hf_cache:/root/.cache/huggingface \
-    -v ./ms_cache:/root/.cache/modelscope \
-    -v ./om_cache:/root/.cache/openmind \
-    -v ./data:/app/data \
-    -v ./output:/app/output \
+docker run -dit --ipc=host --gpus=all \
    -p 7860:7860 \
    -p 8000:8000 \
-    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

@@ -625,19 +693,12 @@ docker exec -it llamafactory bash
 For Ascend NPU users:

 ```bash
-# Choose docker image upon your environment
 docker build -f ./docker/docker-npu/Dockerfile \
-    --build-arg INSTALL_DEEPSPEED=false \
    --build-arg PIP_INDEX=https://pypi.org/simple \
+    --build-arg EXTRAS=torch-npu,metrics \
    -t llamafactory:latest .

-# Change `device` upon your resources
-docker run -dit \
-    -v ./hf_cache:/root/.cache/huggingface \
-    -v ./ms_cache:/root/.cache/modelscope \
-    -v ./om_cache:/root/.cache/openmind \
-    -v ./data:/app/data \
-    -v ./output:/app/output \
+docker run -dit --ipc=host \
    -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 \
@@ -648,7 +709,6 @@ docker run -dit \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
-    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

@@ -659,25 +719,15 @@ 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 \
+    --build-arg EXTRAS=metrics \
    -t llamafactory:latest .

-docker run -dit \
-    -v ./hf_cache:/root/.cache/huggingface \
-    -v ./ms_cache:/root/.cache/modelscope \
-    -v ./om_cache:/root/.cache/openmind \
-    -v ./data:/app/data \
-    -v ./output:/app/output \
-    -v ./saves:/app/saves \
+docker run -dit --ipc=host \
    -p 7860:7860 \
    -p 8000:8000 \
    --device /dev/kfd \
    --device /dev/dri \
-    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

@@ -686,12 +736,14 @@ docker exec -it llamafactory bash

 </details>

-<details><summary>Details about volume</summary>
+<details><summary>Use Docker volumes</summary>

- `hf_cache`: Utilize Hugging Face cache on the host machine. Reassignable if a cache already exists in a different directory.
- `ms_cache`: Similar to Hugging Face cache but for ModelScope users.
- `om_cache`: Similar to Hugging Face cache but for Modelers users.
- `data`: Place datasets on this dir of the host machine so that they can be selected on LLaMA Board GUI.
+You can uncomment `VOLUME [ "/root/.cache/huggingface", "/app/shared_data", "/app/output" ]` in the Dockerfile to use data volumes.
+
+When building the Docker image, use `-v ./hf_cache:/root/.cache/huggingface` argument to mount the local directory to the container. The following data volumes are available.
+
+- `hf_cache`: Utilize Hugging Face cache on the host machine.
+- `shared_data`: The directionary to store datasets on the host machine.
 - `output`: Set export dir to this location so that the merged result can be accessed directly on the host machine.

 </details>
@@ -699,7 +751,7 @@ docker exec -it llamafactory bash
 ### Deploy with OpenAI-style API and vLLM

 ```bash
-API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
+API_PORT=8000 llamafactory-cli api examples/inference/llama3.yaml infer_backend=vllm vllm_enforce_eager=true
 ```

 > [!TIP]
@@ -841,6 +893,7 @@ If you have a project that should be incorporated, please contact via email or c
 1. Xia et al. Using Pre-trained Language Model for Accurate ESG Prediction. FinNLP 2024. [[paper]](https://aclanthology.org/2024.finnlp-2.1/)
 1. Liang et al. I-SHEEP: Self-Alignment of LLM from Scratch through an Iterative Self-Enhancement Paradigm. 2024. [[arxiv]](https://arxiv.org/abs/2408.08072)
 1. Bai et al. Aligning Large Language Model with Direct Multi-Preference Optimization for Recommendation. CIKM 2024. [[paper]](https://dl.acm.org/doi/10.1145/3627673.3679611)
+1. Zhang et al. CPsyCoun: A Report-based Multi-turn Dialogue Reconstruction and Evaluation Framework for Chinese Psychological Counseling. ACL 2024. [[paper]](https://aclanthology.org/2024.findings-acl.830.pdf)
 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.
@@ -854,14 +907,15 @@ If you have a project that should be incorporated, please contact via email or c
 1. **[RAG-Retrieval](https://github.com/NLPJCL/RAG-Retrieval)**: A full pipeline for RAG retrieval model fine-tuning, inference, and distillation. [[blog]](https://zhuanlan.zhihu.com/p/987727357)
 1. **[360-LLaMA-Factory](https://github.com/Qihoo360/360-LLaMA-Factory)**: A modified library that supports long sequence SFT & DPO using ring attention.
 1. **[Sky-T1](https://novasky-ai.github.io/posts/sky-t1/)**: An o1-like model fine-tuned by NovaSky AI with very small cost.
-
+1. **[WeClone](https://github.com/xming521/WeClone)**: One-stop solution for creating your digital avatar from chat logs.
+1. **[EmoLLM](https://github.com/SmartFlowAI/EmoLLM)**: A project about large language models (LLMs) and mental health.
 </details>

 ## License

 This repository is licensed under the [Apache-2.0 License](LICENSE).

-Please follow the model licenses to use the corresponding model weights: [Baichuan 2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [GPT-2](https://github.com/openai/gpt-2/blob/master/LICENSE) / [Granite](LICENSE) / [Index](https://huggingface.co/IndexTeam/Index-1.9B/blob/main/LICENSE) / [InternLM](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [MiniCPM](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) / [Mistral/Mixtral/Pixtral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3/Phi-4](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [Skywork](https://huggingface.co/Skywork/Skywork-13B-base/blob/main/Skywork%20Community%20License.pdf) / [StarCoder 2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [TeleChat2](https://huggingface.co/Tele-AI/telechat-7B/blob/main/TeleChat%E6%A8%A1%E5%9E%8B%E7%A4%BE%E5%8C%BA%E8%AE%B8%E5%8F%AF%E5%8D%8F%E8%AE%AE.pdf) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yi-1.5](LICENSE) / [Yuan 2](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+Please follow the model licenses to use the corresponding model weights: [Baichuan 2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [GPT-2](https://github.com/openai/gpt-2/blob/master/LICENSE) / [Granite](LICENSE) / [Index](https://huggingface.co/IndexTeam/Index-1.9B/blob/main/LICENSE) / [InternLM](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [Llama 4](https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE) / [MiniCPM](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) / [Mistral/Mixtral/Pixtral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3/Phi-4](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [Skywork](https://huggingface.co/Skywork/Skywork-13B-base/blob/main/Skywork%20Community%20License.pdf) / [StarCoder 2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [TeleChat2](https://huggingface.co/Tele-AI/telechat-7B/blob/main/TeleChat%E6%A8%A1%E5%9E%8B%E7%A4%BE%E5%8C%BA%E8%AE%B8%E5%8F%AF%E5%8D%8F%E8%AE%AE.pdf) / [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

--- a/README_zh.md
+++ b/README_zh.md
@@ -5,8 +5,8 @@
 [![GitHub contributors](https://img.shields.io/github/contributors/hiyouga/LLaMA-Factory?color=orange)](https://github.com/hiyouga/LLaMA-Factory/graphs/contributors)
 [![GitHub workflow](https://github.com/hiyouga/LLaMA-Factory/actions/workflows/tests.yml/badge.svg)](https://github.com/hiyouga/LLaMA-Factory/actions/workflows/tests.yml)
 [![PyPI](https://img.shields.io/pypi/v/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-349-green)](https://scholar.google.com/scholar?cites=12620864006390196564)
-[![GitHub pull request](https://img.shields.io/badge/PRs-welcome-blue)](https://github.com/hiyouga/LLaMA-Factory/pulls)
+[![Citation](https://img.shields.io/badge/citation-614-green)](https://scholar.google.com/scholar?cites=12620864006390196564)
+[![Docker Pulls](https://img.shields.io/docker/pulls/hiyouga/llamafactory)](https://hub.docker.com/r/hiyouga/llamafactory/tags)

 [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
 [![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
@@ -14,36 +14,50 @@

 [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1d5KQtbemerlSDSxZIfAaWXhKr30QypiK?usp=sharing)
 [![Open in DSW](https://gallery.pai-ml.com/assets/open-in-dsw.svg)](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory)
-[![Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
-[![Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
-[![SageMaker](https://img.shields.io/badge/SageMaker-Open%20in%20AWS-blue)](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
+[![Open in Alaya](assets/alaya_new.svg)](https://docs.alayanew.com/docs/documents/newActivities/llamafactory/?utm_source=LLaMA-Factory)
+[![Open in Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
+[![Open in Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Open in Novita](https://img.shields.io/badge/Novita-Deploy%20Template-blue)](https://novita.ai/templates-library/105981?sharer=88115474-394e-4bda-968e-b88e123d0c47)

-<h3 align="center">
-    使用零代码<a href="#快速开始">命令行</a>与 <a href="#llama-board-可视化微调由-gradio-驱动">Web UI</a> 轻松微调百余种大模型
-</h3>
-<p align="center">
-    <picture>
-        <img alt="Github trend" src="https://trendshift.io/api/badge/repositories/4535">
-    </picture>
-</p>
+### 获得[亚马逊](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)、[英伟达](https://developer.nvidia.cn/rtx/ai-toolkit)、[阿里云](https://help.aliyun.com/zh/pai/use-cases/fine-tune-a-llama-3-model-with-llama-factory)等的应用。

+<div align="center" markdown="1">

-👋 加入我们的[微信群](assets/wechat.jpg)或 [NPU 用户群](assets/wechat_npu.jpg)。
+### 赞助商 ❤️
+
+<a href="https://warp.dev/llama-factory">
+    <img alt="Warp sponsorship" width="400" src="https://github.com/user-attachments/assets/ab8dd143-b0fd-4904-bdc5-dd7ecac94eae">
+</a>
+
+#### [Warp，面向开发者的智能终端](https://warp.dev/llama-factory)
+
+[适用于 MacOS、Linux 和 Windows](https://warp.dev/llama-factory)
+
+----
+
+### 使用零代码[命令行](#快速开始)与 [Web UI](#llama-board-可视化微调由-gradio-驱动) 轻松微调百余种大模型
+
+![GitHub Trend](https://trendshift.io/api/badge/repositories/4535)
+
+</div>
+
+👋 加入我们的[微信群](assets/wechat.jpg)、[NPU 用户群](assets/wechat_npu.jpg)或 [九章智算云算力优惠群](assets/wechat_alaya.png)。

 \[ [English](README.md) | 中文 \]

 **微调大模型可以像这样轻松…**

-https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
+https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc

 选择你的打开方式：

 - **入门教程**：https://zhuanlan.zhihu.com/p/695287607
 - **框架文档**：https://llamafactory.readthedocs.io/zh-cn/latest/
+- **框架文档（昇腾 NPU）**：https://ascend.github.io/docs/sources/llamafactory/
 - **Colab（免费）**：https://colab.research.google.com/drive/1d5KQtbemerlSDSxZIfAaWXhKr30QypiK?usp=sharing
 - **本地机器**：请见[如何使用](#如何使用)
- **PAI-DSW（免费试用）**：[Llama3 案例](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory) | [Qwen2-VL 案例](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_qwen2vl) | [DeepSeek-R1-Distill 案例](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_deepseek_r1_distill_7b)
- **Amazon SageMaker**：[博客](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)
+- **PAI-DSW（免费试用）**：https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory
+- **九章智算云（算力优惠活动）**：https://docs.alayanew.com/docs/documents/useGuide/LLaMAFactory/mutiple/?utm_source=LLaMA-Factory

 > [!NOTE]
 > 除上述链接以外的其他网站均为未经许可的第三方网站，请小心甄别。
@@ -51,7 +65,7 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 ## 目录

 - [项目特色](#项目特色)
- [性能指标](#性能指标)
+- [官方博客](#官方博客)
 - [更新日志](#更新日志)
 - [模型](#模型)
 - [训练方法](#训练方法)
@@ -78,36 +92,54 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 - **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Qwen2-VL、DeepSeek、Yi、Gemma、ChatGLM、Phi 等等。
 - **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO 训练、DPO 训练、KTO 训练、ORPO 训练等等。
 - **多种精度**：16 比特全参数微调、冻结微调、LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ 的 2/3/4/5/6/8 比特 QLoRA 微调。
- **先进算法**：[GaLore](https://github.com/jiaweizzhao/GaLore)、[BAdam](https://github.com/Ledzy/BAdam)、[APOLLO](https://github.com/zhuhanqing/APOLLO)、[Adam-mini](https://github.com/zyushun/Adam-mini)、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ 和 PiSSA。
+- **先进算法**：[GaLore](https://github.com/jiaweizzhao/GaLore)、[BAdam](https://github.com/Ledzy/BAdam)、[APOLLO](https://github.com/zhuhanqing/APOLLO)、[Adam-mini](https://github.com/zyushun/Adam-mini)、[Muon](https://github.com/KellerJordan/Muon)、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ 和 PiSSA。
 - **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。
 - **广泛任务**：多轮对话、工具调用、图像理解、视觉定位、视频识别和语音理解等等。
- **实验监控**：LlamaBoard、TensorBoard、Wandb、MLflow、SwanLab 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。
+- **实验监控**：LlamaBoard、TensorBoard、Wandb、MLflow、[SwanLab](https://github.com/SwanHubX/SwanLab) 等等。
+- **极速推理**：基于 [vLLM](https://github.com/vllm-project/vllm) 或 [SGLang](https://github.com/sgl-project/sglang) 的 OpenAI 风格 API、浏览器界面和命令行接口。

 ### 最新模型的 Day-N 微调适配

-| 适配时间      | 模型名称                                                    |
-| ------------ | ---------------------------------------------------------- |
-| Day 0        | Qwen2.5 / Qwen2-VL / QwQ / QvQ / InternLM3 / MiniCPM-o-2.6 |
-| Day 1        | Llama 3 / GLM-4 / Mistral Small / PaliGemma2               |
+| 适配时间      | 模型名称                                                       |
+| ------------ | ------------------------------------------------------------ |
+| Day 0        | Qwen3 / Qwen2.5-VL / Gemma 3 / InternLM 3 / MiniCPM-o-2.6    |
+| Day 1        | Llama 3 / GLM-4 / Mistral Small / PaliGemma2 / Llama 4       |

-## 性能指标
+## 官方博客

-与 ChatGLM 官方的 [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning) 微调相比，LLaMA Factory 的 LoRA 微调提供了 **3.7 倍**的加速比，同时在广告文案生成任务上取得了更高的 Rouge 分数。结合 4 比特量化技术，LLaMA Factory 的 QLoRA 微调进一步降低了 GPU 显存消耗。
+- [使用 LLaMA-Factory 微调 Qwen2.5-VL 实现自动驾驶场景微调](https://docs.alayanew.com/docs/documents/useGuide/LLaMAFactory/mutiple/?utm_source=LLaMA-Factory)（中文）
+- [通过亚马逊 SageMaker HyperPod 上的 LLaMA-Factory 增强多模态模型银行文档的视觉信息提取](https://aws.amazon.com/cn/blogs/machine-learning/how-apoidea-group-enhances-visual-information-extraction-from-banking-documents-with-multimodal-models-using-llama-factory-on-amazon-sagemaker-hyperpod/)（英文）
+- [Easy Dataset × LLaMA Factory: 让大模型高效学习领域知识](https://buaa-act.feishu.cn/wiki/KY9xwTGs1iqHrRkjXBwcZP9WnL9)（中文）

-![benchmark](assets/benchmark.svg)
+<details><summary>全部博客</summary>

-<details><summary>变量定义</summary>
-
- **Training Speed**: 训练阶段每秒处理的样本数量。（批处理大小=4，截断长度=1024）
- **Rouge Score**: [广告文案生成](https://aclanthology.org/D19-1321.pdf)任务验证集上的 Rouge-2 分数。（批处理大小=4，截断长度=1024）
- **GPU Memory**: 4 比特量化训练的 GPU 显存峰值。（批处理大小=1，截断长度=1024）
- 我们在 ChatGLM 的 P-Tuning 中采用 `pre_seq_len=128`，在 LLaMA Factory 的 LoRA 微调中采用 `lora_rank=32`。
+- [LLaMA Factory：微调 DeepSeek-R1-Distill-Qwen-7B 模型实现新闻标题分类器](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_deepseek_r1_distill_7b)（中文）
+- [基于 Amazon SageMaker 和 LLaMA-Factory 打造一站式无代码模型微调部署平台 Model Hub](https://aws.amazon.com/cn/blogs/china/a-one-stop-code-free-model-fine-tuning-deployment-platform-based-on-sagemaker-and-llama-factory/)（中文）
+- [LLaMA Factory 多模态微调实践：微调 Qwen2-VL 构建文旅大模型](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory_qwen2vl)（中文）
+- [LLaMA Factory：微调LLaMA3模型实现角色扮演](https://gallery.pai-ml.com/#/preview/deepLearning/nlp/llama_factory)（中文）

 </details>

 ## 更新日志

+[25/04/28] 我们支持了 **[Qwen3](https://qwenlm.github.io/blog/qwen3/)** 系列模型的微调。
+
+[25/04/21] 我们支持了 **[Muon](https://github.com/KellerJordan/Muon)** 优化器。详细用法请参照 [examples](examples/README_zh.md)。感谢 [@tianshijing](https://github.com/tianshijing) 的 PR。
+
+[25/04/16] 我们支持了 **[InternVL3](https://huggingface.co/OpenGVLab/InternVL3-8B)** 模型的微调。查看 [PR #7258](https://github.com/hiyouga/LLaMA-Factory/pull/7258) 以使用。
+
+[25/04/14] 我们支持了 **[GLM-Z1](https://huggingface.co/THUDM/GLM-Z1-9B-0414)** 和 **[Kimi-VL](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct)** 模型的微调。
+
+[25/04/06] 我们支持了 **[Llama 4](https://ai.meta.com/blog/llama-4-multimodal-intelligence/)** 模型的微调。查看 [PR #7611](https://github.com/hiyouga/LLaMA-Factory/pull/7611) 以使用。
+
+<details><summary>展开日志</summary>
+
+[25/03/31] 我们支持了 **[Qwen2.5 Omni](https://qwenlm.github.io/blog/qwen2.5-omni/)** 模型的微调。查看 [PR #7537](https://github.com/hiyouga/LLaMA-Factory/pull/7537) 以使用。
+
+[25/03/15] 我们支持了 **[SGLang](https://github.com/sgl-project/sglang)** 推理后端，请使用 `infer_backend: sglang` 启用。
+
+[25/03/12] 我们支持了 **[Gemma 3](https://huggingface.co/blog/gemma3)** 模型的微调。
+
 [25/02/24] 我们宣布开源 **[EasyR1](https://github.com/hiyouga/EasyR1)**，一个高效可扩展的多模态强化学习框架，支持高效的 GRPO 训练。

 [25/02/11] 我们支持了在导出模型时保存 **[Ollama](https://github.com/ollama/ollama)** 配置文件。详细用法请参照 [examples](examples/README_zh.md)。
@@ -116,13 +148,11 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272

 [25/01/31] 我们支持了 **[DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)** 和 **[Qwen2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)** 模型的微调。

-<details><summary>展开日志</summary>
-
 [25/01/15] 我们支持了 **[APOLLO](https://arxiv.org/abs/2412.05270)** 优化器。详细用法请参照 [examples](examples/README_zh.md)。

 [25/01/14] 我们支持了 **[MiniCPM-o-2.6](https://huggingface.co/openbmb/MiniCPM-o-2_6)** 和 **[MiniCPM-V-2.6](https://huggingface.co/openbmb/MiniCPM-V-2_6)** 模型的微调。 感谢 [@BUAADreamer](https://github.com/BUAADreamer) 的 PR.

-[25/01/14] 我们支持了 **[InternLM3](https://huggingface.co/collections/internlm/)** 模型的微调。感谢 [@hhaAndroid](https://github.com/hhaAndroid) 的 PR。
+[25/01/14] 我们支持了 **[InternLM 3](https://huggingface.co/collections/internlm/)** 模型的微调。感谢 [@hhaAndroid](https://github.com/hhaAndroid) 的 PR。

 [25/01/10] 我们支持了 **[Phi-4](https://huggingface.co/microsoft/phi-4)** 模型的微调。

@@ -218,6 +248,9 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272

 </details>

+> [!TIP]
+> 如果您无法使用最新的功能，请尝试重新拉取代码并再次安装 LLaMA-Factory。
+
 ## 模型

 | 模型名                                                             | 参数量                            | Template            |
@@ -228,22 +261,28 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 | [Command R](https://huggingface.co/CohereForAI)                   | 35B/104B                         | cohere              |
 | [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)         | 7B/16B/67B/236B                  | deepseek            |
 | [DeepSeek 2.5/3](https://huggingface.co/deepseek-ai)              | 236B/671B                        | deepseek3           |
-| [DeepSeek R1 (Distill)](https://huggingface.co/deepseek-ai)       | 1.5B/7B/8B/14B/32B/70B/671B      | deepseek3           |
+| [DeepSeek R1 (Distill)](https://huggingface.co/deepseek-ai)       | 1.5B/7B/8B/14B/32B/70B/671B      | deepseekr1          |
 | [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                |
+| [Gemma 3](https://huggingface.co/google)                          | 1B/4B/12B/27B                    | gemma3/gemma (1B)   |
+| [GLM-4/GLM-4-0414/GLM-Z1](https://huggingface.co/THUDM)           | 9B/32B                           | glm4/glmz1          |
 | [GPT-2](https://huggingface.co/openai-community)                  | 0.1B/0.4B/0.8B/1.5B              | -                   |
-| [Granite 3.0-3.1](https://huggingface.co/ibm-granite)             | 1B/2B/3B/8B                      | granite3            |
+| [Granite 3.0-3.3](https://huggingface.co/ibm-granite)             | 1B/2B/3B/8B                      | granite3            |
+| [Hunyuan](https://huggingface.co/tencent/)                        | 7B                               | hunyuan             |
 | [Index](https://huggingface.co/IndexTeam)                         | 1.9B                             | index               |
 | [InternLM 2-3](https://huggingface.co/internlm)                   | 7B/8B/20B                        | intern2             |
+| [InternVL 2.5-3](https://huggingface.co/OpenGVLab)                | 1B/2B/8B/14B/38B/78B             | intern_vl           |
+| [Kimi-VL](https://huggingface.co/moonshotai)                      | 16B                              | kimi_vl             |
 | [Llama](https://github.com/facebookresearch/llama)                | 7B/13B/33B/65B                   | -                   |
 | [Llama 2](https://huggingface.co/meta-llama)                      | 7B/13B/70B                       | llama2              |
 | [Llama 3-3.3](https://huggingface.co/meta-llama)                  | 1B/3B/8B/70B                     | llama3              |
+| [Llama 4](https://huggingface.co/meta-llama)                      | 109B/402B                        | llama4              |
 | [Llama 3.2 Vision](https://huggingface.co/meta-llama)             | 11B/90B                          | mllama              |
 | [LLaVA-1.5](https://huggingface.co/llava-hf)                      | 7B/13B                           | llava               |
 | [LLaVA-NeXT](https://huggingface.co/llava-hf)                     | 7B/8B/13B/34B/72B/110B           | llava_next          |
 | [LLaVA-NeXT-Video](https://huggingface.co/llava-hf)               | 7B/34B                           | llava_next_video    |
-| [MiniCPM](https://huggingface.co/openbmb)                         | 1B/2B/4B                         | cpm/cpm3            |
+| [MiMo](https://huggingface.co/XiaomiMiMo)                         | 7B                               | mimo                |
+| [MiniCPM](https://huggingface.co/openbmb)                         | 0.5B/1B/2B/4B/8B                 | cpm/cpm3/cpm4       |
 | [MiniCPM-o-2.6/MiniCPM-V-2.6](https://huggingface.co/openbmb)     | 8B                               | minicpm_o/minicpm_v |
 | [Ministral/Mistral-Nemo](https://huggingface.co/mistralai)        | 8B/12B                           | ministral           |
 | [Mistral/Mixtral](https://huggingface.co/mistralai)               | 7B/8x7B/8x22B                    | mistral             |
@@ -255,9 +294,12 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 | [Phi-3-small](https://huggingface.co/microsoft)                   | 7B                               | phi_small           |
 | [Phi-4](https://huggingface.co/microsoft)                         | 14B                              | phi4                |
 | [Pixtral](https://huggingface.co/mistralai)                       | 12B                              | pixtral             |
-| [Qwen/QwQ (1-2.5) (Code/Math/MoE)](https://huggingface.co/Qwen)   | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen                |
+| [Qwen (1-2.5) (Code/Math/MoE/QwQ)](https://huggingface.co/Qwen)   | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen                |
+| [Qwen3 (MoE)](https://huggingface.co/Qwen)                        | 0.6B/1.7B/4B/8B/14B/32B/235B     | qwen3               |
 | [Qwen2-Audio](https://huggingface.co/Qwen)                        | 7B                               | qwen2_audio         |
-| [Qwen2-VL/Qwen2.5-VL/QVQ](https://huggingface.co/Qwen)            | 2B/3B/7B/72B                     | qwen2_vl            |
+| [Qwen2.5-Omni](https://huggingface.co/Qwen)                       | 3B/7B                            | qwen2_omni          |
+| [Qwen2-VL/Qwen2.5-VL/QVQ](https://huggingface.co/Qwen)            | 2B/3B/7B/32B/72B                 | qwen2_vl            |
+| [Seed Coder](https://huggingface.co/ByteDance-Seed)               | 8B                               | seed_coder          |
 | [Skywork o1](https://huggingface.co/Skywork)                      | 8B                               | skywork_o1          |
 | [StarCoder 2](https://huggingface.co/bigcode)                     | 3B/7B/15B                        | -                   |
 | [TeleChat2](https://huggingface.co/Tele-AI)                       | 3B/7B/35B/115B                   | telechat2           |
@@ -270,6 +312,10 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 > 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**。
 >
 > 请务必在训练和推理时采用**完全一致**的模板。
+>
+> \*：您需要从 main 分支安装 `transformers` 并使用 `DISABLE_VERSION_CHECK=1` 来跳过版本检查。
+>
+> \*\*：您需要安装特定版本的 `transformers` 以使用该模型。

 项目所支持模型的完整列表请参阅 [constants.py](src/llamafactory/extras/constants.py)。

@@ -373,8 +419,10 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272

 - [DPO mixed (en&zh)](https://huggingface.co/datasets/hiyouga/DPO-En-Zh-20k)
 - [UltraFeedback (en)](https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized)
+- [COIG-P (zh)](https://huggingface.co/datasets/m-a-p/COIG-P)
 - [RLHF-V (en)](https://huggingface.co/datasets/openbmb/RLHF-V-Dataset)
 - [VLFeedback (en)](https://huggingface.co/datasets/Zhihui/VLFeedback)
+- [RLAIF-V (en)](https://huggingface.co/datasets/openbmb/RLAIF-V-Dataset)
 - [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)
@@ -392,23 +440,24 @@ huggingface-cli login

 ## 软硬件依赖

-| 必需项       | 至少     | 推荐      |
+| 必需项        | 至少     | 推荐      |
 | ------------ | ------- | --------- |
 | python       | 3.9     | 3.10      |
-| torch        | 1.13.1  | 2.5.1     |
-| transformers | 4.41.2  | 4.49.0    |
+| torch        | 2.0.0   | 2.6.0     |
+| torchvision  | 0.15.0  | 0.21.0    |
+| transformers | 4.45.0  | 4.50.0    |
 | datasets     | 2.16.0  | 3.2.0     |
 | accelerate   | 0.34.0  | 1.2.1     |
-| peft         | 0.11.1  | 0.12.0    |
+| peft         | 0.14.0  | 0.15.1    |
 | trl          | 0.8.6   | 0.9.6     |

-| 可选项       | 至少     | 推荐      |
+| 可选项        | 至少     | 推荐      |
 | ------------ | ------- | --------- |
 | CUDA         | 11.6    | 12.2      |
 | deepspeed    | 0.10.0  | 0.16.4    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.7.3     |
-| flash-attn   | 2.3.0   | 2.7.2     |
+| vllm         | 0.4.3   | 0.8.2     |
+| flash-attn   | 2.5.6   | 2.7.2     |

 ### 硬件依赖

@@ -430,16 +479,27 @@ huggingface-cli login
 > [!IMPORTANT]
 > 此步骤为必需。

+#### 从源码安装
+
 ```bash
 git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
 cd LLaMA-Factory
-pip install -e ".[torch,metrics]"
+pip install -e ".[torch,metrics]" --no-build-isolation
 ```

-可选的额外依赖项：torch、torch-npu、metrics、deepspeed、liger-kernel、bitsandbytes、hqq、eetq、gptq、awq、aqlm、vllm、galore、apollo、badam、adam-mini、qwen、minicpm_v、modelscope、openmind、swanlab、quality
+可选的额外依赖项：torch、torch-npu、metrics、deepspeed、liger-kernel、bitsandbytes、hqq、eetq、gptq、aqlm、vllm、sglang、galore、apollo、badam、adam-mini、qwen、minicpm_v、modelscope、openmind、swanlab、dev

-> [!TIP]
-> 遇到包冲突时，可使用 `pip install --no-deps -e .` 解决。
+#### 从镜像安装
+
+```bash
+docker run -it --rm --gpus=all --ipc=host hiyouga/llamafactory:latest
+```
+
+该镜像基于 Ubuntu 22.04（x86\_64）、CUDA 12.4、Python 3.11、PyTorch 2.6.0 和 Flash-attn 2.7.4 构建。
+
+查看全部镜像：https://hub.docker.com/r/hiyouga/llamafactory/tags
+
+请参阅[构建 Docker](#构建-docker) 来重新构建镜像。

 <details><summary>使用 <b>uv</b> 构建虚拟环境</summary>

@@ -457,9 +517,22 @@ uv run --prerelease=allow llamafactory-cli train examples/train_lora/llama3_lora

 </details>

-
 <details><summary>Windows 用户指南</summary>

+#### 安装 PyTorch
+
+Windows 平台需要额外手动安装 GPU 版本的 PyTorch 依赖包，您可以参考[官方网站](https://pytorch.org/get-started/locally/)和以下命令安装并测试 PyTorch 是否正确安装。
+
+```bash
+pip uninstall torch torchvision torchaudio
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
+python -c "import torch; print(torch.cuda.is_available())"
+```
+
+如果看到 `True` 则说明安装成功。
+
+若遇到类似 `Can't pickle local object` 的报错，请设置 `dataloader_num_workers: 0`。
+
 #### 安装 BitsAndBytes

 如果要在 Windows 平台上开启量化 LoRA（QLoRA），需要安装预编译的 `bitsandbytes` 库, 支持 CUDA 11.1 到 12.2, 请根据您的 CUDA 版本情况选择适合的[发布版本](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels)。
@@ -498,6 +571,7 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch        | 2.1.0   | 2.4.0          |
 | torch-npu    | 2.1.0   | 2.4.0.post2    |
 | deepspeed    | 0.13.2  | 0.13.2         |
+| vllm-ascend  | -       | 0.7.3          |

 请使用 `ASCEND_RT_VISIBLE_DEVICES` 而非 `CUDA_VISIBLE_DEVICES` 来指定运算设备。

@@ -548,6 +622,8 @@ pip install .
 > [!NOTE]
 > 使用自定义数据集时，请更新 `data/dataset_info.json` 文件。

+您也可以使用 **[Easy Dataset](https://github.com/ConardLi/easy-dataset)** 或 **[GraphGen](https://github.com/open-sciencelab/GraphGen)** 构建用于微调的合成数据。
+
 ### 快速开始

 下面三行命令分别对 Llama3-8B-Instruct 模型进行 LoRA **微调**、**推理**和**合并**。
@@ -603,22 +679,13 @@ 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 \
+    --build-arg EXTRAS=metrics \
    -t llamafactory:latest .

-docker run -dit --gpus=all \
-    -v ./hf_cache:/root/.cache/huggingface \
-    -v ./ms_cache:/root/.cache/modelscope \
-    -v ./om_cache:/root/.cache/openmind \
-    -v ./data:/app/data \
-    -v ./output:/app/output \
+docker run -dit --ipc=host --gpus=all \
    -p 7860:7860 \
    -p 8000:8000 \
-    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

@@ -628,19 +695,12 @@ docker exec -it llamafactory bash
 昇腾 NPU 用户：

 ```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 \
+    --build-arg EXTRAS=torch-npu,metrics \
    -t llamafactory:latest .

-# 根据您的资源更改 `device`
-docker run -dit \
-    -v ./hf_cache:/root/.cache/huggingface \
-    -v ./ms_cache:/root/.cache/modelscope \
-    -v ./om_cache:/root/.cache/openmind \
-    -v ./data:/app/data \
-    -v ./output:/app/output \
+docker run -dit --ipc=host \
    -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 \
@@ -651,7 +711,6 @@ docker run -dit \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
-    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

@@ -662,25 +721,15 @@ 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 \
+    --build-arg EXTRAS=metrics \
    -t llamafactory:latest .

-docker run -dit \
-    -v ./hf_cache:/root/.cache/huggingface \
-    -v ./ms_cache:/root/.cache/modelscope \
-    -v ./om_cache:/root/.cache/openmind \
-    -v ./data:/app/data \
-    -v ./output:/app/output \
-    -v ./saves:/app/saves \
+docker run -dit --ipc=host \
    -p 7860:7860 \
    -p 8000:8000 \
    --device /dev/kfd \
    --device /dev/dri \
-    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

@@ -689,12 +738,14 @@ docker exec -it llamafactory bash

 </details>

-<details><summary>数据卷详情</summary>
+<details><summary>使用数据卷</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。
- `ms_cache`：类似 Hugging Face 缓存文件夹，为 ModelScope 用户提供。
- `om_cache`：类似 Hugging Face 缓存文件夹，为 Modelers 用户提供。
- `data`：宿主机中存放数据集的文件夹路径。
+您可以通过移除 Dockerfile 中 `VOLUME [ "/root/.cache/huggingface", "/app/shared_data", "/app/output" ]` 的注释来使用数据卷。
+
+在构建 Docker 时使用参数 `-v ./hf_cache:/root/.cache/huggingface` 来挂载数据卷。各个数据卷的含义表示如下。
+
+- `hf_cache`：使用宿主机的 Hugging Face 缓存文件夹。
+- `shared_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>
@@ -702,7 +753,7 @@ docker exec -it llamafactory bash
 ### 利用 vLLM 部署 OpenAI API

 ```bash
-API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
+API_PORT=8000 llamafactory-cli api examples/inference/llama3.yaml infer_backend=vllm vllm_enforce_eager=true
 ```

 > [!TIP]
@@ -857,6 +908,7 @@ swanlab_run_name: test_run # 可选
 1. **[RAG-Retrieval](https://github.com/NLPJCL/RAG-Retrieval)**：一个全链路 RAG 检索模型微调、推理和蒸馏代码库。[[blog]](https://zhuanlan.zhihu.com/p/987727357)
 1. **[360-LLaMA-Factory](https://github.com/Qihoo360/360-LLaMA-Factory)**：一个魔改后的代码库，通过 Ring Attention 支持长序列的 SFT 和 DPO 训练。
 1. **[Sky-T1](https://novasky-ai.github.io/posts/sky-t1/)**：由 NovaSky AI 微调的低成本类 o1 长推理模型。
+1. **[WeClone](https://github.com/xming521/WeClone)**：从聊天记录创造数字分身的一站式解决方案。

 </details>

@@ -864,7 +916,7 @@ swanlab_run_name: test_run # 可选

 本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源。

-使用模型权重时，请遵循对应的模型协议：[Baichuan 2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [GPT-2](https://github.com/openai/gpt-2/blob/master/LICENSE) / [Granite](LICENSE) / [Index](https://huggingface.co/IndexTeam/Index-1.9B/blob/main/LICENSE) / [InternLM](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [MiniCPM](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) / [Mistral/Mixtral/Pixtral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3/Phi-4](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [Skywork](https://huggingface.co/Skywork/Skywork-13B-base/blob/main/Skywork%20Community%20License.pdf) / [StarCoder 2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [TeleChat2](https://huggingface.co/Tele-AI/telechat-7B/blob/main/TeleChat%E6%A8%A1%E5%9E%8B%E7%A4%BE%E5%8C%BA%E8%AE%B8%E5%8F%AF%E5%8D%8F%E8%AE%AE.pdf) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yi-1.5](LICENSE) / [Yuan 2](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+使用模型权重时，请遵循对应的模型协议：[Baichuan 2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [Command R](https://cohere.com/c4ai-cc-by-nc-license) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [GPT-2](https://github.com/openai/gpt-2/blob/master/LICENSE) / [Granite](LICENSE) / [Index](https://huggingface.co/IndexTeam/Index-1.9B/blob/main/LICENSE) / [InternLM](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [Llama 4](https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE) / [MiniCPM](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) / [Mistral/Mixtral/Pixtral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3/Phi-4](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [Skywork](https://huggingface.co/Skywork/Skywork-13B-base/blob/main/Skywork%20Community%20License.pdf) / [StarCoder 2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [TeleChat2](https://huggingface.co/Tele-AI/telechat-7B/blob/main/TeleChat%E6%A8%A1%E5%9E%8B%E7%A4%BE%E5%8C%BA%E8%AE%B8%E5%8F%AF%E5%8D%8F%E8%AE%AE.pdf) / [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

--- a/assets/alaya_new.svg
+++ b/assets/alaya_new.svg
--- a/assets/benchmark.svg
+++ b/assets/benchmark.svg
--- a/data/README.md
+++ b/data/README.md
@@ -1,12 +1,15 @@
 The [dataset_info.json](dataset_info.json) contains all available datasets. If you are using a custom dataset, please **make sure** to add a *dataset description* in `dataset_info.json` and specify `dataset: dataset_name` before training to use it.

-Currently we support datasets in **alpaca** and **sharegpt** format.
+The `dataset_info.json` file should be put in the `dataset_dir` directory. You can change `dataset_dir` to use another directory. The default value is `./data`.
+
+Currently we support datasets in **alpaca** and **sharegpt** format. Allowed file types include json, jsonl, csv, parquet, arrow.

 ```json
 "dataset_name": {
-  "hf_hub_url": "the name of the dataset repository on the Hugging Face hub. (if specified, ignore script_url and file_name)",
-  "ms_hub_url": "the name of the dataset repository on the Model Scope hub. (if specified, ignore script_url and file_name)",
-  "script_url": "the name of the directory containing a dataset loading script. (if specified, ignore file_name)",
+  "hf_hub_url": "the name of the dataset repository on the Hugging Face hub. (if specified, ignore script_url, file_name and cloud_file_name)",
+  "ms_hub_url": "the name of the dataset repository on the Model Scope hub. (if specified, ignore script_url, file_name and cloud_file_name)",
+  "script_url": "the name of the directory containing a dataset loading script. (if specified, ignore file_name and cloud_file_name)",
+  "cloud_file_name": "the name of the dataset file in s3/gcs cloud storage. (if specified, ignore file_name)",
  "file_name": "the name of the dataset folder or dataset file in this directory. (required if above are not specified)",
  "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)",
@@ -47,7 +50,9 @@ Currently we support datasets in **alpaca** and **sharegpt** format.

 * [Example dataset](alpaca_en_demo.json)

-In supervised fine-tuning, the `instruction` column will be concatenated with the `input` column and used as the human prompt, then the human prompt would be `instruction\ninput`. The `output` column represents the model response.
+In supervised fine-tuning, the `instruction` column will be concatenated with the `input` column and used as the user prompt, then the user prompt would be `instruction\ninput`. The `output` column represents the model response.
+
+For reasoning models, if the dataset contains chain-of-thought (CoT), the CoT needs to be placed in the model responses, such as `<think>cot</think>output`.

 The `system` column will be used as the system prompt if specified.

@@ -56,13 +61,13 @@ The `history` column is a list consisting of string tuples representing prompt-r
 ```json
 [
  {
-    "instruction": "human instruction (required)",
-    "input": "human input (optional)",
+    "instruction": "user instruction (required)",
+    "input": "user input (optional)",
    "output": "model response (required)",
    "system": "system prompt (optional)",
    "history": [
-      ["human instruction in the first round (optional)", "model response in the first round (optional)"],
-      ["human instruction in the second round (optional)", "model response in the second round (optional)"]
+      ["user instruction in the first round (optional)", "model response in the first round (optional)"],
+      ["user instruction in the second round (optional)", "model response in the second round (optional)"]
    ]
  }
 ]
@@ -83,9 +88,14 @@ Regarding the above dataset, the *dataset description* in `dataset_info.json` sh
 }
 ```

+> [!TIP]  
+> If the model has reasoning capabilities (e.g. Qwen3) but the dataset does not contain chain-of-thought (CoT), LLaMA-Factory will automatically add empty CoT to the data. When `enable_thinking` is `True` (slow thinking, by default), the empty CoT will be added to the model responses and loss computation will be considered; otherwise (fast thinking), it will be added to the user prompts and loss computation will be ignored. Please keep the `enable_thinking` parameter consistent during training and inference.
+>
+> If you want to train data containing CoT with slow thinking and data without CoT with fast thinking, you can set `enable_thinking` to `None`. However, this feature is relatively complicated and should be used with caution.
+
 ### Pre-training Dataset

- [Example dataset](c4_demo.json)
+- [Example dataset](c4_demo.jsonl)

 In pre-training, only the `text` column will be used for model learning.

@@ -116,8 +126,8 @@ It requires a better response in `chosen` column and a worse response in `reject
 ```json
 [
  {
-    "instruction": "human instruction (required)",
-    "input": "human input (optional)",
+    "instruction": "user instruction (required)",
+    "input": "user input (optional)",
    "chosen": "chosen answer (required)",
    "rejected": "rejected answer (required)"
  }
@@ -171,7 +181,7 @@ Note that the human and observation should appear in odd positions, while gpt an
    "conversations": [
      {
        "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
      },
      {
        "from": "function_call",
@@ -222,7 +232,7 @@ Preference datasets in sharegpt format also require a better message in `chosen`
    "conversations": [
      {
        "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
      },
      {
        "from": "gpt",
@@ -230,7 +240,7 @@ Preference datasets in sharegpt format also require a better message in `chosen`
      },
      {
        "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
      }
    ],
    "chosen": {
@@ -272,7 +282,7 @@ KTO datasets require a extra `kto_tag` column containing the boolean human feedb
    "conversations": [
      {
        "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
      },
      {
        "from": "gpt",
@@ -311,7 +321,7 @@ The number of images should be identical to the `<image>` tokens in the conversa
    "conversations": [
      {
        "from": "human",
-        "value": "<image>human instruction"
+        "value": "<image>user instruction"
      },
      {
        "from": "gpt",
@@ -352,7 +362,7 @@ The number of videos should be identical to the `<video>` tokens in the conversa
    "conversations": [
      {
        "from": "human",
-        "value": "<video>human instruction"
+        "value": "<video>user instruction"
      },
      {
        "from": "gpt",
@@ -393,7 +403,7 @@ The number of audios should be identical to the `<audio>` tokens in the conversa
    "conversations": [
      {
        "from": "human",
-        "value": "<audio>human instruction"
+        "value": "<audio>user instruction"
      },
      {
        "from": "gpt",
@@ -434,7 +444,7 @@ The openai format is simply a special case of the sharegpt format, where the fir
      },
      {
        "role": "user",
-        "content": "human instruction"
+        "content": "user instruction"
      },
      {
        "role": "assistant",
--- a/data/README_zh.md
+++ b/data/README_zh.md
@@ -1,6 +1,8 @@
 [dataset_info.json](dataset_info.json) 包含了所有可用的数据集。如果您希望使用自定义数据集，请**务必**在 `dataset_info.json` 文件中添加*数据集描述*，并通过修改 `dataset: 数据集名称` 配置来使用数据集。

-目前我们支持 **alpaca** 格式和 **sharegpt** 格式的数据集。
+其中 `dataset_info.json` 文件应放置在 `dataset_dir` 目录下。您可以通过修改 `dataset_dir` 参数来使用其他目录。默认值为 `./data`。
+
+目前我们支持 **alpaca** 格式和 **sharegpt** 格式的数据集。允许的文件类型包括 json、jsonl、csv、parquet 和 arrow。

 ```json
 "数据集名称": {
@@ -47,7 +49,9 @@

 - [样例数据集](alpaca_zh_demo.json)

-在指令监督微调时，`instruction` 列对应的内容会与 `input` 列对应的内容拼接后作为人类指令，即人类指令为 `instruction\ninput`。而 `output` 列对应的内容为模型回答。
+在指令监督微调时，`instruction` 列对应的内容会与 `input` 列对应的内容拼接后作为提示词，即提示词为 `instruction\ninput`。而 `output` 列对应的内容为模型回答。
+
+对于推理类模型的微调，如果数据集包含思维链，则需要把思维链放在模型回答中，例如 `<think>cot</think>output`。

 如果指定，`system` 列对应的内容将被作为系统提示词。

@@ -56,8 +60,8 @@
 ```json
 [
  {
-    "instruction": "人类指令（必填）",
-    "input": "人类输入（选填）",
+    "instruction": "用户指令（必填）",
+    "input": "用户输入（选填）",
    "output": "模型回答（必填）",
    "system": "系统提示词（选填）",
    "history": [
@@ -83,9 +87,14 @@
 }
 ```

+> [!TIP]
+> 如果模型本身具备推理能力（如 Qwen3）而数据集不包含思维链，LLaMA-Factory 会自动为数据添加空思维链。当 `enable_thinking` 为 `True` 时（慢思考，默认），空思维链会添加到模型回答中并且计算损失，否则会添加到用户指令中并且不计算损失（快思考）。请在训练和推理时保持 `enable_thinking` 参数一致。
+>
+> 如果您希望训练包含思维链的数据时使用慢思考，训练不包含思维链的数据时使用快思考，可以设置 `enable_thinking` 为 `None`。但该功能较为复杂，请谨慎使用。
+
 ### 预训练数据集

- [样例数据集](c4_demo.json)
+- [样例数据集](c4_demo.jsonl)

 在预训练时，只有 `text` 列中的内容会用于模型学习。

@@ -116,8 +125,8 @@
 ```json
 [
  {
-    "instruction": "人类指令（必填）",
-    "input": "人类输入（选填）",
+    "instruction": "用户指令（必填）",
+    "input": "用户输入（选填）",
    "chosen": "优质回答（必填）",
    "rejected": "劣质回答（必填）"
  }
@@ -171,7 +180,7 @@ KTO 数据集需要提供额外的 `kto_tag` 列。详情请参阅 [sharegpt](#s
    "conversations": [
      {
        "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
      },
      {
        "from": "function_call",
@@ -222,7 +231,7 @@ Sharegpt 格式的偏好数据集同样需要在 `chosen` 列中提供更优的
    "conversations": [
      {
        "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
      },
      {
        "from": "gpt",
@@ -230,7 +239,7 @@ Sharegpt 格式的偏好数据集同样需要在 `chosen` 列中提供更优的
      },
      {
        "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
      }
    ],
    "chosen": {
@@ -272,7 +281,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
    "conversations": [
      {
        "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
      },
      {
        "from": "gpt",
@@ -311,7 +320,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
    "conversations": [
      {
        "from": "human",
-        "value": "<image>人类指令"
+        "value": "<image><image>用户指令"
      },
      {
        "from": "gpt",
@@ -319,6 +328,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
      }
    ],
    "images": [
+      "图像路径（必填）",
      "图像路径（必填）"
    ]
  }
@@ -352,7 +362,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
    "conversations": [
      {
        "from": "human",
-        "value": "<video>人类指令"
+        "value": "<video><video>用户指令"
      },
      {
        "from": "gpt",
@@ -360,6 +370,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
      }
    ],
    "videos": [
+      "视频路径（必填）",
      "视频路径（必填）"
    ]
  }
@@ -393,7 +404,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
    "conversations": [
      {
        "from": "human",
-        "value": "<audio>人类指令"
+        "value": "<audio><audio>用户指令"
      },
      {
        "from": "gpt",
@@ -401,6 +412,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
      }
    ],
    "audios": [
+      "音频路径（必填）",
      "音频路径（必填）"
    ]
  }
@@ -435,7 +447,7 @@ OpenAI 格式仅仅是 sharegpt 格式的一种特殊情况，其中第一条消
      },
      {
        "role": "user",
-        "content": "人类指令"
+        "content": "用户指令"
      },
      {
        "role": "assistant",
--- a/data/belle_multiturn/belle_multiturn.py
+++ b/data/belle_multiturn/belle_multiturn.py
@@ -1,3 +1,18 @@
+# Copyright 2025 the LlamaFactory team.
+# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
+#
+# 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

@@ -10,7 +25,7 @@ _DESCRIPTION = "BELLE multiturn chat dataset."

 _CITATION = """\
@article{belle2023exploring,
-  title={Exploring the Impact of Instruction Data Scaling on Large Language Models: An Empirical Study on Real-World Use Cases},
+  title={Exploring the Impact of Instruction Data Scaling on Large Language Models},
  author={Yunjie Ji, Yong Deng, Yan Gong, Yiping Peng, Qiang Niu, Lei Zhang, Baochang Ma, Xiangang Li},
  journal={arXiv preprint arXiv:2303.14742},
  year={2023}
--- a/data/c4_demo.jsonl
+++ b/data/c4_demo.jsonl
--- a/data/hh_rlhf_en/hh_rlhf_en.py
+++ b/data/hh_rlhf_en/hh_rlhf_en.py
@@ -1,6 +1,20 @@
+# Copyright 2025 the LlamaFactory team.
+# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 import os
-from typing import List

 import datasets

@@ -50,7 +64,7 @@ class HhRlhfEn(datasets.GeneratorBasedBuilder):
            datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"filepaths": file_path["test"]}),
        ]

-    def _generate_examples(self, filepaths: List[str]):
+    def _generate_examples(self, filepaths: list[str]):
        key = 0
        for filepath in filepaths:
            with open(filepath, encoding="utf-8") as f:
--- a/data/mllm_demo_data/4.mp3
+++ b/data/mllm_demo_data/4.mp3
--- a/data/mllm_demo_data/4.mp4
+++ b/data/mllm_demo_data/4.mp4
--- a/data/ultra_chat/ultra_chat.py
+++ b/data/ultra_chat/ultra_chat.py
@@ -1,6 +1,20 @@
+# Copyright 2025 the LlamaFactory team.
+# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 import os
-from typing import List

 import datasets

@@ -11,7 +25,7 @@ _DESCRIPTION = "UltraChat: Large-scale, Informative, and Diverse Multi-round Dia

 _CITATION = """\
@misc{UltraChat,
-  author = {Ding, Ning and Chen, Yulin and Xu, Bokai and Hu, Shengding and Qin, Yujia and Liu, Zhiyuan and Sun, Maosong and Zhou, Bowen},
+  author = {Ding, Ning and Chen, Yulin and Xu, Bokai and Hu, Shengding and others},
  title = {UltraChat: A Large-scale Auto-generated Multi-round Dialogue Data},
  year = {2023},
  publisher = {GitHub},
@@ -40,7 +54,7 @@ class UltraChat(datasets.GeneratorBasedBuilder):
        file_paths = [dl_manager.download(_BASE_DATA_URL.format(idx=idx)) for idx in range(10)]  # multiple shards
        return [datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepaths": file_paths})]

-    def _generate_examples(self, filepaths: List[str]):
+    def _generate_examples(self, filepaths: list[str]):
        for filepath in filepaths:
            with open(filepath, encoding="utf-8") as f:
                for row in f:
@@ -49,7 +63,7 @@ class UltraChat(datasets.GeneratorBasedBuilder):
                    except Exception:
                        continue
                    key: int = data["id"]
-                    content: List[str] = data["data"]
+                    content: list[str] = data["data"]
                    if len(content) % 2 == 1:
                        content.pop(-1)
                    if len(content) < 2:
--- a/docker/docker-cuda/Dockerfile
+++ b/docker/docker-cuda/Dockerfile
@@ -1,101 +1,66 @@
-# Default use the NVIDIA official image with PyTorch 2.3.0
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/index.html
-ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:24.02-py3
+# https://hub.docker.com/r/hiyouga/pytorch/tags
+ARG BASE_IMAGE=hiyouga/pytorch:th2.6.0-cu124-flashattn2.7.4-cxx11abi0-devel
 FROM ${BASE_IMAGE}

+# Installation arguments
+ARG PIP_INDEX=https://pypi.org/simple
+ARG EXTRAS=metrics
+ARG INSTALL_FLASHATTN=false
+ARG HTTP_PROXY=""
+
 # Define environments
-ENV MAX_JOBS=4
+ENV MAX_JOBS=16
 ENV FLASH_ATTENTION_FORCE_BUILD=TRUE
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+ENV DEBIAN_FRONTEND=noninteractive
+ENV NODE_OPTIONS=""
+ENV PIP_ROOT_USER_ACTION=ignore
+ENV http_proxy="${HTTP_PROXY}"
+ENV https_proxy="${HTTP_PROXY}"

-# Define installation arguments
-ARG INSTALL_BNB=false
-ARG INSTALL_VLLM=false
-ARG INSTALL_DEEPSPEED=false
-ARG INSTALL_FLASHATTN=false
-ARG INSTALL_LIGER_KERNEL=false
-ARG INSTALL_HQQ=false
-ARG INSTALL_EETQ=false
-ARG PIP_INDEX=https://pypi.org/simple
-ARG HTTP_PROXY=
+# Use Bash instead of default /bin/sh
+SHELL ["/bin/bash", "-c"]

 # Set the working directory
 WORKDIR /app

-# Set http proxy
-RUN if [ -n "$HTTP_PROXY" ]; then \
-        echo "Configuring proxy..."; \
-        export http_proxy=$HTTP_PROXY; \
-        export https_proxy=$HTTP_PROXY; \
-    fi
+# Change pip source
+RUN pip config set global.index-url "${PIP_INDEX}" && \
+    pip config set global.extra-index-url "${PIP_INDEX}" && \
+    pip install --no-cache-dir --upgrade pip packaging wheel setuptools

 # 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 && \
-    if [ -n "$HTTP_PROXY" ]; then \
-        python -m pip install --proxy=$HTTP_PROXY -r requirements.txt; \
-    else \
-        python -m pip install -r requirements.txt; \
-    fi
+RUN pip install --no-cache-dir -r requirements.txt

 # Copy the rest of the application into the image
 COPY . /app

-# Install the LLaMA Factory
-RUN EXTRA_PACKAGES="metrics"; \
-    if [ "$INSTALL_BNB" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},bitsandbytes"; \
-    fi; \
-    if [ "$INSTALL_VLLM" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},vllm"; \
-    fi; \
-    if [ "$INSTALL_DEEPSPEED" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},deepspeed"; \
-    fi; \
-    if [ "$INSTALL_LIGER_KERNEL" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},liger-kernel"; \
-    fi; \
-    if [ "$INSTALL_HQQ" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},hqq"; \
-    fi; \
-    if [ "$INSTALL_EETQ" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},eetq"; \
-    fi; \
-    if [ -n "$HTTP_PROXY" ]; then \
-        pip install --proxy=$HTTP_PROXY -e ".[$EXTRA_PACKAGES]"; \
-    else \
-        pip install -e ".[$EXTRA_PACKAGES]"; \
-    fi
+# Install LLaMA Factory
+RUN pip install --no-cache-dir -e ".[${EXTRAS}]" --no-build-isolation

 # Rebuild flash attention
-RUN pip uninstall -y transformer-engine flash-attn && \
-    if [ "$INSTALL_FLASHATTN" == "true" ]; then \
+RUN if [ "${INSTALL_FLASHATTN}" == "true" ]; then \
        pip uninstall -y ninja && \
-        if [ -n "$HTTP_PROXY" ]; then \
-            pip install --proxy=$HTTP_PROXY ninja && \
-            pip install --proxy=$HTTP_PROXY --no-cache-dir flash-attn --no-build-isolation; \
-        else \
-            pip install ninja && \
-            pip install --no-cache-dir flash-attn --no-build-isolation; \
-        fi; \
-    fi
-
-
-# Unset http proxy
-RUN if [ -n "$HTTP_PROXY" ]; then \
-        unset http_proxy; \
-        unset https_proxy; \
+        pip install --no-cache-dir 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" ]
+# VOLUME [ "/root/.cache/huggingface", "/app/shared_data", "/app/output" ]

-# Expose port 7860 for the LLaMA Board
-ENV GRADIO_SERVER_PORT 7860
+# Expose port 7860 for LLaMA Board
+ENV GRADIO_SERVER_PORT=7860
 EXPOSE 7860

-# Expose port 8000 for the API service
-ENV API_PORT 8000
+# Expose port 8000 for API service
+ENV API_PORT=8000
 EXPOSE 8000
+
+# unset proxy
+ENV http_proxy=
+ENV https_proxy=
+
+# Reset pip config
+RUN pip config unset global.index-url && \
+    pip config unset global.extra-index-url
--- a/docker/docker-cuda/Dockerfile.base
+++ b/docker/docker-cuda/Dockerfile.base
@@ -0,0 +1,55 @@
+# Start from the pytorch official image (ubuntu-22.04 + cuda-12.4.1 + python-3.11)
+# https://hub.docker.com/r/pytorch/pytorch/tags
+FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel
+
+# Define environments
+ENV MAX_JOBS=16
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+ENV DEBIAN_FRONTEND=noninteractive
+ENV NODE_OPTIONS=""
+ENV PIP_ROOT_USER_ACTION=ignore
+
+# Define installation arguments
+ARG APT_SOURCE=https://mirrors.tuna.tsinghua.edu.cn/ubuntu/
+ARG PIP_INDEX=https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+# Set apt source
+RUN cp /etc/apt/sources.list /etc/apt/sources.list.bak && \
+    { \
+    echo "deb ${APT_SOURCE} jammy main restricted universe multiverse"; \
+    echo "deb ${APT_SOURCE} jammy-updates main restricted universe multiverse"; \
+    echo "deb ${APT_SOURCE} jammy-backports main restricted universe multiverse"; \
+    echo "deb ${APT_SOURCE} jammy-security main restricted universe multiverse"; \
+    } > /etc/apt/sources.list
+
+# Install systemctl and wget
+RUN apt-get update && \
+    apt-get install -y -o Dpkg::Options::="--force-confdef" systemd wget && \
+    apt-get clean
+
+# Install git and vim
+RUN apt-get update && \
+    apt-get install -y git vim && \
+    apt-get clean
+
+# Install gcc and g++
+RUN apt-get update && \
+    apt-get install -y gcc g++ && \
+    apt-get clean
+
+# Change pip source
+RUN pip config set global.index-url "${PIP_INDEX}" && \
+    pip config set global.extra-index-url "${PIP_INDEX}" && \
+    pip install --no-cache-dir --upgrade pip packaging wheel setuptools
+
+# Install flash-attn-2.7.4.post1 (cxx11abi=False)
+RUN wget -nv https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp311-cp311-linux_x86_64.whl && \
+    pip install --no-cache-dir flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp311-cp311-linux_x86_64.whl
+
+# Install flashinfer-0.2.2.post1+cu124 (cxx11abi=False)
+RUN wget -nv https://github.com/flashinfer-ai/flashinfer/releases/download/v0.2.2.post1/flashinfer_python-0.2.2.post1+cu124torch2.6-cp38-abi3-linux_x86_64.whl && \
+    pip install --no-cache-dir flashinfer_python-0.2.2.post1+cu124torch2.6-cp38-abi3-linux_x86_64.whl
+
+# Reset pip config
+RUN pip config unset global.index-url && \
+    pip config unset global.extra-index-url
--- a/docker/docker-cuda/docker-compose.yml
+++ b/docker/docker-cuda/docker-compose.yml
@@ -4,27 +4,15 @@ services:
      dockerfile: ./docker/docker-cuda/Dockerfile
      context: ../..
      args:
-        INSTALL_BNB: "false"
-        INSTALL_VLLM: "false"
-        INSTALL_DEEPSPEED: "false"
-        INSTALL_FLASHATTN: "false"
-        INSTALL_LIGER_KERNEL: "false"
-        INSTALL_HQQ: "false"
-        INSTALL_EETQ: "false"
        PIP_INDEX: https://pypi.org/simple
+        EXTRAS: metrics
    container_name: llamafactory
-    volumes:
-      - ../../hf_cache:/root/.cache/huggingface
-      - ../../ms_cache:/root/.cache/modelscope
-      - ../../om_cache:/root/.cache/openmind
-      - ../../data:/app/data
-      - ../../output:/app/output
    ports:
      - "7860:7860"
      - "8000:8000"
    ipc: host
    tty: true
-    shm_size: "16gb"
+    # shm_size: "16gb"  # ipc: host is set
    stdin_open: true
    command: bash
    deploy:
@@ -33,5 +21,5 @@ services:
          devices:
          - driver: nvidia
            count: "all"
-            capabilities: [gpu]
+            capabilities: [ gpu ]
    restart: unless-stopped
--- a/docker/docker-npu/Dockerfile
+++ b/docker/docker-npu/Dockerfile
@@ -1,67 +1,58 @@
-# 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:8.0.0-910b-ubuntu22.04-py3.10
-# FROM ascendai/cann:8.0.rc1-910-openeuler22.03-py3.8
-# FROM ascendai/cann:8.0.rc1-910b-openeuler22.03-py3.8
+# https://hub.docker.com/r/ascendai/cann/tags
+ARG BASE_IMAGE=ascendai/cann:8.0.0-910b-ubuntu22.04-py3.11
+FROM ${BASE_IMAGE}
+
+# Installation arguments
+ARG PIP_INDEX=https://pypi.org/simple
+ARG EXTRAS=torch-npu,metrics
+ARG HTTP_PROXY=""

 # Define environments
+ENV MAX_JOBS=16
+ENV FLASH_ATTENTION_FORCE_BUILD=TRUE
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 ENV DEBIAN_FRONTEND=noninteractive
+ENV NODE_OPTIONS=""
+ENV PIP_ROOT_USER_ACTION=ignore
+ENV http_proxy="${HTTP_PROXY}"
+ENV https_proxy="${HTTP_PROXY}"

-# Define installation arguments
-ARG INSTALL_DEEPSPEED=false
-ARG PIP_INDEX=https://pypi.org/simple
-ARG TORCH_INDEX=https://download.pytorch.org/whl/cpu
-ARG HTTP_PROXY=
+# Use Bash instead of default /bin/sh
+SHELL ["/bin/bash", "-c"]

 # Set the working directory
 WORKDIR /app

-# Set http proxy
-RUN if [ -n "$HTTP_PROXY" ]; then \
-        echo "Configuring proxy..."; \
-        export http_proxy=$HTTP_PROXY; \
-        export https_proxy=$HTTP_PROXY; \
-    fi
+# Change pip source
+RUN pip config set global.index-url "${PIP_INDEX}" && \
+    pip config set global.extra-index-url "${PIP_INDEX}" && \
+    pip install --no-cache-dir --upgrade pip packaging wheel setuptools

 # 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 && \
-    if [ -n "$HTTP_PROXY" ]; then \
-        python -m pip install --proxy=$HTTP_PROXY -r requirements.txt; \
-    else \
-        python -m pip install -r requirements.txt; \
-    fi
+RUN pip install --no-cache-dir -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; \
-    if [ -n "$HTTP_PROXY" ]; then \
-        pip install --proxy=$HTTP_PROXY -e ".[$EXTRA_PACKAGES]"; \
-    else \
-        pip install -e ".[$EXTRA_PACKAGES]"; \
-    fi
-
-# Unset http proxy
-RUN if [ -n "$HTTP_PROXY" ]; then \
-        unset http_proxy; \
-        unset https_proxy; \
-    fi
+# Install LLaMA Factory
+RUN pip install --no-cache-dir -e ".[${EXTRAS}]" --no-build-isolation

 # Set up volumes
-VOLUME [ "/root/.cache/huggingface", "/root/.cache/modelscope", "/app/data", "/app/output" ]
+# VOLUME [ "/root/.cache/huggingface", "/app/shared_data", "/app/output" ]

-# Expose port 7860 for the LLaMA Board
-ENV GRADIO_SERVER_PORT 7860
+# Expose port 7860 for LLaMA Board
+ENV GRADIO_SERVER_PORT=7860
 EXPOSE 7860

-# Expose port 8000 for the API service
-ENV API_PORT 8000
+# Expose port 8000 for API service
+ENV API_PORT=8000
 EXPOSE 8000
+
+# unset proxy
+ENV http_proxy=
+ENV https_proxy=
+
+# Reset pip config
+RUN pip config unset global.index-url && \
+    pip config unset global.extra-index-url
--- a/docker/docker-npu/docker-compose.yml
+++ b/docker/docker-npu/docker-compose.yml
@@ -4,15 +4,10 @@ services:
      dockerfile: ./docker/docker-npu/Dockerfile
      context: ../..
      args:
-        INSTALL_DEEPSPEED: "false"
        PIP_INDEX: https://pypi.org/simple
+        EXTRAS: torch-npu,metrics
    container_name: llamafactory
    volumes:
-      - ../../hf_cache:/root/.cache/huggingface
-      - ../../ms_cache:/root/.cache/modelscope
-      - ../../om_cache:/root/.cache/openmind
-      - ../../data:/app/data
-      - ../../output:/app/output
      - /usr/local/dcmi:/usr/local/dcmi
      - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
      - /usr/local/Ascend/driver:/usr/local/Ascend/driver
@@ -22,7 +17,7 @@ services:
      - "8000:8000"
    ipc: host
    tty: true
-    shm_size: "16gb"
+    # shm_size: "16gb"  # ipc: host is set
    stdin_open: true
    command: bash
    devices:
--- a/docker/docker-rocm/Dockerfile
+++ b/docker/docker-rocm/Dockerfile
@@ -1,93 +1,71 @@
-FROM hardandheavy/transformers-rocm:2.2.0
+# https://hub.docker.com/r/rocm/pytorch/tags
+ARG BASE_IMAGE=rocm/pytorch:rocm6.4.1_ubuntu22.04_py3.10_pytorch_release_2.6.0
+FROM ${BASE_IMAGE}
+
+# Installation arguments
+ARG PIP_INDEX=https://pypi.org/simple
+ARG EXTRAS=metrics
+ARG INSTALL_FLASHATTN=false
+ARG HTTP_PROXY=""
+ARG PYTORCH_INDEX=https://download.pytorch.org/whl/rocm6.3

 # Define environments
-ENV MAX_JOBS=4
+ENV MAX_JOBS=16
 ENV FLASH_ATTENTION_FORCE_BUILD=TRUE
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+ENV DEBIAN_FRONTEND=noninteractive
+ENV NODE_OPTIONS=""
+ENV PIP_ROOT_USER_ACTION=ignore
+ENV http_proxy="${HTTP_PROXY}"
+ENV https_proxy="${HTTP_PROXY}"

-# Define installation arguments
-ARG INSTALL_BNB=false
-ARG INSTALL_VLLM=false
-ARG INSTALL_DEEPSPEED=false
-ARG INSTALL_FLASHATTN=false
-ARG INSTALL_LIGER_KERNEL=false
-ARG INSTALL_HQQ=false
-ARG PIP_INDEX=https://pypi.org/simple
-ARG HTTP_PROXY=
+# Use Bash instead of default /bin/sh
+SHELL ["/bin/bash", "-c"]

 # Set the working directory
 WORKDIR /app

-# Set http proxy
-RUN if [ -n "$HTTP_PROXY" ]; then \
-        echo "Configuring proxy..."; \
-        export http_proxy=$HTTP_PROXY; \
-        export https_proxy=$HTTP_PROXY; \
-    fi
+# Change pip source
+RUN pip config set global.index-url "${PIP_INDEX}" && \
+    pip config set global.extra-index-url "${PIP_INDEX}" && \
+    pip install --no-cache-dir --upgrade pip packaging wheel setuptools
+
+# Reinstall pytorch rocm
+RUN pip uninstall -y torch torchvision torchaudio && \
+    pip install --no-cache-dir --pre torch torchvision torchaudio --index-url "${PYTORCH_INDEX}"

 # 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 && \
-    if [ -n "$HTTP_PROXY" ]; then \
-        python -m pip install --proxy=$HTTP_PROXY -r requirements.txt; \
-    else \
-        python -m pip install -r requirements.txt; \
-    fi
+RUN pip install --no-cache-dir -r requirements.txt

 # Copy the rest of the application into the image
 COPY . /app

-# Install the LLaMA Factory
-RUN EXTRA_PACKAGES="metrics"; \
-    if [ "$INSTALL_BNB" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},bitsandbytes"; \
-    fi; \
-    if [ "$INSTALL_VLLM" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},vllm"; \
-    fi; \
-    if [ "$INSTALL_DEEPSPEED" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},deepspeed"; \
-    fi; \
-    if [ "$INSTALL_LIGER_KERNEL" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},liger-kernel"; \
-    fi; \
-    if [ "$INSTALL_HQQ" == "true" ]; then \
-        EXTRA_PACKAGES="${EXTRA_PACKAGES},hqq"; \
-    fi; \
-    if [ -n "$HTTP_PROXY" ]; then \
-        pip install --proxy=$HTTP_PROXY -e ".[$EXTRA_PACKAGES]"; \
-    else \
-        pip install -e ".[$EXTRA_PACKAGES]"; \
-    fi
+# Install LLaMA Factory
+RUN pip install --no-cache-dir -e ".[${EXTRAS}]" --no-build-isolation

 # Rebuild flash attention
-RUN pip uninstall -y transformer-engine flash-attn && \
-    if [ "$INSTALL_FLASHATTN" == "true" ]; then \
+RUN if [ "${INSTALL_FLASHATTN}" == "true" ]; then \
        pip uninstall -y ninja && \
-        if [ -n "$HTTP_PROXY" ]; then \
-            pip install --proxy=$HTTP_PROXY ninja && \
-            pip install --proxy=$HTTP_PROXY --no-cache-dir flash-attn --no-build-isolation; \
-        else \
-            pip install ninja && \
-            pip install --no-cache-dir flash-attn --no-build-isolation; \
-        fi; \
-    fi
-
-# Unset http proxy
-RUN if [ -n "$HTTP_PROXY" ]; then \
-        unset http_proxy; \
-        unset https_proxy; \
+        pip install --no-cache-dir 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" ]
+# VOLUME [ "/root/.cache/huggingface", "/app/shared_data", "/app/output" ]

-# Expose port 7860 for the LLaMA Board
-ENV GRADIO_SERVER_PORT 7860
+# Expose port 7860 for LLaMA Board
+ENV GRADIO_SERVER_PORT=7860
 EXPOSE 7860

-# Expose port 8000 for the API service
-ENV API_PORT 8000
+# Expose port 8000 for API service
+ENV API_PORT=8000
 EXPOSE 8000
+
+# unset proxy
+ENV http_proxy=
+ENV https_proxy=
+
+# Reset pip config
+RUN pip config unset global.index-url && \
+    pip config unset global.extra-index-url
--- a/docker/docker-rocm/docker-compose.yml
+++ b/docker/docker-rocm/docker-compose.yml
@@ -4,27 +4,15 @@ services:
      dockerfile: ./docker/docker-rocm/Dockerfile
      context: ../..
      args:
-        INSTALL_BNB: "false"
-        INSTALL_VLLM: "false"
-        INSTALL_DEEPSPEED: "false"
-        INSTALL_FLASHATTN: "false"
-        INSTALL_LIGER_KERNEL: "false"
-        INSTALL_HQQ: "false"
        PIP_INDEX: https://pypi.org/simple
+        EXTRAS: metrics
    container_name: llamafactory
-    volumes:
-      - ../../hf_cache:/root/.cache/huggingface
-      - ../../ms_cache:/root/.cache/modelscope
-      - ../../om_cache:/root/.cache/openmind
-      - ../../data:/app/data
-      - ../../output:/app/output
-      - ../../saves:/app/saves
    ports:
      - "7860:7860"
      - "8000:8000"
    ipc: host
    tty: true
-    shm_size: "16gb"
+    # shm_size: "16gb"  # ipc: host is set
    stdin_open: true
    command: bash
    devices:
--- a/evaluation/ceval/ceval.py
+++ b/evaluation/ceval/ceval.py
@@ -1,3 +1,4 @@
+# Copyright 2025 the LlamaFactory team.
 # Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -21,14 +22,15 @@ import pandas as pd
 _CITATION = """\
@article{huang2023ceval,
  title={C-Eval: A Multi-Level Multi-Discipline Chinese Evaluation Suite for Foundation Models},
-  author={Huang, Yuzhen and Bai, Yuzhuo and Zhu, Zhihao and Zhang, Junlei and Zhang, Jinghan and Su, Tangjun and Liu, Junteng and Lv, Chuancheng and Zhang, Yikai and Lei, Jiayi and Fu, Yao and Sun, Maosong and He, Junxian},
+  author={Huang, Yuzhen and Bai, Yuzhuo and Zhu, Zhihao and others},
  journal={arXiv preprint arXiv:2305.08322},
  year={2023}
 }
 """

 _DESCRIPTION = """\
-C-Eval is a comprehensive Chinese evaluation suite for foundation models. It consists of 13948 multi-choice questions spanning 52 diverse disciplines and four difficulty levels.
+C-Eval is a comprehensive Chinese evaluation suite for foundation models.
+It consists of 13948 multi-choice questions spanning 52 diverse disciplines and four difficulty levels.
 """

 _HOMEPAGE = "https://cevalbenchmark.com"
--- a/evaluation/cmmlu/cmmlu.py
+++ b/evaluation/cmmlu/cmmlu.py
@@ -1,3 +1,4 @@
+# Copyright 2025 the LlamaFactory team.
 # Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -21,14 +22,15 @@ import pandas as pd
 _CITATION = """\
@article{li2023cmmlu,
  title={CMMLU: Measuring massive multitask language understanding in Chinese},
-  author={Haonan Li and Yixuan Zhang and Fajri Koto and Yifei Yang and Hai Zhao and Yeyun Gong and Nan Duan and Timothy Baldwin},
+  author={Haonan Li and Yixuan Zhang and Fajri Koto and Yifei Yang and others,
  journal={arXiv preprint arXiv:2306.09212},
  year={2023}
 }
 """

 _DESCRIPTION = """\
-CMMLU is a comprehensive Chinese assessment suite specifically designed to evaluate the advanced knowledge and reasoning abilities of LLMs within the Chinese language and cultural context.
+CMMLU is a comprehensive Chinese assessment suite specifically designed to evaluate the advanced knowledge
+and reasoning abilities of LLMs within the Chinese language and cultural context.
 """

 _HOMEPAGE = "https://github.com/haonan-li/CMMLU"
--- a/evaluation/mmlu/mmlu.py
+++ b/evaluation/mmlu/mmlu.py
@@ -1,3 +1,4 @@
+# Copyright 2025 the LlamaFactory team.
 # Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -21,14 +22,15 @@ import pandas as pd
 _CITATION = """\
@article{hendryckstest2021,
  title={Measuring Massive Multitask Language Understanding},
-  author={Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt},
+  author={Dan Hendrycks and Collin Burns and others},
  journal={Proceedings of the International Conference on Learning Representations (ICLR)},
  year={2021}
 }
 """

 _DESCRIPTION = """\
-Measuring Massive Multitask Language Understanding by Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt (ICLR 2021).
+Measuring Massive Multitask Language Understanding by Dan Hendrycks, Collin Burns, Steven Basart,
+Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt (ICLR 2021).
 """

 _HOMEPAGE = "https://github.com/hendrycks/test"
--- a/examples/README.md
+++ b/examples/README.md
@@ -15,6 +15,24 @@ Use `CUDA_VISIBLE_DEVICES` (GPU) or `ASCEND_RT_VISIBLE_DEVICES` (NPU) to choose

 By default, LLaMA-Factory uses all visible computing devices.

+Basic usage:
+
+```bash
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+```
+
+Advanced usage:
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml \
+    learning_rate=1e-5 \
+    logging_steps=1
+```
+
+```bash
+bash examples/train_lora/llama3_lora_sft.sh
+```
+
 ## Examples

 ### LoRA Fine-Tuning
@@ -34,8 +52,7 @@ llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 #### Multimodal Supervised Fine-Tuning

 ```bash
-llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
-llamafactory-cli train examples/train_lora/qwen2vl_lora_sft.yaml
+llamafactory-cli train examples/train_lora/qwen2_5vl_lora_sft.yaml
 ```

 #### DPO/ORPO/SimPO Training
@@ -47,7 +64,7 @@ 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
+llamafactory-cli train examples/train_lora/qwen2_5vl_lora_dpo.yaml
 ```

 #### Reward Modeling
@@ -148,10 +165,18 @@ FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500
 FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
 ```

+### Elastic and Fault-Tolerant Supervised Fine-Tuning on Multiple Nodes
+
+To launch an elastic job with `MAX_RESTARTS` failures retries, run the following on at least `MIN_NNODES` nodes and at most `MAX_NNODES` nodes. `RDZV_ID` should be set as a unique job id (shared by all nodes participating in the job). See also [torchrun](https://docs.pytorch.org/docs/stable/elastic/run.html).
+
+```bash
+FORCE_TORCHRUN=1 MIN_NNODES=1 MAX_NNODES=3 MAX_RESTARTS=3 RDZV_ID=llamafactory MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
+```
+
 #### Multimodal Supervised Fine-Tuning

 ```bash
-FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/qwen2vl_full_sft.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/qwen2_5vl_full_sft.yaml
 ```

 ### Merging LoRA Adapters and Quantization
@@ -178,10 +203,11 @@ llamafactory-cli export examples/merge_lora/llama3_full_sft.yaml

 ### Inferring LoRA Fine-Tuned Models

-#### Batch Generation using vLLM Tensor Parallel
+#### Evaluation using vLLM's Multi-GPU Inference

 ```
-python scripts/vllm_infer.py --model_name_or_path path_to_merged_model --dataset alpaca_en_demo
+python scripts/vllm_infer.py --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct --template llama3 --dataset alpaca_en_demo
+python scripts/eval_bleu_rouge.py generated_predictions.jsonl
 ```

 #### Use CLI ChatBox
@@ -228,6 +254,12 @@ llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
 llamafactory-cli train examples/extras/adam_mini/qwen2_full_sft.yaml
 ```

+#### Full-Parameter Fine-Tuning using Muon
+
+```bash
+llamafactory-cli train examples/extras/muon/qwen2_full_sft.yaml
+```
+
 #### LoRA+ Fine-Tuning

 ```bash
@@ -258,9 +290,3 @@ llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
 ```bash
 bash examples/extras/fsdp_qlora/train.sh
 ```
-
-#### Computing BLEU and ROUGE Scores
-
-```bash
-llamafactory-cli train examples/extras/nlg_eval/llama3_lora_predict.yaml
-```
--- a/examples/README_zh.md
+++ b/examples/README_zh.md
@@ -15,6 +15,24 @@

 LLaMA-Factory 默认使用所有可见的计算设备。

+基础用法：
+
+```bash
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+```
+
+高级用法：
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml \
+    learning_rate=1e-5 \
+    logging_steps=1
+```
+
+```bash
+bash examples/train_lora/llama3_lora_sft.sh
+```
+
 ## 示例

 ### LoRA 微调
@@ -34,8 +52,7 @@ llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 #### 多模态指令监督微调

 ```bash
-llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
-llamafactory-cli train examples/train_lora/qwen2vl_lora_sft.yaml
+llamafactory-cli train examples/train_lora/qwen2_5vl_lora_sft.yaml
 ```

 #### DPO/ORPO/SimPO 训练
@@ -47,7 +64,7 @@ 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
+llamafactory-cli train examples/train_lora/qwen2_5vl_lora_dpo.yaml
 ```

 #### 奖励模型训练
@@ -89,6 +106,14 @@ FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500
 FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```

+### 支持弹性和容错的多机指令监督微调
+
+要启动一个支持弹性节点和容错的多机指令微调，在每个节点上执行以下命令。弹性节点数量范围为 `MIN_NNODES:MAX_NNODES`，每个节点最多允许因为错误重启 `MAX_RESTARTS` 次。`RDZV_ID` 应设置为一个唯一的作业 ID（由参与该作业的所有节点共享）。更多新可以参考官方文档 [torchrun](https://docs.pytorch.org/docs/stable/elastic/run.html)。
+
+```bash
+FORCE_TORCHRUN=1 MIN_NNODES=1 MAX_NNODES=3 MAX_RESTARTS=3 RDZV_ID=llamafactory MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
+```
+
 #### 使用 DeepSpeed ZeRO-3 平均分配显存

 ```bash
@@ -151,7 +176,7 @@ FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500
 #### 多模态指令监督微调

 ```bash
-FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/qwen2vl_full_sft.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/qwen2_5vl_full_sft.yaml
 ```

 ### 合并 LoRA 适配器与模型量化
@@ -178,10 +203,11 @@ llamafactory-cli export examples/merge_lora/llama3_full_sft.yaml

 ### 推理 LoRA 模型

-#### 使用 vLLM+TP 批量推理
+#### 使用 vLLM 多卡推理评估

 ```
-python scripts/vllm_infer.py --model_name_or_path path_to_merged_model --dataset alpaca_en_demo
+python scripts/vllm_infer.py --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct --template llama3 --dataset alpaca_en_demo
+python scripts/eval_bleu_rouge.py generated_predictions.jsonl
 ```

 #### 使用命令行对话框
@@ -228,6 +254,12 @@ llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
 llamafactory-cli train examples/extras/adam_mini/qwen2_full_sft.yaml
 ```

+#### 使用 Muon 进行全参数训练
+
+```bash
+llamafactory-cli train examples/extras/muon/qwen2_full_sft.yaml
+```
+
 #### LoRA+ 微调

 ```bash
@@ -258,9 +290,3 @@ llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
 ```bash
 bash examples/extras/fsdp_qlora/train.sh
 ```
-
-#### 计算 BLEU 和 ROUGE 分数
-
-```bash
-llamafactory-cli train examples/extras/nlg_eval/llama3_lora_predict.yaml
-```
--- a/examples/accelerate/fsdp_config.yaml
+++ b/examples/accelerate/fsdp_config.yaml
@@ -7,16 +7,16 @@ fsdp_config:
  fsdp_backward_prefetch: BACKWARD_PRE
  fsdp_forward_prefetch: false
  fsdp_cpu_ram_efficient_loading: true
-  fsdp_offload_params: true # offload may affect training speed
+  fsdp_offload_params: false
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sync_module_states: true
  fsdp_use_orig_params: true
 machine_rank: 0
 main_training_function: main
-mixed_precision: bf16 # or fp16
-num_machines: 1 # the number of nodes
-num_processes: 2 # the number of GPUs in all nodes
+mixed_precision: bf16  # or fp16
+num_machines: 1  # the number of nodes
+num_processes: 2  # the number of GPUs in all nodes
 rdzv_backend: static
 same_network: true
 tpu_env: []
--- a/examples/accelerate/fsdp_config_offload.yaml
+++ b/examples/accelerate/fsdp_config_offload.yaml
@@ -0,0 +1,25 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_backward_prefetch: BACKWARD_PRE
+  fsdp_forward_prefetch: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_offload_params: true  # offload may affect training speed
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: true
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16  # or fp16
+num_machines: 1  # the number of nodes
+num_processes: 2  # the number of GPUs in all nodes
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
--- a/examples/extras/adam_mini/qwen2_full_sft.yaml
+++ b/examples/extras/adam_mini/qwen2_full_sft.yaml
@@ -15,6 +15,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/qwen2-1_5b/full/sft
@@ -22,6 +23,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/extras/apollo/llama3_full_sft.yaml
+++ b/examples/extras/apollo/llama3_full_sft.yaml
@@ -20,6 +20,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/full/sft
@@ -27,6 +28,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/extras/badam/llama3_full_sft.yaml
+++ b/examples/extras/badam/llama3_full_sft.yaml
@@ -20,6 +20,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/full/sft
@@ -27,6 +28,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/extras/fsdp_qlora/llama3_lora_sft.yaml
+++ b/examples/extras/fsdp_qlora/llama3_lora_sft.yaml
@@ -17,6 +17,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -24,6 +25,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/extras/galore/llama3_full_sft.yaml
+++ b/examples/extras/galore/llama3_full_sft.yaml
@@ -19,6 +19,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/full/sft
@@ -26,6 +27,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/extras/llama_pro/llama3_freeze_sft.yaml
+++ b/examples/extras/llama_pro/llama3_freeze_sft.yaml
@@ -17,6 +17,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b-pro/freeze/sft
@@ -24,6 +25,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/extras/loraplus/llama3_lora_sft.yaml
+++ b/examples/extras/loraplus/llama3_lora_sft.yaml
@@ -17,6 +17,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -24,6 +25,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/extras/mod/llama3_full_sft.yaml
+++ b/examples/extras/mod/llama3_full_sft.yaml
@@ -15,6 +15,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b-mod/full/sft
@@ -22,6 +23,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llava1_5_lora_sft.yaml
+++ b/examples/train_lora/llava1_5_lora_sft.yaml
@@ -1,17 +1,16 @@
 ### model
-model_name_or_path: llava-hf/llava-1.5-7b-hf
+model_name_or_path: Qwen/Qwen2-1.5B-Instruct
 trust_remote_code: true

 ### method
 stage: sft
 do_train: true
-finetuning_type: lora
-lora_rank: 8
-lora_target: all
+finetuning_type: full
+use_muon: true

 ### dataset
-dataset: mllm_demo
-template: llava
+dataset: identity,alpaca_en_demo
+template: qwen
 cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
@@ -19,23 +18,23 @@ preprocessing_num_workers: 16
 dataloader_num_workers: 4

 ### output
-output_dir: saves/llava1_5-7b/lora/sft
+output_dir: saves/qwen2-1_5b/full/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 8
-learning_rate: 1.0e-4
+learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
-resume_from_checkpoint: null

 ### eval
 # val_size: 0.1
--- a/examples/extras/nlg_eval/llama3_lora_predict.yaml
+++ b/examples/extras/nlg_eval/llama3_lora_predict.yaml
@@ -18,10 +18,12 @@ cutoff_len: 2048
 max_samples: 50
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/predict
 overwrite_output_dir: true
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### eval
 per_device_eval_batch_size: 1
--- a/examples/extras/pissa/llama3_lora_sft.yaml
+++ b/examples/extras/pissa/llama3_lora_sft.yaml
@@ -19,6 +19,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -26,6 +27,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/inference/llama3.yaml
+++ b/examples/inference/llama3.yaml
@@ -1,4 +1,4 @@
 model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 template: llama3
-infer_backend: huggingface  # choices: [huggingface, vllm]
+infer_backend: huggingface  # choices: [huggingface, vllm, sglang]
 trust_remote_code: true
--- a/examples/inference/llama3_full_sft.yaml
+++ b/examples/inference/llama3_full_sft.yaml
@@ -1,4 +1,4 @@
 model_name_or_path: saves/llama3-8b/full/sft
 template: llama3
-infer_backend: huggingface  # choices: [huggingface, vllm]
+infer_backend: huggingface  # choices: [huggingface, vllm, sglang]
 trust_remote_code: true
--- a/examples/inference/llama3_lora_sft.yaml
+++ b/examples/inference/llama3_lora_sft.yaml
@@ -1,5 +1,5 @@
 model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 adapter_name_or_path: saves/llama3-8b/lora/sft
 template: llama3
-infer_backend: huggingface  # choices: [huggingface, vllm]
+infer_backend: huggingface  # choices: [huggingface, vllm, sglang]
 trust_remote_code: true
--- a/examples/inference/llama3_vllm.yaml
+++ b/examples/inference/llama3_vllm.yaml
@@ -1,5 +0,0 @@
-model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
-template: llama3
-infer_backend: vllm
-vllm_enforce_eager: true
-trust_remote_code: true
--- a/examples/inference/llava1_5.yaml
+++ b/examples/inference/llava1_5.yaml
@@ -1,4 +0,0 @@
-model_name_or_path: llava-hf/llava-1.5-7b-hf
-template: llava
-infer_backend: huggingface  # choices: [huggingface, vllm]
-trust_remote_code: true
--- a/examples/inference/qwen2_5vl.yaml
+++ b/examples/inference/qwen2_5vl.yaml
@@ -0,0 +1,4 @@
+model_name_or_path: Qwen/Qwen2.5-VL-7B-Instruct
+template: qwen2_vl
+infer_backend: huggingface  # choices: [huggingface, vllm, sglang]
+trust_remote_code: true
--- a/examples/inference/qwen2_vl.yaml
+++ b/examples/inference/qwen2_vl.yaml
@@ -1,4 +0,0 @@
-model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
-template: qwen2_vl
-infer_backend: huggingface  # choices: [huggingface, vllm]
-trust_remote_code: true
--- a/examples/merge_lora/llama3_full_sft.yaml
+++ b/examples/merge_lora/llama3_full_sft.yaml
@@ -6,5 +6,5 @@ trust_remote_code: true
 ### export
 export_dir: output/llama3_full_sft
 export_size: 5
-export_device: cpu
+export_device: cpu  # choices: [cpu, auto]
 export_legacy_format: false
--- a/examples/merge_lora/llama3_gptq.yaml
+++ b/examples/merge_lora/llama3_gptq.yaml
@@ -6,7 +6,7 @@ trust_remote_code: true
 ### export
 export_dir: output/llama3_gptq
 export_quantization_bit: 4
-export_quantization_dataset: data/c4_demo.json
+export_quantization_dataset: data/c4_demo.jsonl
 export_size: 5
-export_device: cpu
+export_device: cpu  # choices: [cpu, auto]
 export_legacy_format: false
--- a/examples/merge_lora/llama3_lora_sft.yaml
+++ b/examples/merge_lora/llama3_lora_sft.yaml
@@ -9,5 +9,5 @@ trust_remote_code: true
 ### export
 export_dir: output/llama3_lora_sft
 export_size: 5
-export_device: cpu
+export_device: cpu  # choices: [cpu, auto]
 export_legacy_format: false
--- a/examples/merge_lora/qwen2_5vl_lora_sft.yaml
+++ b/examples/merge_lora/qwen2_5vl_lora_sft.yaml
@@ -1,13 +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
+model_name_or_path: Qwen/Qwen2.5-VL-7B-Instruct
+adapter_name_or_path: saves/qwen2_5vl-7b/lora/sft
 template: qwen2_vl
 trust_remote_code: true

 ### export
-export_dir: output/qwen2_vl_lora_sft
+export_dir: output/qwen2_5vl_lora_sft
 export_size: 5
-export_device: cpu
+export_device: cpu  # choices: [cpu, auto]
 export_legacy_format: false
--- a/examples/train_full/llama3_full_sft.yaml
+++ b/examples/train_full/llama3_full_sft.yaml
@@ -24,6 +24,7 @@ save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_full/qwen2_5vl_full_sft.yaml
+++ b/examples/train_full/qwen2_5vl_full_sft.yaml
@@ -0,0 +1,49 @@
+### model
+model_name_or_path: Qwen/Qwen2.5-VL-7B-Instruct
+image_max_pixels: 262144
+video_max_pixels: 16384
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+freeze_vision_tower: true
+freeze_multi_modal_projector: true
+freeze_language_model: false
+deepspeed: examples/deepspeed/ds_z3_config.json
+
+### dataset
+dataset: mllm_demo,identity,alpaca_en_demo
+template: qwen2_vl
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/qwen2_5vl-7b/full/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### 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
+resume_from_checkpoint: null
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500
--- a/examples/train_lora/llama3_lora_dpo.yaml
+++ b/examples/train_lora/llama3_lora_dpo.yaml
@@ -27,6 +27,7 @@ save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama3_lora_kto.yaml
+++ b/examples/train_lora/llama3_lora_kto.yaml
@@ -17,6 +17,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/kto
@@ -24,6 +25,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama3_lora_ppo.yaml
+++ b/examples/train_lora/llama3_lora_ppo.yaml
@@ -17,6 +17,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/ppo
@@ -24,6 +25,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama3_lora_pretrain.yaml
+++ b/examples/train_lora/llama3_lora_pretrain.yaml
@@ -24,6 +24,7 @@ save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama3_lora_reward.yaml
+++ b/examples/train_lora/llama3_lora_reward.yaml
@@ -25,6 +25,7 @@ save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama3_lora_sft.sh
+++ b/examples/train_lora/llama3_lora_sft.sh
@@ -0,0 +1,36 @@
+#!/bin/bash
+
+set -x
+
+MODEL_PATH=meta-llama/Meta-Llama-3-8B-Instruct
+
+llamafactory-cli train \
+    --model_name_or_path ${MODEL_PATH} \
+    --trust_remote_code \
+    --stage sft \
+    --do_train \
+    --finetuning_type lora \
+    --lora_rank 8 \
+    --lora_target all \
+    --dataset identity,alpaca_en_demo \
+    --template llama3 \
+    --cutoff_len 2048 \
+    --max_samples 1000 \
+    --overwrite_cache \
+    --preprocessing_num_workers 16 \
+    --dataloader_num_workers 4 \
+    --output_dir saves/llama3-8b/lora/sft \
+    --logging_steps 10 \
+    --save_steps 500 \
+    --plot_loss \
+    --overwrite_output_dir \
+    --save_only_model false \
+    --report_to none \
+    --per_device_train_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --learning_rate 1e-4 \
+    --num_train_epochs 3.0 \
+    --lr_scheduler_type cosine \
+    --warmup_ratio 0.1 \
+    --bf16 \
+    --ddp_timeout 180000000
--- a/examples/train_lora/llama3_lora_sft.yaml
+++ b/examples/train_lora/llama3_lora_sft.yaml
@@ -25,6 +25,7 @@ save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama3_lora_sft_ds3.yaml
+++ b/examples/train_lora/llama3_lora_sft_ds3.yaml
@@ -26,6 +26,7 @@ save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama3_lora_sft_ray.yaml
+++ b/examples/train_lora/llama3_lora_sft_ray.yaml
@@ -26,14 +26,21 @@ save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### ray
 ray_run_name: llama3_8b_sft_lora
 ray_storage_path: ./saves
-ray_num_workers: 4  # number of GPUs to use
+ray_num_workers: 4  # Number of GPUs to use.
+placement_strategy: PACK
 resources_per_worker:
  GPU: 1
-placement_strategy: PACK
+# ray_init_kwargs:
+#   runtime_env:
+#     env_vars:
+#       <YOUR-ENV-VAR-HERE>: "<YOUR-ENV-VAR-HERE>"
+#     pip:
+#       - emoji

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/llama4_lora_sft_ds3.yaml
+++ b/examples/train_lora/llama4_lora_sft_ds3.yaml
@@ -1,21 +1,20 @@
+# pip install git+https://github.com/hiyouga/transformers.git@llama4_train
+
 ### model
-model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
-image_max_pixels: 262144
-video_max_pixels: 16384
+model_name_or_path: meta-llama/Llama-4-Scout-17B-16E-Instruct
 trust_remote_code: true

 ### method
 stage: sft
 do_train: true
-finetuning_type: full
-freeze_vision_tower: true  # choices: [true, false]
-freeze_multi_modal_projector: true  # choices: [true, false]
-freeze_language_model: false  # choices: [true, false]
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
 deepspeed: examples/deepspeed/ds_z3_config.json  # choices: [ds_z0_config.json, ds_z2_config.json, ds_z3_config.json]

 ### dataset
 dataset: mllm_demo,identity,alpaca_en_demo
-template: qwen2_vl
+template: llama4
 cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
@@ -23,17 +22,18 @@ preprocessing_num_workers: 16
 dataloader_num_workers: 4

 ### output
-output_dir: saves/qwen2_vl-7b/full/sft
+output_dir: saves/llama4-8b/lora/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 2
-learning_rate: 1.0e-5
+learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
@@ -42,6 +42,7 @@ ddp_timeout: 180000000
 resume_from_checkpoint: null

 ### eval
+# eval_dataset: alpaca_en_demo
 # val_size: 0.1
 # per_device_eval_batch_size: 1
 # eval_strategy: steps
--- a/examples/train_lora/qwen2_5vl_lora_dpo.yaml
+++ b/examples/train_lora/qwen2_5vl_lora_dpo.yaml
@@ -1,5 +1,5 @@
 ### model
-model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
+model_name_or_path: Qwen/Qwen2.5-VL-7B-Instruct
 image_max_pixels: 262144
 video_max_pixels: 16384
 trust_remote_code: true
@@ -23,12 +23,13 @@ preprocessing_num_workers: 16
 dataloader_num_workers: 4

 ### output
-output_dir: saves/qwen2_vl-7b/lora/dpo
+output_dir: saves/qwen2_5vl-7b/lora/dpo
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_lora/qwen2_5vl_lora_sft.yaml
+++ b/examples/train_lora/qwen2_5vl_lora_sft.yaml
@@ -1,5 +1,5 @@
 ### model
-model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
+model_name_or_path: Qwen/Qwen2.5-VL-7B-Instruct
 image_max_pixels: 262144
 video_max_pixels: 16384
 trust_remote_code: true
@@ -21,12 +21,13 @@ preprocessing_num_workers: 16
 dataloader_num_workers: 4

 ### output
-output_dir: saves/qwen2_vl-7b/lora/sft
+output_dir: saves/qwen2_5vl-7b/lora/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_qlora/llama3_lora_sft_aqlm.yaml
+++ b/examples/train_qlora/llama3_lora_sft_aqlm.yaml
@@ -16,6 +16,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -23,6 +24,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_qlora/llama3_lora_sft_awq.yaml
+++ b/examples/train_qlora/llama3_lora_sft_awq.yaml
@@ -16,6 +16,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -23,6 +24,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_qlora/llama3_lora_sft_bnb_npu.yaml
+++ b/examples/train_qlora/llama3_lora_sft_bnb_npu.yaml
@@ -1,7 +1,7 @@
 ### model
 model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 quantization_bit: 4
-quantization_method: bitsandbytes
+quantization_method: bnb
 double_quantization: false
 trust_remote_code: true

@@ -19,6 +19,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -26,6 +27,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_qlora/llama3_lora_sft_gptq.yaml
+++ b/examples/train_qlora/llama3_lora_sft_gptq.yaml
@@ -16,6 +16,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -23,6 +24,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/examples/train_qlora/llama3_lora_sft_otfq.yaml
+++ b/examples/train_qlora/llama3_lora_sft_otfq.yaml
@@ -1,7 +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)]
+quantization_bit: 4  # choices: [8 (bnb/hqq/eetq), 4 (bnb/hqq), 3 (hqq), 2 (hqq)]
+quantization_method: bnb  # choices: [bnb, hqq, eetq]
 trust_remote_code: true

 ### method
@@ -18,6 +18,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4

 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -25,6 +26,8 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

 ### train
 per_device_train_batch_size: 1
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -19,13 +19,36 @@ dynamic = [
 ]

 [tool.ruff]
-target-version = "py38"
+target-version = "py39"
 line-length = 119
 indent-width = 4

 [tool.ruff.lint]
-ignore = ["C408", "C901", "E501", "E731", "E741", "W605"]
-select = ["C", "E", "F", "I", "W"]
+ignore = [
+    "C408", # collection
+    "C901", # complex
+    "E501", # line too long
+    "E731", # lambda function
+    "E741", # ambiguous var name
+    "D100", # no doc public module
+    "D101", # no doc public class
+    "D102", # no doc public method
+    "D103", # no doc public function
+    "D104", # no doc public package
+    "D105", # no doc magic method
+    "D107", # no doc __init__
+]
+extend-select = [
+    "C",      # complexity
+    "E",      # error
+    "F",      # pyflakes
+    "I",      # isort
+    "W",      # warning
+    "UP",     # pyupgrade
+    "D",      # pydocstyle
+    "PT009",  # pytest assert
+    "RUF022", # sort __all__
+]

 [tool.ruff.lint.isort]
 lines-after-imports = 2
@@ -38,9 +61,12 @@ known-third-party = [
    "peft",
    "torch",
    "transformers",
-    "trl"
+    "trl",
 ]

+[tool.ruff.lint.pydocstyle]
+convention = "google"
+
 [tool.ruff.format]
 quote-style = "double"
 indent-style = "space"
@@ -56,10 +82,14 @@ conflicts = [
    ],
    [
        { extra = "torch-npu" },
-        { extra = "liger-kernel" },
+        { extra = "vllm" },
    ],
    [
        { extra = "torch-npu" },
+        { extra = "sglang" },
+    ],
+    [
        { extra = "vllm" },
-    ]
+        { extra = "sglang" },
+    ],
 ]
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,26 +1,27 @@
-transformers>=4.41.2,<=4.49.0,!=4.46.*,!=4.47.*,!=4.48.*;python_version<'3.10'
-transformers>=4.41.2,<=4.49.0,!=4.46.*,!=4.47.*,!=4.48.0;python_version>='3.10'
-datasets>=2.16.0,<=3.2.0
-accelerate>=0.34.0,<=1.2.1
-peft>=0.11.1,<=0.12.0
+transformers>=4.45.0,<=4.52.4,!=4.46.*,!=4.47.*,!=4.48.0,!=4.52.0; sys_platform != 'darwin'
+transformers>=4.45.0,<=4.51.3,!=4.46.*,!=4.47.*,!=4.48.0,!=4.52.0; sys_platform == 'darwin'
+datasets>=2.16.0,<=3.6.0
+accelerate>=0.34.0,<=1.7.0
+peft>=0.14.0,<=0.15.2
 trl>=0.8.6,<=0.9.6
-tokenizers>=0.19.0,<=0.21.0
-gradio>=4.38.0,<=5.21.0
-pandas>=2.0.0
+tokenizers>=0.19.0,<=0.21.1
+gradio>=4.38.0,<=5.31.0
 scipy
 einops
 sentencepiece
 tiktoken
 protobuf
 uvicorn
-pydantic
 fastapi
 sse-starlette
 matplotlib>=3.7.0
 fire
+omegaconf
 packaging
 pyyaml
 numpy<2.0.0
+pydantic<=2.10.6
+pandas>=2.0.0
 av
 librosa
 tyro<0.9.0
--- a/scripts/api_example/test_image.py
+++ b/scripts/api_example/test_image.py
@@ -23,8 +23,8 @@ require_version("openai>=1.5.0", "To fix: pip install openai>=1.5.0")

 def main():
    client = OpenAI(
-        api_key="{}".format(os.environ.get("API_KEY", "0")),
-        base_url="http://localhost:{}/v1".format(os.environ.get("API_PORT", 8000)),
+        api_key="{}".format(os.getenv("API_KEY", "0")),
+        base_url="http://localhost:{}/v1".format(os.getenv("API_PORT", 8000)),
    )
    messages = []
    messages.append(
--- a/scripts/api_example/test_toolcall.py
+++ b/scripts/api_example/test_toolcall.py
@@ -14,7 +14,6 @@

 import json
 import os
-from typing import Sequence

 from openai import OpenAI
 from transformers.utils.versions import require_version
@@ -23,7 +22,7 @@ from transformers.utils.versions import require_version
 require_version("openai>=1.5.0", "To fix: pip install openai>=1.5.0")


-def calculate_gpa(grades: Sequence[str], hours: Sequence[int]) -> float:
+def calculate_gpa(grades: list[str], hours: list[int]) -> float:
    grade_to_score = {"A": 4, "B": 3, "C": 2}
    total_score, total_hour = 0, 0
    for grade, hour in zip(grades, hours):
@@ -34,8 +33,8 @@ def calculate_gpa(grades: Sequence[str], hours: Sequence[int]) -> float:

 def main():
    client = OpenAI(
-        api_key="{}".format(os.environ.get("API_KEY", "0")),
-        base_url="http://localhost:{}/v1".format(os.environ.get("API_PORT", 8000)),
+        api_key="{}".format(os.getenv("API_KEY", "0")),
+        base_url="http://localhost:{}/v1".format(os.getenv("API_PORT", 8000)),
    )
    tools = [
        {
--- a/scripts/convert_ckpt/llamafy_baichuan2.py
+++ b/scripts/convert_ckpt/llamafy_baichuan2.py
@@ -15,7 +15,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import Any, Dict
+from typing import Any

 import fire
 import torch
@@ -29,13 +29,13 @@ CONFIG_NAME = "config.json"


 def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool):
-    baichuan2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
+    baichuan2_state_dict: dict[str, torch.Tensor] = OrderedDict()
    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
        if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".bin"):
-            shard_weight = torch.load(os.path.join(input_dir, filepath), map_location="cpu")
+            shard_weight = torch.load(os.path.join(input_dir, filepath), map_location="cpu", weights_only=True)
            baichuan2_state_dict.update(shard_weight)

-    llama_state_dict: Dict[str, torch.Tensor] = OrderedDict()
+    llama_state_dict: dict[str, torch.Tensor] = OrderedDict()
    for key, value in tqdm(baichuan2_state_dict.items(), desc="Convert format"):
        if "W_pack" in key:
            proj_size = value.size(0) // 3
@@ -75,7 +75,7 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso

 def save_config(input_dir: str, output_dir: str):
    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
-        llama2_config_dict: Dict[str, Any] = json.load(f)
+        llama2_config_dict: dict[str, Any] = json.load(f)

    llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
    llama2_config_dict.pop("auto_map", None)
@@ -94,8 +94,8 @@ def llamafy_baichuan2(
    shard_size: str = "2GB",
    save_safetensors: bool = True,
 ):
-    r"""
-    Converts the Baichuan2-7B model in the same format as LLaMA2-7B.
+    r"""Convert the Baichuan2-7B model in the same format as LLaMA2-7B.
+
    Usage: python llamafy_baichuan2.py --input_dir input --output_dir output
    Converted model: https://huggingface.co/hiyouga/Baichuan2-7B-Base-LLaMAfied
    """
--- a/scripts/convert_ckpt/llamafy_qwen.py
+++ b/scripts/convert_ckpt/llamafy_qwen.py
@@ -15,7 +15,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import Any, Dict
+from typing import Any

 import fire
 import torch
@@ -37,14 +37,14 @@ CONFIG_NAME = "config.json"


 def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool) -> str:
-    qwen_state_dict: Dict[str, torch.Tensor] = OrderedDict()
+    qwen_state_dict: dict[str, torch.Tensor] = OrderedDict()
    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
        if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".safetensors"):
            with safe_open(os.path.join(input_dir, filepath), framework="pt", device="cpu") as f:
                for key in f.keys():
                    qwen_state_dict[key] = f.get_tensor(key)

-    llama_state_dict: Dict[str, torch.Tensor] = OrderedDict()
+    llama_state_dict: dict[str, torch.Tensor] = OrderedDict()
    torch_dtype = None
    for key, value in tqdm(qwen_state_dict.items(), desc="Convert format"):
        if torch_dtype is None:
@@ -112,9 +112,9 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso

 def save_config(input_dir: str, output_dir: str, torch_dtype: str):
    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
-        qwen_config_dict: Dict[str, Any] = json.load(f)
+        qwen_config_dict: dict[str, Any] = json.load(f)

-    llama2_config_dict: Dict[str, Any] = OrderedDict()
+    llama2_config_dict: dict[str, Any] = OrderedDict()
    llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
    llama2_config_dict["hidden_act"] = "silu"
    llama2_config_dict["hidden_size"] = qwen_config_dict["hidden_size"]
@@ -147,8 +147,8 @@ def llamafy_qwen(
    shard_size: str = "2GB",
    save_safetensors: bool = False,
 ):
-    r"""
-    Converts the Qwen models in the same format as LLaMA2.
+    r"""Convert the Qwen models in the same format as LLaMA2.
+
    Usage: python llamafy_qwen.py --input_dir input --output_dir output
    Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
    """
--- a/scripts/convert_ckpt/tiny_llama4.py
+++ b/scripts/convert_ckpt/tiny_llama4.py
@@ -0,0 +1,39 @@
+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from transformers import Llama4Config, Llama4ForConditionalGeneration, Llama4TextConfig, Llama4VisionConfig
+
+
+if __name__ == "__main__":
+    vision_config = Llama4VisionConfig(
+        hidden_size=1408,
+        image_size=336,
+        intermediate_size=5632,
+        num_attention_heads=16,
+        num_hidden_layers=4,
+        vision_output_dim=4096,
+    )
+    text_config = Llama4TextConfig(
+        hidden_size=512,
+        intermediate_size=1024,
+        intermediate_size_mlp=1024,
+        num_hidden_layers=4,
+        num_attention_heads=8,
+        num_key_value_heads=2,
+        head_dim=512 // 8,
+        num_local_experts=2,
+    )
+    config = Llama4Config(vision_config=vision_config, text_config=text_config)
+    model = Llama4ForConditionalGeneration._from_config(config)
+    model.save_pretrained("tiny-llama4")
--- a/scripts/eval_bleu_rouge.py
+++ b/scripts/eval_bleu_rouge.py
@@ -0,0 +1,79 @@
+# Copyright 2025 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 logging
+import time
+
+import fire
+from datasets import load_dataset
+
+
+try:
+    import jieba  # type: ignore
+    from nltk.translate.bleu_score import SmoothingFunction, sentence_bleu  # type: ignore
+    from rouge_chinese import Rouge  # type: ignore
+
+    jieba.setLogLevel(logging.CRITICAL)
+    jieba.initialize()
+except ImportError:
+    print("Please install llamafactory with `pip install -e .[metrics]`.")
+    raise
+
+
+def compute_metrics(sample):
+    hypothesis = list(jieba.cut(sample["predict"]))
+    reference = list(jieba.cut(sample["label"]))
+
+    bleu_score = sentence_bleu(
+        [list(sample["label"])],
+        list(sample["predict"]),
+        smoothing_function=SmoothingFunction().method3,
+    )
+
+    if len(" ".join(hypothesis).split()) == 0 or len(" ".join(reference).split()) == 0:
+        result = {"rouge-1": {"f": 0.0}, "rouge-2": {"f": 0.0}, "rouge-l": {"f": 0.0}}
+    else:
+        rouge = Rouge()
+        scores = rouge.get_scores(" ".join(hypothesis), " ".join(reference))
+        result = scores[0]
+
+    metric_result = {}
+    for k, v in result.items():
+        metric_result[k] = round(v["f"] * 100, 4)
+
+    metric_result["bleu-4"] = round(bleu_score * 100, 4)
+
+    return metric_result
+
+
+def main(filename: str):
+    start_time = time.time()
+    dataset = load_dataset("json", data_files=filename, split="train")
+    dataset = dataset.map(compute_metrics, num_proc=8, remove_columns=dataset.column_names)
+    score_dict = dataset.to_dict()
+
+    average_score = {}
+    for task, scores in sorted(score_dict.items(), key=lambda x: x[0]):
+        print(f"{task}: {sum(scores) / len(scores):.4f}")
+        average_score[task] = sum(scores) / len(scores)
+
+    with open("predictions_score.json", "w", encoding="utf-8") as f:
+        json.dump(average_score, f, indent=4)
+
+    print(f"\nDone in {time.time() - start_time:.3f}s.\nScore file saved to predictions_score.json")
+
+
+if __name__ == "__main__":
+    fire.Fire(main)
--- a/scripts/llama_pro.py
+++ b/scripts/llama_pro.py
@@ -18,7 +18,7 @@
 import json
 import os
 from collections import OrderedDict
-from typing import TYPE_CHECKING, Dict
+from typing import TYPE_CHECKING

 import fire
 import torch
@@ -44,11 +44,11 @@ def block_expansion(
    shard_size: str = "5GB",
    save_safetensors: bool = True,
 ):
-    r"""
-    Performs block expansion for LLaMA, Mistral, Qwen2 or Yi models.
+    r"""Perform block expansion for LLaMA, Mistral, Qwen2 or Yi models.
+
    Usage: python llama_pro.py --model_name_or_path meta-llama/Llama-2-7b-hf --output_dir llama2_pro --num_expand 8
    """
-    config: "PretrainedConfig" = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True)
+    config: PretrainedConfig = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True)
    num_layers = getattr(config, "num_hidden_layers")
    if num_layers % num_expand != 0:
        raise ValueError(f"`num_layers` {num_layers} should be divisible by `num_expand` {num_expand}.")
@@ -70,7 +70,7 @@ def block_expansion(
    split = num_layers // num_expand
    layer_cnt = 0
    state_dict = model.state_dict()
-    output_state_dict: Dict[str, "torch.Tensor"] = OrderedDict()
+    output_state_dict: dict[str, torch.Tensor] = OrderedDict()
    for i in range(num_layers):
        for key, value in state_dict.items():
            if f".{i:d}." in key:
--- a/scripts/loftq_init.py
+++ b/scripts/loftq_init.py
@@ -38,8 +38,8 @@ def quantize_loftq(
    lora_target: tuple = ("q_proj", "v_proj"),
    save_safetensors: bool = True,
 ):
-    r"""
-    Initializes LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ)
+    r"""Initialize 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):
@@ -72,7 +72,7 @@ def quantize_loftq(
    print(f"Adapter weights saved in {loftq_dir}")

    # Save base model
-    base_model: "PreTrainedModel" = peft_model.unload()
+    base_model: PreTrainedModel = peft_model.unload()
    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
    tokenizer.save_pretrained(output_dir)
    print(f"Model weights saved in {output_dir}")
--- a/scripts/pissa_init.py
+++ b/scripts/pissa_init.py
@@ -37,8 +37,8 @@ def quantize_pissa(
    lora_target: tuple = ("q_proj", "v_proj"),
    save_safetensors: bool = True,
 ):
-    r"""
-    Initializes LoRA weights with Principal Singular values and Singular vectors Adaptation (PiSSA)
+    r"""Initialize 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):
@@ -67,7 +67,7 @@ def quantize_pissa(
    print(f"Adapter weights saved in {pissa_dir}")

    # Save base model
-    base_model: "PreTrainedModel" = peft_model.unload()
+    base_model: PreTrainedModel = peft_model.unload()
    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
    tokenizer.save_pretrained(output_dir)
    print(f"Model weights saved in {output_dir}")
--- a/scripts/qwen_omni_merge.py
+++ b/scripts/qwen_omni_merge.py
@@ -0,0 +1,136 @@
+# Copyright 2025 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.
+
+"""Why we need this script for qwen_omni?
+
+Because the qwen_omni model is constructed by two parts:
+1. [Thinker]:[audio_encoder, vision_encoder, LLM backbone], which our repository does support to post-training.
+2. [Talker]: [audio_decoder, wave_model], which is not supported to post-training without specific tokenizer.
+When we post-training the model, we exactly train the [Thinker] part, and the [Talker] part is dropped.
+So, to get the complete model, we need to merge the [Talker] part back to the [Thinker] part.
+LoRA mode: [Thinker + LoRA weights] + [Original Talker] -> [Omni model]
+Full mode: [Thinker] + [Original Talker] -> [Omni model]
+For Processor, we do saved the processor from trained model instead of the original model.
+"""
+
+import os
+import shutil
+
+import fire
+from peft import PeftModel
+from transformers import (
+    AutoProcessor,
+    Qwen2_5OmniForConditionalGeneration,  # type: ignore
+    Qwen2_5OmniThinkerForConditionalGeneration,
+)
+
+
+def merge_lora(
+    base_model_path: str,
+    lora_checkpoint_path: str,
+    extra_file: str = "spk_dict.pt",
+    submodule_name: str = "thinker",
+    save_path: str = "./merged_model_checkpoint",
+):
+    """Load the original model, merge the LoRA weights.
+
+    For a specified submodule, and save the final merged model along with its configurations.
+
+    Args:
+        base_model_path (str): Path to the original model directory.
+        lora_checkpoint_path (str): Path to the directory containing LoRA weights.
+        extra_file (str): Name of the extra file to be copied (default: "spk_dict.pt").
+        submodule_name (str): Name of the submodule to merge (default: "thinker").
+        save_path (str): Directory where the merged model and configurations will be saved.
+    """
+    # 1. Load the original model
+    model = Qwen2_5OmniForConditionalGeneration.from_pretrained(base_model_path, torch_dtype="auto", device_map="cpu")
+    print("Successfully loaded the original model.")
+
+    # 2. Extract the submodule to be merged (e.g., model.thinker)
+    if not hasattr(model, submodule_name):
+        raise AttributeError(f"The model does not have a submodule named '{submodule_name}'.")
+
+    base_submodule = getattr(model, submodule_name)
+    print(f"Successfully extracted submodule: {submodule_name}.")
+
+    # 3. Load the LoRA weights onto the extracted submodule
+    lora_model = PeftModel.from_pretrained(base_submodule, lora_checkpoint_path)
+    processor = AutoProcessor.from_pretrained(lora_checkpoint_path)
+    print("LoRA weights and processor loaded successfully.")
+
+    # 4. Merge the LoRA weights into the submodule and unload the LoRA modules
+    merged_submodule = lora_model.merge_and_unload()
+    print("LoRA weights merged successfully.")
+
+    # 5. Replace the original submodule with the merged submodule in the model
+    setattr(model, submodule_name, merged_submodule)
+
+    # 6. Save the final merged model along with the tokenizer and processor configuration
+    model.save_pretrained(save_path)
+    processor.save_pretrained(save_path)
+    print(f"Merged model and tokenizer saved to {save_path}.")
+
+    source_file = os.path.join(base_model_path, extra_file)
+    target_file = os.path.join(save_path, extra_file)
+    if os.path.exists(source_file):
+        shutil.copy(source_file, target_file)
+        print(f"File '{extra_file}' copied from {base_model_path} to {save_path}.")
+    else:
+        print(f"File '{extra_file}' not found in {base_model_path}, skipping copy.")
+
+
+def save_full_model(
+    saved_thinker_path: str,
+    base_model_path: str,
+    save_path: str = "./merged_model_checkpoint",
+    extra_file: str = "spk_dict.pt",
+):
+    """Load the saved thinker module and the original model, replace the thinker in the original model.
+
+    Then save the complete model along with its tokenizer and processor configuration.
+
+    Args:
+        saved_thinker_path (str): Path to the saved thinker weights.
+        base_model_path (str): Directory path of the original model.
+        save_path (str): Directory where the merged model and configurations will be saved.
+        extra_file (str): Name of the extra file to be copied (default: "spk_dict.pt").
+    """
+    # 1. Load the saved thinker module and the original model
+    thinker = Qwen2_5OmniThinkerForConditionalGeneration.from_pretrained(
+        saved_thinker_path, torch_dtype="auto", device_map="cpu"
+    )
+    base_model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
+        base_model_path, torch_dtype="auto", device_map="cpu"
+    )
+    base_model.thinker = thinker
+
+    # 2. Save the complete model along with its tokenizer and processor configuration
+    processor = AutoProcessor.from_pretrained(saved_thinker_path)
+    base_model.save_pretrained(save_path)
+    processor.save_pretrained(save_path)
+    print(f"Merged model and processor saved to {save_path}.")
+
+    # 3. Copy the extra file from the base model directory to the save_path
+    source_file = os.path.join(base_model_path, extra_file)
+    target_file = os.path.join(save_path, extra_file)
+    if os.path.exists(source_file):
+        shutil.copy(source_file, target_file)
+        print(f"File '{extra_file}' copied from {base_model_path} to {save_path}.")
+    else:
+        print(f"File '{extra_file}' not found in {base_model_path}, skipping copy.")
+
+
+if __name__ == "__main__":
+    fire.Fire({"save_full": save_full_model, "merge_lora": merge_lora})
--- a/scripts/stat_utils/cal_flops.py
+++ b/scripts/stat_utils/cal_flops.py
@@ -29,8 +29,8 @@ def calculate_flops(
    seq_length: int = 512,
    flash_attn: str = "auto",
 ):
-    r"""
-    Calculates the flops of pre-trained models.
+    r"""Calculate the flops of pre-trained models.
+
    Usage: python cal_flops.py --model_name_or_path path_to_model --batch_size 1 --seq_length 512
    """
    with get_accelerator().device(0):
--- a/scripts/stat_utils/cal_lr.py
+++ b/scripts/stat_utils/cal_lr.py
@@ -45,8 +45,8 @@ def calculate_lr(
    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.
+    r"""Calculate the optimal learning rate for 7B/13B models using LLaMA's hyper-parameters.
+
    Usage:
    python cal_lr.py --model_name_or_path path_to_model --dataset alpaca_en_demo --cutoff_len 1024 --batch_size 16
    """
@@ -89,9 +89,8 @@ def calculate_lr(
    lr = BASE_LR * math.sqrt(token_batch_size / BASE_BS)  # lr ~ sqrt(batch_size)
    lr = lr / 6.0 if is_mistral_or_gemma else lr
    print(
-        "Optimal learning rate is {:.2e} for valid ratio% {:.2f} and effective token batch size {:.2f}".format(
-            lr, valid_ratio * 100, token_batch_size
-        )
+        f"Optimal learning rate is {lr:.2e} for valid ratio% {valid_ratio * 100:.2f} "
+        f"and effective token batch size {token_batch_size:.2f}"
    )


--- a/scripts/stat_utils/cal_mfu.py
+++ b/scripts/stat_utils/cal_mfu.py
@@ -34,9 +34,7 @@ def compute_model_flops(
    include_recompute: bool = False,
    include_flashattn: bool = False,
 ) -> int:
-    r"""
-    Calculates the FLOPs of model per forward/backward pass.
-    """
+    r"""Calculate 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)
@@ -86,9 +84,7 @@ def compute_model_flops(


 def compute_device_flops(world_size: int) -> float:
-    r"""
-    Calculates the FLOPs of the device capability per second.
-    """
+    r"""Calculate 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
@@ -114,8 +110,8 @@ def calculate_mfu(
    liger_kernel: bool = False,
    unsloth_gc: bool = False,
 ) -> float:
-    r"""
-    Calculates MFU for given model and hyper-params.
+    r"""Calculate 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 = {
--- a/scripts/stat_utils/cal_ppl.py
+++ b/scripts/stat_utils/cal_ppl.py
@@ -14,7 +14,7 @@

 import json
 from dataclasses import dataclass
-from typing import Any, Dict, Literal, Optional, Sequence
+from typing import Any, Literal, Optional

 import fire
 import torch
@@ -30,16 +30,12 @@ from llamafactory.model import load_model, load_tokenizer

@dataclass
 class PairwiseDataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
-    r"""
-    Data collator for pairwise data.
-    """
+    r"""Data collator for pairwise data."""

    train_on_prompt: bool = False

-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
-        r"""
-        Pads batched data to the longest sequence in the batch.
-        """
+    def __call__(self, features: list[dict[str, Any]]) -> dict[str, torch.Tensor]:
+        r"""Pad batched data to the longest sequence in the batch."""
        chosen_features = []
        for feature in features:
            chosen_features.append(
@@ -68,8 +64,8 @@ def calculate_ppl(
    max_samples: Optional[int] = None,
    train_on_prompt: bool = False,
 ):
-    r"""
-    Calculates the ppl on the dataset of the pre-trained models.
+    r"""Calculate the ppl on the dataset of the pre-trained models.
+
    Usage: export CUDA_VISIBLE_DEVICES=0
    python cal_ppl.py --model_name_or_path path_to_model --dataset alpaca_en_demo --save_name ppl.json
    """
@@ -111,17 +107,17 @@ def calculate_ppl(
    criterion = torch.nn.CrossEntropyLoss(reduction="none")
    total_ppl = 0
    perplexities = []
-    batch: Dict[str, "torch.Tensor"]
+    batch: dict[str, torch.Tensor]
    with torch.no_grad():
        for batch in tqdm(dataloader, desc="Computing perplexities"):
            batch = batch.to(model.device)
            outputs = model(**batch)
-            shift_logits: "torch.Tensor" = outputs["logits"][..., :-1, :]
-            shift_labels: "torch.Tensor" = batch["labels"][..., 1:]
+            shift_logits: torch.Tensor = outputs["logits"][..., :-1, :]
+            shift_labels: torch.Tensor = batch["labels"][..., 1:]
            loss_mask = shift_labels != IGNORE_INDEX
            flatten_logits = shift_logits.contiguous().view(shift_labels.size(0) * shift_labels.size(1), -1)
            flatten_labels = shift_labels.contiguous().view(-1)
-            token_logps: "torch.Tensor" = criterion(flatten_logits, flatten_labels)
+            token_logps: torch.Tensor = criterion(flatten_logits, flatten_labels)
            token_logps = token_logps.contiguous().view(shift_logits.size(0), -1)
            sentence_logps = (token_logps * loss_mask).sum(-1) / loss_mask.sum(-1)
            total_ppl += sentence_logps.exp().sum().item()
--- a/scripts/stat_utils/length_cdf.py
+++ b/scripts/stat_utils/length_cdf.py
@@ -29,8 +29,8 @@ def length_cdf(
    template: str = "default",
    interval: int = 1000,
 ):
-    r"""
-    Calculates the distribution of the input lengths in the dataset.
+    r"""Calculate the distribution of the input lengths in the dataset.
+
    Usage: export CUDA_VISIBLE_DEVICES=0
    python length_cdf.py --model_name_or_path path_to_model --dataset alpaca_en_demo --template default
    """
--- a/scripts/vllm_infer.py
+++ b/scripts/vllm_infer.py
@@ -12,15 +12,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+import gc
 import json
 from typing import Optional

 import fire
+from tqdm import tqdm
 from transformers import Seq2SeqTrainingArguments

 from llamafactory.data import get_dataset, get_template_and_fix_tokenizer
 from llamafactory.extras.constants import IGNORE_INDEX
-from llamafactory.extras.misc import check_version, get_device_count
+from llamafactory.extras.misc import get_device_count
 from llamafactory.extras.packages import is_vllm_available
 from llamafactory.hparams import get_infer_args
 from llamafactory.model import load_tokenizer
@@ -47,16 +49,20 @@ def vllm_infer(
    max_new_tokens: int = 1024,
    repetition_penalty: float = 1.0,
    skip_special_tokens: bool = True,
+    default_system: Optional[str] = None,
+    enable_thinking: bool = True,
    seed: Optional[int] = None,
    pipeline_parallel_size: int = 1,
    image_max_pixels: int = 768 * 768,
    image_min_pixels: int = 32 * 32,
+    video_fps: float = 2.0,
+    video_maxlen: int = 128,
+    batch_size: int = 1024,
 ):
-    r"""
-    Performs batch generation using vLLM engine, which supports tensor parallelism.
+    r"""Perform batch generation using vLLM engine, which supports tensor parallelism.
+
    Usage: python vllm_infer.py --model_name_or_path meta-llama/Llama-2-7b-hf --template llama --dataset alpaca_en_demo
    """
-    check_version("vllm>=0.4.3,<=0.7.3")
    if pipeline_parallel_size > get_device_count():
        raise ValueError("Pipeline parallel size should be smaller than the number of gpus.")

@@ -70,6 +76,8 @@ def vllm_infer(
            cutoff_len=cutoff_len,
            max_samples=max_samples,
            preprocessing_num_workers=16,
+            default_system=default_system,
+            enable_thinking=enable_thinking,
            vllm_config=vllm_config,
            temperature=temperature,
            top_p=top_p,
@@ -84,26 +92,28 @@ def vllm_infer(
    tokenizer = tokenizer_module["tokenizer"]
    template_obj = get_template_and_fix_tokenizer(tokenizer, data_args)
    template_obj.mm_plugin.expand_mm_tokens = False  # for vllm generate
+
+    engine_args = {
+        "model": model_args.model_name_or_path,
+        "trust_remote_code": True,
+        "dtype": model_args.infer_dtype,
+        "max_model_len": cutoff_len + max_new_tokens,
+        "tensor_parallel_size": (get_device_count() // pipeline_parallel_size) or 1,
+        "pipeline_parallel_size": pipeline_parallel_size,
+        "disable_log_stats": True,
+        "enable_lora": model_args.adapter_name_or_path is not None,
+    }
+    if template_obj.mm_plugin.__class__.__name__ != "BasePlugin":
+        engine_args["limit_mm_per_prompt"] = {"image": 4, "video": 2, "audio": 2}
+
+    if isinstance(model_args.vllm_config, dict):
+        engine_args.update(model_args.vllm_config)
+
+    llm = LLM(**engine_args)
+
+    # load datasets
    dataset_module = get_dataset(template_obj, model_args, data_args, training_args, "ppo", **tokenizer_module)
-
-    inputs, prompts, labels = [], [], []
-    for sample in dataset_module["train_dataset"]:
-        if sample["images"]:
-            multi_modal_data = {
-                "image": template_obj.mm_plugin._regularize_images(
-                    sample["images"], image_max_pixels=image_max_pixels, image_min_pixels=image_min_pixels
-                )
-            }
-        else:
-            multi_modal_data = None
-
-        inputs.append({"prompt_token_ids": sample["input_ids"], "multi_modal_data": multi_modal_data})
-        prompts.append(tokenizer.decode(sample["input_ids"], skip_special_tokens=skip_special_tokens))
-        labels.append(
-            tokenizer.decode(
-                list(filter(lambda x: x != IGNORE_INDEX, sample["labels"])), skip_special_tokens=skip_special_tokens
-            )
-        )
+    train_dataset = dataset_module["train_dataset"]

    sampling_params = SamplingParams(
        repetition_penalty=generating_args.repetition_penalty or 1.0,  # repetition_penalty must > 0
@@ -120,30 +130,68 @@ def vllm_infer(
    else:
        lora_request = None

-    engine_args = {
-        "model": model_args.model_name_or_path,
-        "trust_remote_code": True,
-        "dtype": model_args.infer_dtype,
-        "max_model_len": cutoff_len + max_new_tokens,
-        "tensor_parallel_size": (get_device_count() // pipeline_parallel_size) or 1,
-        "pipeline_parallel_size": pipeline_parallel_size,
-        "disable_log_stats": True,
-        "enable_lora": model_args.adapter_name_or_path is not None,
-    }
-    if template_obj.mm_plugin.__class__.__name__ != "BasePlugin":
-        engine_args["limit_mm_per_prompt"] = {"image": 4, "video": 2}
+    # Store all results in these lists
+    all_prompts, all_preds, all_labels = [], [], []

-    if isinstance(model_args.vllm_config, dict):
-        engine_args.update(model_args.vllm_config)
+    # Add batch process to avoid the issue of too many files opened
+    for i in tqdm(range(0, len(train_dataset), batch_size), desc="Processing batched inference"):
+        vllm_inputs, prompts, labels = [], [], []
+        batch = train_dataset[i : min(i + batch_size, len(train_dataset))]

-    results = LLM(**engine_args).generate(inputs, sampling_params, lora_request=lora_request)
-    preds = [result.outputs[0].text for result in results]
+        for j in range(len(batch["input_ids"])):
+            if batch["images"][j] is not None:
+                image = batch["images"][j]
+                multi_modal_data = {
+                    "image": template_obj.mm_plugin._regularize_images(
+                        image, image_max_pixels=image_max_pixels, image_min_pixels=image_min_pixels
+                    )["images"]
+                }
+            elif batch["videos"][j] is not None:
+                video = batch["videos"][j]
+                multi_modal_data = {
+                    "video": template_obj.mm_plugin._regularize_videos(
+                        video,
+                        image_max_pixels=image_max_pixels,
+                        image_min_pixels=image_min_pixels,
+                        video_fps=video_fps,
+                        video_maxlen=video_maxlen,
+                    )["videos"]
+                }
+            elif batch["audios"][j] is not None:
+                audio = batch["audios"][j]
+                audio_data = template_obj.mm_plugin._regularize_audios(
+                    audio,
+                    sampling_rate=16000,
+                )
+                multi_modal_data = {"audio": zip(audio_data["audios"], audio_data["sampling_rates"])}
+            else:
+                multi_modal_data = None
+
+            vllm_inputs.append({"prompt_token_ids": batch["input_ids"][j], "multi_modal_data": multi_modal_data})
+            prompts.append(tokenizer.decode(batch["input_ids"][j], skip_special_tokens=skip_special_tokens))
+            labels.append(
+                tokenizer.decode(
+                    list(filter(lambda x: x != IGNORE_INDEX, batch["labels"][j])),
+                    skip_special_tokens=skip_special_tokens,
+                )
+            )
+
+        results = llm.generate(vllm_inputs, sampling_params, lora_request=lora_request)
+        preds = [result.outputs[0].text for result in results]
+
+        # Accumulate results
+        all_prompts.extend(prompts)
+        all_preds.extend(preds)
+        all_labels.extend(labels)
+        gc.collect()
+
+    # Write all results at once outside the loop
    with open(save_name, "w", encoding="utf-8") as f:
-        for text, pred, label in zip(prompts, preds, labels):
+        for text, pred, label in zip(all_prompts, all_preds, all_labels):
            f.write(json.dumps({"prompt": text, "predict": pred, "label": label}, ensure_ascii=False) + "\n")

    print("*" * 70)
-    print(f"{len(prompts)} generated results have been saved at {save_name}.")
+    print(f"{len(all_prompts)} total generated results have been saved at {save_name}.")
    print("*" * 70)


--- a/setup.py
+++ b/setup.py
@@ -14,7 +14,6 @@

 import os
 import re
-from typing import List

 from setuptools import find_packages, setup

@@ -27,14 +26,14 @@ def get_version() -> str:
        return version


-def get_requires() -> List[str]:
+def get_requires() -> list[str]:
    with open("requirements.txt", encoding="utf-8") as f:
        file_content = f.read()
        lines = [line.strip() for line in file_content.strip().split("\n") if not line.startswith("#")]
        return lines


-def get_console_scripts() -> List[str]:
+def get_console_scripts() -> list[str]:
    console_scripts = ["llamafactory-cli = llamafactory.cli:main"]
    if os.getenv("ENABLE_SHORT_CONSOLE", "1").lower() in ["true", "y", "1"]:
        console_scripts.append("lmf = llamafactory.cli:main")
@@ -43,23 +42,22 @@ def get_console_scripts() -> List[str]:


 extra_require = {
-    "torch": ["torch>=1.13.1"],
+    "torch": ["torch>=2.0.0", "torchvision>=0.15.0"],
    "torch-npu": ["torch==2.4.0", "torch-npu==2.4.0.post2", "decorator"],
    "metrics": ["nltk", "jieba", "rouge-chinese"],
-    "deepspeed": ["deepspeed>=0.10.0,<=0.16.4"],
-    "liger-kernel": ["liger-kernel"],
+    "deepspeed": ["deepspeed>=0.10.0,<=0.16.9"],
+    "liger-kernel": ["liger-kernel>=0.5.5"],
    "bitsandbytes": ["bitsandbytes>=0.39.0"],
    "hqq": ["hqq"],
    "eetq": ["eetq"],
-    "gptq": ["optimum>=1.17.0", "auto-gptq>=0.5.0"],
-    "awq": ["autoawq"],
+    "gptq": ["optimum>=1.24.0", "gptqmodel>=2.0.0"],
    "aqlm": ["aqlm[gpu]>=1.1.0"],
-    "vllm": ["vllm>=0.4.3,<=0.7.3"],
+    "vllm": ["vllm>=0.4.3,<=0.8.6"],
+    "sglang": ["sglang[srt]>=0.4.5", "transformers==4.51.1"],
    "galore": ["galore-torch"],
    "apollo": ["apollo-torch"],
    "badam": ["badam>=1.2.1"],
    "adam-mini": ["adam-mini"],
-    "qwen": ["transformers_stream_generator"],
    "minicpm_v": [
        "soundfile",
        "torchvision",
@@ -73,7 +71,7 @@ extra_require = {
    "modelscope": ["modelscope"],
    "openmind": ["openmind"],
    "swanlab": ["swanlab"],
-    "dev": ["pre-commit", "ruff", "pytest"],
+    "dev": ["pre-commit", "ruff", "pytest", "build"],
 }


@@ -82,11 +80,11 @@ def main():
        name="llamafactory",
        version=get_version(),
        author="hiyouga",
-        author_email="hiyouga AT buaa.edu.cn",
-        description="Easy-to-use LLM fine-tuning framework",
+        author_email="hiyouga@buaa.edu.cn",
+        description="Unified Efficient Fine-Tuning of 100+ LLMs",
        long_description=open("README.md", encoding="utf-8").read(),
        long_description_content_type="text/markdown",
-        keywords=["LLaMA", "BLOOM", "Falcon", "LLM", "ChatGPT", "transformer", "pytorch", "deep learning"],
+        keywords=["AI", "LLM", "GPT", "ChatGPT", "Llama", "Transformer", "DeepSeek", "Pytorch"],
        license="Apache 2.0 License",
        url="https://github.com/hiyouga/LLaMA-Factory",
        package_dir={"": "src"},
--- a/src/llamafactory/init.py
+++ b/src/llamafactory/init.py
@@ -12,29 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-r"""
-Efficient fine-tuning of large language models.
+r"""Efficient fine-tuning of large language models.

 Level:
  api, webui > chat, eval, train > data, model > hparams > extras

-Dependency graph:
-  main:
-    transformers>=4.41.2,<=4.49.0,!=4.46.*,!=4.47.*,!=4.48.0
-    datasets>=2.16.0,<=3.2.0
-    accelerate>=0.34.0,<=1.2.1
-    peft>=0.11.1,<=0.12.0
-    trl>=0.8.6,<=0.9.6
-  attention:
-    transformers>=4.42.4 (gemma+fa2)
-  longlora:
-    transformers>=4.41.2,<4.48.0
-  packing:
-    transformers>=4.43.0
-
 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
--- a/src/llamafactory/api/app.py
+++ b/src/llamafactory/api/app.py
@@ -16,9 +16,7 @@ import asyncio
 import os
 from contextlib import asynccontextmanager
 from functools import partial
-from typing import Optional
-
-from typing_extensions import Annotated
+from typing import Annotated, Optional

 from ..chat import ChatModel
 from ..extras.constants import EngineName
--- a/src/llamafactory/api/chat.py
+++ b/src/llamafactory/api/chat.py
@@ -18,11 +18,12 @@ import json
 import os
 import re
 import uuid
-from typing import TYPE_CHECKING, AsyncGenerator, Dict, List, Optional, Tuple
+from collections.abc import AsyncGenerator
+from typing import TYPE_CHECKING, Optional

 from ..data import Role as DataRole
 from ..extras import logging
-from ..extras.constants import IMAGE_PLACEHOLDER
+from ..extras.constants import AUDIO_PLACEHOLDER, IMAGE_PLACEHOLDER, VIDEO_PLACEHOLDER
 from ..extras.misc import is_env_enabled
 from ..extras.packages import is_fastapi_available, is_pillow_available, is_requests_available
 from .common import dictify, jsonify
@@ -55,7 +56,7 @@ if is_requests_available():

 if TYPE_CHECKING:
    from ..chat import ChatModel
-    from ..data.mm_plugin import ImageInput
+    from ..data.mm_plugin import AudioInput, ImageInput, VideoInput
    from .protocol import ChatCompletionRequest, ScoreEvaluationRequest


@@ -71,7 +72,14 @@ ROLE_MAPPING = {

 def _process_request(
    request: "ChatCompletionRequest",
-) -> Tuple[List[Dict[str, str]], Optional[str], Optional[str], Optional[List["ImageInput"]]]:
+) -> tuple[
+    list[dict[str, str]],
+    Optional[str],
+    Optional[str],
+    Optional[list["ImageInput"]],
+    Optional[list["VideoInput"]],
+    Optional[list["AudioInput"]],
+]:
    if is_env_enabled("API_VERBOSE", "1"):
        logger.info_rank0(f"==== request ====\n{json.dumps(dictify(request), indent=2, ensure_ascii=False)}")

@@ -79,7 +87,8 @@ def _process_request(
        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid length")

    if request.messages[0].role == Role.SYSTEM:
-        system = request.messages.pop(0).content
+        content = request.messages.pop(0).content
+        system = content[0].text if isinstance(content, list) else content
    else:
        system = None

@@ -87,7 +96,7 @@ def _process_request(
        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Only supports u/a/u/a/u...")

    input_messages = []
-    images = []
+    images, videos, audios = [], [], []
    for i, message in enumerate(request.messages):
        if i % 2 == 0 and message.role not in [Role.USER, Role.TOOL]:
            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
@@ -106,7 +115,7 @@ def _process_request(
            for input_item in message.content:
                if input_item.type == "text":
                    text_content += input_item.text
-                else:
+                elif input_item.type == "image_url":
                    text_content += IMAGE_PLACEHOLDER
                    image_url = input_item.image_url.url
                    if re.match(r"^data:image\/(png|jpg|jpeg|gif|bmp);base64,(.+)$", image_url):  # base64 image
@@ -117,6 +126,32 @@ def _process_request(
                        image_stream = requests.get(image_url, stream=True).raw

                    images.append(Image.open(image_stream).convert("RGB"))
+                elif input_item.type == "video_url":
+                    text_content += VIDEO_PLACEHOLDER
+                    video_url = input_item.video_url.url
+                    if re.match(r"^data:video\/(mp4|mkv|avi|mov);base64,(.+)$", video_url):  # base64 video
+                        video_stream = io.BytesIO(base64.b64decode(video_url.split(",", maxsplit=1)[1]))
+                    elif os.path.isfile(video_url):  # local file
+                        video_stream = open(video_url, "rb")
+                    else:  # web uri
+                        video_stream = requests.get(video_url, stream=True).raw
+
+                    videos.append(video_stream)
+                elif input_item.type == "audio_url":
+                    text_content += AUDIO_PLACEHOLDER
+                    audio_url = input_item.audio_url.url
+                    if re.match(r"^data:audio\/(mpeg|mp3|wav|ogg);base64,(.+)$", audio_url):  # base64 audio
+                        audio_stream = io.BytesIO(base64.b64decode(audio_url.split(",", maxsplit=1)[1]))
+                    elif os.path.isfile(audio_url):  # local file
+                        audio_stream = open(audio_url, "rb")
+                    else:  # web uri
+                        audio_stream = requests.get(audio_url, stream=True).raw
+
+                    audios.append(audio_stream)
+                else:
+                    raise HTTPException(
+                        status_code=status.HTTP_400_BAD_REQUEST, detail=f"Invalid input type {input_item.type}."
+                    )

            input_messages.append({"role": ROLE_MAPPING[message.role], "content": text_content})
        else:
@@ -131,7 +166,7 @@ def _process_request(
    else:
        tools = None

-    return input_messages, system, tools, images or None
+    return input_messages, system, tools, images or None, videos or None, audios or None


 def _create_stream_chat_completion_chunk(
@@ -150,17 +185,20 @@ async def create_chat_completion_response(
    request: "ChatCompletionRequest", chat_model: "ChatModel"
 ) -> "ChatCompletionResponse":
    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
-    input_messages, system, tools, images = _process_request(request)
+    input_messages, system, tools, images, videos, audios = _process_request(request)
    responses = await chat_model.achat(
        input_messages,
        system,
        tools,
        images,
+        videos,
+        audios,
        do_sample=request.do_sample,
        temperature=request.temperature,
        top_p=request.top_p,
        max_new_tokens=request.max_tokens,
        num_return_sequences=request.n,
+        repetition_penalty=request.presence_penalty,
        stop=request.stop,
    )

@@ -201,7 +239,7 @@ async def create_stream_chat_completion_response(
    request: "ChatCompletionRequest", chat_model: "ChatModel"
 ) -> AsyncGenerator[str, None]:
    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
-    input_messages, system, tools, images = _process_request(request)
+    input_messages, system, tools, images, videos, audios = _process_request(request)
    if tools:
        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Cannot stream function calls.")

@@ -216,10 +254,13 @@ async def create_stream_chat_completion_response(
        system,
        tools,
        images,
+        videos,
+        audios,
        do_sample=request.do_sample,
        temperature=request.temperature,
        top_p=request.top_p,
        max_new_tokens=request.max_tokens,
+        repetition_penalty=request.presence_penalty,
        stop=request.stop,
    ):
        if len(new_token) != 0:
--- a/src/llamafactory/api/common.py
+++ b/src/llamafactory/api/common.py
@@ -13,14 +13,14 @@
 # limitations under the License.

 import json
-from typing import TYPE_CHECKING, Any, Dict
+from typing import TYPE_CHECKING, Any


 if TYPE_CHECKING:
    from pydantic import BaseModel


-def dictify(data: "BaseModel") -> Dict[str, Any]:
+def dictify(data: "BaseModel") -> dict[str, Any]:
    try:  # pydantic v2
        return data.model_dump(exclude_unset=True)
    except AttributeError:  # pydantic v1
--- a/src/llamafactory/api/protocol.py
+++ b/src/llamafactory/api/protocol.py
@@ -14,7 +14,7 @@

 import time
 from enum import Enum, unique
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Optional, Union

 from pydantic import BaseModel, Field
 from typing_extensions import Literal
@@ -45,7 +45,7 @@ class ModelCard(BaseModel):

 class ModelList(BaseModel):
    object: Literal["list"] = "list"
-    data: List[ModelCard] = []
+    data: list[ModelCard] = []


 class Function(BaseModel):
@@ -56,7 +56,7 @@ class Function(BaseModel):
 class FunctionDefinition(BaseModel):
    name: str
    description: str
-    parameters: Dict[str, Any]
+    parameters: dict[str, Any]


 class FunctionAvailable(BaseModel):
@@ -70,38 +70,42 @@ class FunctionCall(BaseModel):
    function: Function


-class ImageURL(BaseModel):
+class URL(BaseModel):
    url: str
+    detail: Literal["auto", "low", "high"] = "auto"


 class MultimodalInputItem(BaseModel):
-    type: Literal["text", "image_url"]
+    type: Literal["text", "image_url", "video_url", "audio_url"]
    text: Optional[str] = None
-    image_url: Optional[ImageURL] = None
+    image_url: Optional[URL] = None
+    video_url: Optional[URL] = None
+    audio_url: Optional[URL] = None


 class ChatMessage(BaseModel):
    role: Role
-    content: Optional[Union[str, List[MultimodalInputItem]]] = None
-    tool_calls: Optional[List[FunctionCall]] = None
+    content: Optional[Union[str, list[MultimodalInputItem]]] = None
+    tool_calls: Optional[list[FunctionCall]] = None


 class ChatCompletionMessage(BaseModel):
    role: Optional[Role] = None
    content: Optional[str] = None
-    tool_calls: Optional[List[FunctionCall]] = None
+    tool_calls: Optional[list[FunctionCall]] = None


 class ChatCompletionRequest(BaseModel):
    model: str
-    messages: List[ChatMessage]
-    tools: Optional[List[FunctionAvailable]] = None
+    messages: list[ChatMessage]
+    tools: Optional[list[FunctionAvailable]] = None
    do_sample: Optional[bool] = None
    temperature: Optional[float] = None
    top_p: Optional[float] = None
    n: int = 1
+    presence_penalty: Optional[float] = None
    max_tokens: Optional[int] = None
-    stop: Optional[Union[str, List[str]]] = None
+    stop: Optional[Union[str, list[str]]] = None
    stream: bool = False


@@ -128,7 +132,7 @@ class ChatCompletionResponse(BaseModel):
    object: Literal["chat.completion"] = "chat.completion"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
-    choices: List[ChatCompletionResponseChoice]
+    choices: list[ChatCompletionResponseChoice]
    usage: ChatCompletionResponseUsage


@@ -137,12 +141,12 @@ class ChatCompletionStreamResponse(BaseModel):
    object: Literal["chat.completion.chunk"] = "chat.completion.chunk"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
-    choices: List[ChatCompletionStreamResponseChoice]
+    choices: list[ChatCompletionStreamResponseChoice]


 class ScoreEvaluationRequest(BaseModel):
    model: str
-    messages: List[str]
+    messages: list[str]
    max_length: Optional[int] = None


@@ -150,4 +154,4 @@ class ScoreEvaluationResponse(BaseModel):
    id: str
    object: Literal["score.evaluation"] = "score.evaluation"
    model: str
-    scores: List[float]
+    scores: list[float]
--- a/src/llamafactory/chat/base_engine.py
+++ b/src/llamafactory/chat/base_engine.py
@@ -13,8 +13,9 @@
 # limitations under the License.

 from abc import ABC, abstractmethod
+from collections.abc import AsyncGenerator
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, List, Literal, Optional, Sequence, Union
+from typing import TYPE_CHECKING, Any, Literal, Optional, Union


 if TYPE_CHECKING:
@@ -36,8 +37,7 @@ class Response:


 class BaseEngine(ABC):
-    r"""
-    Base class for inference engine of chat models.
+    r"""Base class for inference engine of chat models.

    Must implements async methods: chat(), stream_chat() and get_scores().
    """
@@ -47,7 +47,7 @@ class BaseEngine(ABC):
    tokenizer: "PreTrainedTokenizer"
    can_generate: bool
    template: "Template"
-    generating_args: Dict[str, Any]
+    generating_args: dict[str, Any]

    @abstractmethod
    def __init__(
@@ -57,50 +57,42 @@ class BaseEngine(ABC):
        finetuning_args: "FinetuningArguments",
        generating_args: "GeneratingArguments",
    ) -> None:
-        r"""
-        Initializes an inference engine.
-        """
+        r"""Initialize an inference engine."""
        ...

    @abstractmethod
    async def chat(
        self,
-        messages: Sequence[Dict[str, str]],
+        messages: list[dict[str, str]],
        system: Optional[str] = None,
        tools: Optional[str] = None,
-        images: Optional[Sequence["ImageInput"]] = None,
-        videos: Optional[Sequence["VideoInput"]] = None,
-        audios: Optional[Sequence["AudioInput"]] = None,
+        images: Optional[list["ImageInput"]] = None,
+        videos: Optional[list["VideoInput"]] = None,
+        audios: Optional[list["AudioInput"]] = None,
        **input_kwargs,
-    ) -> List["Response"]:
-        r"""
-        Gets a list of responses of the chat model.
-        """
+    ) -> list["Response"]:
+        r"""Get a list of responses of the chat model."""
        ...

    @abstractmethod
    async def stream_chat(
        self,
-        messages: Sequence[Dict[str, str]],
+        messages: list[dict[str, str]],
        system: Optional[str] = None,
        tools: Optional[str] = None,
-        images: Optional[Sequence["ImageInput"]] = None,
-        videos: Optional[Sequence["VideoInput"]] = None,
-        audios: Optional[Sequence["AudioInput"]] = None,
+        images: Optional[list["ImageInput"]] = None,
+        videos: Optional[list["VideoInput"]] = None,
+        audios: Optional[list["AudioInput"]] = None,
        **input_kwargs,
    ) -> AsyncGenerator[str, None]:
-        r"""
-        Gets the response token-by-token of the chat model.
-        """
+        r"""Get the response token-by-token of the chat model."""
        ...

    @abstractmethod
    async def get_scores(
        self,
-        batch_input: List[str],
+        batch_input: list[str],
        **input_kwargs,
-    ) -> List[float]:
-        r"""
-        Gets a list of scores of the reward model.
-        """
+    ) -> list[float]:
+        r"""Get a list of scores of the reward model."""
        ...
--- a/Show More
+++ b/Show More