[model] Add DeepSpeed Z3 leaf module for Qwen3-Next (#10194 )

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
[misc] remove safe_serialization arg for transformers v5 compatibility (#10208 )
2026-02-26 07:45:59 +08:00 · 2026-02-24 19:54:37 +08:00 · 2026-02-24 11:14:19 +08:00 · 2026-02-20 22:02:49 +08:00 · 2026-02-12 22:14:10 +08:00 · 2026-02-12 20:37:41 +08:00
98 changed files with 3748 additions and 79 deletions
--- a/.github/workflows/docs.yml
+++ b/.github/workflows/docs.yml
@@ -0,0 +1,77 @@
+name: Build and Deploy Sphinx Docs
+
+on:
+  push:
+    branches: ["main"]
+    paths:
+      - "docs/**"
+  pull_request:
+    branches: ["main"]
+    paths:
+      - "docs/**"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  pages: write
+  id-token: write
+
+concurrency:
+  group: "pages"
+  cancel-in-progress: false
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+          
+      - name: Install dependencies
+        run: |
+          pip install -r docs/requirements.txt
+          
+      - name: Build Sphinx
+        run: |
+          sphinx-build -b html docs/zh docs/_build/html/zh
+          sphinx-build -b html docs/en docs/_build/html/en
+
+          printf '%s\n' \
+            '<!DOCTYPE html>' \
+            '<html>' \
+            '  <head>' \
+            '    <meta charset="utf-8" />' \
+            '    <meta http-equiv="refresh" content="0; url=zh/index.html" />' \
+            '    <script>window.location.href="zh/index.html"+window.location.search+window.location.hash;</script>' \
+            '    <title>Redirecting...</title>' \
+            '  </head>' \
+            '  <body>' \
+            '    <a href="zh/index.html">Redirecting...</a>' \
+            '  </body>' \
+            '</html>' \
+            > docs/_build/html/index.html
+
+          touch docs/_build/html/.nojekyll
+          
+      - name: Setup Pages
+        uses: actions/configure-pages@v5
+        
+      - name: Upload artifact
+        uses: actions/upload-pages-artifact@v3
+        with:
+          path: docs/_build/html
+
+  deploy:
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    runs-on: ubuntu-latest
+    needs: build
+    steps:
+      - name: Deploy to GitHub Pages
+        id: deployment
+        uses: actions/deploy-pages@v4
--- a/.github/workflows/tests_cuda.yml
+++ b/.github/workflows/tests_cuda.yml
@@ -61,6 +61,7 @@ jobs:
          uv venv
          uv pip install -e .
          uv pip install -r requirements/dev.txt
+          uv pip install -r requirements/bitsandbytes.txt

      - name: Check quality
        run: |
--- a/docs/Makefile
+++ b/docs/Makefile
@@ -0,0 +1,20 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line, and also
+# from the environment for the first two.
+SPHINXOPTS    =
+SPHINXBUILD   = sphinx-build
+SOURCEDIR     = .
+BUILDDIR      = _build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
--- a/docs/_static/css/lang-switcher.css
+++ b/docs/_static/css/lang-switcher.css
@@ -0,0 +1,50 @@
+.lang-switcher {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+}
+
+.lang-switcher__select {
+  appearance: none;
+  -webkit-appearance: none;
+  -moz-appearance: none;
+  padding: 6px 28px 6px 10px;
+  border-radius: 999px;
+  border: 1px solid rgba(0, 0, 0, 0.18);
+  background-color: #ffffff;
+  color: #333333;
+  font-size: 13px;
+  line-height: 18px;
+  box-shadow: 0 1px 2px rgba(0, 0, 0, 0.08);
+  cursor: pointer;
+  background-image: linear-gradient(45deg, transparent 50%, #667085 50%),
+    linear-gradient(135deg, #667085 50%, transparent 50%);
+  background-position: calc(100% - 16px) 50%, calc(100% - 11px) 50%;
+  background-size: 5px 5px, 5px 5px;
+  background-repeat: no-repeat;
+}
+
+.lang-switcher__select:focus {
+  outline: none;
+  border-color: rgba(41, 128, 185, 0.65);
+  box-shadow: 0 0 0 3px rgba(41, 128, 185, 0.18);
+}
+
+.wy-side-nav-search .lang-switcher {
+  margin-top: 10px;
+}
+
+.wy-side-nav-search .lang-switcher__select {
+  border-color: rgba(255, 255, 255, 0.18);
+  background-color: rgba(255, 255, 255, 0.08);
+  color: #ffffff;
+  box-shadow: none;
+  background-image: linear-gradient(45deg, transparent 50%, rgba(255, 255, 255, 0.75) 50%),
+    linear-gradient(135deg, rgba(255, 255, 255, 0.75) 50%, transparent 50%);
+}
+
+.wy-side-nav-search .lang-switcher__select:focus {
+  border-color: rgba(255, 255, 255, 0.45);
+  box-shadow: 0 0 0 3px rgba(255, 255, 255, 0.12);
+}
+
--- a/docs/_static/js/switcher.js
+++ b/docs/_static/js/switcher.js
@@ -0,0 +1,93 @@
+document.addEventListener('DOMContentLoaded', function () {
+  var path = window.location.pathname || '';
+  var isZh = path.indexOf('/zh/') !== -1;
+  var isEn = path.indexOf('/en/') !== -1;
+  if (!isZh && !isEn) return;
+
+  var currentLang = isZh ? 'zh' : 'en';
+
+  function buildSwitcher() {
+    var container = document.createElement('div');
+    container.className = 'lang-switcher';
+
+    var select = document.createElement('select');
+    select.setAttribute('aria-label', 'Language');
+    select.className = 'lang-switcher__select';
+
+    var optionZh = document.createElement('option');
+    optionZh.value = 'zh';
+    optionZh.textContent = 'Simplified Chinese';
+    optionZh.selected = isZh;
+
+    var optionEn = document.createElement('option');
+    optionEn.value = 'en';
+    optionEn.textContent = 'English';
+    optionEn.selected = isEn;
+
+    select.appendChild(optionZh);
+    select.appendChild(optionEn);
+
+    select.addEventListener('change', function () {
+      var nextLang = select.value;
+      if (nextLang === currentLang) return;
+      var targetUrl = path.replace('/' + currentLang + '/', '/' + nextLang + '/');
+      window.location.href = targetUrl + window.location.search + window.location.hash;
+    });
+
+    container.appendChild(select);
+    return container;
+  }
+
+  function hideOtherLanguageToc() {
+    var captions = document.querySelectorAll('p.caption');
+    for (var i = 0; i < captions.length; i++) {
+      var caption = captions[i];
+      var textEl = caption.querySelector('.caption-text');
+      if (!textEl) continue;
+      var label = (textEl.textContent || '').trim().toLowerCase();
+
+      var isCaptionZh = label === '中文' || label === 'chinese' || label === 'zh';
+      var isCaptionEn = label === 'english' || label === 'en';
+
+      if (!isCaptionZh && !isCaptionEn) continue;
+
+      var shouldHide = (currentLang === 'zh' && isCaptionEn) || (currentLang === 'en' && isCaptionZh);
+      var shouldHideCaption = true;
+
+      var next = caption.nextElementSibling;
+      if (next && next.tagName && next.tagName.toLowerCase() === 'ul') {
+        if (shouldHide) {
+          caption.style.display = 'none';
+          next.style.display = 'none';
+        } else if (shouldHideCaption) {
+          caption.style.display = 'none';
+        }
+      } else if (shouldHide) {
+        caption.style.display = 'none';
+      } else if (shouldHideCaption) {
+        caption.style.display = 'none';
+      }
+    }
+  }
+
+  var side = document.querySelector('.wy-side-nav-search');
+  if (side) {
+    var sideSwitcher = buildSwitcher();
+    sideSwitcher.style.marginTop = '8px';
+    sideSwitcher.style.display = 'flex';
+    sideSwitcher.style.justifyContent = 'center';
+    side.appendChild(sideSwitcher);
+  } else {
+    var topRight = buildSwitcher();
+    topRight.style.position = 'fixed';
+    topRight.style.top = '12px';
+    topRight.style.right = '12px';
+    topRight.style.zIndex = '9999';
+    document.body.appendChild(topRight);
+  }
+
+  hideOtherLanguageToc();
+  window.addEventListener('load', hideOtherLanguageToc);
+  setTimeout(hideOtherLanguageToc, 50);
+  setTimeout(hideOtherLanguageToc, 300);
+});
--- a/docs/conf.py
+++ b/docs/conf.py
@@ -0,0 +1,37 @@
+# Configuration file for the Sphinx documentation builder.
+
+import os
+import sys
+
+# Define common settings here
+project = 'LlamaFactory'
+copyright = '2024, LlamaFactory Team'
+author = 'LlamaFactory Team'
+
+extensions = [
+    'sphinx.ext.autodoc',
+    'sphinx.ext.viewcode',
+    'sphinx.ext.napoleon',
+    'myst_parser',
+]
+
+templates_path = ['_templates']
+exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store']
+
+html_theme = 'sphinx_rtd_theme'
+
+html_static_path = ['_static']
+
+html_js_files = [
+    'js/switcher.js',
+]
+
+html_css_files = [
+    'css/lang-switcher.css',
+]
+
+myst_enable_extensions = [
+    "colon_fence",
+    "deflist",
+]
+myst_heading_anchors = 3
--- a/docs/en/advanced/custom-kernels/custom-kernels.md
+++ b/docs/en/advanced/custom-kernels/custom-kernels.md
@@ -0,0 +1,3 @@
+# Custom Kernels
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/advanced/custom-kernels/fused-operators.md
+++ b/docs/en/advanced/custom-kernels/fused-operators.md
@@ -0,0 +1,3 @@
+# Fused Operators
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/advanced/custom-kernels/triton.md
+++ b/docs/en/advanced/custom-kernels/triton.md
@@ -0,0 +1,3 @@
+# Triton
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/advanced/distributed/deepspeed.md
+++ b/docs/en/advanced/distributed/deepspeed.md
@@ -0,0 +1,3 @@
+# DeepSpeed
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/advanced/distributed/fsdp.md
+++ b/docs/en/advanced/distributed/fsdp.md
@@ -0,0 +1,3 @@
+# FSDP
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/advanced/distributed/parallel-dp-tp-ep-sp-cp.md
+++ b/docs/en/advanced/distributed/parallel-dp-tp-ep-sp-cp.md
@@ -0,0 +1,3 @@
+# Parallel (DP, TP, EP, SP, CP)
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/advanced/lora-and-quantization/lora.md
+++ b/docs/en/advanced/lora-and-quantization/lora.md
@@ -0,0 +1,3 @@
+# LoRA
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/advanced/lora-and-quantization/quantization.md
+++ b/docs/en/advanced/lora-and-quantization/quantization.md
@@ -0,0 +1,3 @@
+# Quantization
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/conf.py
+++ b/docs/en/conf.py
@@ -0,0 +1,20 @@
+import os
+import sys
+
+# Add parent dir to path to allow importing conf.py
+sys.path.insert(0, os.path.abspath('..'))
+
+from conf import *
+
+# Language settings
+language = 'en'
+html_search_language = 'en'
+
+# Static files
+# Point to the root _static directory
+html_static_path = ['../_static']
+
+# Add custom JS for language switcher
+html_js_files = [
+    'js/switcher.js',
+]
--- a/docs/en/data-preparation/data-processing.md
+++ b/docs/en/data-preparation/data-processing.md
@@ -0,0 +1,3 @@
+# Data Processing
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/dev-guide/core/data-engine.md
+++ b/docs/en/dev-guide/core/data-engine.md
@@ -0,0 +1,3 @@
+# DataEngine
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/dev-guide/core/model-engine.md
+++ b/docs/en/dev-guide/core/model-engine.md
@@ -0,0 +1,3 @@
+# ModelEngine
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/dev-guide/core/trainer.md
+++ b/docs/en/dev-guide/core/trainer.md
@@ -0,0 +1,3 @@
+# Trainer
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/dev-guide/plugins/data-plugins.md
+++ b/docs/en/dev-guide/plugins/data-plugins.md
@@ -0,0 +1,3 @@
+# Data Plugins
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/dev-guide/plugins/model-plugins/initialization.md
+++ b/docs/en/dev-guide/plugins/model-plugins/initialization.md
@@ -0,0 +1,3 @@
+# Initialization
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/dev-guide/plugins/model-plugins/kernels.md
+++ b/docs/en/dev-guide/plugins/model-plugins/kernels.md
@@ -0,0 +1,3 @@
+# Kernels
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/dev-guide/plugins/model-plugins/rendering.md
+++ b/docs/en/dev-guide/plugins/model-plugins/rendering.md
@@ -0,0 +1,3 @@
+# Rendering
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/getting-started.md
+++ b/docs/en/getting-started.md
@@ -0,0 +1,3 @@
+# Getting Started
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/hyperparameters/data-argument.md
+++ b/docs/en/hyperparameters/data-argument.md
@@ -0,0 +1,3 @@
+# Data Argument
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/hyperparameters/model-argument.md
+++ b/docs/en/hyperparameters/model-argument.md
@@ -0,0 +1,3 @@
+# Model Argument
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/hyperparameters/sample-argument.md
+++ b/docs/en/hyperparameters/sample-argument.md
@@ -0,0 +1,3 @@
+# Sample Argument
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/hyperparameters/training-argument.md
+++ b/docs/en/hyperparameters/training-argument.md
@@ -0,0 +1,3 @@
+# Training Argument
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/index.rst
+++ b/docs/en/index.rst
@@ -0,0 +1,62 @@
+LlamaFactory Docs
+=================
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Getting Started
+
+   getting-started
+   installation
+   llamaboard-web-ui
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Data Preparation
+
+   data-preparation/data-processing
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Training
+
+   training/sft
+   training/dpo
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Inference
+
+   inference/deploy
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Advanced
+
+   advanced/lora-and-quantization/lora
+   advanced/lora-and-quantization/quantization
+   advanced/distributed/fsdp
+   advanced/distributed/deepspeed
+   advanced/distributed/parallel-dp-tp-ep-sp-cp
+   advanced/custom-kernels/triton
+   advanced/custom-kernels/fused-operators
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Hyperparameters
+
+   hyperparameters/data-argument
+   hyperparameters/model-argument
+   hyperparameters/sample-argument
+   hyperparameters/training-argument
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Dev Guide
+
+   dev-guide/core/data-engine
+   dev-guide/core/model-engine
+   dev-guide/core/trainer
+   dev-guide/plugins/data-plugins
+   dev-guide/plugins/model-plugins/initialization
+   dev-guide/plugins/model-plugins/kernels
+   dev-guide/plugins/model-plugins/rendering
--- a/docs/en/inference/deploy.md
+++ b/docs/en/inference/deploy.md
@@ -0,0 +1,3 @@
+# Deploy
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/installation.md
+++ b/docs/en/installation.md
@@ -0,0 +1,3 @@
+# Installation
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/llamaboard-web-ui.md
+++ b/docs/en/llamaboard-web-ui.md
@@ -0,0 +1,3 @@
+# LlamaBoard Web UI
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/training/dpo.md
+++ b/docs/en/training/dpo.md
@@ -0,0 +1,3 @@
+# DPO
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/en/training/sft.md
+++ b/docs/en/training/sft.md
@@ -0,0 +1,3 @@
+# SFT
+
+This page is not yet available in English. Use the language switcher to view Simplified Chinese.
--- a/docs/make.bat
+++ b/docs/make.bat
@@ -0,0 +1,35 @@
+@ECHO OFF
+
+pushd %~dp0
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+	set SPHINXBUILD=sphinx-build
+)
+set SOURCEDIR=.
+set BUILDDIR=_build
+
+if "%1" == "" goto help
+
+%SPHINXBUILD% >NUL 2>NUL
+if errorlevel 9009 (
+	echo.
+	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
+	echo.installed, then set the SPHINXBUILD environment variable to point
+	echo.to the full path of the 'sphinx-build' executable. Alternatively you
+	echo.may add the Sphinx directory to your PATH.
+	echo.
+	echo.If you don't have Sphinx installed, grab it from
+	echo.http://sphinx-doc.org/
+	exit /b 1
+)
+
+%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+goto end
+
+:help
+%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+
+:end
+popd
--- a/docs/requirements.txt
+++ b/docs/requirements.txt
@@ -0,0 +1,3 @@
+sphinx>=6.0.0
+sphinx-rtd-theme>=1.2.0
+myst-parser>=2.0.0
--- a/docs/zh/advanced/custom-kernels/custom-kernels.md
+++ b/docs/zh/advanced/custom-kernels/custom-kernels.md
@@ -0,0 +1,93 @@
+# LLaMA-Factory Kernels 系统
+
+## 概述
+
+LLaMA-Factory Kernels 系统用于管理不同硬件设备提供的高性能计算内核（kernel）实现，该系统通过替换模型中的关键模块（如 RMSNorm、SwiGLU、RoPE、MoE 等）为硬件优化的版本，从而显著提升模型训练和推理的性能。
+
+Kernels 系统采用基于注册表的自动发现机制，能够根据当前运行环境自动检测可用的硬件设备（NPU、CUDA 等），并使能相应的高性能 kernels。这种设计使得用户无需关心底层实现细节，只需简单调用接口即可获得性能提升。
+
+## 核心特性
+
+- **自动注册机制**：基于 `@register_kernel` 装饰器实现自动注册系统。系统启动时会自动扫描 `ops` 目录下的 kernel 实现，并将其注册到全局注册表中。
+
+- **设备适配感知**：自动检测当前硬件设备（NPU、CUDA 等）并应用相应的优化。系统会跳过不支持的设备，确保在不同环境下都能正常工作。
+
+- **模块化设计**：每个 kernel 独立实现，互不干扰。可以单独应用某个 kernel，也可以批量应用所有默认的 kernels。
+
+- **后向兼容**：kernel 替换不修改模型权重，保持数值一致性。优化后的实现与原始实现保持精度一致（在浮点误差范围内）。
+
+- **灵活扩展**：通过继承 `BaseKernel` 基类并使用装饰器，可以轻松添加新的 kernel 实现，支持新的硬件设备或优化算法。
+
+## 使用方式
+
+### 1. 通过训练 YAML 配置文件使用
+
+要在训练过程中使能 kernels，只需在配置文件中增加如下配置，即可自动使能所有默认可用的 kernels：
+
+```yaml
+...
+kernel_config:
+    name: auto
+    include_kernels: auto # choice: null/true/false/auto/kernel_id1,kernel_id2,kernel_id3, default is null
+
+...
+```
+
+### 2. 调用 API 使能
+
+#### 2.1 apply_default_kernels 使能所有默认 kernels
+
+`apply_default_kernels` API 能够自动应用当前设备上所有默认注册的 kernels：
+
+```python
+from transformers import AutoModelForCausalLM
+from llamafactory.v1.plugins.model_plugins.kernels import apply_default_kernels
+
+# 加载模型
+model = AutoModelForCausalLM.from_pretrained("qwen/qwen2.5-0.5B")
+
+# 自动应用所有默认 kernels
+model = apply_default_kernels(model, include_kernels="auto")
+```
+
+#### 2.2 apply_kernel 使能特定 kernel
+
+如果需要更精细的控制，例如在某些场合单独应用某个 kernel，可以手动调用 `apply_kernel` 函数并传入 kernel ID：
+
+```python
+from transformers import AutoModelForCausalLM
+from llamafactory.v1.plugins.model_plugins.kernels import apply_kernel
+
+# 加载模型
+model = AutoModelForCausalLM.from_pretrained("qwen/qwen2.5-0.5B")
+
+# 手动应用各个 kernels
+# 注意：kernel ID 必须与定义时的 _kernel_id 一致
+model = apply_kernel("npu_fused_rope", model=model)
+model = apply_kernel("npu_fused_rmsnorm", model=model)
+model = apply_kernel("npu_fused_swiglu", model=model)
+model = apply_kernel("npu_fused_moe", model=model)
+
+### 3. 查询已注册的可用 kernels
+
+可以通过 `get_default_kernels` 获取当前环境中所有已注册且可用的默认 kernel ID：
+
+```python
+from llamafactory.v1.plugins.model_plugins.kernels import get_default_kernels
+
+# 获取默认 kernel 列表
+available_kernels = get_default_kernels()
+print(f"Available kernels: {available_kernels}")
+# 输出示例: ['npu_fused_rmsnorm', 'npu_fused_swiglu', 'npu_fused_rope', 'npu_fused_moe']
+```
+
+### 当前已实现的 kernels
+
+| Kernel ID | 功能 | 支持的设备 | 备注 |
+|-----------|------|-----------|------|
+| [npu_fused_rmsnorm](./fused-operators.md/#npufusedrmsnorm) | RMSNorm 融合算子 | NPU | NPU 设备的高性能 RMSNorm 实现 |
+| [npu_fused_swiglu](./fused-operators.md/#npufusedswiglu) | SwiGLU 融合算子 | NPU | NPU 设备的高性能 SwiGLU 实现 |
+| [npu_fused_rope](./fused-operators.md/#npufusedrope) | RoPE 融合算子 | NPU | NPU 设备的高性能 RoPE 实现 |
+| [npu_fused_moe](./fused-operators.md/#npufusedmoe) | MoE 融合算子 | NPU | MoE 融合算子，适配 Qwen3-MoE 等模型 |
+
+我们会持续适配更多的 kernels，如果您需要自己开发新的 kernels，请参考我们的 [Kernel 开发文档](../../dev-guide/plugins/model-plugins/kernels.md)，欢迎您向 LLaMA-Factory 贡献代码。
--- a/docs/zh/advanced/custom-kernels/fused-operators.md
+++ b/docs/zh/advanced/custom-kernels/fused-operators.md
@@ -0,0 +1,104 @@
+# Fused Operators
+
+LLaMA-Factory 提供了一系列针对特定硬件优化的融合算子。这些算子位于 `src/llamafactory/v1/plugins/model_plugins/kernels/ops` 目录下。
+
+系统启动时，`scan_all_kernels` 函数会自动扫描该目录，注册所有可用的算子。您可以通过 `apply_default_kernels(model, include_kernels="auto")` 一键启用它们，或者使用 `apply_kernel` 单独启用。
+
+以下是当前支持的融合算子详情：
+
+## NpuFusedRMSNorm
+RMSNorm（Root Mean Square Layer Normalization）是一种常用于大模型的归一化方法。在推理或训练中，RMSNorm 融合算子 将bias、residual等操作进行融合，可以减少显存访问次数，加速计算。
+
+Ascend npu 通过 `torch_npu.npu_rms_norm` 接口提供 RMSNorm 融合算子调用接口，支持 float16, bfloat16, float 等数据格式。RMSNorm 算子常见于Qwen等LLM模型中，由于torch侧没有提供 RMSNorm 算子的接口，因此在模型中通常是以自定义类的形式出现，通过替换 RMSNorm 类的 `forward` 方法即可使能。
+
+```python
+def _npu_rms_forward(self, hidden_states):
+    """NPU forward implementation for RMSNorm.
+
+    Args:
+        self: RMSNorm module instance with `weight` and `variance_epsilon`.
+        hidden_states: Input hidden states tensor, same shape as the baseline.
+
+    Returns:
+        Normalized tensor consistent with the baseline RMSNorm behavior.
+    """
+
+    return torch_npu.npu_rms_norm(hidden_states, self.weight, epsilon=self.variance_epsilon)[0]
+```
+
+ 在 LlamaFactory 中，通过 `NpuRMSNormKernel` 提供使能该融合算子的入口，只需要调用 `apply_kernel("npu_fused_rmsnorm", model=model)` 即可针对已适配的模型使能 npu RMSNorm 融合算子。
+
+## NpuFusedSwiGlu
+SwiGLU（Swish-Gated Linear Unit）是一种结合了Swish激活函数和门控线性单元（GLU）的混合激活函数，其主要功能是对输入张量进行门控线性变换，近年来被广泛应用于 LLM 模型中的 MLP 层。SwiGLU 融合算子将分割、激活、矩阵乘等多个操作融合为单一硬件指令，避免多次内核启动开销。
+
+Ascend npu 通过 `torch_npu.npu_swiglu` 接口提供 SwiGLU 融合算子调用接口，支持 float16，bfloat16，float SwiGLU 算子常见于Qwen等LLM模型中，由于torch侧没有提供 SwiGLU 算子的接口，因此在模型中通常是以自定义类的形式出现，通过替换 SwiGLU 类的 `forward` 方法即可使能。替换过程可参考如下示例：
+
+```python
+# 原始 MLP forward 方法：
+def forward(self, x):
+    down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+    return down_proj
+
+# 替换后的 forward 方法：
+def _npu_swiglu_forward(self, hidden_state):
+    return self.down_proj(
+        torch_npu.npu_swiglu(torch.cat((self.gate_proj(hidden_state), self.up_proj(hidden_state)), dim=-1), dim=-1)
+    )
+```
+
+ 在 LLaMA-Factory 中，通过 `NpuSwiGluKernel` 提供使能该融合算子的入口，只需要调用 `apply_kernel("npu_fused_swiglu", model=model)` 即可针对已适配的模型使能 npu SwiGLU 融合算子。对于未适配的模型，如有需要，您可根据示例以及[开发者文档](../../dev-guide/plugins/model-plugins/kernels.md)自行适配。
+
+
+## NpuFusedRoPE
+RoPE（Rotary Positional Embedding，旋转式位置嵌入） 是一种位置编码技术，广泛应用于 Qwen 等 LLM 模型中，用于有效编码文本序列的位置信息。它结合了绝对位置编码的稳定性与相对位置编码的灵活性，同时具备优秀的长度泛化能力。传统 RoPE 算子通常在 LLM 等模型结构中通过自定义函数的形式实现。RoPE 融合算子将原计算流程合并为单个硬件优化算子，从而提升性能。
+
+Ascend npu 通过 `torch_npu.npu_rotary_mul` 提供 RoPE 融合算子调用接口，支持 float16，bfloat16，float32 等数据格式。以 Qwen3 系列模型为例，通过替换其 `apply_rotary_pos_emb` 函数即可实现 RoPE融合算子使能：
+
+```python
+# 原始 apply_rotary_pos_emb：
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+# 替换 RoPE 融合算子后：
+def _apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = torch_npu.npu_rotary_mul(q, cos, sin)
+    k_embed = torch_npu.npu_rotary_mul(k, cos, sin)
+    return q_embed, k_embed
+```
+
+ 在 LLaMA-Factory 中，通过 `NpuRoPEKernel` 提供使能该融合算子的入口，只需要调用 `apply_kernel("npu_fused_rope", model=model)` 即可针对已适配的模型使能 npu RoPE 融合算子。对于未适配的模型，如有需要，您可根据示例以及[开发者文档](../../dev-guide/plugins/model-plugins/kernels.md)自行适配。
+
+
+## NpuFusedMoE
+MoE（Mixture of Experts）模型通过稀疏激活扩展容量。在原生 Transformers 实现中，使用串行循环遍历专家，导致内核启动开销大、硬件利用率低。
+
+**MoE 融合算子** 利用 **GMM（Grouped Matrix Multiplication，分组矩阵乘）** 技术，支持在单个硬件指令内并行处理多组不同形状（行数不一）的矩阵乘法，消减循环开销，同时无需额外的显存复制，显著提升训练性能。
+
+Ascend npu 通过 `torch_npu.npu_grouped_matmul` 等接口提供底层支持，通过替换模型中的 MoE Block forward 方法，即可利用 NPU 的分组矩阵乘能力。
+
+核心逻辑替换如下（简化示意）：
+
+```python
+def _npu_moe_forward(self, hidden_states, routing_weights, router_indices):
+    # 1. 排序：将乱序的 Token 按指派的专家归类，并生成索引映射
+    permuted_states, row_map = torch_npu.npu_moe_token_permute(hidden_states, router_indices)
+
+    # 2. 统计：计算每个专家需要处理的 Token 数量
+    tokens_per_expert = torch.histc(router_indices, bins=self.num_experts, min=0, max=self.num_experts)
+
+    # 3. 计算 (GMM)：一次性并行计算所有专家的权重，自动适配不同专家的输入长度
+    inter_states = torch_npu.npu_grouped_matmul(permuted_states, self.gate_up_proj_weights, split_sizes=tokens_per_expert, ...)
+    inter_states = torch_npu.npu_swiglu(inter_states)
+    output = torch_npu.npu_grouped_matmul(inter_states, self.down_proj_weights, split_sizes=tokens_per_expert, ...)
+
+    # 4. 还原：将结果恢复成原始 Token 顺序并应用路由权重
+    return torch_npu.npu_moe_token_unpermute(output, row_map, routing_weights)
+```
+
+在 LLaMA-Factory 中，通过 `NpuFusedMoEKernel` 提供使能该融合算子的入口。只需要调用 `apply_kernel("npu_fused_moe", model=model)` 即可针对已适配的模型使能 NPU MoE 融合算子。对于未适配的模型，您也可以参考上述示例代码以及[开发者文档](../../dev-guide/plugins/model-plugins/kernels.md)自行适配。
--- a/docs/zh/advanced/custom-kernels/triton.md
+++ b/docs/zh/advanced/custom-kernels/triton.md
@@ -0,0 +1 @@
+# Triton
--- a/docs/zh/advanced/distributed/deepspeed.md
+++ b/docs/zh/advanced/distributed/deepspeed.md
@@ -0,0 +1 @@
+# DeepSpeed
--- a/docs/zh/advanced/distributed/fsdp.md
+++ b/docs/zh/advanced/distributed/fsdp.md
@@ -0,0 +1 @@
+# FSDP
--- a/docs/zh/advanced/distributed/parallel-dp-tp-ep-sp-cp.md
+++ b/docs/zh/advanced/distributed/parallel-dp-tp-ep-sp-cp.md
@@ -0,0 +1 @@
+# Parallel(DP, TP, EP, SP, CP)
--- a/docs/zh/advanced/lora-and-quantization/lora.md
+++ b/docs/zh/advanced/lora-and-quantization/lora.md
@@ -0,0 +1,3 @@
+# Lora
+
+参数管理（二级参数形式）
--- a/docs/zh/advanced/lora-and-quantization/quantization.md
+++ b/docs/zh/advanced/lora-and-quantization/quantization.md
@@ -0,0 +1 @@
+# Quantization
--- a/docs/zh/conf.py
+++ b/docs/zh/conf.py
@@ -0,0 +1,20 @@
+import os
+import sys
+
+# Add parent dir to path to allow importing conf.py
+sys.path.insert(0, os.path.abspath('..'))
+
+from conf import *
+
+# Language settings
+language = 'zh_CN'
+html_search_language = 'zh'
+
+# Static files
+# Point to the root _static directory
+html_static_path = ['../_static']
+
+# Add custom JS for language switcher
+html_js_files = [
+    'js/switcher.js',
+]
--- a/docs/zh/data-preparation/data-processing.md
+++ b/docs/zh/data-preparation/data-processing.md
@@ -0,0 +1,479 @@
+# LLaMA-Factory v1 数据预处理
+
+## 总览
+
+LLaMA-Factory `v1` 采用了全新的数据处理架构，主要包含以下核心组件：
+
+- **DataEngine**：数据引擎，负责数据集的加载、索引和转换等各种插件的接入和调用，并提供数据访问接口
+- **DataConverterPlugin**：数据转换器，将非标准格式转换为统一的标准格式
+- **DataLoaderPlugin**：数据加载插件，支持多种文件格式的加载
+- **DataIndexPlugin**：数据索引插件，支持数据集的采样和权重调整
+- **DataSelectorPlugin**：数据选择插件，支持灵活的数据访问方式
+
+与 LLaMA-Factory `v0` 版本相比，`v1` 版本采用了统一的数据格式（Messages Format），所有数据都会被转换为标准的对话消息列表；此外，`v1` 版本通过 DataEngine 与 Plugin 机制，提供了自定义数据处理流的接口，具有更好的可扩展性和一致性。
+
+---
+
+## 目录
+
+- [基本用法](#基本用法)
+- [标准数据格式](#标准数据格式)
+- [数据集配置文件](#数据集配置文件)
+- [完整示例](#完整示例)
+
+---
+
+## 基本用法
+
+### 在训练配置文件，可以通过如下方式配置数据集：
+
+<details open>
+<summary>方式 1：使用 HF Hub Repo ID</summary>
+
+直接指定 HF Hub 上的数据集 Repo ID，DataEngine 会自动从 HF Hub 下载并加载数据集。
+
+注：使用 Repo ID 直接加载的数据集需要为标准格式
+
+**训练配置文件示例：**
+
+```yaml
+# example_sft.yaml
+
+...
+
+dataset: llamafactory/v1-sft-demo  # HF Hub Repo ID
+
+...
+```
+
+</details>
+
+<details>
+<summary>方式 2：使用 HF Hub 上的 YAML 配置文件</summary>
+
+`dataset`字段指定 HF Hub 上的 `dataset_info.yaml` 的 URI，DataEngine 会自动下载该配置文件并根据其中的配置加载数据集。
+
+**训练配置文件示例：**
+
+```yaml
+# example_sft.yaml
+
+...
+
+dataset: llamafactory/v1-sft-demo/dataset_info.yaml  # 远程 dataset_info.yaml 路径
+
+...
+
+```
+
+</details>
+
+<details>
+<summary>方式 3：使用本地 HF 数据集文件路径</summary>
+
+`dataset`字段指定本地的数据集文件路径（`.json`、`.jsonl` 等）
+
+注：直接指定数据集文件路径，要求该数据文件的格式已为标准格式
+
+**训练配置文件示例：**
+
+```yaml
+# example_sft.yaml
+
+...
+
+dataset: ~/data/v1_sft_demo.jsonl   # 本地数据集文件绝对路径
+
+...
+```
+
+</details>
+
+<details>
+<summary>方式 4：使用本地 YAML 配置文件路径</summary>
+
+`dataset`字段指定本地的 `dataset_info.yaml` 配置文件路径，DataEngine 会根据该配置加载其中的数据集。
+
+**训练配置文件示例：**
+
+```yaml
+# example_sft.yaml
+
+...
+
+dataset: ~/data/dataset_info.yaml    # 本地 dataset_info.yaml 文件路径
+
+...
+```
+
+</details>
+
+
+---
+
+
+
+## 标准数据格式
+
+v1 使用统一的 **Messages 格式**作为标准数据格式。每个样本都是一个包含 `messages` 字段的 JSON 对象。
+
+针对alpaca、sharegpt、以及dpo等格式的数据，可以通过内置的`DataConverterPlugin`插件，自动将其转化为标准格式，对于其他自定义格式的数据，用户也可通过自定义`DataConverterPlugin`来实现数据格式标准化，这部分内容参见[`DataConverterPlugin`](../dev-guide/plugins/data-plugins.md/#data-converter-plugin)
+
+### 1. SFT（监督微调）样本格式
+
+
+```json
+{
+  "messages": [
+    {
+      "role": "system",
+      "content": [{"type": "text", "value": "You are a helpful assistant."}],
+      "loss_weight": 0.0
+    },
+    {
+      "role": "user",
+      "content": [{"type": "text", "value": "Hello, who are you?"}],
+      "loss_weight": 0.0
+    },
+    {
+      "role": "assistant",
+      "content": [{"type": "text", "value": "I am an AI assistant."}],
+      "loss_weight": 1.0
+    }
+  ]
+}
+```
+
+#### 字段说明：
+
+- **messages**: 消息列表，包含一轮或多轮对话
+  - **role**: 消息角色，可选值：
+    - `"system"`: 系统提示
+    - `"user"`: 用户输入
+    - `"assistant"`: 模型回复
+  - **content**: 内容列表，每个元素包含：
+    - **type**: 内容类型，可选值：
+      - `"text"`: 文本内容
+      - `"image_url"`: 图像 URL（多模态）
+      - `"audio_url"`: 音频 URL（多模态）
+      - `"video_url"`: 视频 URL（多模态）
+      - `"tools"`: 工具描述
+      - `"tool_calls"`: 工具调用
+      - `"reasoning"`: 推理过程
+    - **value**: 具体内容（字符串）
+  - **loss_weight**: 损失权重（浮点数）
+    - `0.0`: 不计算损失（用于提示词部分）
+    - `1.0`: 完全计算损失（用于回复部分）
+    - 可设置为其他值以调整不同部分的学习权重
+
+- **_dataset_name** (可选): 数据集名称，由 DataEngine 自动添加
+- **extra_info** (可选): 额外信息字段
+
+### 2. DPO（偏好对齐）样本格式
+
+```json
+{
+  "chosen_messages": [
+    {
+      "role": "user",
+      "content": [{"type": "text", "value": "用户提问"}],
+      "loss_weight": 0.0
+    },
+    {
+      "role": "assistant",
+      "content": [{"type": "text", "value": "更优的回答"}],
+      "loss_weight": 1.0
+    }
+  ],
+  "rejected_messages": [
+    {
+      "role": "user",
+      "content": [{"type": "text", "value": "用户提问"}],
+      "loss_weight": 0.0
+    },
+    {
+      "role": "assistant",
+      "content": [{"type": "text", "value": "较差的回答"}],
+      "loss_weight": 1.0
+    }
+  ]
+}
+```
+
+### 3. 多模态支持
+
+对于多模态数据，可以在 `content` 列表中添加非文本类型的内容：
+
+```json
+{
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {"type": "text", "value": "这张图片里有什么？"},
+        {"type": "image_url", "value": "path/to/image.jpg"}
+      ],
+      "loss_weight": 0.0
+    },
+    {
+      "role": "assistant",
+      "content": [{"type": "text", "value": "图片中有一只猫。"}],
+      "loss_weight": 1.0
+    }
+  ]
+}
+```
+
+**说明**：`image_url`、`audio_url`、`video_url` 的路径可以是相对路径或绝对路径，具体加载方式由 `DataLoaderPlugin` 决定。
+
+---
+
+## 数据集配置文件
+
+### 1. dataset_info.yaml 配置文件格式
+
+`dataset_info.yaml` 支持同时配置多个数据集，支持分别从 HF Hub 和本地获取数据集，数据集默认会混合并打乱顺序。
+
+**示例配置文件：`data/dataset_info.yaml`**
+
+```yaml
+# 数据集 1：使用本地文件 + Alpaca 转换器
+identity:
+  file_name: ~/data/identity.json            #本地数据集文件绝对路径
+  converter: alpaca                           # 使用 alpaca 转换器
+
+# 数据集 2：指定自定义数据集目录
+alpaca_en_demo:
+  file_name:  ~/data/alpaca_en_demo.json     # 数据集文件名
+  converter: alpaca                           # 转换器插件
+  size: 500                                   # 只使用 500 个样本
+  weight: 0.5                                 # 数据集权重，用于控制该数据集的采样频率
+  split: train                                # 数据集划分，默认为 train
+  streaming: false                            # 是否流式加载，默认为 false
+
+# 数据集 3：从 Hugging Face Hub 加载
+hf_dataset:
+  hf_hub_url: llamafactory/v1-sft-demo  # HF repo ID
+  streaming: false  
+
+# 数据集 4：已经是标准格式，无需转换器
+standard:
+  file_name: ~/data/v1_sft_demo.jsonl   # 本地标准数据集文件路径
+
+# 数据集 5：自定义数据集和 converter 插件
+custom_dataset:
+  file_name: custom_data.json
+  converter: custom_converter
+  weight: 1.0  
+```
+
+### 2. 配置字段说明
+
+#### 数据源配置（二者必选其一）：
+
+- **hf_hub_url** (str): Hugging Face Hub 数据集仓库 ID
+  - 示例：`"llamafactory/v1-sft-demo"`
+  - 如果指定，则从 HF Hub 加载数据集
+
+- **file_name** (str): 本地文件路径
+  - 支持格式：`.json`、`.jsonl`、`.csv`、`.parquet`、`.arrow`、`.txt`
+
+#### 可选配置：
+
+- **split** (str): 数据集划分，默认为 `"train"`
+- **converter** (str): 数据转换器名称
+  - 可选值：`"alpaca"`（更多转换器持续添加中，也可在 data_plugin 中添加自定义 converter）
+  - 如果不指定，则假定数据已是标准格式
+- **size** (int): 使用的样本数量，默认使用全部
+- **weight** (float): 数据集权重，用于混合数据集时的采样频率，默认为 1.0
+- **streaming** (bool): 是否流式加载，默认为 `False`
+
+---
+
+
+## 完整示例
+
+### 1. 基础使用示例
+
+```python
+from llamafactory.v1.config.data_args import DataArguments
+from llamafactory.v1.core.data_engine import DataEngine
+
+# 使用本地 YAML 配置
+data_args = DataArguments(
+    dataset="~/data/v1_sft_demo.jsonl",
+    cutoff_len=2048
+)
+
+# 初始化 DataEngine
+engine = DataEngine(data_args=data_args)
+
+# 查看数据集信息
+print(f"数据集总样本数: {len(engine)}")
+print(f"数据集列表: {list(engine.datasets.keys())}")
+
+# 访问数据样本
+sample = engine[0]
+print(f"样本格式: {sample.keys()}")
+print(f"消息列表: {sample['messages']}")
+
+# 批量访问
+batch = engine[0:10]
+print(f"批量样本数: {len(batch)}")
+```
+
+### 2. 输出示例
+
+**查看数据集信息输出：**
+
+```
+数据集总样本数: 500
+数据集列表: ['default']
+样本格式: dict_keys(['_dataset_name', 'messages'])
+消息列表: [{'role': 'user', 'content': [{'type': 'text', 'value': 'hi'}], 'loss_weight': 0.0}, {'role': 'assistant', 'content': [{'type': 'text', 'value': 'Hello! I am {{name}}, an AI assistant developed by {{author}}. How can I assist you today?'}], 'loss_weight': 1.0}]
+批量样本数: 10
+```
+
+**访问单个样本输出：**
+
+```python
+{
+  '_dataset_name': 'alpaca_en_demo',
+  'messages': [
+    {
+      'role': 'user',
+      'content': [{'type': 'text', 'value': 'What is the capital of France?'}],
+      'loss_weight': 0.0
+    },
+    {
+      'role': 'assistant',
+      'content': [{'type': 'text', 'value': 'The capital of France is Paris.'}],
+      'loss_weight': 1.0
+    }
+  ]
+}
+```
+
+### 3. 混合多数据集配置文件示例
+
+**配置文件：`data/mixed_datasets.yaml`**
+
+```yaml
+dataset_1:
+  file_name: alpaca_en_demo.json
+  converter: alpaca
+  weight: 1.0
+
+dataset_2:
+  file_name: identity.json
+  converter: alpaca
+  weight: 2.0
+
+dataset_3:
+  hf_hub_url: llamafactory/v1-sft-demo
+  weight: 1.5
+```
+
+
+### 4. 多模态数据示例
+
+**数据文件：`data/multimodal_demo.jsonl`**
+
+标准化后数据示例：
+
+```json
+[
+  {
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "value": "Who are they?"},
+          {"type": "image_url", "value": "mllm_demo_data/1.jpg"}
+        ],
+        "loss_weight": 0.0
+      },
+      {
+        "role": "assistant",
+        "content": [
+          {"type": "text", "value": "They're Kane and Gretzka from Bayern Munich."}
+        ],
+        "loss_weight": 1.0
+      },
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "value": "What are they doing?"},
+          {"type": "image_url", "value": "mllm_demo_data/1.jpg"}
+        ],
+        "loss_weight": 0.0
+      },
+      {
+        "role": "assistant",
+        "content": [
+          {"type": "text", "value": "They are celebrating on the soccer field."}
+        ],
+        "loss_weight": 1.0
+      }
+    ]
+  },
+  {
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "value": "Who is he?"},
+          {"type": "image_url", "value": "mllm_demo_data/2.jpg"}
+        ],
+        "loss_weight": 0.0
+      },
+      {
+        "role": "assistant",
+        "content": [
+          {"type": "text", "value": "He's Thomas Muller from Bayern Munich."}
+        ],
+        "loss_weight": 1.0
+      },
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "value": "Why is he on the ground?"}
+        ],
+        "loss_weight": 0.0
+      },
+      {
+        "role": "assistant",
+        "content": [
+          {"type": "text", "value": "Because he's sliding on his knees to celebrate."}
+        ],
+        "loss_weight": 1.0
+      }
+    ]
+  }
+]
+```
+
+```python
+from llamafactory.v1.config.data_args import DataArguments
+from llamafactory.v1.core.data_engine import DataEngine
+
+data_args = DataArguments(dataset="data/multimodal_demo.jsonl")
+engine = DataEngine(data_args=data_args)
+
+# 访问多模态样本
+sample = engine[0]
+print("用户消息内容：")
+for content_item in sample['messages'][0]['content']:
+    print(f"  类型: {content_item['type']}, 值: {content_item['value']}")
+```
+
+---
+
+**注意事项**：
+
+1. 所有数据最终都会转换为标准的 Messages 格式
+2. 通过 `converter` 插件可以支持多种数据格式
+3. 通过 `weight` 和 `size` 参数可以灵活控制数据分布
+4. 支持同时使用本地数据集和 HuggingFace Hub 数据集
+5. 多模态数据通过在 `content` 中添加不同类型的元素来支持
+6. 更多细节信息请参考我们的 [API REFERENCE](../dev-guide/core/data-engine.md/#data-engine)
--- a/docs/zh/dev-guide/core/data-engine.md
+++ b/docs/zh/dev-guide/core/data-engine.md
@@ -0,0 +1,253 @@
+# DataEngine
+
+## 1. DataEngine 简介
+
+
+`DataEngine` 是 LLaMA-Factory v1 数据处理的核心类，继承自 PyTorch 的 `Dataset`，负责各种插件的接入，其他功能（如数据格式转换、数据加载等）均通过插件的形式实现并接入 `DataEngine`。
+
+`DataEngine`接受一个唯一入参：`DataArguments` 实例，所有的元数据集信息均通过该参数配置传入。
+
+## 2. DataEngine 与 DataArguments 接口定义
+
+```python
+@dataclass
+class DataArguments:
+    """ `DataEngine`初始化入参
+
+    args:
+        dataset (str): 数据集路径，远程数据集 repo id / dataset_info.yaml 路径，或本地数据集路径/dataset_info.yaml路径
+        cutoff_len (int): 数据集截止长度，即数据集最大样本采样数量
+    """
+    ...
+
+
+class DataEngine(Dataset):
+    """数据引擎（DataEngine）
+
+    `DataEngine` 负责数据集的加载与统一管理，支持：
+        - 从本地路径或 Hugging Face Hub 加载数据
+        - 通过插件机制加载自定义数据
+        - 构建统一的数据索引
+        - 支持流式（streaming）与非流式数据访问
+
+    attr:
+        args (DataArguments): 数据参数配置
+        datasets (dict[str, HFDataset]): 数据集名称到数据对象的映射
+        dataset_infos (dict[str, DatasetInfo]): 数据集名称到元信息的映射
+        data_index (list[tuple[str, int]]): 数据索引列表，每项为 (dataset_name, sample_index)
+        streaming (bool): 是否为流式数据集
+    """
+
+    def __init__(self, data_args: DataArguments) -> None:
+        """初始化 `DataEngine`
+
+        初始化时自动执行以下步骤：
+            1. 调用 `get_dataset_info`， 从 `data_args` 读取并解析数据集元信息
+            2. 调用 `load_dataset`，根据配置加载数据集
+            3. 调用 `build_data_index`，构建统一的索引列表
+
+        args:
+            data_args (DataArguments): 数据参数配置对象
+        """
+        ...
+
+    def get_dataset_info(self) -> None:
+        """从配置文件或远程仓库加载数据集元信息
+
+        根据 `self.args.dataset` 确定数据源，数据源支持如下选项：
+            - 本地 YAML 配置文件路径
+            - Hugging Face Hub 上的 YAML 配置文件路径
+            - 本地数据集文件路径
+            - Hugging Face Hub 数据集 repo id
+
+        """
+        ...
+
+    def load_dataset(self) -> None:
+        """根据数据集元信息加载所有数据集
+
+        每个数据集条目可以包含以下字段：
+            - `hf_hub_url`: 使用 `datasets.load_dataset` 加载
+            - 本地数据文件：通过 `DataLoaderPlugin` 插件加载
+            - `streaming`: 是否启用流式模式
+
+        更新:
+            self.datasets (dict): 数据集名称到已加载数据对象的映射
+            self.streaming (bool): 如果任一数据集为流式模式，则设置为 True
+        """
+        ...
+
+    def build_data_index(self) -> None:
+        """构建统一的数据索引
+
+        为所有数据集创建全局索引列表 `(dataset_name, sample_index)`
+
+        当启用流式模式时，生成固定长度（例如 1000）的占位索引；
+        否则，为每条样本建立索引。
+
+        插件 `DataIndexPlugin` 可根据数据集大小或权重调整索引分布
+        """
+        ...
+
+    def _convert_data_sample(self, raw_sample: dict[str, Any], dataset_name: str) -> Sample:
+        """将原始样本转换为统一格式
+
+        根据 `dataset_info` 中的 `converter` 字段，调用对应的转换插件，
+        将原始样本标准化为统一的数据结构。
+
+        args:
+            raw_sample (dict[str, Any]): 原始数据样本
+            dataset_name (str): 样本所属的数据集名称
+
+        return:
+            Sample: 转换后的标准化格式样本
+        """
+        ...
+
+    def __len__(self) -> int:
+        """返回数据集的总样本数
+
+        return:
+            int: 数据集长度
+                如果为流式数据集，返回 `-1`
+        """
+        ...
+
+    def __getitem__(self, index: Union[int, Any]) -> Union[Sample, list[Sample]]:
+        """根据索引或选择器获取样本
+
+        args:
+            index (Union[int, Any]): 数据索引，int 或 list[int]
+
+        return:
+            Union[Sample, list[Sample]]: 单个样本或样本列表
+        """
+        ...
+
+    def __iter__(self) -> Iterable:
+        """返回数据集迭代器
+
+        用于非流式数据集的顺序或随机访问
+        流式模式下需要实现异步加载逻辑
+
+        return:
+            Iterable: 数据集迭代器。
+        """
+        ...
+
+    async def __aiter__(self) -> AsyncIterable:
+        """返回异步数据集迭代器
+
+        用于流式数据集或异步数据加载场景
+        允许在异步环境中以流的方式读取样本
+
+        return:
+            AsyncIterable: 异步迭代器，按顺序产出样本
+        """
+        ...
+
+
+```
+
+`DataArguments`  参数说明：
+
+`dataset`: 数据集路径，支持本地或远程，当传入本地数据集文件路径时，需要满足该数据集为标准格式；否则需要传入 `dataset_info.yaml` 来配置数据集的 `converter` 等元信息，以告知 `DataEngine` 应当如何处理该数据。
+
+`cutoff_len`: 数据集的截止长度，即该数据集的最大样本数量。
+
+---
+
+## 3. DataEngine 核心方法
+
+### 3.1 `get_dataset_info`：加载数据元信息
+
+根据 `dataset` 参数加载数据集配置，获取数据位置、数据格式、插件配置等所有数据元信息，在实例化 `DataEngine` 时会自动调用此方法。
+
+### 3.2 加载数据集：`load_dataset`
+
+遍历所有数据源，根据不同的数据源加载数据，在实例化 `DataEngine` 时会自动调用此方法。
+
+```python
+for key, value in self.dataset_infos.items():
+    split = value.get("split", "train")
+    streaming = value.get("streaming", False)
+
+    if "hf_hub_url" in value:
+        # 从 HF Hub 加载
+        dataset = load_dataset(value["hf_hub_url"], split=split, streaming=streaming)
+    else:
+        # 使用 DataLoaderPlugin 加载本地文件
+        dataset = DataLoaderPlugin(args=self.args).auto_load_data(value)
+
+    self.datasets[key] = dataset
+```
+
+### 3.3 `build_data_index`：构建数据索引
+
+为每个数据集创建索引列表 `[(dataset_name, sample_index), ...]`, `DataIndexPlugin`插件在此处被调用，可控制各数据集的采样频率、采样方式等，在实例化`DataEngine`时会自动调用此方法。
+
+```python
+for dataset_name, dataset in self.datasets.items():
+    # 创建基础索引
+    data_index = [(dataset_name, idx) for idx in range(len(dataset))]
+
+    # 根据 size 和 weight 调整索引
+    size = self.dataset_infos[dataset_name].get("size")
+    weight = self.dataset_infos[dataset_name].get("weight")
+    if size or weight:
+        data_index = DataIndexPlugin().adjust_data_index(data_index, size, weight)
+
+    self.data_index.extend(data_index)
+```
+
+### 3.4 `_convert_data_sample`：数据格式标准化
+
+将原始数据转换为标准格式，`DataConverterPlugin`插件在此处被调用，具体调用的插件由 `get_dataset_info` 方法获取的 `converter` 信息指定，若 `converter` 为空则假定数据集为标准格式，此方法由`DataEngine`的 `__getitem__` 方法调用。
+
+```python
+def _convert_data_sample(self, raw_sample: dict, dataset_name: str) -> Sample:
+    converter = self.dataset_infos[dataset_name].get("converter")
+    if converter is not None:
+        # 使用指定的转换器
+        from ..plugins.data_plugins.converter import get_converter
+        return {"_dataset_name": dataset_name, **get_converter(converter)(raw_sample)}
+    else:
+        # 已经是标准格式
+        return {"_dataset_name": dataset_name, **raw_sample}
+```
+
+---
+
+## 4. 初始化
+
+`DataEngine` 初始化过程只需传入一个构建好的 `DataArguments` 即可，后续可通过该 `DataEngine` 访问数据集中的数据。
+
+```python
+from llamafactory.v1.config.data_args import DataArguments
+from llamafactory.v1.core.data_engine import DataEngine
+
+# 1. 创建数据参数
+data_args = DataArguments(
+    dataset="~/data/v1_sft_demo.jsonl",
+    cutoff_len=2048
+)
+
+# 2. 初始化 Data Engine
+data_engine = DataEngine(data_args=data_args)
+
+# 3. 访问数据
+sample = data_engine[0]  # 获取第一个样本
+```
+
+## 5. 数据访问方式
+
+实例化后的`DataEngine`支持整数索引、列表索引、以及切片等访问方式，其数据读取用法可等价于 Python 列表。
+
+```python
+sample = data_engine[0]  # 获取第一个样本
+
+sample = data_engine[0:10]  # 获取前 10 个样本
+
+sample = data_engine[[0, 5, 10]]  # 获取指定索引的样本
+
+```
--- a/docs/zh/dev-guide/core/model-engine.md
+++ b/docs/zh/dev-guide/core/model-engine.md
@@ -0,0 +1 @@
+# ModelEngine
--- a/docs/zh/dev-guide/core/trainer.md
+++ b/docs/zh/dev-guide/core/trainer.md
@@ -0,0 +1 @@
+# Trainer
--- a/docs/zh/dev-guide/plugins/data-plugins.md
+++ b/docs/zh/dev-guide/plugins/data-plugins.md
@@ -0,0 +1,467 @@
+# Data Plugins
+
+## 1. Data Plugins 简介
+
+## DataConverterPlugin
+
+### 1. DataConverterPlugin 简介
+
+DataConverter 负责将非标准格式的数据集转换为 v1 的标准 Messages 格式。这使得用户可以继续使用现有的数据集（如 Alpaca 格式），而无需手动转换。针对自定义格式的数据集，用户也可以通过构建对应的自定义 DataConverter 插件，来负责其数据格式标准化。
+
+当前，LLaMA-Factory 已内置了 `Alpaca Converter` 和 `Pair Converter`，这两类数据集可以直接使用对应的 converter 进行标准化，无需自定义转换器。
+
+
+### 2. Alpaca Converter 详解
+
+#### 2.1 Alpaca 格式
+
+Alpaca 格式是一种常见的指令微调数据格式：
+
+```json
+{
+  "system": "You are a helpful assistant.",
+  "instruction": "Describe a process of making crepes.",
+  "input": "",
+  "output": "Making crepes is an easy and delicious process..."
+}
+```
+
+#### 2.2 Alpaca Converter 接口定义
+
+```python
+class AlpacaSample(TypedDict, total=False):
+    """Alpaca 格式数据样本结构
+
+    attr:
+        system (str, 可选): 系统提示信息（system prompt），用于设定对话背景或模型行为。
+        instruction (str, 可选): 用户指令（user instruction），通常为任务描述。
+        input (str, 可选): 额外的输入内容（input text），可与 instruction 拼接。
+        output (str, 可选): 模型生成的目标输出（expected response）。
+    """
+    ...
+
+
+def alpaca_converter(raw_sample: AlpacaSample) -> SFTSample:
+    """将 Alpaca 样本转换为 SFT（Supervised Fine-Tuning）标准样本格式
+
+    `alpaca_converter` 将 Alpaca 数据集中一条样本转换为通用的 `SFTSample` 格式
+    该格式用于监督微调（SFT）或多轮对话建模
+
+    转换逻辑:
+        - 若存在 `system` 字段，则生成一条系统消息，loss_weight = 0.0
+        - 若存在 `instruction` 或 `input` 字段，则合并为一条用户消息，loss_weight = 0.0
+        - 若存在 `output` 字段，则生成一条助手机器人回复消息，loss_weight = 1.0
+
+    args:
+        raw_sample (AlpacaSample): 原始 Alpaca 数据样本
+
+    return:
+        SFTSample: 转换后的标准化样本，格式如下:
+
+            {
+                "messages": [
+                    {"role": "system", "content": [{"type": "text", "value": "..."}], "loss_weight": 0.0},
+                    {"role": "user", "content": [{"type": "text", "value": "..."}], "loss_weight": 0.0},
+                    {"role": "assistant", "content": [{"type": "text", "value": "..."}], "loss_weight": 1.0},
+                ]
+            }
+
+    example:
+        >>> raw = {"instruction": "请将以下句子翻译成英文：", "input": "你好", "output": "Hello"}
+        >>> alpaca_converter(raw)
+        {
+            "messages": [
+                {"role": "user", "content": [{"type": "text", "value": "请将以下句子翻译成英文：你好"}], "loss_weight": 0.0},
+                {"role": "assistant", "content": [{"type": "text", "value": "Hello"}], "loss_weight": 1.0}
+            ]
+        }
+    """
+
+```
+
+#### 2.3 转换过程
+
+`alpaca_converter` 函数将 Alpaca 格式转换为标准格式，转换逻辑如下：
+
+```python
+def alpaca_converter(raw_sample: AlpacaSample) -> SFTSample:
+    messages = []
+
+    # 1. 添加系统提示词（如果存在）
+    if "system" in raw_sample:
+        messages.append({
+            "role": "system",
+            "content": [{"type": "text", "value": raw_sample["system"]}],
+            "loss_weight": 0.0
+        })
+
+    # 2. 添加用户输入（instruction + input）
+    if "instruction" in raw_sample or "input" in raw_sample:
+        user_content = raw_sample.get("instruction", "") + raw_sample.get("input", "")
+        messages.append({
+            "role": "user",
+            "content": [{"type": "text", "value": user_content}],
+            "loss_weight": 0.0
+        })
+
+    # 3. 添加模型回复
+    if "output" in raw_sample:
+        messages.append({
+            "role": "assistant",
+            "content": [{"type": "text", "value": raw_sample["output"]}],
+            "loss_weight": 1.0
+        })
+
+    return {"messages": messages}
+```
+
+#### 2.4 转换示例
+
+**输入（Alpaca 格式）：**
+
+```json
+{
+  "instruction": "What is the capital of France?",
+  "input": "",
+  "output": "The capital of France is Paris."
+}
+```
+
+**输出（标准格式）：**
+
+```json
+{
+  "messages": [
+    {
+      "role": "user",
+      "content": [{"type": "text", "value": "What is the capital of France?"}],
+      "loss_weight": 0.0
+    },
+    {
+      "role": "assistant",
+      "content": [{"type": "text", "value": "The capital of France is Paris."}],
+      "loss_weight": 1.0
+    }
+  ]
+}
+```
+
+### 3. 自定义转换器
+
+#### 3.1 创建自定义转换器
+
+如果用户有自己的数据格式，可以轻松添加自定义转换器将其标准化，实现过程可参考如下示例：
+
+```python
+# src/llamafactory/v1/plugins/data_plugins/converter.py
+
+from typing import TypedDict, NotRequired
+from ...extras.types import SFTSample
+
+# 1. 定义输入格式的类型
+class MyCustomSample(TypedDict, total=False):
+    question: str
+    answer: str
+    context: NotRequired[str]
+
+# 2. 实现转换逻辑
+def custom_converter(raw_sample: MyCustomSample) -> SFTSample:
+    messages = []
+
+    # 构建用户消息
+    user_text = raw_sample["question"]
+    if "context" in raw_sample:
+        user_text = f"Context: {raw_sample['context']}\n\nQuestion: {user_text}"
+
+    messages.append({
+        "role": "user",
+        "content": [{"type": "text", "value": user_text}],
+        "loss_weight": 0.0
+    })
+
+    # 构建助手消息
+    messages.append({
+        "role": "assistant",
+        "content": [{"type": "text", "value": raw_sample["answer"]}],
+        "loss_weight": 1.0
+    })
+
+    return {"messages": messages}
+
+# 3. 注册 custom_converter
+#src/llamafactory/v1/plugins/data_plugins/converter.py: CONVERTERS
+CONVERTERS = {
+    "alpaca": alpaca_converter,
+    "custom": custom_converter,  # 添加自定义转换器
+}
+```
+
+#### 3.2 使用自定义转换器
+
+在 YAML 配置中指定转换器名称：
+
+```yaml
+my_dataset:
+  file_name: custom_data.json
+  converter: custom
+```
+
+---
+
+## DataLoaderPlugin
+
+### 1. DataLoaderPlugin 简介
+
+`DataLoaderPlugin` 负责从本地文件加载数据集，当前支持如下文件格式：
+
+- **JSON**: `.json`
+- **JSONL**: `.jsonl`
+- **CSV**: `.csv`
+- **Parquet**: `.parquet`
+- **Arrow**: `.arrow`
+- **Text**: `.txt`
+
+### 2. DataLoaderPlugin 接口定义
+
+```python
+@dataclass
+class DataLoaderPlugin:
+    """数据加载插件（DataLoaderPlugin）
+
+    负责根据数据集信息（`DatasetInfo`）自动加载本地或远程数据集。  
+    支持多种文件格式（如 CSV、JSON、Parquet、Text、Arrow），并可选择是否以流式方式加载。
+
+    通常由 `DataEngine` 调用，用于统一封装数据加载逻辑。
+    """
+
+    args: DataArguments
+    """数据参数对象，包含数据目录、缓存路径、分片等配置信息。"""
+
+
+    def _get_builder_name(self, path: str) -> Literal["arrow", "csv", "json", "parquet", "text"]:
+        """获取数据集文件格式
+
+        根据输入文件路径自动判断应使用的 HuggingFace `load_dataset` 构建器类型。
+        通过文件扩展名推断数据类型，例如 `.csv`、`.jsonl`、`.parquet`、`.txt` 等。
+
+        args:
+            path (str): 数据集文件路径，用于识别文件类型。
+
+        return:
+            Literal["arrow", "csv", "json", "parquet", "text"]:
+                数据构建器名称，用于 `datasets.load_dataset()`。
+
+        example:
+            >>> _get_builder_name("data/train.jsonl")
+            "json"
+        """
+        ...
+
+
+    def auto_load_data(self, dataset_info: DatasetInfo) -> HFDataset:
+        """根据传入的 `dataset_info` 自动选择合适的加载方式
+
+        args:
+            dataset_info (DatasetInfo): 数据集元信息，通常包含：
+                - `file_name`: 数据文件路径
+                - `split`: 数据划分（如 "train"、"test"）；
+                - `streaming`: 是否启用流式加载
+
+        return:
+            HFDataset: 加载完成的 Hugging Face 数据集对象。
+
+        example:
+            >>> plugin = DataLoaderPlugin(args)
+            >>> ds = plugin.auto_load_data({"file_name": "~/data.json", "split": "train"})
+        """
+        ...
+
+
+    def load_data_from_file(self, filepath: str, split: str, streaming: bool) -> HFDataset:
+        """从文件或目录加载数据集
+
+        根据输入路径自动识别文件类型（CSV、JSON、Parquet、Text 等），  
+        并通过 `datasets.load_dataset()` 加载数据集。  
+        若 `streaming=True`，则将结果转换为迭代式数据集。
+
+        args:
+            filepath (str): 文件路径或目录路径。
+            split (str): 数据划分名称（如 "train"、"validation"）。
+            streaming (bool): 是否启用流式加载模式。
+
+        return:
+            HFDataset: 加载后的数据集对象。
+
+        example:
+            >>> plugin.load_data_from_file("data/train.json", "train", False)
+        """
+        ...
+
+```
+
+---
+
+## DataIndexPlugin
+
+### 1. DataIndexPlugin 简介
+
+`DataIndexPlugin` 负责调整数据索引，支持通过配置 `size`, `weight` 等参数控制数据集样本数量和采样频率。
+
+- 使用 `size` 参数 限制使用的样本数量：
+
+```yaml
+my_dataset:
+  file_name: large_dataset.json
+  size: 1000  # 只使用前 1000 个样本
+```
+
+- 使用 `weight` 参数调整数据集在混合数据中的采样频率：
+
+```yaml
+dataset_a:
+  file_name: data_a.json
+  weight: 1.0
+
+dataset_b:
+  file_name: data_b.json
+  weight: 2.0  # dataset_b 的样本出现频率是 dataset_a 的 2 倍
+```
+
+**说明**：`weight` 参数适用于在多个数据集混合训练时，调整不同数据集的的采样频率
+
+- 当 `weight=1.0` 时，数据集按原始比例采样
+- 当 `weight=2.0` 时，该数据集的索引会复制 2 倍，使其样本出现频率翻倍
+
+### 2. DataIndexPlugin 接口定义
+
+```python
+@dataclass
+class DataIndexPlugin:
+    """数据索引插件（DataIndexPlugin）
+
+    根据 `size` 和 `weight` 调整数据索引列表，控制数据集的样本数量和采样频率  
+    通常在多数据集混合训练时使用，以控制不同数据集在总体样本中的占比。
+
+    在 `DataEngine.build_data_index` 中被自动调用，用于实现样本重采样或加权分布。
+    """
+
+    def adjust_data_index(
+        self, data_index: list[tuple[str, int]], size: Optional[int], weight: Optional[float]
+    ) -> list[tuple[str, int]]:
+        """调整数据索引列表
+
+        根据 `size` 或 `weight` 参数对输入的数据索引进行采样、扩展或缩减。  
+        若两个参数同时存在，将依次执行基于大小和基于权重的调整。
+
+        args:
+            data_index (list[tuple[str, int]]):  
+                数据索引列表，每个元素为 `(dataset_name, sample_index)`。  
+            size (Optional[int]):  
+                目标样本数量，若指定则根据该数量裁剪或重复样本。  
+            weight (Optional[float]):  
+                数据集权重，用于控制数据集在混合训练中的采样比例。
+
+        return:
+            list[tuple[str, int]]:  
+                调整后的数据索引列表。
+
+        example:
+            >>> plugin = DataIndexPlugin()
+            >>> adjusted = plugin.adjust_data_index([("ds1", i) for i in range(100)], size=50, weight=None)
+            >>> len(adjusted)
+            50
+        """
+        ...
+
+
+    def adjust_by_size(self, data_index: list[tuple[str, int]], size: int) -> list[tuple[str, int]]:
+        """根据目标大小调整数据索引
+
+        通过裁剪或重复样本，使索引总数等于 `size`。  
+        常用于统一不同数据集的样本数量。
+
+        args:
+            data_index (list[tuple[str, int]]):  
+                原始数据索引列表。  
+            size (int):  
+                目标样本数量。
+
+        return:
+            list[tuple[str, int]]:  
+                调整后长度等于 `size` 的数据索引列表。
+
+        example:
+            >>> plugin.adjust_by_size([("ds1", i) for i in range(10)], 20)
+        """
+        ...
+
+
+    def adjust_by_weight(self, data_index: list[tuple[str, int]], weight: float) -> list[tuple[str, int]]:
+        """根据权重调整数据索引
+
+        通过加权采样或重复样本，使数据集样本出现频率符合指定权重。  
+        常用于多数据源训练中按比例平衡样本。
+
+        args:
+            data_index (list[tuple[str, int]]):  
+                原始数据索引列表。  
+            weight (float):  
+                数据集权重（相对比例，可与其他数据集共同归一化）。
+
+        return:
+            list[tuple[str, int]]:  
+                调整后的加权数据索引列表。
+
+        example:
+            >>> plugin.adjust_by_weight([("ds1", i) for i in range(10)], 0.5)
+        """
+        ...
+
+```
+---
+
+## DataSelectorPlugin
+
+### 1. DataSelectorPlugin 简介
+
+`DataSelectorPlugin` 为 `DataEngine`提供基于索引访问数据的功能，由 `DataEngine` 的 `__getitem__` 方法自动调用。
+
+
+### 2. DataSelectorPlugin 接口定义
+
+```python
+@dataclass
+class DataSelectorPlugin:
+    """根据索引选择数据集样本。
+
+    配合 `DataEngine` 使用，通过统一的 `data_index` 结构（包含数据集名与样本索引）来实现灵活的数据选择
+
+    """
+
+    data_index: list[tuple[str, int]]
+    """数据索引列表，每个元素为 (dataset_name, sample_index)。"""
+
+
+    def select(self, index: Union[slice, list[int], Any]) -> Union[tuple[str, int], list[tuple[str, int]]]:
+        """选择数据集样本
+
+        根据输入类型从 `data_index` 中选择对应的样本索引  
+        支持三种索引方式：
+            - 切片（slice）：返回对应范围内的样本
+            - 索引列表（list[int]）：返回指定索引处的多个样本
+            - 其他类型输入将触发异常。
+
+        args:
+            index (Union[slice, list[int], Any]): 数据样本索引
+                可以是切片（`slice`）或索引列表
+
+        return:
+            Union[tuple[str, int], list[tuple[str, int]]]:
+                - 若为单个索引：返回一个 `(dataset_name, sample_index)`
+                - 若为多个索引或切片：返回多个样本的列表
+
+        except:
+        Raises:
+            ValueError: 当输入索引类型不受支持时抛出。
+        ...
+```
--- a/docs/zh/dev-guide/plugins/model-plugins/initialization.md
+++ b/docs/zh/dev-guide/plugins/model-plugins/initialization.md
--- a/docs/zh/dev-guide/plugins/model-plugins/kernels.md
+++ b/docs/zh/dev-guide/plugins/model-plugins/kernels.md
@@ -0,0 +1,197 @@
+# Kernels plugins
+
+## 概览
+LLaMA-Factory 通过 Kernels plugins 系统，依据不同硬件设备提供高性能计算内核（kernel）实现。该系统通过注册表机制管理所有 kernel，通过 `@register_kernel` 装饰器实现 kernel 定义后自动注册，由 `apply_kernel` 方法来使能指定的 kernel，`apply_default_kernels` 可使能注册表中当前环境所有可用的默认 kernels。
+
+## 架构设计
+
+### 核心组件
+
+#### 1. Registry（注册表）
+
+`Registry` 是一个用于管理所有 kernel 实现的静态类。它维护一个字典结构：`{kernel_id: KernelClass}`。
+
+```python
+# 注册表结构示例
+{
+    "npu_fused_rmsnorm": NpuRMSNormKernel,
+    "npu_fused_swiglu": NpuSwiGluKernel,
+    ...
+}
+```
+
+#### 2. register_kernel (装饰器)
+
+`@register_kernel` 是 `Registry.register` 的别名。所有 kernel 类均应使用该装饰器进行注册。
+
+**注册机制**：
+- 装饰器检查类是否继承自 `BaseKernel`。
+- 检查类是否定义了 `_kernel_id` 和 `_device` 属性。
+- 检查 `_device` 是否与当前运行环境的加速器类型匹配。如果不匹配，则跳过注册。
+- 如果一切符合要求，将 kernel 类注册到全局注册表中。
+
+#### 3. BaseKernel（基类）
+
+所有 kernel 的实现都必须继承自 `BaseKernel` 抽象基类。`BaseKernel` 定义了 kernel 的基本属性和接口。
+
+#### 4. 标识系统
+
+**Kernel ID** (`_kernel_id`)：
+每个 kernel 必须拥有一个唯一的字符串标识符，例如 `"npu_fused_rmsnorm"`。
+
+**Device Type** (`_device`)：
+kernel 必须声明其支持的设备类型，例如 `DeviceType.NPU` 或 `DeviceType.CUDA`。
+
+## Kernel 系统 API 设计
+
+### **Registry**：全局 kernel 注册表
+
+`Registry` 类提供了注册和获取 kernel 的接口：
+
+```python
+class Registry:
+    @classmethod
+    def register(cls, kernel_cls: type[BaseKernel]) -> type[BaseKernel] | None:
+        """注册一个 kernel 类"""
+        ...
+
+    @classmethod
+    def get(cls, kernel_id: str) -> type[BaseKernel] | None:
+        """根据 ID 获取 kernel 类"""
+        ...
+```
+
+### **BaseKernel**
+
+`BaseKernel` 定义了所有 kernel 必须实现的协议：
+
+- `_kernel_id`: 类属性，kernel 的唯一标识符。
+- `_device`: 类属性，kernel 支持的设备类型。
+- `check_deps()`: 类方法，检查 kernel 的依赖项是否满足（如 `torch_npu` 是否安装）。
+- `apply(**kwargs)`: 抽象类方法，实现 kernel 的具体应用逻辑。
+
+```python
+class BaseKernel(ABC):
+    _kernel_id: Any = ""
+    _device: DeviceType = DeviceType.CPU
+
+    @classmethod
+    def check_deps(cls) -> bool:
+        """检查依赖项"""
+        ...
+
+    @classmethod
+    @abstractmethod
+    def apply(cls, **kwargs) -> HFModel:
+        """应用 kernel 到模型"""
+        ...
+```
+
+### **scan_all_kernels**
+
+`scan_all_kernels` 函数会自动扫描 `ops` 目录下的所有 `.py` 文件并导入它们，从而触发 `@register_kernel` 装饰器完成自动注册。
+
+### **apply_kernel**
+
+对模型使能指定的 kernel。
+
+```python
+def apply_kernel(kernel_id: str, **kwargs) -> HFModel:
+    """应用指定的 kernel 到模型
+
+    Args:
+        kernel_id: 目标 kernel 的 ID
+        **kwargs: 传递给 kernel.apply 的参数，通常包含 model
+    """
+```
+
+**用法示例**：
+```python
+from llamafactory.v1.plugins.model_plugins.kernels import apply_kernel
+
+model = apply_kernel("npu_fused_rmsnorm", model=model)
+```
+
+### **apply_default_kernels**
+
+对模型使能所有默认注册的 kernel。这是一个高级 API，通常在模型加载流程中自动调用。
+
+```python
+def apply_default_kernels(model: HFModel, include_kernels: str = None) -> HFModel:
+    """应用所有默认 kernel
+
+    Args:
+        model: HFModel 实例
+        include_kernels: 包含的 kernel ID 列表（逗号分隔字符串），或者 "auto"/True 表示全部
+    """
+```
+
+## 扩展 Kernels
+
+如果用户有针对特定模型或者设备的 kernel，可以按照下述步骤去实现并接入 LLaMA-Factory。
+
+### 创建新 Kernel 的步骤
+
+#### 1. 创建 Kernel 实现文件
+
+在 `src/llamafactory/v1/plugins/model_plugins/kernels/ops` 下的相应子目录中创建新的 kernel 实现文件，例如 `mlp/cuda_swiglu.py`：
+
+```python
+import torch
+from ......accelerator.helper import DeviceType
+from ......utils.types import HFModel
+from ...base import BaseKernel
+from ...registry import register_kernel
+
+# 实现具体的 kernel 函数
+def _cuda_swiglu_forward(self, hidden_state):
+    # ... CUDA 优化实现 ...
+    pass
+
+@register_kernel
+class CudaSwiGluKernel(BaseKernel):
+    _kernel_id = "cuda_fused_swiglu"
+    _device = DeviceType.CUDA
+
+    @classmethod
+    def apply(cls, **kwargs) -> HFModel:
+        model = kwargs.get("model")
+        if model is None:
+            raise ValueError("model is required")
+
+        if not cls.check_deps():
+            raise RuntimeError("Dependencies not met")
+
+        # 遍历模型并替换 forward 方法
+        for name, module in model.named_modules():
+            # ... 匹配和替换逻辑 ...
+            pass
+
+        return model
+```
+
+#### 2. 自动发现
+
+由于 `scan_all_kernels` 会自动扫描 `ops` 目录，只要文件位于该目录下且没有语法错误，系统启动时会自动导入并注册，无需手动修改注册表代码。
+
+#### 3. 测试 Kernel
+
+创建测试用例验证 kernel 的正确性：
+
+```python
+from llamafactory.v1.plugins.model_plugins.kernels import apply_kernel
+
+# ... 加载模型 ...
+model = apply_kernel("cuda_fused_swiglu", model=model)
+# ... 验证 forward 是否被替换 ...
+```
+
+## 异常处理
+
+### 依赖不可用
+
+`BaseKernel.check_deps()` 默认会检查当前设备类型是否匹配。子类可以重写此方法以添加额外的依赖检查（如检查特定的库是否安装）。如果 `check_deps()` 返回 `False`，`apply()` 方法应当抛出异常或进行相应处理。
+
+### Kernel ID 未找到
+
+如果调用 `apply_kernel` 时传入了不存在的 `kernel_id`，会抛出 `ValueError`。
--- a/docs/zh/dev-guide/plugins/model-plugins/rendering.md
+++ b/docs/zh/dev-guide/plugins/model-plugins/rendering.md
--- a/docs/zh/getting-started.md
+++ b/docs/zh/getting-started.md
@@ -0,0 +1,71 @@
+# Getting Started
+
+
+## 训练方法
+
+|          方法          |     全参数训练      |    部分参数训练     |       LoRA         |       QLoRA        |
+|:---------------------:| ------------------ | ------------------ | ------------------ | ------------------ |
+|      指令监督微调       | :white_check_mark: |  |  | |
+|      奖励模型训练       |  |  |  | |
+|        DPO 训练        |  |  |  | |
+
+
+
+
+## 软件依赖
+
+|          必需项          | 至少     | 推荐     |
+|:---------------------:|--------|--------|
+|        python         | 3.11   | 3.12   |
+|         torch         | 2.7.1  | 2.7.1  |
+| torch-npu(Ascend NPU) | 2.7.1  | 2.7.1  |
+|      torchvision      | 0.22.1 | 0.22.1 |
+|     transformers      | 5.0.0  | 5.0.0  |
+|       datasets        | 3.2.0  | 4.0.0  |
+|         peft          | 0.18.1 | 0.18.1 |
+
+
+|       可选项        | 至少     | 推荐     |
+|:----------------:|--------|--------|
+| CUDA(NVIDIA GPU) | 11.6   | 12.2   |
+|    deepspeed     | 0.18.4 | 0.18.4 |
+|   flash-attn(NVIDIA GPU)   | 2.5.6  | 2.7.2  |
+
+
+## 如何使用
+
+### 安装 LLaMA Factory
+
+> [!IMPORTANT]
+> 此步骤为必需。
+
+#### 从源码安装
+
+```bash
+git clone --depth 1 https://github.com/hiyouga/LlamaFactory.git
+cd LlamaFactory
+pip install -e .
+```
+
+
+### 数据准备
+
+关于数据集文件的格式，请参考 [data-preparation/README.md](data-preparation/README.md) 的内容。你可以使用 HuggingFace / ModelScope 上的数据集或加载本地数据集。
+
+> [!NOTE]
+> 使用自定义数据集或自定义数据集格式时，请参照 [data-preparation/README.md](data-preparation/README.md) 进行配置，如有必要，请重新实现自定义数据集的数据处理逻辑，包括对应的`converter`。
+
+您也可以使用 **[Easy Dataset](https://github.com/ConardLi/easy-dataset)**、**[DataFlow](https://github.com/OpenDCAI/DataFlow)** 和 **[GraphGen](https://github.com/open-sciencelab/GraphGen)** 构建用于微调的合成数据。
+
+### 快速开始
+
+下面的命令展示了对 Qwen3-0.6B 模型使用 FSDP2 进行 全参**微调**，两行命令等价。
+
+```bash
+export USE_V1=1
+llamafactory-cli sft examples/v1/train_full/train_full_fsdp2.yaml
+llamafactory-cli train examples/v1/train_full/train_full_fsdp2.yaml
+
+```
+
+高级用法请参考 [advanced](./advanced/README.md)（包括多卡多机微调、分布式、Lora、量化、以及各种加速特性等）。
--- a/docs/zh/hyperparameters/data-argument.md
+++ b/docs/zh/hyperparameters/data-argument.md
@@ -0,0 +1 @@
+# Data Argument
--- a/docs/zh/hyperparameters/model-argument.md
+++ b/docs/zh/hyperparameters/model-argument.md
--- a/docs/zh/hyperparameters/sample-argument.md
+++ b/docs/zh/hyperparameters/sample-argument.md
--- a/docs/zh/hyperparameters/training-argument.md
+++ b/docs/zh/hyperparameters/training-argument.md
--- a/docs/zh/index.rst
+++ b/docs/zh/index.rst
@@ -0,0 +1,62 @@
+LlamaFactory 文档
+=================
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Getting Started
+
+   getting-started
+   installation
+   llamaboard-web-ui
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Data Preparation
+
+   data-preparation/data-processing
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Training
+
+   training/sft
+   training/dpo
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Inference
+
+   inference/deploy
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Advanced
+
+   advanced/lora-and-quantization/lora
+   advanced/lora-and-quantization/quantization
+   advanced/distributed/fsdp
+   advanced/distributed/deepspeed
+   advanced/distributed/parallel-dp-tp-ep-sp-cp
+   advanced/custom-kernels/triton
+   advanced/custom-kernels/fused-operators
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Hyperparameters
+
+   hyperparameters/data-argument
+   hyperparameters/model-argument
+   hyperparameters/sample-argument
+   hyperparameters/training-argument
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Dev Guide
+
+   dev-guide/core/data-engine
+   dev-guide/core/model-engine
+   dev-guide/core/trainer
+   dev-guide/plugins/data-plugins
+   dev-guide/plugins/model-plugins/initialization
+   dev-guide/plugins/model-plugins/kernels
+   dev-guide/plugins/model-plugins/rendering
--- a/docs/zh/inference/deploy.md
+++ b/docs/zh/inference/deploy.md
@@ -0,0 +1 @@
+# Deploy
--- a/docs/zh/installation.md
+++ b/docs/zh/installation.md
@@ -0,0 +1 @@
+# Installation
--- a/docs/zh/llamaboard-web-ui.md
+++ b/docs/zh/llamaboard-web-ui.md
@@ -0,0 +1 @@
+# LlamaBoard Web UI
--- a/docs/zh/training/dpo.md
+++ b/docs/zh/training/dpo.md
@@ -0,0 +1 @@
+# DPO
--- a/docs/zh/training/sft.md
+++ b/docs/zh/training/sft.md
@@ -0,0 +1 @@
+# SFT
--- a/examples/extras/asft/llama2_full_asft.yaml
+++ b/examples/extras/asft/llama2_full_asft.yaml
@@ -0,0 +1,45 @@
+### model
+model_name_or_path: models/Llama-2-7b
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z0_config.json
+use_asft_loss: true
+asft_alpha: 0.1
+
+### dataset
+dataset: med
+template: llama2
+cutoff_len: 2048
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/llama2-7b/full/asft2
+logging_steps: 1
+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: 4
+gradient_accumulation_steps: 8
+learning_rate: 2.0e-5
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500
--- a/examples/extras/asft/qwen2_full_asft.yaml
+++ b/examples/extras/asft/qwen2_full_asft.yaml
@@ -0,0 +1,45 @@
+### model
+model_name_or_path: models/Qwen2.5-7B
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z0_config.json
+use_asft_loss: true
+asft_alpha: 0.05
+
+### dataset
+dataset: math
+template: qwen
+cutoff_len: 2048
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/qwen2-7b/full/asft
+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: 4
+gradient_accumulation_steps: 8
+learning_rate: 5.0e-5
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500
--- a/examples/v1/train_freeze/train_freeze_sft.yaml
+++ b/examples/v1/train_freeze/train_freeze_sft.yaml
@@ -0,0 +1,38 @@
+model: Qwen/Qwen3-4B
+trust_remote_code: true
+model_class: llm
+
+template: qwen3_nothink
+
+# Freeze Configuration
+peft_config:
+  name: freeze
+  freeze_trainable_layers: 2      # Train the last 2 layers
+  freeze_trainable_modules: all       # In these layers, train specific modules
+  freeze_extra_modules: null      # Extra modules to train (e.g. embed_tokens, lm_head)
+
+# Kernel Config
+kernel_config:
+  name: auto
+  include_kernels: auto
+
+# FSDP Config
+dist_config:
+  name: fsdp2
+  dcp_path: null
+
+### data
+train_dataset: data/v1_sft_demo.yaml
+
+### training
+output_dir: ./outputs/test_freeze
+micro_batch_size: 1
+global_batch_size: 4
+cutoff_len: 2048
+learning_rate: 2.0e-5
+bf16: false
+max_steps: 10
+
+### sample
+sample_backend: hf
+max_new_tokens: 128
--- a/examples/v1/train_full/train_full_deepspeed.yaml
+++ b/examples/v1/train_full/train_full_deepspeed.yaml
@@ -0,0 +1,25 @@
+model: Qwen/Qwen3-0.6B
+
+model_class: llm
+
+template: qwen3_nothink
+
+kernel_config:
+    name: auto
+    include_kernels: auto 
+
+dist_config:
+    name: deepspeed
+    config_file: examples/deepspeed/ds_z3_config.json
+
+### data
+train_dataset: data/v1_sft_demo.yaml 
+
+### training
+output_dir: outputs/Qwen3-0.6B-deepspeed
+micro_batch_size: 1
+cutoff_len: 2048
+learning_rate: 1.0e-4
+bf16: true
+max_steps: 10
+
--- a/examples/v1/train_lora/export_lora.yaml
+++ b/examples/v1/train_lora/export_lora.yaml
@@ -0,0 +1,7 @@
+model: Qwen/Qwen3-4B
+peft_config:
+  name: lora
+  adapter_name_or_path: ./outputs/test_lora
+  export_dir: ./merge_lora_model
+  export_size: 5
+  infer_dtype: auto
--- a/examples/v1/train_lora/train_lora_sft.yaml
+++ b/examples/v1/train_lora/train_lora_sft.yaml
@@ -0,0 +1,39 @@
+model: Qwen/Qwen3-4B
+trust_remote_code: true
+model_class: llm
+
+template: qwen3_nothink
+
+# PEFT Configuration
+peft_config:
+  name: lora
+  r: 16
+  lora_alpha: 32
+  lora_dropout: 0.05
+  target_modules: all
+
+# Kernel Config
+kernel_config:
+  name: auto
+  include_kernels: auto
+
+# FSDP Config
+dist_config:
+  name: fsdp2
+  dcp_path: null
+
+### data
+train_dataset: data/v1_sft_demo.yaml
+
+### training
+output_dir: ./outputs/test_lora
+micro_batch_size: 1
+global_batch_size: 4
+cutoff_len: 2048
+learning_rate: 1.0e-4
+bf16: true
+max_steps: 10
+
+### sample
+sample_backend: hf
+max_new_tokens: 128
--- a/examples/v1/train_qlora/quantization.yaml
+++ b/examples/v1/train_qlora/quantization.yaml
@@ -0,0 +1,43 @@
+model: Qwen/Qwen3-0.6B
+trust_remote_code: true
+model_class: llm
+
+template: qwen3_nothink
+
+# PEFT Configuration
+peft_config:
+  name: lora
+  r: 16
+  lora_alpha: 32
+  lora_dropout: 0.05
+  target_modules: all
+
+# Kernel Config
+kernel_config:
+  name: auto
+  include_kernels: auto
+
+# FSDP Config
+dist_config:
+  name: fsdp2
+  dcp_path: null
+
+# Quantization Config
+quant_config:
+  name: bnb # choice: auto/bnb if auto is selected, the quantization method will be automatically selected based on the model and environment.
+  quantization_bit: 4 # choice: 8/4(bnb)
+
+### data
+train_dataset: data/v1_sft_demo.yaml
+
+### training
+output_dir: outputs/test_quantization
+micro_batch_size: 1
+cutoff_len: 2048
+learning_rate: 1.0e-4
+bf16: false
+max_steps: 10
+
+### sample
+sample_backend: hf
+max_new_tokens: 128
--- a/requirements/liger-kernel.txt
+++ b/requirements/liger-kernel.txt
@@ -1 +1 @@
-liger-kernel>=0.5.5
+liger-kernel>=0.6.3
--- a/src/llamafactory/data/collator.py
+++ b/src/llamafactory/data/collator.py
@@ -213,6 +213,7 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
            and getattr(self.model.config, "model_type", None)
            in [
                "glm4v",
+                "glm_ocr",
                "Keye",
                "qwen2_vl",
                "qwen2_5_vl",
--- a/src/llamafactory/data/template.py
+++ b/src/llamafactory/data/template.py
@@ -459,6 +459,18 @@ class ReasoningTemplate(Template):
        return [(encoded_messages[i], encoded_messages[i + 1]) for i in range(0, len(encoded_messages), 2)]


+@dataclass
+class Glm47ReasoningTemplate(ReasoningTemplate):
+    r"""GLM-4.7 uses only the closing </think> tag for empty thinking blocks."""
+
+    @override
+    def add_thought(self, content: str = "") -> str:
+        if not content:
+            return self.thought_words[1]
+
+        return self.thought_words[0] + content + self.thought_words[1]
+
+
 TEMPLATES: dict[str, "Template"] = {}


@@ -1049,6 +1061,39 @@ register_template(
 )


+# copied from glm4 template
+register_template(
+    name="glm_ocr",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>"]),
+    format_assistant=StringFormatter(slots=["\n{{content}}"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}"]),
+    format_function=FunctionFormatter(slots=["{{content}}"], tool_format="glm4"),
+    format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>"]),
+    format_tools=ToolFormatter(tool_format="glm4"),
+    format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
+    stop_words=["<|user|>", "<|observation|>"],
+    efficient_eos=True,
+    mm_plugin=get_mm_plugin(name="glm4v", image_token="<|image|>", video_token="<|video|>"),
+)
+
+
+# copied from glm4_moe template
+register_template(
+    name="glm4_7",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>"]),
+    format_assistant=StringFormatter(slots=["\n{{content}}"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}"]),
+    format_function=FunctionFormatter(slots=["{{content}}"], tool_format="glm4_moe"),
+    format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>"]),
+    format_tools=ToolFormatter(tool_format="glm4_moe"),
+    format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
+    stop_words=["<|user|>", "<|observation|>"],
+    thought_words=("<think>", "</think>"),
+    efficient_eos=True,
+    template_class=Glm47ReasoningTemplate,
+)
+
+
 # copied from glm4 template
 register_template(
    name="glmz1",
--- a/src/llamafactory/extras/constants.py
+++ b/src/llamafactory/extras/constants.py
@@ -65,6 +65,7 @@ MCA_SUPPORTED_MODELS = {
    "qwen2_vl",
    "qwen2_5_vl",
    "qwen3_vl",
+    "qwen3_vl_moe",
    "qwen3",
    "qwen3_moe",
    "qwen3_next",
@@ -939,6 +940,29 @@ register_model_group(
 )


+register_model_group(
+    models={
+        "GLM-4.7-Flash": {
+            DownloadSource.DEFAULT: "zai-org/GLM-4.7-Flash",
+            DownloadSource.MODELSCOPE: "ZhipuAI/GLM-4.7-Flash",
+        },
+    },
+    template="glm4_7",
+)
+
+
+register_model_group(
+    models={
+        "GLM-OCR": {
+            DownloadSource.DEFAULT: "zai-org/GLM-OCR",
+            DownloadSource.MODELSCOPE: "ZhipuAI/GLM-OCR",
+        },
+    },
+    template="glm_ocr",
+    multimodal=True,
+)
+
+
 register_model_group(
    models={
        "GLM-Z1-0414-9B-Chat": {
--- a/src/llamafactory/hparams/finetuning_args.py
+++ b/src/llamafactory/hparams/finetuning_args.py
@@ -490,6 +490,14 @@ class FinetuningArguments(
        default=False,
        metadata={"help": "Whether to use the DFT loss."},
    )
+    use_asft_loss: bool = field(
+        default=False,
+        metadata={"help": "Whether to use the ASFT loss."},
+    )
+    asft_alpha: float = field(
+        default=0.1,
+        metadata={"help": "The alpha parameter for ASFT loss to control the power of adaptive weight."},
+    )
    use_eaft_loss: bool = field(
        default=False,
        metadata={"help": "Whether to use the EAFT loss."},
--- a/src/llamafactory/model/model_utils/liger_kernel.py
+++ b/src/llamafactory/model/model_utils/liger_kernel.py
@@ -77,6 +77,8 @@ def apply_liger_kernel(
        from liger_kernel.transformers import apply_liger_kernel_to_qwen3 as apply_liger_kernel
    elif model_type == "qwen3_moe":
        from liger_kernel.transformers import apply_liger_kernel_to_qwen3_moe as apply_liger_kernel
+    elif model_type == "qwen3_next":
+        from liger_kernel.transformers import apply_liger_kernel_to_qwen3_next as apply_liger_kernel
    elif model_type == "gpt_oss":
        try:
            from liger_kernel.transformers import apply_liger_kernel_to_gpt_oss as apply_liger_kernel
--- a/src/llamafactory/model/model_utils/moe.py
+++ b/src/llamafactory/model/model_utils/moe.py
@@ -77,6 +77,11 @@ def add_z3_leaf_module(model: "PreTrainedModel") -> None:

        _set_z3_leaf_modules(model, [Glm4MoeMoE])

+    if model_type == "glm4_moe_lite":
+        from transformers.models.glm4_moe_lite.modeling_glm4_moe_lite import Glm4MoeLiteMoE
+
+        _set_z3_leaf_modules(model, [Glm4MoeLiteMoE])
+
    if model_type == "glm4v_moe":
        from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeTextMoE

@@ -137,6 +142,10 @@ def add_z3_leaf_module(model: "PreTrainedModel") -> None:

        _set_z3_leaf_modules(model, [Qwen3OmniMoeThinkerTextSparseMoeBlock])

+    if model_type == "qwen3_next":
+        from transformers.models.qwen3_next.modeling_qwen3_next import Qwen3NextSparseMoeBlock
+
+        _set_z3_leaf_modules(model, [Qwen3NextSparseMoeBlock])

 def configure_moe(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
    if not is_trainable or not model_args.moe_aux_loss_coef:
--- a/src/llamafactory/model/model_utils/visual.py
+++ b/src/llamafactory/model/model_utils/visual.py
@@ -239,6 +239,15 @@ _register_composite_model(
 )


+_register_composite_model(
+    model_type="glm_ocr",
+    projector_key="visual.merger",
+    vision_model_keys=["visual.patch_embed", "visual.blocks"],
+    language_model_keys=["language_model", "lm_head"],
+    lora_conflict_keys=["patch_embed"],
+)
+
+
 _register_composite_model(
    model_type="internvl",
 )
--- a/src/llamafactory/train/mca/workflow.py
+++ b/src/llamafactory/train/mca/workflow.py
@@ -82,9 +82,34 @@ def _check_model_support(model_args: "ModelArguments"):
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
    )
    if config.model_type not in MCA_SUPPORTED_MODELS:
-        raise ValueError(f"Model {config.model_type} is not supported by MCA.")
+        raise ValueError(
+            f"Model {config.model_type} is not supported by mcore_adapter."
+            "You can try to upgrade mcore_adapter to the latest version for more supported models."
+        )


+def _freeze_model_parameters(model: Any, finetuning_args: "FinetuningArguments"):
+    """Freeze model parameters for qwen_vl series models based on finetuning arguments."""
+    if getattr(model.config, "hf_model_type", None) not in ["qwen2_vl", "qwen2_5_vl", "qwen3_vl", "qwen3_vl_moe"]:
+        return
+
+    params_to_freeze = []
+    if finetuning_args.freeze_vision_tower:
+        params_to_freeze.extend(["vision_model.blocks", "vision_model.patch_embed"])
+        if getattr(model.config, "hf_model_type", None) in ["qwen3_vl", "qwen3_vl_moe"]:
+            params_to_freeze.extend(["vision_model.pos_embed"])
+
+    if finetuning_args.freeze_multi_modal_projector:
+        params_to_freeze.extend(["multi_modal_projector"])
+
+    if finetuning_args.freeze_language_model:
+        params_to_freeze.extend(["embedding", "decoder", "output_layer"])
+
+    if params_to_freeze:
+        for name, p in model.named_parameters():
+            if any(name.startswith(k) for k in params_to_freeze):
+                p.requires_grad_(False)
+
 def run_pt(
    model_args: "ModelArguments",
    data_args: "DataArguments",
@@ -161,22 +186,8 @@ def run_sft(
    _check_model_support(model_args)
    model = AutoModel.from_pretrained(model_args.model_name_or_path, training_args)

-    # optional freezing for qwen2_vl, qwen2_5_vl
-    if getattr(model.config, "hf_model_type", None) in ["qwen2_vl", "qwen2_5_vl", "qwen3_vl"]:
-        params_to_freeze = []
-        if finetuning_args.freeze_vision_tower:
-            params_to_freeze.extend(["vision_model.blocks", "vision_model.patch_embed"])
-
-        if finetuning_args.freeze_multi_modal_projector:
-            params_to_freeze.extend(["multi_modal_projector"])
-
-        if finetuning_args.freeze_language_model:
-            params_to_freeze.extend(["embedding", "decoder", "output_layer"])
-
-        if params_to_freeze:
-            for name, p in model.named_parameters():
-                if any(name.startswith(k) for k in params_to_freeze):
-                    p.requires_grad_(False)
+    # optional freezing for qwen_vl series
+    _freeze_model_parameters(model, finetuning_args)

    pad_to_max = training_args.expert_model_parallel_size is not None and training_args.expert_model_parallel_size > 1
    data_collator = SFTDataCollatorWith4DAttentionMask(
@@ -229,6 +240,8 @@ def run_dpo(
    _check_model_support(model_args)
    model = AutoModel.from_pretrained(model_args.model_name_or_path, training_args)

+    _freeze_model_parameters(model, finetuning_args)
+
    if finetuning_args.use_ref_model:
        ref_config = AutoConfig.from_pretrained(model_args.model_name_or_path, training_args)
        ref_model = AutoModel.from_config(ref_config)
--- a/src/llamafactory/train/sft/trainer.py
+++ b/src/llamafactory/train/sft/trainer.py
@@ -17,6 +17,7 @@

 import json
 import os
+from functools import partial
 from types import MethodType
 from typing import TYPE_CHECKING, Any, Optional, Union

@@ -52,6 +53,7 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
        processor: Optional["ProcessorMixin"],
        model_args: Optional["ModelArguments"] = None,
        gen_kwargs: Optional[dict[str, Any]] = None,
+        ref_model: Optional["torch.nn.Module"] = None,
        **kwargs,
    ) -> None:
        kwargs["processing_class"] = kwargs.pop("tokenizer")
@@ -82,6 +84,27 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
            self.add_callback(BAdamCallback)

+        self.ref_model = ref_model
+
+        if ref_model is not None:
+            from trl.models.utils import prepare_deepspeed, prepare_fsdp
+
+            if getattr(self.accelerator.state, "deepspeed_plugin", None) is not None:
+                if not (
+                    getattr(ref_model, "is_loaded_in_8bit", False) or getattr(ref_model, "is_loaded_in_4bit", False)
+                ):  # quantized models are already set on the correct device
+                    self.ref_model = prepare_deepspeed(self.ref_model, self.accelerator)
+            elif getattr(self.accelerator.state, "fsdp_plugin", None) is not None:
+                if self.accelerator.is_fsdp2:
+                    from accelerate.utils.fsdp_utils import fsdp2_prepare_model
+
+                    self.ref_model = fsdp2_prepare_model(self.accelerator, self.ref_model)
+                else:
+                    self.ref_model = prepare_fsdp(self.ref_model, self.accelerator)
+            else:
+                self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
+                self.ref_model.eval()
+
        if finetuning_args.use_dft_loss:
            from ..trainer_utils import dft_loss_func

@@ -93,6 +116,13 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
            self.compute_loss_func = lambda outputs, labels, num_items_in_batch=None: eaft_loss_func(
                outputs, labels, num_items_in_batch, finetuning_args.eaft_alpha
            )
+        elif finetuning_args.use_asft_loss:
+            from ..trainer_utils import asft_loss_func
+
+            self.compute_loss_func = partial(
+                asft_loss_func,
+                asft_alpha=finetuning_args.asft_alpha,
+            )

        if training_args.fp8 and hasattr(self, "accelerator"):  # verify FP8 status after trainer initialization
            verify_fp8_status(self.accelerator, training_args)
@@ -119,7 +149,17 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):

    @override
    def compute_loss(self, model, inputs, *args, **kwargs):
-        return super().compute_loss(model, inputs, *args, **kwargs)
+        if self.finetuning_args.use_asft_loss:
+            with torch.no_grad():
+                ref_outputs = self.ref_model(
+                    input_ids=inputs["input_ids"],
+                    attention_mask=inputs.get("attention_mask", None),
+                )
+                ref_logits = ref_outputs.logits
+            outputs = model(**inputs)
+            return self.compute_loss_func(outputs, inputs["labels"], ref_logits)
+        else:
+            return super().compute_loss(model, inputs, *args, **kwargs)

    @override
    def prediction_step(
--- a/src/llamafactory/train/sft/workflow.py
+++ b/src/llamafactory/train/sft/workflow.py
@@ -24,7 +24,7 @@ from ...extras.misc import calculate_tps
 from ...extras.packages import is_transformers_version_greater_than
 from ...extras.ploting import plot_loss
 from ...model import load_model, load_tokenizer
-from ..trainer_utils import create_modelcard_and_push
+from ..trainer_utils import create_modelcard_and_push, create_ref_model
 from .metric import ComputeAccuracy, ComputeSimilarity, eval_logit_processor
 from .trainer import CustomSeq2SeqTrainer

@@ -52,6 +52,10 @@ def run_sft(
    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="sft", **tokenizer_module)
    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)

+    ref_model = None
+    if finetuning_args.use_asft_loss:
+        ref_model = create_ref_model(model_args, finetuning_args)
+
    if getattr(model, "is_quantized", False) and not training_args.do_train:
        setattr(model, "_hf_peft_config_loaded", True)  # hack here: make model compatible with prediction

@@ -124,6 +128,7 @@ def run_sft(
            data_collator=data_collator,
            callbacks=callbacks,
            gen_kwargs=gen_kwargs,
+            ref_model=ref_model,
            **dataset_module,
            **tokenizer_module,
            **metric_module,
--- a/src/llamafactory/train/trainer_utils.py
+++ b/src/llamafactory/train/trainer_utils.py
@@ -23,6 +23,7 @@ from collections.abc import Callable, Mapping
 from typing import TYPE_CHECKING, Any, Optional, Union

 import torch
+import torch.nn.functional as F
 from transformers import Trainer
 from transformers.integrations import is_deepspeed_zero3_enabled
 from transformers.modeling_utils import is_fsdp_enabled
@@ -681,6 +682,88 @@ def _dft_cross_entropy(
    return loss


+def asft_loss_func(
+    outputs,
+    labels: torch.Tensor,
+    ref_logits: torch.Tensor,
+    asft_alpha: float = 0.1,
+    ignore_index: int = -100,
+) -> torch.Tensor:
+    logits = outputs.get("logits")
+    if logits is None:
+        return outputs.get("loss", torch.tensor(0.0))
+
+    logits = logits.float()
+
+    # shift for causal LM
+    shift_logits = logits[..., :-1, :].contiguous()
+    shift_labels = labels[..., 1:].contiguous()
+    shift_ref_logits = ref_logits[..., :-1, :].contiguous()
+
+    vocab_size = shift_logits.size(-1)
+
+    # flatten
+    shift_logits = shift_logits.view(-1, vocab_size)
+    shift_ref_logits = shift_ref_logits.view(-1, vocab_size)
+    shift_labels = shift_labels.view(-1).to(shift_logits.device)
+
+    return _asft_cross_entropy(
+        policy_logits=shift_logits,
+        policy_labels=shift_labels,
+        ref_logits=shift_ref_logits,
+        asft_alpha=asft_alpha,
+        ignore_index=ignore_index,
+    )
+
+
+def _asft_cross_entropy(
+    policy_logits: torch.Tensor,
+    policy_labels: torch.Tensor,
+    ref_logits: torch.Tensor,
+    asft_alpha: float = 0.1,
+    ignore_index: int = -100,
+) -> torch.Tensor:
+    dft_loss = _dft_cross_entropy(
+        policy_logits,
+        policy_labels,
+        ignore_index=ignore_index,
+    )
+
+    kl_loss = _kl_divergence(
+        policy_logits,
+        ref_logits,
+        policy_labels,
+        ignore_index=ignore_index,
+    )
+
+    return dft_loss + asft_alpha * kl_loss
+
+
+def _kl_divergence(
+    policy_logits: torch.Tensor,
+    ref_logits: torch.Tensor,
+    labels: torch.Tensor,
+    ignore_index: int = -100,
+) -> torch.Tensor:
+    # log p(y|x)
+    log_p = F.log_softmax(policy_logits, dim=-1)
+
+    # q(y|x)
+    q = F.softmax(ref_logits, dim=-1)
+
+    # token-wise KL
+    kl = F.kl_div(
+        log_p,
+        q,
+        reduction="none",
+    ).sum(dim=-1)  # [N]
+
+    # mask padding tokens
+    mask = (labels != ignore_index).float()
+
+    return (kl * mask).sum() / mask.sum()
+
+
 def eaft_loss_func(
    outputs: "torch.Tensor",
    labels: "torch.Tensor",
--- a/src/llamafactory/train/tuner.py
+++ b/src/llamafactory/train/tuner.py
@@ -24,7 +24,7 @@ from ..data import get_template_and_fix_tokenizer
 from ..extras import logging
 from ..extras.constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from ..extras.misc import find_available_port, get_device_name, get_torch_device, infer_optim_dtype
-from ..extras.packages import is_mcore_adapter_available, is_ray_available
+from ..extras.packages import is_mcore_adapter_available, is_ray_available, is_transformers_version_greater_than
 from ..hparams import RayArguments, get_infer_args, get_ray_args, get_train_args, read_args
 from ..model import load_model, load_tokenizer
 from .callbacks import LogCallback, PissaConvertCallback, ReporterCallback
@@ -160,17 +160,28 @@ def export_model(args: Optional[dict[str, Any]] = None) -> None:
        model = model.to(output_dtype)
        logger.info_rank0(f"Convert model dtype to: {output_dtype}.")

-    model.save_pretrained(
-        save_directory=model_args.export_dir,
-        max_shard_size=f"{model_args.export_size}GB",
-        safe_serialization=(not model_args.export_legacy_format),
-    )
+    # Prepare save arguments (safe_serialization removed in transformers v5.0.0)
+    save_kwargs = {
+        "save_directory": model_args.export_dir,
+        "max_shard_size": f"{model_args.export_size}GB",
+    }
+    if not is_transformers_version_greater_than("5.0.0"):
+        save_kwargs["safe_serialization"] = not model_args.export_legacy_format
+
+    model.save_pretrained(**save_kwargs)
+
    if model_args.export_hub_model_id is not None:
+        # Prepare push arguments (safe_serialization removed in transformers v5.0.0)
+        push_kwargs = {
+            "max_shard_size": f"{model_args.export_size}GB",
+        }
+        if not is_transformers_version_greater_than("5.0.0"):
+            push_kwargs["safe_serialization"] = not model_args.export_legacy_format
+
        model.push_to_hub(
            model_args.export_hub_model_id,
            token=model_args.hf_hub_token,
-            max_shard_size=f"{model_args.export_size}GB",
-            safe_serialization=(not model_args.export_legacy_format),
+            **push_kwargs,
        )

    if finetuning_args.stage == "rm":
--- a/src/llamafactory/v1/core/base_trainer.py
+++ b/src/llamafactory/v1/core/base_trainer.py
@@ -76,19 +76,28 @@ class BaseTrainer:
        if self.args.enable_activation_checkpointing:
            self.model.gradient_checkpointing_enable({"use_reentrant": False})

-        if self.args.dist_config is not None:
-            shard_need_optimizer = self.args.dist_config.name == "deepspeed"
-        else:
-            shard_need_optimizer = False
+        self._accelerate_engine = None
+        dist_name = self.args.dist_config.name if self.args.dist_config is not None else None

-        if shard_need_optimizer:
+        if dist_name == "deepspeed":
+            from ..plugins.trainer_plugins.distributed.hub import DistributedPlugin
+
+            self._deepspeed_engine = DistributedPlugin("deepspeed")(
+                self.model,
+                self.args.dist_config,
+                num_micro_batch=self.train_batch_generator.num_micro_batch,
+                micro_batch_size=self.args.micro_batch_size,
+            )
            self._init_optimizer()
-            self._shard_model()
+            self._init_lr_scheduler()
+            self.model, self.optimizer, self.lr_scheduler = self._deepspeed_engine.prepare(
+                self.model, self.optimizer, self.lr_scheduler
+            )
        else:
+            # fsdp2 / DDP / no dist
            self._shard_model()
            self._init_optimizer()
-
-        self._init_lr_scheduler()
+            self._init_lr_scheduler()

    def _create_batch_generator(self) -> None:
        self.train_batch_generator = BatchGenerator(
@@ -171,25 +180,35 @@ class BaseTrainer:
                step_loss = 0
                step_valid_tokens = compute_valid_tokens(micro_batches)
                step_valid_tokens = DistributedInterface().all_reduce(step_valid_tokens, op=ReduceOp.SUM)
-                for micro_batch in micro_batches:
+                num_micro = len(micro_batches)
+                for i, micro_batch in enumerate(micro_batches):
                    loss = self.compute_loss(micro_batch)
                    mini_step_valid_tokens = compute_valid_tokens([micro_batch])
                    # fsdp uses mean reduction so we need to scale the loss by dp_size
                    loss = loss * mini_step_valid_tokens * self.dp_size / (step_valid_tokens + 1e-6)

-                    loss.backward()
+                    if self._deepspeed_engine is not None:
+                        # deepspeed: set sync_gradients so engine.step() only fires on last micro-batch
+                        self._deepspeed_engine.accelerator.sync_gradients = i == num_micro - 1
+                        self._deepspeed_engine.backward(loss)
+                    else:
+                        loss.backward()
                    step_loss += loss.item()

-                grad_norm = torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args.max_grad_norm).item()
-
-                # isfinite(): argument 'input' (position 1) must be Tensor, not float
-                if not torch.isfinite(torch.tensor(grad_norm)):  # type: ignore # pyright: ignore [reportUnknownReturnType]
-                    logger.warning_rank0(f"Gradient norm is not finite: {grad_norm}")
+                if self._deepspeed_engine is not None:
+                    # deepspeed: engine.step() already ran inside backward at the sync boundary
+                    grad_norm = self._deepspeed_engine.get_grad_norm()
                else:
-                    self.optimizer.step()
+                    grad_norm = torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args.max_grad_norm).item()

-                self.lr_scheduler.step()
-                self.optimizer.zero_grad()
+                    # isfinite(): argument 'input' (position 1) must be Tensor, not float
+                    if not torch.isfinite(torch.tensor(grad_norm)):  # type: ignore # pyright: ignore [reportUnknownReturnType]
+                        logger.warning_rank0(f"Gradient norm is not finite: {grad_norm}")
+                    else:
+                        self.optimizer.step()
+
+                    self.lr_scheduler.step()
+                    self.optimizer.zero_grad()

                step_loss, grad_norm = DistributedInterface().all_reduce([step_loss, grad_norm])
                DistributedInterface().sync()
@@ -203,7 +222,14 @@ class BaseTrainer:

    def save_model(self) -> None:
        """Save the model."""
-        model_to_save = self.model.module if hasattr(self.model, "module") else self.model
-        model_to_save.save_pretrained(self.args.output_dir)
-        self.renderer.processor.save_pretrained(self.args.output_dir)
-        logger.info_rank0(f"Model saved to {self.args.output_dir}")
+        if self.args.dist_config is not None and self.args.dist_config.name in ("deepspeed", "fsdp2"):
+            from ..plugins.trainer_plugins.distributed.hub import DistributedPlugin
+
+            DistributedPlugin(self.args.dist_config.name).save_model(
+                self.model, self.args.output_dir, self.renderer.processor
+            )
+        else:
+            model_to_save = self.model.module if hasattr(self.model, "module") else self.model
+            model_to_save.save_pretrained(self.args.output_dir, max_shard_size="4GB")
+            self.renderer.processor.save_pretrained(self.args.output_dir, max_shard_size="4GB")
+            logger.info_rank0(f"Model saved to {self.args.output_dir}")
--- a/src/llamafactory/v1/core/model_engine.py
+++ b/src/llamafactory/v1/core/model_engine.py
@@ -90,6 +90,26 @@ class ModelEngine:
        Transformers can choose the proper model init context.
        https://github.com/huggingface/transformers/blob/v5.0.0rc0/src/transformers/modeling_utils.py#L3538
        """
+        if self.args.init_config is not None:
+            from ..plugins.model_plugins.initialization import InitPlugin
+
+            init_device = InitPlugin(self.args.init_config.name)()
+        else:
+            init_device = DistributedInterface().current_device
+
+        init_kwargs = {"device_map": init_device}
+
+        if self.args.quant_config is not None:
+            from ..plugins.model_plugins.quantization import QuantizationPlugin
+
+            init_kwargs = QuantizationPlugin(self.args.quant_config.name)(
+                init_kwargs=init_kwargs,
+                config=self.model_config,
+                tokenizer=self.processor,
+                model_args=self.args,
+                is_trainable=self.is_train,
+            )
+
        if self.args.model_class == ModelClass.LLM:
            from transformers import AutoModelForCausalLM, AutoModelForImageTextToText

@@ -107,14 +127,8 @@ class ModelEngine:

            AutoClass = AutoModel

-        if self.args.init_config is not None:
-            from ..plugins.model_plugins.initialization import InitPlugin
-
-            init_device = InitPlugin(self.args.init_config.name)()
-        else:
-            init_device = DistributedInterface().current_device
-
        if init_device.type == DeviceType.META:
+            assert self.args.quant_config is None, "Quantization is not supported with meta device."
            with init_empty_weights():
                model = AutoClass.from_config(self.model_config)
        else:
@@ -122,8 +136,8 @@ class ModelEngine:
                self.args.model,
                config=self.model_config,
                dtype="auto",
-                device_map=init_device,
                trust_remote_code=self.args.trust_remote_code,
+                **init_kwargs,
            )

        if self.args.peft_config is None:
--- a/src/llamafactory/v1/launcher.py
+++ b/src/llamafactory/v1/launcher.py
@@ -125,6 +125,11 @@ def launch():

        run_chat()

+    elif command == "merge":
+        from llamafactory.v1.plugins.model_plugins.peft import merge_and_export_model
+
+        merge_and_export_model()
+
    elif command == "env":
        raise NotImplementedError("Environment information is not implemented yet.")

--- a/src/llamafactory/v1/plugins/model_plugins/peft.py
+++ b/src/llamafactory/v1/plugins/model_plugins/peft.py
@@ -12,14 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-from typing import Literal, TypedDict
+import re
+from typing import Literal, TypedDict, Union

-from peft import LoraConfig, PeftModel, get_peft_model
+import torch
+from peft import LoraConfig, PeftModel, TaskType, get_peft_model

+from ...config import InputArgument, get_args
+from ...core.model_engine import ModelEngine
+from ...utils import logging
 from ...utils.plugin import BasePlugin
 from ...utils.types import HFModel


+logger = logging.get_logger(__name__)
+
+
 class LoraConfigDict(TypedDict, total=False):
    name: Literal["lora"]
    """Plugin name."""
@@ -27,8 +35,28 @@ class LoraConfigDict(TypedDict, total=False):
    """Lora rank."""
    lora_alpha: int
    """Lora alpha."""
-    target_modules: list[str]
+    lora_dropout: float
+    """Lora dropout."""
+    target_modules: Union[list[str], str]
    """Target modules."""
+    use_rslora: bool
+    """Use RS-LoRA."""
+    use_dora: bool
+    """Use DoRA."""
+    modules_to_save: list[str]
+    """Modules to save."""
+    adapter_name_or_path: Union[list[str], str]
+    """Path to the adapter(s)."""
+    export_dir: str
+    """Path to the export directory."""
+    export_size: int
+    """Shard size for the export model."""
+    export_hub_model_id: str
+    """Hub model ID for the export model."""
+    infer_dtype: Literal["auto", "float16", "float32", "bfloat16"]
+    """Inference data type for the export model."""
+    export_legacy_format: bool
+    """Use legacy format for the export model."""


 class FreezeConfigDict(TypedDict, total=False):
@@ -36,22 +64,280 @@ class FreezeConfigDict(TypedDict, total=False):
    """Plugin name."""
    freeze_trainable_layers: int
    """Freeze trainable layers."""
-    freeze_trainable_modules: list[str] | None
+    freeze_trainable_modules: Union[list[str], str]
    """Freeze trainable modules."""
+    freeze_extra_modules: list[str]
+    """Freeze extra modules."""
+    cast_trainable_params_to_fp32: bool
+    """Cast trainable params to fp32."""


 class PeftPlugin(BasePlugin):
    def __call__(self, model: HFModel, config: dict, is_train: bool) -> HFModel:
-        return super().__call__(model, config)
+        return super().__call__(model, config, is_train)
+
+
+def _find_all_linear_modules(model: HFModel) -> list[str]:
+    r"""Find all available modules to apply LoRA."""
+    forbidden_modules = {"lm_head", "output_layer", "output"}
+    module_names = set()
+    for name, module in model.named_modules():
+        if any(forbidden_module in name for forbidden_module in forbidden_modules):
+            continue
+
+        if "Linear" in module.__class__.__name__ and "Embedding" not in module.__class__.__name__:
+            module_names.add(name.split(".")[-1])
+
+    return list(module_names)
+
+
+def merge_adapters(model: HFModel, adapter_name_or_path: Union[list[str], str]) -> HFModel:
+    if not isinstance(adapter_name_or_path, list):
+        adapter_name_or_path = [adapter_name_or_path]
+
+    for adapter_path in adapter_name_or_path:
+        model = PeftModel.from_pretrained(model, adapter_path)
+        model = model.merge_and_unload()
+        logger.info_rank0(f"Merged adapter from {adapter_path}")
+
+    return model
+
+
+def load_adapter(model: HFModel, adapter_name_or_path: Union[list[str], str], is_train: bool) -> HFModel:
+    r"""Loads adapter(s) into the model.
+
+    Determine adapter usage based on mode:
+    - Training: Load the single adapter for continued training.
+    - Inference: Merge all adapters to clean up the model.
+    - Unmergeable: Keep the single adapter active without merging.
+    """
+    if not isinstance(adapter_name_or_path, list):
+        adapter_name_or_path = [adapter_name_or_path]
+
+    # TODO
+    # Adapters fix for deepspeed and quant
+    # Adapters fix for vision
+
+    if is_train and len(adapter_name_or_path) > 1:
+        raise ValueError(
+            "When `adapter_name_or_path` is provided for training, only a single LoRA adapter is supported. "
+            "Training will continue on the specified adapter. "
+            "Please merge multiple adapters before starting a new LoRA adapter."
+        )
+
+    if is_train:
+        adapter_to_merge = []
+        adapter_to_resume = adapter_name_or_path[0]
+    else:
+        adapter_to_merge = adapter_name_or_path
+        adapter_to_resume = None
+
+    if adapter_to_merge:
+        model = merge_adapters(model, adapter_to_merge)
+
+    if adapter_to_resume is not None:
+        model = PeftModel.from_pretrained(model, adapter_to_resume, is_trainable=is_train)
+        if is_train:
+            logger.info_rank0(
+                f"Resuming training from existing LoRA adapter at {adapter_to_resume}. "
+                "LoRA hyperparameters will be loaded from the adapter itself; "
+                "the current LoRA configuration will be ignored. "
+                "Merge the adapter into the base model before training if you want to start a new adapter."
+            )
+
+    return model


@PeftPlugin("lora").register()
-def get_lora_model(model: HFModel, config: LoraConfigDict, is_train: bool) -> PeftModel:
-    peft_config = LoraConfig(**config)
+def get_lora_model(model: HFModel, config: LoraConfigDict, is_train: bool = False) -> HFModel:
+    adapter_name_or_path = config.get("adapter_name_or_path")
+
+    if adapter_name_or_path:
+        return load_adapter(model, adapter_name_or_path, is_train)
+
+    logger.info_rank0("Fine-tuning method: LoRA")
+
+    target_modules = config.get("target_modules", "all")
+
+    # Handle target modules
+    if target_modules == "all":
+        target_modules = _find_all_linear_modules(model)
+    elif isinstance(target_modules, str):
+        target_modules = [target_modules]
+
+    logger.info_rank0(f"LoRA target modules: {target_modules}")
+
+    peft_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        inference_mode=not is_train,
+        r=config.get("r", 8),
+        lora_alpha=config.get("lora_alpha", 16),
+        lora_dropout=config.get("lora_dropout", 0.05),
+        use_rslora=config.get("use_rslora", False),
+        use_dora=config.get("use_dora", False),
+        target_modules=target_modules,
+        modules_to_save=config.get("modules_to_save", None),
+    )
+
    model = get_peft_model(model, peft_config)
+
+    if is_train:
+        model.print_trainable_parameters()
+
    return model


@PeftPlugin("freeze").register()
-def get_freeze_model(model: HFModel, config: FreezeConfigDict, is_train: bool) -> HFModel:
-    raise NotImplementedError()
+def get_freeze_model(model: HFModel, config: FreezeConfigDict, is_train: bool = False) -> HFModel:
+    logger.info_rank0("Fine-tuning method: Freeze")
+
+    if not is_train:
+        return model
+
+    freeze_trainable_layers = config.get("freeze_trainable_layers", 2)
+    freeze_trainable_modules = config.get("freeze_trainable_modules", ["all"])
+    freeze_extra_modules = config.get("freeze_extra_modules", [])
+    cast_trainable_params_to_fp32 = config.get("cast_trainable_params_to_fp32", True)
+
+    if isinstance(freeze_trainable_modules, str):
+        freeze_trainable_modules = [module.strip() for module in freeze_trainable_modules.split(",")]
+
+    if isinstance(freeze_extra_modules, str):
+        freeze_extra_modules = [module.strip() for module in freeze_extra_modules.split(",")]
+
+    # Get number of layers
+    num_layers = (
+        getattr(model.config, "num_hidden_layers", None)
+        or getattr(model.config, "num_layers", None)
+        or getattr(model.config, "n_layer", None)
+    )
+
+    if not num_layers:
+        raise ValueError("Current model does not support freeze tuning.")
+
+    if freeze_trainable_layers > 0:
+        # last n layers
+        trainable_layer_ids = range(max(0, num_layers - freeze_trainable_layers), num_layers)
+    else:
+        # first n layers
+        trainable_layer_ids = range(min(-freeze_trainable_layers, num_layers))
+
+    # Identify hidden and non-hidden modules
+    hidden_modules = set()
+    non_hidden_modules = set()
+    for name, _ in model.named_parameters():
+        if ".0." in name:
+            hidden_modules.add(name.split(".0.")[-1].split(".")[0])
+        elif ".1." in name:
+            hidden_modules.add(name.split(".1.")[-1].split(".")[0])
+
+        if re.search(r"\.\d+\.", name) is None:
+            non_hidden_modules.add(name.split(".")[-2])
+
+    # Build list of trainable layer patterns
+    trainable_layers = []
+    for module_name in freeze_trainable_modules:
+        if module_name == "all":
+            for idx in trainable_layer_ids:
+                trainable_layers.append(f".{idx:d}.")
+        elif module_name in hidden_modules:
+            for idx in trainable_layer_ids:
+                trainable_layers.append(f".{idx:d}.{module_name}")
+        else:
+            raise ValueError(f"Module {module_name} not found in hidden modules: {hidden_modules}")
+
+    # Add extra modules
+    if freeze_extra_modules:
+        for module_name in freeze_extra_modules:
+            if module_name in non_hidden_modules:
+                trainable_layers.append(module_name)
+            else:
+                raise ValueError(f"Module {module_name} not found in non-hidden modules: {non_hidden_modules}")
+
+    # TODO
+    # Multi-modal special handling
+
+    # Set requires_grad
+    forbidden_modules = {"quant_state", "quantization_weight", "qweight", "qzeros", "scales"}
+    for name, param in model.named_parameters():
+        if any(trainable_layer in name for trainable_layer in trainable_layers) and not any(
+            forbidden_module in name for forbidden_module in forbidden_modules
+        ):
+            param.requires_grad_(True)
+            if cast_trainable_params_to_fp32:
+                param.data = param.data.to(torch.float32)  # Cast to fp32 for stability
+        else:
+            param.requires_grad_(False)
+
+    logger.info_rank0(f"Set trainable layers: {trainable_layers}")
+
+    # Count trainable params for verification
+    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    all_params = sum(p.numel() for p in model.parameters())
+    logger.info_rank0(
+        f"trainable params: {trainable_params} || all params: {all_params} || trainable%: {100 * trainable_params / all_params:.4f}"
+    )
+
+    return model
+
+
+def merge_and_export_model(args: InputArgument = None):
+    model_args, _, _, _ = get_args(args)
+
+    export_config = model_args.peft_config
+    if export_config is None:
+        raise ValueError("Please specify peft_config to merge and export model.")
+
+    export_dir = export_config.get("export_dir")
+    if export_dir is None:
+        raise ValueError("Please specify export_dir.")
+
+    export_size = export_config.get("export_size", 5)
+    export_hub_model_id = export_config.get("export_hub_model_id")
+    infer_dtype = export_config.get("infer_dtype", "auto")
+    export_legacy_format = export_config.get("export_legacy_format", False)
+
+    adapters = None
+    if export_config.get("name") == "lora":
+        adapters = export_config.get("adapter_name_or_path")
+    else:
+        raise ValueError("Currently merge and export model function is only supported for lora.")
+
+    if adapters is None:
+        raise ValueError("Please set adapter_name_or_path to merge adapters into base model.")
+
+    logger.info_rank0("Loading model for export...")
+    model_engine = ModelEngine(model_args, is_train=False)
+    model = model_engine.model
+    tokenizer = model_engine.processor
+
+    if infer_dtype == "auto":
+        if model.config.torch_dtype == torch.float32 and torch.cuda.is_bf16_supported():
+            model = model.to(torch.bfloat16)
+            logger.info_rank0("Converted model to bfloat16.")
+    else:
+        target_dtype = getattr(torch, infer_dtype)
+        model = model.to(target_dtype)
+        logger.info_rank0(f"Converted model to {infer_dtype}.")
+
+    logger.info_rank0(f"Exporting model to {export_dir}...")
+    model.save_pretrained(
+        export_dir,
+        max_shard_size=f"{export_size}GB",
+        safe_serialization=not export_legacy_format,
+    )
+    if tokenizer is not None:
+        try:
+            if hasattr(tokenizer, "padding_side"):
+                tokenizer.padding_side = "left"
+            tokenizer.save_pretrained(export_dir)
+        except Exception as e:
+            logger.warning(f"Failed to save tokenizer: {e}")
+
+    if export_hub_model_id:
+        logger.info_rank0(f"Pushing to hub: {export_hub_model_id}...")
+        model.push_to_hub(export_hub_model_id)
+        if tokenizer is not None:
+            tokenizer.push_to_hub(export_hub_model_id)
+
+    logger.info_rank0("Model exported successfully.")
--- a/src/llamafactory/v1/plugins/model_plugins/quantization.py
+++ b/src/llamafactory/v1/plugins/model_plugins/quantization.py
@@ -0,0 +1,122 @@
+# Copyright 2025 HuggingFace Inc., the KVCache.AI team, Approaching AI, and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/language-modeling/run_clm.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, Any
+
+import torch
+from transformers import BitsAndBytesConfig
+
+from ...accelerator.helper import get_current_device
+from ...config.model_args import ModelArguments
+from ...utils import logging
+from ...utils.packages import check_version
+from ...utils.plugin import BasePlugin
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedTokenizer
+
+logger = logging.get_logger(__name__)
+
+
+class QuantizationPlugin(BasePlugin):
+    r"""Plugin for model quantization."""
+
+    def __call__(
+        self,
+        init_kwargs: dict[str, Any] = None,
+        config: "PretrainedConfig" = None,
+        tokenizer: "PreTrainedTokenizer" = None,
+        model_args: "ModelArguments" = None,
+        is_trainable: bool = False,
+    ) -> dict[str, Any]:
+        return super().__call__(
+            init_kwargs, config=config, tokenizer=tokenizer, model_args=model_args, is_trainable=is_trainable
+        )
+
+
+@QuantizationPlugin("auto").register()
+def quantization_auto(
+    init_kwargs: dict[str, Any],
+    **kwargs,
+) -> dict[str, Any]:
+    """Automatic quantization selection, only support bnb currently.
+
+    Args:
+        init_kwargs (dict[str, Any]): The kwargs for model initialization.
+        **kwargs: Keyword arguments containing the model.
+
+    Returns:
+        dict[str, Any]: The updated kwargs for model initialization.
+    """
+    model_args: ModelArguments = kwargs.get("model_args", None)
+    quant_config = model_args.quant_config
+
+    quantization_bit = quant_config.get("quantization_bit", None)
+    if quantization_bit is not None:
+        logger.info_rank0(f"Loading {quantization_bit}-bit quantized model.")
+        if quantization_bit in [8, 4]:
+            return quantization_with_bnb(init_kwargs, **kwargs)
+        else:
+            raise ValueError(f"Unsupported quantization bit: {quantization_bit} for auto quantization.")
+    logger.warning_rank0("No quantization method applied.")
+    return init_kwargs
+
+
+@QuantizationPlugin("bnb").register()
+def quantization_with_bnb(
+    init_kwargs: dict[str, Any],
+    model_args: "ModelArguments" = None,
+    **kwargs,
+) -> dict[str, Any]:
+    r"""Quantization with BNB."""
+    logger.info_rank0("Using Bitsandbytes quantization.")
+    quantization_bit = model_args.quant_config.get("quantization_bit", None)
+    if quantization_bit is None:
+        logger.warning_rank0("quantization_bit is not specified, default to 8-bit quantization.")
+        quantization_bit = 4
+    assert quantization_bit in [8, 4], "Bitsandbytes only accepts 4-bit or 8-bit quantization."
+    if quantization_bit == 8:
+        check_version("bitsandbytes>=0.37.0", mandatory=True)
+        init_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_8bit=True)
+    elif quantization_bit == 4:
+        check_version("bitsandbytes>=0.39.0", mandatory=True)
+        init_kwargs["quantization_config"] = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=model_args.quant_config.get("compute_dtype", torch.float16),
+            bnb_4bit_use_double_quant=model_args.quant_config.get("double_quantization", True),
+            bnb_4bit_quant_type=model_args.quant_config.get("quantization_type", "nf4"),
+            bnb_4bit_quant_storage=model_args.quant_config.get(
+                "compute_dtype", torch.float16
+            ),  # crucial for fsdp+qlora
+        )
+    else:
+        raise ValueError("Bitsandbytes only accepts 4-bit or 8-bit quantization.")
+
+    # TODO: improve deepspeed zero3 and fsdp detection.
+    if kwargs.get("is_trainable", False):
+        logger.info_rank0("Detected inference mode, setting device_map for bitsandbytes quantization.")
+        init_kwargs["device_map"] = {"": get_current_device()}  # change auto device map for inference
+    else:
+        logger.info_rank0("Detected training mode, skip setting device_map for bitsandbytes quantization.")
+        if model_args.quant_config.get("quantization_bit") != 4:
+            raise ValueError("Only 4-bit quantized model can use fsdp+qlora or auto device map.")
+
+        check_version("bitsandbytes>=0.43.0", mandatory=True)
+
+    logger.info_rank0(f"Quantizing model to {model_args.quant_config.get('quantization_bit')} bit with bitsandbytes.")
+    return init_kwargs
--- a/src/llamafactory/v1/plugins/trainer_plugins/distributed/deepspeed.py
+++ b/src/llamafactory/v1/plugins/trainer_plugins/distributed/deepspeed.py
@@ -0,0 +1,129 @@
+# 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.
+
+"""DeepSpeed integration via accelerate's built-in capabilities.
+
+Instead of manually calling deepspeed.initialize() and syncing config,
+this module leverages accelerate's Accelerator + DeepSpeedPlugin to handle
+initialization, backward, gradient accumulation, and model saving.
+"""
+
+from typing import Any, Optional
+
+import torch
+from accelerate import Accelerator
+from accelerate.utils import DeepSpeedPlugin
+
+from ....utils.logging import get_logger
+from ....utils.types import HFModel, Processor
+
+
+logger = get_logger(__name__)
+
+
+class DeepSpeedEngine:
+    """DeepSpeed integration using accelerate's built-in capabilities.
+
+    This replaces the manual DeepSpeedConfigHelper / DeepSpeedEngine approach
+    with accelerate's Accelerator + DeepSpeedPlugin, which handles:
+    - Config syncing (auto values, batch size, lr, etc.)
+    - deepspeed.initialize() call
+    - Optimizer / LR scheduler wrapping
+    - Backward + gradient accumulation boundary
+    - ZeRO-3 parameter gathering for saving
+    """
+
+    def __init__(self, dist_config: dict[str, Any], num_micro_batch: int = 1, micro_batch_size: int = 1):
+        config_file = dist_config.get("config_file")
+        if not config_file:
+            raise ValueError("DeepSpeed config_file is required in dist_config")
+
+        ds_plugin = DeepSpeedPlugin(hf_ds_config=config_file)
+
+        self.accelerator = Accelerator(
+            deepspeed_plugin=ds_plugin,
+            gradient_accumulation_steps=num_micro_batch,
+        )
+
+        # Resolve "auto" for train_micro_batch_size_per_gpu so that
+        # accelerate.prepare() does not require a DataLoader to infer it.
+        ds_config = self.accelerator.state.deepspeed_plugin.deepspeed_config
+        if ds_config.get("train_micro_batch_size_per_gpu") in (None, "auto"):
+            ds_config["train_micro_batch_size_per_gpu"] = micro_batch_size
+
+        logger.info_rank0(f"DeepSpeedEngine initialized with config: {config_file}")
+
+    def shard_model(self, model: HFModel) -> "DeepSpeedEngine":
+        """No-op shard — actual model wrapping happens in prepare().
+
+        Returns self so the caller gets the engine instance via the hub interface.
+        """
+        return self
+
+    def prepare(
+        self,
+        model: HFModel,
+        optimizer: torch.optim.Optimizer,
+        lr_scheduler: Optional[Any] = None,
+    ) -> tuple[HFModel, torch.optim.Optimizer, Any]:
+        """Prepare model, optimizer, and lr_scheduler using accelerate.
+
+        Internally calls deepspeed.initialize() and wraps the returned objects.
+        """
+        if lr_scheduler is not None:
+            model, optimizer, lr_scheduler = self.accelerator.prepare(model, optimizer, lr_scheduler)
+        else:
+            model, optimizer = self.accelerator.prepare(model, optimizer)
+
+        model._accelerator = self.accelerator  # type: ignore[assignment]
+
+        logger.info_rank0("Model, optimizer, and lr_scheduler prepared via accelerate")
+        return model, optimizer, lr_scheduler
+
+    def backward(self, loss: torch.Tensor) -> None:
+        """Backward pass using accelerate.
+
+        Delegates to DeepSpeedEngineWrapper.backward() which respects
+        sync_gradients to control gradient accumulation boundaries.
+        When sync_gradients=True: engine.backward(loss) + engine.step()
+        When sync_gradients=False: engine.backward(loss) only
+        """
+        self.accelerator.backward(loss)
+
+    def get_grad_norm(self) -> float:
+        """Get the global gradient norm from the DeepSpeed engine."""
+        engine_wrapper = getattr(self.accelerator, "deepspeed_engine_wrapped", None)
+        if engine_wrapper is not None:
+            return engine_wrapper.engine.get_global_grad_norm() or 0.0
+        return 0.0
+
+
+def save_model(model: HFModel, output_dir: str, processor: Processor) -> None:
+    """Save model using accelerate's built-in ZeRO-aware utilities.
+
+    Expects model._accelerator to be set during prepare().
+    Handles ZeRO-3 parameter gathering automatically via
+    accelerator.get_state_dict().
+    """
+    accelerator: Accelerator = model._accelerator  # type: ignore[union-attr]
+
+    unwrapped_model = accelerator.unwrap_model(model)
+    state_dict = accelerator.get_state_dict(model)
+
+    if accelerator.is_main_process:
+        unwrapped_model.save_pretrained(output_dir, state_dict=state_dict, max_shard_size="4GB")
+        processor.save_pretrained(output_dir, max_shard_size="4GB")
+
+    accelerator.wait_for_everyone()
+    logger.info_rank0(f"Model saved to {output_dir}")
--- a/src/llamafactory/v1/plugins/trainer_plugins/distributed/fsdp2.py
+++ b/src/llamafactory/v1/plugins/trainer_plugins/distributed/fsdp2.py
@@ -17,23 +17,24 @@ import os

 import torch
 import torch.nn as nn
+from peft.tuners.lora import LoraLayer
 from torch.distributed.checkpoint.state_dict import StateDictOptions, get_model_state_dict, set_model_state_dict
 from torch.distributed.fsdp import (
    CPUOffloadPolicy,
    MixedPrecisionPolicy,
    fully_shard,
 )
-from transformers import PreTrainedModel

 from ....accelerator.helper import get_current_accelerator
 from ....accelerator.interface import DistributedInterface
 from ....utils.logging import get_logger
+from ....utils.types import HFModel, Processor


 logger = get_logger(__name__)


-def get_transformer_layer_cls(model: PreTrainedModel) -> type[nn.Module] | None:
+def get_transformer_layer_cls(model: HFModel) -> type[nn.Module] | None:
    no_split_modules = getattr(model, "_no_split_modules", None)
    if no_split_modules:
        if isinstance(no_split_modules, (list, tuple)):
@@ -49,6 +50,20 @@ def get_transformer_layer_cls(model: PreTrainedModel) -> type[nn.Module] | None:
    return None


+def save_model(model: HFModel, output_dir: str, processor: Processor) -> None:
+    if DistributedInterface().get_rank() == 0:
+        logger.info("Gathering state dict for saving...")
+
+    options = StateDictOptions(full_state_dict=True, cpu_offload=True)
+    state_dict = get_model_state_dict(model, options=options)
+
+    if DistributedInterface().get_rank() == 0:
+        model_to_save = model.module if hasattr(model, "module") else model
+        model_to_save.save_pretrained(output_dir, state_dict=state_dict, max_shard_size="4GB")
+        processor.save_pretrained(output_dir, max_shard_size="4GB")
+        logger.info(f"Model saved to {output_dir}")
+
+
 class FSDP2Engine:
    def __init__(self, dist_config: dict):
        self.dist_interface = DistributedInterface()
@@ -94,7 +109,10 @@ class FSDP2Engine:
            cast_forward_inputs=True,
        )

-    def prepare_model(self, model: PreTrainedModel) -> PreTrainedModel:
+    def is_lora_module_wrap(self, model) -> bool:
+        return any(isinstance(module, LoraLayer) for module in model.modules())
+
+    def prepare_model(self, model: HFModel) -> HFModel:
        if self.fsdp_mesh is None:
            logger.warning("No FSDP Mesh available, skipping FSDP wrapping.")
            return model
@@ -111,6 +129,25 @@ class FSDP2Engine:
            logger.info(f"Applying per-layer FSDP to {layer_cls.__name__}")
            transformer_layer_cls_to_wrap = {layer_cls}

+        if self.is_lora_module_wrap(model):
+            lora_modules = []
+            for module in model.modules():
+                if len(list(module.children())) != 0:
+                    continue
+                if any(param.requires_grad for param in module.parameters(recurse=False)):
+                    lora_modules.append(module)
+
+            for module in lora_modules:
+                fully_shard(
+                    module,
+                    mesh=self.fsdp_mesh,
+                    reshard_after_forward=self.reshard_after_forward,
+                    mp_policy=mp_policy,
+                    offload_policy=CPUOffloadPolicy(pin_memory=self.pin_memory) if self.offload_params else None,
+                )
+
+            logger.info("Applying FSDP wrap for LoRA layer separately.")
+
        for name, module in model.named_modules():
            should_wrap = False

@@ -156,7 +193,7 @@ class FSDP2Engine:
        return model

    @torch.no_grad()
-    def materialize_and_load(self, model: PreTrainedModel, hf_model_path: str, dcp_path: str = None):
+    def materialize_and_load(self, model: HFModel, hf_model_path: str, dcp_path: str = None):
        if self.rank == 0:
            logger.info("Materializing sharded model params...")

@@ -176,7 +213,7 @@ class FSDP2Engine:

        return model

-    def shard_model(self, model: PreTrainedModel) -> PreTrainedModel:
+    def shard_model(self, model: HFModel) -> HFModel:
        if model.device.type == "meta":
            model = self.prepare_model(model)
            model = self.materialize_and_load(model, hf_model_path=model.config.name_or_path, dcp_path=self.dcp_path)
@@ -184,7 +221,7 @@ class FSDP2Engine:
            model = self.prepare_model(model)
        return model

-    def _load_from_dcp(self, model: PreTrainedModel, dcp_path: str):
+    def _load_from_dcp(self, model: HFModel, dcp_path: str):
        import torch.distributed.checkpoint as dcp

        try:
@@ -203,7 +240,7 @@ class FSDP2Engine:
            logger.error(f"Failed to load from DCP: {e}")
            raise e

-    def _load_weights_from_hf_checkpoint(self, model, hf_model_path):
+    def _load_weights_from_hf_checkpoint(self, model: HFModel, hf_model_path: str):
        import glob
        import json

--- a/src/llamafactory/v1/plugins/trainer_plugins/distributed/hub.py
+++ b/src/llamafactory/v1/plugins/trainer_plugins/distributed/hub.py
@@ -12,9 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
 from ....config.arg_utils import PluginConfig
 from ....utils.plugin import BasePlugin
-from ....utils.types import HFModel
+
+
+if TYPE_CHECKING:
+    from ....utils.types import HFModel, Processor


 class DistributedPlugin(BasePlugin):
@@ -23,12 +30,32 @@ class DistributedPlugin(BasePlugin):


@DistributedPlugin("fsdp2").register()
-def shard_model_fsdp2(model: HFModel, dist_config: PluginConfig) -> HFModel:
+def shard_model_fsdp2(model: HFModel, dist_config: PluginConfig, **kwargs) -> HFModel:
    from .fsdp2 import FSDP2Engine

    return FSDP2Engine(dist_config).shard_model(model)


+@DistributedPlugin("fsdp2").register("save_model")
+def save_model_fsdp2(model: HFModel, output_dir: str, processor: Processor) -> None:
+    from .fsdp2 import save_model
+
+    return save_model(model, output_dir, processor)
+
+
@DistributedPlugin("deepspeed").register()
-def shard_model_deepspeed(model: HFModel, dist_config: PluginConfig) -> HFModel:
-    return model
+def shard_model_deepspeed(model: HFModel, dist_config: PluginConfig, **kwargs) -> HFModel:
+    from .deepspeed import DeepSpeedEngine
+
+    return DeepSpeedEngine(
+        dist_config,
+        num_micro_batch=kwargs.get("num_micro_batch"),
+        micro_batch_size=kwargs.get("micro_batch_size"),
+    ).shard_model(model)
+
+
+@DistributedPlugin("deepspeed").register("save_model")
+def save_model_deepspeed(model: HFModel, output_dir: str, processor: Processor) -> None:
+    from .deepspeed import save_model
+
+    return save_model(model, output_dir, processor)
--- a/src/llamafactory/v1/trainers/sft_trainer.py
+++ b/src/llamafactory/v1/trainers/sft_trainer.py
@@ -33,7 +33,7 @@ def run_sft(args: InputArgument = None):
    model_args, data_args, training_args, _ = get_args(args)
    DistributedInterface(training_args.dist_config)
    train_dataset = DataEngine(data_args.train_dataset)
-    model_engine = ModelEngine(model_args)
+    model_engine = ModelEngine(model_args, is_train=True)
    trainer = SFTTrainer(
        args=training_args,
        model=model_engine.model,
--- a/src/llamafactory/v1/utils/packages.py
+++ b/src/llamafactory/v1/utils/packages.py
@@ -21,6 +21,13 @@ from functools import lru_cache
 from typing import TYPE_CHECKING

 from packaging import version
+from transformers.utils.versions import require_version
+
+from . import logging
+from .env import is_env_enabled
+
+
+logger = logging.get_logger(__name__)


 if TYPE_CHECKING:
@@ -41,3 +48,22 @@ def _get_package_version(name: str) -> "Version":
@lru_cache
 def is_transformers_version_greater_than(content: str):
    return _get_package_version("transformers") >= version.parse(content)
+
+
+def check_version(requirement: str, mandatory: bool = False) -> None:
+    r"""Optionally check the package version."""
+    if is_env_enabled("DISABLE_VERSION_CHECK") and not mandatory:
+        logger.warning_rank0_once("Version checking has been disabled, may lead to unexpected behaviors.")
+        return
+
+    if "gptqmodel" in requirement or "autoawq" in requirement:
+        pip_command = f"pip install {requirement} --no-build-isolation"
+    else:
+        pip_command = f"pip install {requirement}"
+
+    if mandatory:
+        hint = f"To fix: run `{pip_command}`."
+    else:
+        hint = f"To fix: run `{pip_command}` or set `DISABLE_VERSION_CHECK=1` to skip this check."
+
+    require_version(requirement, hint)
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -166,3 +166,33 @@ def _manage_distributed_env(request: FixtureRequest, monkeypatch: MonkeyPatch) -
 def fix_valuehead_cpu_loading():
    """Fix valuehead model loading."""
    patch_valuehead_model()
+
+
+@pytest.fixture(scope="session", autouse=True)
+def bypass_mistral_regex_check():
+    """Disable Mistral regex network check.
+
+    Monkey-patch TokenizersBackend._patch_mistral_regex into a no-op.
+    """
+    try:
+        from transformers.tokenization_utils_fast import TokenizersBackend
+    except ImportError:
+        # Very old transformers, nothing to patch
+        yield
+        return
+
+    if not hasattr(TokenizersBackend, "_patch_mistral_regex"):
+        # Method does not exist in this version
+        yield
+        return
+
+    # Backup original method
+    original = TokenizersBackend._patch_mistral_regex
+
+    # Replace with no-op
+    TokenizersBackend._patch_mistral_regex = lambda cls, tokenizer, *args, **kwargs: tokenizer
+
+    yield
+
+    # Restore original method
+    TokenizersBackend._patch_mistral_regex = original
--- a/tests_v1/conftest.py
+++ b/tests_v1/conftest.py
@@ -172,3 +172,33 @@ def _manage_distributed_env(request: FixtureRequest, monkeypatch: MonkeyPatch) -
            monkeypatch.setattr(torch.cuda, "device_count", lambda: 1)
        elif CURRENT_DEVICE == "npu":
            monkeypatch.setattr(torch.npu, "device_count", lambda: 1)
+
+
+@pytest.fixture(scope="session", autouse=True)
+def bypass_mistral_regex_check():
+    """Disable Mistral regex network check.
+
+    Monkey-patch TokenizersBackend._patch_mistral_regex into a no-op.
+    """
+    try:
+        from transformers.tokenization_utils_fast import TokenizersBackend
+    except ImportError:
+        # Very old transformers, nothing to patch
+        yield
+        return
+
+    if not hasattr(TokenizersBackend, "_patch_mistral_regex"):
+        # Method does not exist in this version
+        yield
+        return
+
+    # Backup original method
+    original = TokenizersBackend._patch_mistral_regex
+
+    # Replace with no-op
+    TokenizersBackend._patch_mistral_regex = lambda cls, tokenizer, *args, **kwargs: tokenizer
+
+    yield
+
+    # Restore original method
+    TokenizersBackend._patch_mistral_regex = original
--- a/tests_v1/plugins/model_plugins/test_peft.py
+++ b/tests_v1/plugins/model_plugins/test_peft.py
@@ -0,0 +1,156 @@
+# 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 pytest
+from peft import LoraConfig, PeftModel, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from llamafactory.v1.plugins.model_plugins import peft as peft_module
+from llamafactory.v1.plugins.model_plugins.peft import merge_and_export_model
+
+
+TINY_MODEL = "llamafactory/tiny-random-qwen3"
+
+
+@pytest.fixture(scope="module")
+def model_path():
+    return TINY_MODEL
+
+
+@pytest.fixture(scope="function")
+def model(model_path):
+    return AutoModelForCausalLM.from_pretrained(model_path)
+
+
+@pytest.fixture(scope="function")
+def tokenizer(model_path):
+    return AutoTokenizer.from_pretrained(model_path)
+
+
+@pytest.fixture(scope="function")
+def adapter_path(tmp_path):
+    # Create a dummy adapter
+    lora_config = LoraConfig(
+        r=8,
+        lora_alpha=16,
+        target_modules=["q_proj", "v_proj"],
+        lora_dropout=0.05,
+        bias="none",
+        task_type="CAUSAL_LM",
+    )
+
+    base_model = AutoModelForCausalLM.from_pretrained(TINY_MODEL)
+    peft_model = get_peft_model(base_model, lora_config)
+    save_path = tmp_path / "test_adapter"
+    peft_model.save_pretrained(save_path)
+    return str(save_path)
+
+
+def test_find_all_linear_modules(model):
+    """Verify linear modules are discoverable and include q_proj / v_proj for tiny-random-qwen3."""
+    modules = peft_module._find_all_linear_modules(model)
+    expected_subset = {"q_proj", "v_proj"}
+    assert expected_subset.issubset(set(modules))
+
+
+def test_get_lora_model(model):
+    """Verify a PeftModel is returned and LoRA config takes effect."""
+    config = {"name": "lora", "r": 8, "target_modules": "all", "lora_alpha": 16}
+    model = peft_module.get_lora_model(model, config, is_train=True)
+    assert isinstance(model, PeftModel)
+    assert model.peft_config["default"].r == 8
+    assert "q_proj" in model.peft_config["default"].target_modules
+
+
+def test_get_freeze_model_layers(model):
+    """Verify layer-wise freezing: only the last layer stays trainable."""
+    # Freeze all but last layer
+    config = {"name": "freeze", "freeze_trainable_layers": 1, "freeze_trainable_modules": "all"}
+
+    # Ensure we start with something known
+    model = peft_module.get_freeze_model(model, config, is_train=True)
+
+    num_layers = model.config.num_hidden_layers
+    assert num_layers > 0
+
+    for name, param in model.named_parameters():
+        if f"layers.{num_layers - 1}" in name:
+            assert param.requires_grad, f"{name} should be trainable"
+        elif "layers.0" in name and num_layers > 1:
+            assert not param.requires_grad, f"{name} should be frozen"
+
+
+def test_get_freeze_model_modules(model):
+    """Verify module-wise freezing: only last-layer self_attn is trainable."""
+    # Freeze specific modules (e.g. only self_attn)
+    config = {"name": "freeze", "freeze_trainable_layers": 1, "freeze_trainable_modules": "self_attn"}
+    model = peft_module.get_freeze_model(model, config, is_train=True)
+
+    num_layers = model.config.num_hidden_layers
+
+    for name, param in model.named_parameters():
+        if f"layers.{num_layers - 1}" in name and "self_attn" in name:
+            assert param.requires_grad, f"{name} should be trainable"
+        else:
+            assert not param.requires_grad, f"{name} should be frozen"
+
+
+def test_load_adapter_single_for_inference(model, adapter_path):
+    """Verify single adapter is merged+unloaded in inference mode."""
+    # Test loading single adapter for inference (merge and unload)
+    model_result = peft_module.load_adapter(model, adapter_path, is_train=False)
+    assert not isinstance(model_result, PeftModel)
+
+
+def test_load_adapter_resume_train(model, adapter_path):
+    """Verify training mode returns a trainable PeftModel."""
+    # Test loading for training
+    model_result = peft_module.load_adapter(model, adapter_path, is_train=True)
+    assert isinstance(model_result, PeftModel)
+
+
+def test_load_adapter_train_multiple_disallowed(model, adapter_path):
+    """Verify multiple adapters are rejected in training mode."""
+    with pytest.raises(ValueError, match="only a single LoRA adapter"):
+        peft_module.load_adapter(model, [adapter_path, adapter_path], is_train=True)
+
+
+def test_load_adapter_infer_multiple_merges(model, adapter_path):
+    """Verify multiple adapters are merged in inference mode."""
+    # Test merging multiple adapters
+    model_result = peft_module.load_adapter(model, [adapter_path, adapter_path], is_train=False)
+    assert not isinstance(model_result, PeftModel)
+
+
+def test_merge_and_export_model(tmp_path, adapter_path):
+    """Verify merge_and_export_model produces export artifacts."""
+    export_dir = tmp_path / "export"
+
+    args_dict = {
+        "model": TINY_MODEL,
+        "peft_config": {
+            "name": "lora",
+            "adapter_name_or_path": adapter_path,
+            "export_dir": str(export_dir),
+            "export_size": 1,
+            "infer_dtype": "float16",
+        },
+    }
+
+    merge_and_export_model(args_dict)
+
+    assert export_dir.exists()
+    assert (export_dir / "config.json").exists()
+    assert (export_dir / "model.safetensors").exists()
+    assert (export_dir / "tokenizer_config.json").exists()
--- a/tests_v1/plugins/model_plugins/test_quantization_plugin.py
+++ b/tests_v1/plugins/model_plugins/test_quantization_plugin.py
@@ -0,0 +1,51 @@
+# 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 pytest
+
+from llamafactory.v1.config.model_args import ModelArguments
+from llamafactory.v1.core.model_engine import ModelEngine
+
+
+bitsandbytes = pytest.importorskip("bitsandbytes")
+
+
+def check_quantization_status(model):
+    quantized_info = {"bnb": []}
+
+    for name, module in model.named_modules():
+        # check BitsAndBytes quantization
+        if isinstance(module, bitsandbytes.nn.modules.Linear8bitLt) or isinstance(
+            module, bitsandbytes.nn.modules.Linear4bit
+        ):
+            quantized_info["bnb"].append(name)
+
+    return quantized_info
+
+
+@pytest.mark.runs_on(["cuda"])
+@pytest.mark.parametrize("name, quantization_bit", [("bnb", 4), ("auto", 4)])
+def test_quantization_plugin(name, quantization_bit):
+    model_args = ModelArguments(
+        model="llamafactory/tiny-random-qwen3",
+        quant_config={
+            "name": name,
+            "quantization_bit": quantization_bit,
+        },
+    )
+
+    model_engine = ModelEngine(model_args=model_args)
+    quantized_info = check_quantization_status(model_engine.model)
+    print(f"Quantized weights for method {name} with {quantization_bit} bit: {quantized_info}")
+    assert any(v for v in quantized_info.values()), "model is not quantized properly."
Author	SHA1	Message	Date
Shanay Mehta	aab9b400bb	[model] Add DeepSpeed Z3 leaf module for Qwen3-Next (#10194 ) Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>	2026-02-24 19:54:37 +08:00
P. Clawmogorov	50599c719b	[misc] remove safe_serialization arg for transformers v5 compatibility (#10208 ) Co-authored-by: P. Clawmogorov <262173731+Alm0stSurely@users.noreply.github.com>	2026-02-24 11:14:19 +08:00
Kingsley	a0f3ad0cee	[mca] update supported models (#10196 )	2026-02-20 22:02:49 +08:00
jiaqiw09	f80e15dbb4	[ci] fix ut huggingface hub 429 error when transformers>=5.0.0 (#10155 )	2026-02-12 22:14:10 +08:00
sunyi0505	991267fd3b	[v1] support quantization (#10161 )	2026-02-12 20:37:41 +08:00
浮梦	5c52afa30d	[v1] support deepspeed (#10181 )	2026-02-12 17:24:30 +08:00
Junyou Su	675ce8cc7f	[algo] add ASFT (#10174 )	2026-02-12 13:12:14 +08:00
jiaqiw09	ab073f4c13	[v1] add LoRA/Freeze support and merge workflow (#10157 )	2026-02-12 13:02:09 +08:00
Shanay Mehta	184304b5b4	[model] add liger kernel support for Qwen3-Next (#10176 ) Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>	2026-02-10 21:47:48 +08:00
Xue Yadong	d3ebd5678d	[model] support GLM-OCR SFT (#10183 )	2026-02-10 21:41:01 +08:00
浮梦	1d5e8ebcd0	[v1] init commit for v1 docs (#10145 ) Co-authored-by: frozenleaves <frozen@Mac.local> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: jiaqiw09 <jiaqiw960714@gmail.com> Co-authored-by: jiaqiw09 <60021713+jiaqiw09@users.noreply.github.com> Co-authored-by: Yaowei Zheng <hiyouga@buaa.edu.cn>	2026-02-09 19:43:55 +08:00
Shanay Mehta	ea644d04ec	[model] support GLM-4.7-Flash SFT (#10173 ) Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>	2026-02-09 10:40:44 +08:00