docs: update and optimize README and project documentation system

- Update README.md with improved formatting, logo, badges, and comprehensive documentation - Add README.en.md with English translation of the documentation - Include detailed quick start guide, CLI reference, and supported models table - Add changelog information highlighting recent features - Add script directory and adjust script file locations
2026-02-06 16:13:50 +08:00
parent 978b74ec13
commit c582c4cd0a
22 changed files with 5613 additions and 412 deletions
@@ -1,2 +1,19 @@
+.gitignore
+
 /target
 .idea
+/models
+
+/ope*
+AG*
+CLA*
+
+.claude
+README.md.bak
+test.mp3
+docs/MODEL_DEVELOPMENT_GUIDE.md
+.DS_Store
+.gitignore
+
+bak
+.gitignore
@@ -1,112 +0,0 @@
-# Changelog
-
-## [Unreleased] - 2025-02-04
-
-### Added
- Support for Qwen3-ASR model
-
-## [Unreleased]
-
-### Added
-
- **CLI `run` Subcommand**: Direct model inference from CLI without HTTP service overhead:
-  - `aha run` - Run model inference directly
-  - `-m, --model <MODEL>` - Specify which model to use
-  - `-in, --input <INPUT>` - Input text or file path (model-specific interpretation)
-  - `-out, --output <OUTPUT>` - Output file path (optional, auto-generated if not specified)
-  - `--weight-path <WEIGHT_PATH>` - Local model weight path (required)
-
-
-### Changed
-
- **CLI Structure**: Refactored CLI to use clap's Subcommand feature while maintaining backward compatibility
- **Backward Compatibility**: Commands without subcommand now default to `cli` subcommand:
-  - `aha -m qwen3vl-2b` is equivalent to `aha cli -m qwen3vl-2b`
-  - All existing parameter options and defaults remain unchanged
-
-### Technical Details
-
-**Subcommand Parameters:**
-
-`aha cli`:
- `-a, --address <ADDRESS>` - Server address (default: 127.0.0.1)
- `-p, --port <PORT>` - Server port (default: 10100)
- `-m, --model <MODEL>` - Model to use (required)
- `--weight-path <WEIGHT_PATH>` - Local model weight path (optional)
- `--save-dir <SAVE_DIR>` - Directory to save downloaded model (optional)
- `--download-retries <DOWNLOAD_RETRIES>` - Download retry attempts (default: 3)
-
-`aha serv`:
- `-a, --address <ADDRESS>` - Server address (default: 127.0.0.1)
- `-p, --port <PORT>` - Server port (default: 10100)
- `-m, --model <MODEL>` - Model to use (required)
- `--weight-path <WEIGHT_PATH>` - Local model weight path (required)
-
-`aha download`:
- `-m, --model <MODEL>` - Model to download (required)
- `-s, --save-dir <SAVE_DIR>` - Directory to save downloaded model (optional)
- `--download-retries <DOWNLOAD_RETRIES>` - Download retry attempts (default: 3)
-
-**Code Changes:**
- Modified `src/main.rs` only
- Extracted common functions: `get_model_id()`, `start_http_server()`
- Reused existing `download_model()` and `init()` functions
- No changes to other modules or dependencies
-
-## [0.1.8] - 2025-01-20
-
-### Added
-
- Support for Fun-ASR-Nano-2512 model
- Support for Qwen3-0.6B model
-
-## [0.1.7] - 2024-XX-XX
-
-### Added
-
- Support for GLM-ASR-Nano-2512 model
-
-## [0.1.6] - 2024-XX-XX
-
-### Added
-
- Support for RMBG-2.0 model (background removal)
-
-## [0.1.5] - 2024-XX-XX
-
-### Added
-
- Support for VoxCPM1.5 model
-
-## [0.1.4] - 2024-XX-XX
-
-### Added
-
- Support for PaddleOCR-VL model
-
-## [0.1.3] - 2024-XX-XX
-
-### Added
-
- Support for Hunyuan-OCR model
-
-## [0.1.2] - 2024-XX-XX
-
-### Added
-
- Support for DeepSeek-OCR model
-
-## [0.1.1] - 2024-XX-XX
-
-### Added
-
- Support for Qwen3VL model family (2B, 4B, 8B, 32B)
-
-## [0.1.0] - 2024-XX-XX
-
-### Added
-
- Initial release
- Support for Qwen2.5VL models (3B, 7B)
- Support for MiniCPM4-0.5B model
- Support for VoxCPM-0.5B model
@@ -0,0 +1,228 @@
+<p align="center">
+  <img src="assets/img/logo.png" alt="aha logo" width="100"/>
+</p>
+
+<p align="center">
+  <!-- <a href="https://github.com/jhqxxx/aha/releases">
+    <img src="https://img.shields.io/github/v/release/jhqxxx/aha" alt="GitHub release (latest by date)">
+  </a>
+  <a href="https://github.com/jhqxxx/aha/actions">
+    <img src="https://img.shields.io/github/actions/workflow/status/jhqxxx/aha/ci.yml" alt="GitHub Actions Workflow Status">
+  </a> -->
+  <a href="https://github.com/jhqxxx/aha/blob/main/LICENSE">
+    <img src="https://img.shields.io/github/license/jhqxxx/aha" alt="GitHub License">
+  </a>
+  <a href="https://github.com/jhqxxx/aha/stargazers">
+    <img src="https://img.shields.io/github/stars/jhqxxx/aha" alt="GitHub Stars">
+  </a>
+  <a href="https://github.com/jhqxxx/aha/issues">
+    <img src="https://img.shields.io/github/issues/jhqxxx/aha" alt="GitHub Issues">
+  </a>
+</p>
+
+<p align="center">
+  <a href="README.md">简体中文</a> | <strong>English</strong>
+</p>
+
+# aha
+
+**Lightweight AI Inference Engine — All-in-one Solution for Text, Vision, Speech, and OCR**
+
+aha is a high-performance, cross-platform AI inference engine built with Rust and the Candle framework. It brings state-of-the-art AI models to your local machine—no API keys, no cloud dependencies, just pure, fast AI running directly on your hardware.
+
+## Changelog
+
+### v0.2.0 (2026-02-05)
+- Added Qwen3-ASR speech recognition model
+
+### v0.1.9 (2026-01-31)
+- Added CLI `list` subcommand to show supported models
+- Added CLI subcommand structure support (`cli`, `serv`, `download`, `run`)
+- Fixed Qwen3VL thinking startswith bug
+- Fixed `aha run` multiple inputs bug
+
+### v0.1.8 (2026-01-17)
+- Added Qwen3 text model support
+- Added Fun-ASR-Nano-2512 speech recognition model
+- Fixed ModelScope Fun-ASR-Nano model load error
+- Updated audio resampling with rubato
+
+### v0.1.7 (2026-01-07)
+- Added GLM-ASR-Nano-2512 speech recognition model
+- Merged Metal (GPU) support for Apple Silicon
+- Added dynamic home directory and model download script
+
+**[View full changelog](docs/changelog.md)** →
+
+## Quick Start
+
+### Installation
+
+```bash
+git clone https://github.com/jhqxxx/aha.git
+cd aha
+cargo build --release
+```
+
+**Optional Features:**
+
+```bash
+# CUDA (NVIDIA GPU acceleration)
+cargo build --release --features cuda
+
+# Metal (Apple GPU acceleration for macOS)
+cargo build --release --features metal
+
+# Flash Attention (faster inference)
+cargo build --release --features flash-attn
+
+# FFmpeg (multimedia processing)
+cargo build --release --features ffmpeg
+
+# Combine multiple features
+cargo build --release --features "cuda,flash-attn"
+```
+
+### CLI Quick Reference
+
+```bash
+
+# List all supported models
+aha list
+
+# Download model only
+aha download -m qwen3asr-0.6b
+
+# Download model and start service
+aha -m qwen3asr-0.6b
+
+# Run inference directly (without starting service)
+aha run -m qwen3asr-0.6b -i "audio.wav"
+
+# Start service only (model already downloaded)
+aha serv -m qwen3asr-0.6b -p 10100
+
+```
+
+### Chat
+
+```bash
+aha serv -m qwen3-0.6b -p 10100
+```
+
+Then use the unified (OpenAI-compatible) API:
+
+```bash
+curl http://localhost:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [{"role": "user", "content": "Hello!"}]
+  }
+'
+```
+
+### Supported Models
+
+| Category | Models |
+|----------|--------|
+| **Text** | Qwen3, MiniCPM4 |
+| **Vision** | Qwen2.5-VL, Qwen3-VL |
+| **OCR** | DeepSeek-OCR, Hunyuan-OCR, PaddleOCR-VL |
+| **ASR** | GLM-ASR-Nano, Fun-ASR-Nano, Qwen3-ASR |
+| **Audio** | VoxCPM, VoxCPM1.5 |
+| **Image** | RMBG-2.0 (background removal) |
+
+## Documentation
+
+| Document | Description |
+|----------|-------------|
+| [Getting Started](docs/getting-started.md) | First steps with aha |
+| [Installation](docs/installation.md) | Detailed installation guide |
+| [CLI Reference](docs/cli.md) | Command-line interface |
+| [API Documentation](docs/api.md) | Library & REST API |
+| [Supported Models](docs/supported-models.md) | Available AI models |
+| [Concepts](docs/concepts.md) | Architecture & design |
+| [Development](docs/development.md) | Contributing guide |
+| [Changelog](docs/changelog.md) | Version history |
+
+## Why aha?
+- **🚀 High-Performance Inference** - Powered by Candle framework for efficient tensor computation and model inference
+- **🔧 Unified Interface** — One tool for text, vision, speech, and OCR
+- **📦 Local-First** — All processing runs locally, no data leaves your machine
+- **🎯 Cross-Platform** — Works on Linux, macOS, and Windows
+- **⚡ GPU Accelerated** — Optional CUDA support for faster inference
+- **🛡️ Memory Safe** — Built with Rust for reliability
+- **🧠 Attention Optimization** - Optional Flash Attention support for optimized long sequence processing
+
+## Development
+
+### Using aha as a Library
+> cargo add aha
+
+```bash
+# VoxCPM example
+use aha::models::voxcpm::generate::VoxCPMGenerate;
+use aha::utils::audio_utils::save_wav;
+use anyhow::Result;
+
+fn main() -> Result<()> {
+    let model_path = "xxx/openbmb/VoxCPM-0.5B/";
+
+    let mut voxcpm_generate = VoxCPMGenerate::init(model_path, None, None)?;
+
+    let generate = voxcpm_generate.generate(
+        "The sun is shining bright, flowers smile at me, birds say early early early".to_string(),
+        None,
+        None,
+        2,
+        100,
+        10,
+        2.0,
+        false,
+        6.0,
+    )?;
+
+    let _ = save_wav(&generate, "voxcpm.wav")?;
+    Ok(())
+}
+```
+
+### Extending New Models
+
+- Create new model file in src/models/
+- Export in src/models/mod.rs
+- Add support for CLI model inference in src/exec/
+- Add tests and examples in tests/
+
+## Features
+
+- High-performance inference via Candle framework
+- Multi-modal model support (vision, language, speech)
+- Clean, easy-to-use API design
+- Minimal dependencies, compact binaries
+- Flash Attention support for long sequences
+- FFmpeg support for multimedia processing
+
+## License
+
+Apache-2.0 &mdash; See [LICENSE](LICENSE) for details.
+
+## Acknowledgments
+
+- [Candle](https://github.com/huggingface/candle) - Excellent Rust ML framework
+- All model authors and contributors
+
+---
+
+<p align="center">
+  <sub>Built with ❤️ by the aha team</sub>
+</p>
+
+<p align="center">
+  <sub>We're continuously expanding our model support. Contributions are welcome!</sub>
+</p>
+
+<p align="center">
+  <sub>If this project helps you, please consider giving us a ⭐ Star!</sub>
+</p>
@@ -1,241 +1,167 @@
+<p align="center">
+  <img src="assets/img/logo.png" alt="aha logo" width="100"/>
+</p>
+
+<p align="center">
+  <!-- <a href="https://github.com/jhqxxx/aha/releases">
+    <img src="https://img.shields.io/github/v/release/jhqxxx/aha" alt="GitHub release (latest by date)">
+  </a>
+  <a href="https://github.com/jhqxxx/aha/actions">
+    <img src="https://img.shields.io/github/actions/workflow/status/jhqxxx/aha/ci.yml" alt="GitHub Actions Workflow Status">
+  </a> -->
+  <a href="https://github.com/jhqxxx/aha/blob/main/LICENSE">
+    <img src="https://img.shields.io/github/license/jhqxxx/aha" alt="GitHub License">
+  </a>
+  <a href="https://github.com/jhqxxx/aha/stargazers">
+    <img src="https://img.shields.io/github/stars/jhqxxx/aha" alt="GitHub Stars">
+  </a>
+  <a href="https://github.com/jhqxxx/aha/issues">
+    <img src="https://img.shields.io/github/issues/jhqxxx/aha" alt="GitHub Issues">
+  </a>
+</p>
+
+<p align="center">
+  <a href="README.en.md">English</a> | <strong>简体中文</a>
+</p>
+
 # aha
-一个基于 Candle 框架的 Rust 模型推理库，提供高效、易用的多模态模型推理能力。

-## 特性
-* 🚀 高性能推理 - 基于 Candle 框架，提供高效的张量计算和模型推理
-* 🎯 多模型支持 - 集成视觉、语言和语音多模态模型
-* 🔧 易于使用 - 简洁的 API 设计，快速上手
-* 🛡️ 内存安全 - 得益于 Rust 的所有权系统，确保内存安全
-* 📦 轻量级 - 最小化依赖，编译产物小巧
-* ⚡ GPU 加速 - 可选 CUDA 支持
-* 🧠 注意力优化 - 可选 Flash Attention 支持，优化长序列处理
+**轻量 AI 推理引擎 —— 文本、视觉、语音与 OCR 一站式解决方案**

-## 支持的模型
-### 当前已实现
-* [Qwen2.5VL](https://huggingface.co/collections/Qwen/qwen25-vl) - 阿里通义千问 2.5 多模态大语言模型
-    - 模型：[Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) 开源协议：[Qwen RESEARCH LICENSE AGREEMENT](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/blob/main/LICENSE)
-    - 模型：[Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-* [MiniCPM4](https://huggingface.co/collections/openbmb/minicpm4) - 面壁智能 MiniCPM 系列语言模型
-    - 模型：
-    	- huggingface: [MiniCPM4-0.5B](https://huggingface.co/openbmb/MiniCPM4-0.5B)  开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-        - AtomGit: [MiniCPM4-0.5B](https://ai.atomgit.com/OpenBMB/MiniCPM4-0.5B)
-* [VoxCPM](https://huggingface.co/collections/openbmb/voxcpm) - 面壁智能语音生成模型
-    - 模型：
-    	- huggingface: [VoxCPM-0.5B](https://huggingface.co/openbmb/VoxCPM-0.5B) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-        - AtomGit: [VoxCPM-0.5B](https://ai.atomgit.com/OpenBMB/VoxCPM-0.5B)
-* [Qwen3VL](https://huggingface.co/collections/Qwen/qwen3-vl) - 阿里通义千问 3 多模态大语言模型
-    - 模型：[Qwen3-VL-2B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-    - 模型：[Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-    - 模型：[Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-    - 模型：[Qwen3-VL-32B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-* DeepSeek-OCR - 深度求索光学文字识别模型
-    - 模型：[DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) 开源协议：[MIT](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/mit.md)
-* Hunyuan-OCR - 腾讯混元光学文字识别模型
-    - 模型：
-    	- huggingface: [HunyuanOCR](https://huggingface.co/tencent/HunyuanOCR) 开源协议：[TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT](https://huggingface.co/tencent/HunyuanOCR/blob/main/LICENSE)
-        - AtomGit: [HunyuanOCR](https://ai.atomgit.com/tencent_hunyuan/HunyuanOCR)
-* [PaddleOCR-VL](https://huggingface.co/collections/PaddlePaddle/paddleocr-vl) - 百度飞桨光学文字识别模型
-    - 模型：
-    	- huggingface: [PaddleOCR-VL](https://huggingface.co/PaddlePaddle/PaddleOCR-VL) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-        - AtomGit: [PaddleOCR-VL](https://ai.atomgit.com/paddlepaddle/PaddleOCR-VL)
-* [VoxCPM1.5](https://huggingface.co/collections/openbmb/voxcpm) - 面壁智能语音生成模型1.5版本
-    - 模型：
-    	- huggingface: [VoxCPM1.5](https://huggingface.co/openbmb/VoxCPM1.5) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-        - AtomGit: [VoxCPM1.5](https://ai.atomgit.com/OpenBMB/VoxCPM1.5)
-* [RMBG2.0](https://huggingface.co/collections/briaai/rmbg) - RMBGv2.0由BRIA AI开发，供非商业用途使用。
-    - 模型：[RMBG2.0](https://huggingface.co/briaai/RMBG-2.0) 开源协议：[Attribution-NonCommercial 4.0 International](https://creativecommons.org/licenses/by-nc/4.0/deed.en)
-* GLM-ASR-Nano-2512 - 智谱AI语音识别模型
-    - 模型：
-    	- huggingface: [GLM-ASR-Nano-2512](https://huggingface.co/zai-org/GLM-ASR-Nano-2512) 开源协议：[MIT](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/mit.md)
-        - AtomGit: [GLM-ASR-Nano-2512](https://ai.atomgit.com/zai-org/GLM-ASR-Nano-2512)
-* Fun-ASR-Nano-2512 - 通义百聆语音识别模型
-    - 模型:[Fun-ASR-Nano-2512](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512) 开源协议未标明
-* [Qwen3](https://huggingface.co/collections/Qwen/qwen3) - 通义千问 Qwen3系列语言模型
-    - 模型:[Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) 开源协议: [Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-* [Qwen3-ASR](https://huggingface.co/collections/Qwen/qwen3-asr) - 通义千问 Qwen3语音识别模型
-    - 模型:[Qwen3-ASR-0.6B](https://huggingface.co/Qwen/Qwen3-ASR-0.6B) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)    
-    - 模型:[Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B) 开源协议：[Apache license 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
+aha 是一款基于 Rust 和 Candle 框架构建的高性能跨平台 AI 推理引擎。将最先进的 AI 模型带到您的本地机器——无需 API 密钥，无需云依赖，纯粹、快速的 AI 直接在您的硬件上运行。

-## 计划支持
-我们持续扩展支持的模型列表，欢迎贡献！
+## 更新日志

-⭐ 如果这个项目对你有帮助，请给我们一个 Star！
+### v0.2.0 (2026-02-05)
+- 新增 Qwen3-ASR 语音识别模型
+
+### v0.1.9 (2026-01-31)
+- 新增 CLI `list` 子命令，显示支持的模型
+- 新增 CLI 子命令结构支持（`cli`、`serv`、`download`、`run`）
+- 修复 Qwen3VL thinking startswith bug
+- 修复 `aha run` 多输入 bug
+
+### v0.1.8 (2026-01-17)
+- 新增 Qwen3 文本模型支持
+- 新增 Fun-ASR-Nano-2512 语音识别模型
+- 修复 ModelScope Fun-ASR-Nano 模型加载错误
+- 使用 rubato 更新音频重采样
+
+### v0.1.7 (2026-01-07)
+- 新增 GLM-ASR-Nano-2512 语音识别模型
+- 合并 Metal (GPU) 支持，适用于 Apple Silicon
+- 新增动态主目录和模型下载脚本
+
+**[查看完整更新日志](docs/changelog.zh-CN.md)** →
+
+## 快速开始
+
+### 安装

-## 环境依赖
-* 启用ffmpeg的feature时：
-    * ubuntu/WSL
 ```bash
-    sudo apt-get update
-    sudo apt-get install -y clang pkg-config ffmpeg libavutil-dev libavcodec-dev libavformat-dev libavfilter-dev libavdevice-dev libswresample-dev libswscale-dev
-    ```
-    * windows参考： https://github.com/zmwangx/rust-ffmpeg/wiki/Notes-on-building
-
-## 功能特性
-项目提供了几个可选的功能特性，您可以根据需要启用它们：
-* flash-attn: 启用 Flash Attention 支持以提升模型推理性能：
-```bash
-cargo build -r --features flash-attn
+git clone https://github.com/jhqxxx/aha.git
+cd aha
+cargo build --release
 ```

-* cuda: 为 candle 核心组件启用 CUDA 支持，实现 Nvidia GPU 加速计算：
-```bash
-cargo build -r --features cuda
-```
-
-* cuda: 为 candle 核心组件启用 Metal 支持，利用 Apple GPU 加速计算（适用于 macOS/iOS 平台）：
-```bash
-cargo build -r --features metal
-```
-
-* ffmpeg: 启用 FFmpeg 支持，提供多媒体处理功能：
-```bash
-cargo build -r --features ffmpeg
-```
-* 组合使用功能特性
+**可选特性：**

 ```bash
-# 同时启用 CUDA 和 Flash Attention 以获得最佳性能
-cargo build -r --features "cuda,flash-attn"
-```
-## 命令行工具 (CLI)
-项目提供命令行工具，支持模型下载、服务启动和直接推理等多种操作。
+# CUDA (NVIDIA GPU 加速)
+cargo build --release --features cuda
+
+# Metal (Apple GPU 加速，适用于 macOS)
+cargo build --release --features metal
+
+# Flash Attention (更快推理)
+cargo build --release --features flash-attn
+
+# FFmpeg (多媒体处理)
+cargo build --release --features ffmpeg
+
+# 组合多个特性
+cargo build --release --features "cuda,flash-attn"
+```
+
+### CLI 快速参考

-### 快速开始
 ```bash
+
+# 列出所有支持的模型
+aha list
+
+# 仅下载模型
+aha download -m qwen3asr-0.6b
+
 # 下载模型并启动服务
 aha -m qwen3asr-0.6b

 # 直接运行推理（无需启动服务）
-aha run -m qwen3asr-0.6b -i "audio.wav" --weight-path /path/to/model
+aha run -m qwen3asr-0.6b -i "audio.wav"
+
+# 仅启动服务（模型已下载）
+aha serv -m qwen3asr-0.6b -p 10100

-# 仅下载模型
-aha download -m qwen3asr-0.6b
 ```
-### 主要功能
-* 模型管理 - 自动下载、本地模型加载
-* 服务模式 - 启动 HTTP API 服务
-* 直接推理 - 命令行直接执行推理任务
-* 多模型支持 - 支持aha已实现的模型类型

-### 详细使用说明
-请参阅 [CLI_USAGE.md](./docs/CLI_USAGE.md) 获取完整的命令行工具使用说明、参数详解和支持的模型列表。
+### 对话

-## 安装及使用
-
-### 从源码构建部署
 ```bash
-git clone https://github.com/jhqxxx/aha.git
-cd aha
+aha serv -m qwen3-0.6b -p 10100
 ```

-#### cargo run 运行参数说明
-##### 基本用法
+然后使用统一(兼容 OpenAI)的 API：
+
 ```bash
-cargo run -F cuda -r -- [参数]
+curl http://localhost:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [{"role": "user", "content": "你好！"}]
+  }'
 ```
-##### 参数详解
-1. 端口设置
-----
-    -p, --port <PORT>
-* 设置HTTP服务监听的端口号
-* 默认值：10100
-* 示例：--port 8080 或 -p 8080

-2. 模型选择（必选）
-----
-    -m, --model <MODEL>
-* 指定要加载的模型类型
-* 可选值：
-    * minicpm4-0.5b：OpenBMB/MiniCPM4-0.5B 模型
-    * qwen2.5vl-3b：Qwen/Qwen2.5-VL-3B-Instruct 模型
-    * qwen2.5vl-7b：Qwen/Qwen2.5-VL-7B-Instruct 模型
-    * qwen3-0.6b: Qwen/Qwen3-0.6B 模型
-    * qwen3asr-0.6b: Qwen/Qwen3-ASR-0.6B 模型
-    * qwen3asr-1.7b: Qwen/Qwen3-ASR-1.7B 模型
-    * qwen3vl-2b：Qwen/Qwen3-VL-2B-Instruct 模型
-    * qwen3vl-4b：Qwen/Qwen3-VL-4B-Instruct 模型
-    * qwen3vl-8b：Qwen/Qwen3-VL-8B-Instruct 模型
-    * qwen3vl-32b：Qwen/Qwen3-VL-32B-Instruct 模型
-    * deepseek-ocr: deepseek-ai/DeepSeek-OCR 模型
-    * hunyuan-ocr: Tencent-Hunyuan/HunyuanOCR 模型
-    * paddleocr-vl: PaddlePaddle/PaddleOCR-VL 模型
-    * rmbg2.0: AI-ModelScope/RMBG-2.0 模型
-    * voxcpm: OpenBMB/VoxCPM-0.5B 模型
-    * voxcpm1.5: OpenBMB/VoxCPM1.5 模型
-    * glm-asr-nano-2512: ZhipuAI/GLM-ASR-Nano-2512 模型
-    * fun-asr-nano-2512: FunAudioLLM/Fun-ASR-Nano-2512 模型
-* 示例：--model deepseek-ocr 或 -m qwen3vl-2b

-3. 权重路径
-----
-    --weight-path <WEIGHT_PATH>
-* 指定本地模型权重文件路径
-* 如果指定此参数，则跳过模型下载步骤
-* 示例：--weight-path /path/to/model/dir
+### 支持的模型

-4. 保存路径
-----
-    --save-dir <SAVE_DIR>
-* 指定模型下载保存的目录
-* 默认保存在用户主目录下的 .aha 文件夹中
-* 示例：--save-dir /custom/model/path
+| 类别 | 模型 |
+|------|------|
+| **文本** | Qwen3, MiniCPM4 |
+| **视觉** | Qwen2.5-VL, Qwen3-VL |
+| **OCR** | DeepSeek-OCR, Hunyuan-OCR, PaddleOCR-VL |
+| **ASR** | GLM-ASR-Nano, Fun-ASR-Nano,Qwen3-ASR |
+| **音频** | VoxCPM, VoxCPM1.5 |
+| **图像** | RMBG-2.0 (背景移除) |

-5. 下载重试次数
----- 
-    --download-retries <DOWNLOAD_RETRIES>
-* 设置模型下载失败时的最大重试次数
-* 默认值：3次
-* 示例：--download-retries 5
+## 文档

-##### 注意事项
-* 参数前需要使用双横线 -- 分隔 cargo 命令和应用程序参数
-* 模型参数 (--model 或 -m) 是必需的
-* 如果未指定 --weight-path，程序会自动下载指定模型
-* 下载的模型默认保存在 ~/.aha/ 目录下（除非指定了 --save-dir）
+| 文档 | 描述 |
+|------|------|
+| [快速入门](docs/getting-started.zh-CN.md) | aha 入门指南 |
+| [安装指南](docs/installation.zh-CN.md) | 详细安装说明 |
+| [CLI 参考](docs/cli.zh-CN.md) | 命令行界面 |
+| [API 文档](docs/api.zh-CN.md) | 库与 REST API |
+| [支持的模型](docs/supported-models.zh-CN.md) | 可用的 AI 模型 |
+| [核心概念](docs/concepts.zh-CN.md) | 架构与设计 |
+| [开发指南](docs/development.zh-CN.md) | 贡献指南 |
+| [更新日志](docs/changelog.zh-CN.md) | 版本历史 |

-### API接口介绍
-项目提供基于 OpenAI API 兼容的 RESTful 接口，支持多种模型推理任务。
+## 为什么选择 aha？
+- **🚀 高性能推理** - 基于 Candle 框架，提供高效的张量计算和模型推理
+- **🔧 统一接口** — 一个工具搞定文本、视觉、语音和 OCR
+- **📦 本地优先** — 所有处理在本地运行，数据不离境
+- **🎯 跨平台** — 支持 Linux、macOS 和 Windows
+- **⚡ GPU 加速** — 可选 CUDA 支持以获得更快推理
+- **🛡️ 内存安全** — Rust 构建，稳定可靠
+- **🧠 注意力优化** - 可选 Flash Attention 支持，优化长序列处理

-#### 接口列表
-1. 对话接口
- **端点**: `POST /chat/completions`
- **功能**: 多模态对话和文本生成
- **支持模型**: Qwen2.5VL, Qwen3, Qwen3VL, DeepSeekOCR, GLM-ASR-Nano-2512, Fun-ASR-Nano-2512 等
- **请求格式**: OpenAI Chat Completion 格式
- **响应格式**: OpenAI Chat Completion 格式
- **流式支持**: 支持
+## 开发

-2. 图像处理接口
- **端点**: `POST /images/remove_background`
- **功能**: 图像背景移除
- **支持模型**: RMBG-2.0
- **请求格式**: OpenAI Chat Completion 格式
- **响应格式**: OpenAI Chat Completion 格式
- **流式支持**: 不支持
+### aha 作为库使用
+> cargo add aha

-3. 语音生成接口
- **端点**: `POST /audio/speech`
- **功能**: 语音合成和生成
- **支持模型**: VoxCPM,VoxCPM1.5
- **请求格式**: OpenAI Chat Completion 格式
- **响应格式**: OpenAI Chat Completion 格式
- **流式支持**: 不支持
-
-### 作为库使用
-* cargo add aha
-* 或者在Cargo.toml中添加
-```toml
-[dependencies]
-aha = { git = "https://github.com/jhqxxx/aha.git" }
-
-# 启用 CUDA 支持（可选）
-aha = { git = "https://github.com/jhqxxx/aha.git", features = ["cuda"] }
-
-# 启用Flash Attention 支持（可选）
-aha = { git = "https://github.com/jhqxxx/aha.git", features = ["cuda", "flash-attn"] }
-```
-#### VoxCPM示例
-```rust
+```bash
+# VoxCPM示例
 use aha::models::voxcpm::generate::VoxCPMGenerate;
 use aha::utils::audio_utils::save_wav;
 use anyhow::Result;
@@ -262,108 +188,43 @@ fn main() -> Result<()> {
 }
 ```

-### 从源码构建运行测试
-```bash
-git clone https://github.com/jhqxxx/aha.git
-cd aha
-# 修改测试用例中模型路径
-# 运行 PaddleOCR-Vl 示例
-cargo test -F cuda paddleocr_vl_generate -r -- --nocapture

-# 运行 Hunyuan-OCR 示例
-cargo test -F cuda hunyuan_ocr_generate -r -- --nocapture
+### 扩展新的模型 

-# 运行 DeepSeek-OCR 示例
-cargo test -F cuda deepseek_ocr_generate -r -- --nocapture
+- 在src/models/创建新模型文件
+- 在src/models/mod.rs中导出
+- 在src/exec/中添加支持cli运行模型推理
+- 在tests/中添加测试和示例

-# 运行 Qwen3VL 示例
-cargo test -F cuda qwen3vl_generate -r -- --nocapture

-# 运行 MiniCPM4 示例  
-cargo test -F cuda minicpm_generate -r -- --nocapture
+## 特性

-# 运行 VoxCPM 示例
-cargo test -F cuda voxcpm_generate -r -- --nocapture
-```
-
-## 开发
-### 项目结构
-```text
-.
-├── Cargo.toml
-├── README.md
-├── src
-│   ├── chat_template
-│   ├── models
-│   │   ├── common
-│   │   ├── deepseek_ocr
-│   │   ├── hunyuan_ocr
-│   │   ├── minicpm4
-│   │   ├── paddleocr_vl
-│   │   ├── qwen2_5vl
-│   │   ├── qwen3vl
-│   │   ├── voxcpm
-│   │   └── mod.rs
-│   ├── position_embed
-│   ├── tokenizer
-│   ├── utils
-│   ├── api.rs
-│   └── lib.rs
-└── tests
-    ├── test_hunyuan_ocr.rs
-    ├── test_deepseek_ocr.rs
-    ├── test_minicpm4.rs
-    ├── test_paddleocr_vl.rs
-    ├── test_qwen2_5vl.rs
-    └── test_voxcpm.rs
-```
-
-### 添加新模型
-* 在src/models/创建新模型文件
-* 在src/models/mod.rs中导出
-* 在tests/中添加测试和示例
+- 基于 Candle 框架的高性能推理
+- 多模态模型支持（视觉、语言、语音）
+- 简洁易用的 API 设计
+- 最小化依赖，紧凑的二进制文件
+- Flash Attention 支持长序列处理
+- FFmpeg 支持多媒体处理

 ## 许可证
-本项目采用 Apache License, Version 2.0 许可证 - 查看 [LICENSE](./LICENSE) 文件了解详情。
+
+Apache-2.0 &mdash; 详见 [LICENSE](LICENSE)

 ## 致谢
-* [Candle](https://github.com/huggingface/candle) - 优秀的 Rust 机器学习框架
-* 所有模型的原作者和贡献者

-## 支持
-#### 如果你遇到问题：
-1. 查看 Issues 是否已有解决方案
-2. 提交新的 Issue，包含详细描述和复现步骤
+- [Candle](https://github.com/huggingface/candle) - 优秀的 Rust 机器学习框架
+- 所有模型作者和贡献者

-## 更新日志
-### [Unreleased] - 2025-02-04
-* 支持Qwen3-ASR 模型
-### v0.1.8
-* 支持Fun-ASR-Nano-2512, Qwen3 模型
+---

-### v0.1.7
-* 支持GLM-ASR-Nano-2512 模型
-
-### v0.1.6
-* 支持RMGB2.0 模型
-
-### v0.1.5
-* 支持VoxCPM1.5 模型
-
-### v0.1.4
-* 添加PaddleOCR-VL 模型
-
-### v0.1.3
-* 添加 Hunyuan-OCR 模型
-
-### v0.1.2
-* 添加 DeepSeek-OCR 模型
-
-### v0.1.1
-* 添加 Qwen3VL 模型
-
-### v0.1.0
-* 初始版本发布
-* 支持 Qwen2.5VL, MiniCPM4, VoxCPM 模型
+<p align="center">
+  <sub>由 aha 团队用 ❤️ 构建</sub>
+</p>

+<p align="center">
+  <sub>我们持续扩展支持的模型列表，欢迎贡献！</sub>
+</p>

+<p align="center">
+  <sub>如果这个项目对你有帮助，请给我们一个 ⭐ Star！</sub>
+</p>
@@ -0,0 +1,461 @@
+# API Reference
+
+Complete reference for the AHA REST API.
+
+## Overview
+
+AHA provides an OpenAI-compatible REST API for running AI model inference. All endpoints follow standard HTTP conventions and return JSON responses.
+
+### Base URL
+
+By default, the API server runs on:
+```
+http://127.0.0.1:10100
+```
+
+You can customize this when starting the service:
+```bash
+aha cli -m qwen3-0.6b -a 0.0.0.0 -p 8080
+```
+
+### Authentication
+
+Currently, AHA does not require authentication. All endpoints are publicly accessible on the configured address/port.
+
+**Security Note**: If you expose the API to external networks, consider implementing authentication through a reverse proxy (e.g., nginx, traefik).
+
+### Content Types
+
+All requests should use:
+```
+Content-Type: application/json
+```
+
+### Response Format
+
+Success responses follow this structure:
+```json
+{
+  "data": { ... },
+  "model": "model-name",
+  "usage": {
+    "prompt_tokens": 10,
+    "completion_tokens": 20,
+    "total_tokens": 30
+  }
+}
+```
+
+Error responses:
+```json
+{
+  "error": {
+    "message": "Error description",
+    "type": "error_type",
+    "code": "error_code"
+  }
+}
+```
+
+## Endpoints
+
+### Chat Completions
+
+Generate chat completions or text responses.
+
+#### Endpoint
+```
+POST /chat/completions
+```
+
+#### Request Body
+
+| Parameter | Type | Required | Description |
+|-----------|------|----------|-------------|
+| `model` | string | Yes | Model identifier (e.g., "qwen3-0.6b") |
+| `messages` | array | Yes | Array of message objects |
+| `temperature` | number | No | Sampling temperature (0-2, default: 1) |
+| `top_p` | number | No | Nucleus sampling (0-1, default: 1) |
+| `max_tokens` | integer | No | Maximum tokens to generate |
+| `stream` | boolean | No | Enable streaming (default: false) |
+
+#### Message Object
+
+| Field | Type | Required | Description |
+|-------|------|----------|-------------|
+| `role` | string | Yes | "system", "user", or "assistant" |
+| `content` | string/array | Yes | Message content (string or multimodal array) |
+
+#### Multimodal Content
+
+For vision/audio models, content can be an array:
+
+```json
+{
+  "role": "user",
+  "content": [
+    {"type": "text", "text": "Describe this image"},
+    {"type": "image_url", "image_url": {"url": "file:///path/to/image.jpg"}}
+  ]
+}
+```
+
+Supported content types:
+- `text` - Text content
+- `image_url` - Image file (file://, base64://, or http://)
+- `audio_url` - Audio file (file:// or base64://)
+
+#### Examples
+
+**Simple Chat:**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "Hello!"}
+    ]
+  }'
+```
+
+**With System Message:**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "system", "content": "You are a helpful assistant."},
+      {"role": "user", "content": "Explain Rust in one sentence."}
+    ],
+    "max_tokens": 50,
+    "temperature": 0.7
+  }'
+```
+
+**Vision Understanding:**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3vl-2b",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "What is in this image?"},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/image.jpg"}}
+        ]
+      }
+    ]
+  }'
+```
+
+**OCR (Text Extraction):**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "deepseek-ocr",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "Extract all text"},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/document.png"}}
+        ]
+      }
+    ]
+  }'
+```
+
+**ASR (Speech Recognition):**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "glm-asr-nano-2512",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "Transcribe this audio"},
+          {"type": "audio_url", "audio_url": {"url": "file:///path/to/audio.wav"}}
+        ]
+      }
+    ]
+  }'
+```
+
+**Streaming Response:**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "Tell me a story"}
+    ],
+    "stream": true
+  }'
+```
+
+Streaming responses are sent as Server-Sent Events (SSE):
+```
+data: {"id": "1", "choices": [{"delta": {"content": "Once"}}]}
+
+data: {"id": "1", "choices": [{"delta": {"content": " upon"}}]}
+
+data: [DONE]
+```
+
+#### Response
+
+**Non-streaming:**
+
+```json
+{
+  "id": "chatcmpl-123",
+  "object": "chat.completion",
+  "created": 1677652288,
+  "model": "qwen3-0.6b",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "Hello! How can I help you today?"
+      },
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {
+    "prompt_tokens": 10,
+    "completion_tokens": 9,
+    "total_tokens": 19
+  }
+}
+```
+
+#### Supported Models
+
+- Text: `qwen3-0.6b`, `minicpm4-0.5b`
+- Vision: `qwen2.5vl-3b`, `qwen2.5vl-7b`, `qwen3vl-2b`, `qwen3vl-4b`, `qwen3vl-8b`, `qwen3vl-32b`
+- OCR: `deepseek-ocr`, `hunyuan-ocr`, `paddleocr-vl`
+- ASR: `glm-asr-nano-2512`, `fun-asr-nano-2512`, `qwen3asr-0.6b`, `qwen3asr-1.7b`
+
+### Audio Speech
+
+Generate speech from text (Text-to-Speech).
+
+#### Endpoint
+```
+POST /audio/speech
+```
+
+#### Request Body
+
+| Parameter | Type | Required | Description |
+|-----------|------|----------|-------------|
+| `model` | string | Yes | Model identifier (e.g., "voxcpm1.5") |
+| `input` | string | Yes | Text to convert to speech |
+| `voice` | string | No | Voice selection (default: "default") |
+
+#### Example
+
+```bash
+curl http://127.0.0.1:10100/audio/speech \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "voxcpm1.5",
+    "input": "你好，世界！",
+    "voice": "default"
+  }' \
+  --output speech.wav
+```
+
+#### Response
+
+Returns audio data in WAV format.
+
+#### Supported Models
+
+- `voxcpm`, `voxcpm1.5`
+
+### Images Remove Background
+
+Remove background from images.
+
+#### Endpoint
+```
+POST /images/remove_background
+```
+
+#### Request Body
+
+| Parameter | Type | Required | Description |
+|-----------|------|----------|-------------|
+| `model` | string | Yes | Model identifier (e.g., "rmbg2.0") |
+| `image` | string | Yes | Image file path (file://) or base64 data |
+
+#### Example
+
+**From File:**
+
+```bash
+curl http://127.0.0.1:10100/images/remove_background \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "rmbg2.0",
+    "image": "file:///path/to/photo.png"
+  }' \
+  --output no-background.png
+```
+
+**From Base64:**
+
+```bash
+curl http://127.0.0.1:10100/images/remove_background \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "rmbg2.0",
+    "image": "base64://$(base64 -w 0 photo.png)"
+  }' \
+  --output no-background.png
+```
+
+#### Response
+
+Returns the processed image in PNG format.
+
+#### Supported Models
+
+- `rmbg2.0`
+
+## Error Handling
+
+### Error Codes
+
+| Code | Description |
+|------|-------------|
+| 400 | Bad Request - Invalid parameters |
+| 404 | Not Found - Model or endpoint not found |
+| 500 | Internal Server Error - Model inference error |
+| 503 | Service Unavailable - Model not loaded |
+
+### Error Response Format
+
+```json
+{
+  "error": {
+    "message": "Model 'unknown-model' not found",
+    "type": "invalid_request_error",
+    "code": "model_not_found"
+  }
+}
+```
+
+## Rate Limiting
+
+Currently, AHA does not implement rate limiting. The server can handle concurrent requests limited only by system resources.
+
+## File Upload Limits
+
+- String data: 5 MB
+- File uploads: 100 MB
+
+## OpenAI Compatibility
+
+AHA's API is designed to be compatible with OpenAI's API format. This means you can use existing OpenAI client libraries with minimal changes:
+
+### Python Example
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    base_url="http://127.0.0.1:10100",
+    api_key="dummy"  # Not used but required by library
+)
+
+response = client.chat.completions.create(
+    model="qwen3-0.6b",
+    messages=[
+        {"role": "user", "content": "Hello!"}
+    ]
+)
+
+print(response.choices[0].message.content)
+```
+
+### JavaScript Example
+
+```javascript
+import OpenAI from 'openai';
+
+const client = new OpenAI({
+  baseURL: 'http://127.0.0.1:10100',
+  apiKey: 'dummy'  // Not used but required
+});
+
+const response = await client.chat.completions.create({
+  model: 'qwen3-0.6b',
+  messages: [{ role: 'user', content: 'Hello!' }]
+});
+
+console.log(response.choices[0].message.content);
+```
+
+## Best Practices
+
+### 1. Use Streaming for Long Responses
+
+For long text generation, use streaming to get responses incrementally:
+
+```bash
+curl ... -d '{"stream": true, ...}'
+```
+
+### 2. Set Appropriate Token Limits
+
+Prevent excessively long responses:
+
+```json
+{
+  "max_tokens": 500
+}
+```
+
+### 3. Adjust Temperature
+
+Control response creativity:
+- `0.0-0.3`: Deterministic, focused
+- `0.4-0.7`: Balanced (default: 1.0)
+- `0.8-2.0`: Creative, varied
+
+### 4. Use System Messages
+
+Set behavior with system messages:
+
+```json
+{
+  "messages": [
+    {"role": "system", "content": "You are a technical writer."},
+    {"role": "user", "content": "..."}
+  ]
+}
+```
+
+## See Also
+
+- [Getting Started](./getting-started.md) - Quick start guide
+- [CLI Reference](./cli.md) - Command-line usage
+- [Installation](./installation.md) - Installation guide
+- [Development](./development.md) - Contributing guide
@@ -0,0 +1,461 @@
+# API 参考
+
+AHA REST API 的完整参考。
+
+## 概述
+
+AHA 提供了 OpenAI 兼容的 REST API 用于运行 AI 模型推理。所有端点遵循标准 HTTP 约定并返回 JSON 响应。
+
+### 基础 URL
+
+默认情况下，API 服务器运行在：
+```
+http://127.0.0.1:10100
+```
+
+您可以在启动服务时自定义：
+```bash
+aha cli -m qwen3-0.6b -a 0.0.0.0 -p 8080
+```
+
+### 身份验证
+
+目前，AHA 不需要身份验证。所有端点在配置的地址/端口上公开访问。
+
+**安全提示**：如果您将 API 暴露到外部网络，请考虑通过反向代理（如 nginx、traefik）实现身份验证。
+
+### 内容类型
+
+所有请求应使用：
+```
+Content-Type: application/json
+```
+
+### 响应格式
+
+成功响应遵循此结构：
+```json
+{
+  "data": { ... },
+  "model": "model-name",
+  "usage": {
+    "prompt_tokens": 10,
+    "completion_tokens": 20,
+    "total_tokens": 30
+  }
+}
+```
+
+错误响应：
+```json
+{
+  "error": {
+    "message": "错误描述",
+    "type": "error_type",
+    "code": "error_code"
+  }
+}
+```
+
+## 端点
+
+### 对话补全
+
+生成对话补全或文本响应。
+
+#### 端点
+```
+POST /chat/completions
+```
+
+#### 请求体
+
+| 参数 | 类型 | 必需 | 描述 |
+|------|------|------|------|
+| `model` | string | 是 | 模型标识符（如 "qwen3-0.6b"） |
+| `messages` | array | 是 | 消息对象数组 |
+| `temperature` | number | 否 | 采样温度（0-2，默认：1） |
+| `top_p` | number | 否 | 核采样（0-1，默认：1） |
+| `max_tokens` | integer | 否 | 要生成的最大令牌数 |
+| `stream` | boolean | 否 | 启用流式传输（默认：false） |
+
+#### 消息对象
+
+| 字段 | 类型 | 必需 | 描述 |
+|------|------|------|------|
+| `role` | string | 是 | "system"、"user" 或 "assistant" |
+| `content` | string/array | 是 | 消息内容（字符串或多模态数组） |
+
+#### 多模态内容
+
+对于视觉/音频模型，内容可以是数组：
+
+```json
+{
+  "role": "user",
+  "content": [
+    {"type": "text", "text": "描述这张图片"},
+    {"type": "image_url", "image_url": {"url": "file:///path/to/image.jpg"}}
+  ]
+}
+```
+
+支持的内容类型：
+- `text` - 文本内容
+- `image_url` - 图像文件（file://、base64:// 或 http://）
+- `audio_url` - 音频文件（file:// 或 base64://）
+
+#### 示例
+
+**简单对话：**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "你好！"}
+    ]
+  }'
+```
+
+**带系统消息：**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "system", "content": "你是一个有用的助手。"},
+      {"role": "user", "content": "用一句话解释 Rust。"}
+    ],
+    "max_tokens": 50,
+    "temperature": 0.7
+  }'
+```
+
+**视觉理解：**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3vl-2b",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "这张图片里有什么？"},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/image.jpg"}}
+        ]
+      }
+    ]
+  }'
+```
+
+**OCR（文本提取）：**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "deepseek-ocr",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "提取所有文本"},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/document.png"}}
+        ]
+      }
+    ]
+  }'
+```
+
+**ASR（语音识别）：**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "glm-asr-nano-2512",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "转写这段音频"},
+          {"type": "audio_url", "audio_url": {"url": "file:///path/to/audio.wav"}}
+        ]
+      }
+    ]
+  }'
+```
+
+**流式响应：**
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "给我讲个故事"}
+    ],
+    "stream": true
+  }'
+```
+
+流式响应作为服务器发送事件（SSE）发送：
+```
+data: {"id": "1", "choices": [{"delta": {"content": "从前"}}]}
+
+data: {"id": "1", "choices": [{"delta": {"content": "有"}}]}
+
+data: [DONE]
+```
+
+#### 响应
+
+**非流式：**
+
+```json
+{
+  "id": "chatcmpl-123",
+  "object": "chat.completion",
+  "created": 1677652288,
+  "model": "qwen3-0.6b",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "你好！今天我能帮你什么？"
+      },
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {
+    "prompt_tokens": 10,
+    "completion_tokens": 9,
+    "total_tokens": 19
+  }
+}
+```
+
+#### 支持的模型
+
+- 文本：`qwen3-0.6b`、`minicpm4-0.5b`
+- 视觉：`qwen2.5vl-3b`、`qwen2.5vl-7b`、`qwen3vl-2b`、`qwen3vl-4b`、`qwen3vl-8b`、`qwen3vl-32b`
+- OCR：`deepseek-ocr`、`hunyuan-ocr`、`paddleocr-vl`
+- ASR：`glm-asr-nano-2512`、`fun-asr-nano-2512`、`qwen3asr-0.6b`、`qwen3asr-1.7b`
+
+### 语音生成
+
+从文本生成语音（文本转语音）。
+
+#### 端点
+```
+POST /audio/speech
+```
+
+#### 请求体
+
+| 参数 | 类型 | 必需 | 描述 |
+|------|------|------|------|
+| `model` | string | 是 | 模型标识符（如 "voxcpm1.5"） |
+| `input` | string | 是 | 要转换为语音的文本 |
+| `voice` | string | 否 | 语音选择（默认："default"） |
+
+#### 示例
+
+```bash
+curl http://127.0.0.1:10100/audio/speech \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "voxcpm1.5",
+    "input": "你好，世界！",
+    "voice": "default"
+  }' \
+  --output speech.wav
+```
+
+#### 响应
+
+以 WAV 格式返回音频数据。
+
+#### 支持的模型
+
+- `voxcpm`、`voxcpm1.5`
+
+### 图像背景移除
+
+从图像中移除背景。
+
+#### 端点
+```
+POST /images/remove_background
+```
+
+#### 请求体
+
+| 参数 | 类型 | 必需 | 描述 |
+|------|------|------|------|
+| `model` | string | 是 | 模型标识符（如 "rmbg2.0"） |
+| `image` | string | 是 | 图像文件路径（file://）或 base64 数据 |
+
+#### 示例
+
+**从文件：**
+
+```bash
+curl http://127.0.0.1:10100/images/remove_background \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "rmbg2.0",
+    "image": "file:///path/to/photo.png"
+  }' \
+  --output no-background.png
+```
+
+**从 Base64：**
+
+```bash
+curl http://127.0.0.1:10100/images/remove_background \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "rmbg2.0",
+    "image": "base64://$(base64 -w 0 photo.png)"
+  }' \
+  --output no-background.png
+```
+
+#### 响应
+
+以 PNG 格式返回处理后的图像。
+
+#### 支持的模型
+
+- `rmbg2.0`
+
+## 错误处理
+
+### 错误代码
+
+| 代码 | 描述 |
+|------|------|
+| 400 | 错误请求 - 无效参数 |
+| 404 | 未找到 - 模型或端点未找到 |
+| 500 | 内部服务器错误 - 模型推理错误 |
+| 503 | 服务不可用 - 模型未加载 |
+
+### 错误响应格式
+
+```json
+{
+  "error": {
+    "message": "未找到模型 'unknown-model'",
+    "type": "invalid_request_error",
+    "code": "model_not_found"
+  }
+}
+```
+
+## 速率限制
+
+目前，AHA 不实现速率限制。服务器可以处理并发请求，仅受系统资源限制。
+
+## 文件上传限制
+
+- 字符串数据：5 MB
+- 文件上传：100 MB
+
+## OpenAI 兼容性
+
+AHA 的 API 设计为与 OpenAI 的 API 格式兼容。这意味着您可以使用现有的 OpenAI 客户端库，只需最少的更改：
+
+### Python 示例
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    base_url="http://127.0.0.1:10100",
+    api_key="dummy"  # 不使用但库需要
+)
+
+response = client.chat.completions.create(
+    model="qwen3-0.6b",
+    messages=[
+        {"role": "user", "content": "你好！"}
+    ]
+)
+
+print(response.choices[0].message.content)
+```
+
+### JavaScript 示例
+
+```javascript
+import OpenAI from 'openai';
+
+const client = new OpenAI({
+  baseURL: 'http://127.0.0.1:10100',
+  apiKey: 'dummy'  // 不使用但需要
+});
+
+const response = await client.chat.completions.create({
+  model: 'qwen3-0.6b',
+  messages: [{ role: 'user', content: '你好！' }]
+});
+
+console.log(response.choices[0].message.content);
+```
+
+## 最佳实践
+
+### 1. 对长响应使用流式传输
+
+对于长文本生成，使用流式传输增量获取响应：
+
+```bash
+curl ... -d '{"stream": true, ...}'
+```
+
+### 2. 设置适当的令牌限制
+
+防止过长的响应：
+
+```json
+{
+  "max_tokens": 500
+}
+```
+
+### 3. 调整温度
+
+控制响应创造性：
+- `0.0-0.3`：确定性、专注
+- `0.4-0.7`：平衡（默认：1.0）
+- `0.8-2.0`：创造性、多样
+
+### 4. 使用系统消息
+
+使用系统消息设置行为：
+
+```json
+{
+  "messages": [
+    {"role": "system", "content": "你是一个技术写作助手。"},
+    {"role": "user", "content": "..."}
+  ]
+}
+```
+
+## 另见
+
+- [快速入门](./getting-started.zh-CN.md) - 快速入门指南
+- [CLI 参考](./cli.zh-CN.md) - 命令行使用
+- [安装](./installation.zh-CN.md) - 安装指南
+- [开发](./development.zh-CN.md) - 贡献指南
@@ -0,0 +1,90 @@
+# Changelog
+
+All notable changes to aha will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [0.2.0] - 2026-02-05
+
+### Added
+- Qwen3-ASR speech recognition model
+
+## [0.1.9] - 2026-01-31
+
+### Added
+- CLI `list` subcommand to show supported models
+- CLI subcommand structure support (`cli`, `serv`, `download`, `run`)
+- Direct model inference via new `run` subcommand
+
+### Fixed
+- Qwen3VL thinking startswith bug
+- `aha run` multiple inputs bug
+
+## [0.1.8] - 2026-01-17
+
+### Added
+- Qwen3 text model support
+- Fun-ASR-Nano-2512 speech recognition model
+
+### Fixed
+- ModelScope Fun-ASR-Nano model load error
+
+### Changed
+- Updated audio resampling with rubato
+
+## [0.1.7] - 2026-01-07
+
+### Added
+- GLM-ASR-Nano-2512 speech recognition model
+- Metal (GPU) support for Apple Silicon
+- Dynamic home directory and model download script
+
+## [0.1.6] - 2025-12-23
+
+### Added
+- RMBG-2.0 background removal model
+- Image and audio API endpoints
+
+### Changed
+- Performance optimizations for RMBG2.0 image processing
+
+## [0.1.5] - 2025-12-11
+
+### Added
+- VoxCPM1.5 voice generation model
+- PaddleOCR-VL text recognition model
+
+## [0.1.4] - 2025-12-09
+
+### Added
+- PaddleOCR-VL model support
+- FFmpeg feature for multimedia processing
+
+## [0.1.3] - 2025-12-03
+
+### Added
+- Hunyuan-OCR model support
+
+## [0.1.2] - 2025-11-23
+
+### Added
+- DeepSeek-OCR model support
+
+## [0.1.1] - 2025-11-12
+
+### Added
+- Qwen3-VL models (2B, 4B, 8B, 32B)
+
+### Fixed
+- Added serde default for tie_word_embeddings in Qwen3VL
+
+## [0.1.0] - 2025-10-10
+
+### Added
+- Initial release
+- Qwen2.5-VL model support
+- MiniCPM4 model support
+- VoxCPM voice generation model
+- OpenAI-compatible REST API
+- CLI interface for all model types
@@ -0,0 +1,90 @@
+# 更新日志
+
+所有 aha 的重大更改都将记录在此文件中。
+
+格式基于 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.0.0/)，
+本项目遵循 [语义化版本](https://semver.org/lang/zh-CN/spec/v2.0.0.html)。
+
+## [0.2.0] - 2026-02-05
+
+### 新增
+- Qwen3-ASR 语音识别模型
+
+## [0.1.9] - 2026-01-31
+
+### 新增
+- CLI `list` 子命令，显示支持的模型
+- CLI 子命令结构支持（`cli`、`serv`、`download`、`run`）
+- 通过新的 `run` 子命令直接进行模型推理
+
+### 修复
+- Qwen3VL thinking startswith bug
+- `aha run` 多输入 bug
+
+## [0.1.8] - 2026-01-17
+
+### 新增
+- Qwen3 文本模型支持
+- Fun-ASR-Nano-2512 语音识别模型
+
+### 修复
+- ModelScope Fun-ASR-Nano 模型加载错误
+
+### 变更
+- 使用 rubato 更新音频重采样
+
+## [0.1.7] - 2026-01-07
+
+### 新增
+- GLM-ASR-Nano-2512 语音识别模型
+- Metal (GPU) 支持，适用于 Apple Silicon
+- 动态主目录和模型下载脚本
+
+## [0.1.6] - 2025-12-23
+
+### 新增
+- RMBG-2.0 背景移除模型
+- 图像和音频 API 端点
+
+### 变更
+- RMBG2.0 图像处理性能优化
+
+## [0.1.5] - 2025-12-11
+
+### 新增
+- VoxCPM1.5 语音生成模型
+- PaddleOCR-VL 文字识别模型
+
+## [0.1.4] - 2025-12-09
+
+### 新增
+- PaddleOCR-VL 模型支持
+- FFmpeg 多媒体处理功能
+
+## [0.1.3] - 2025-12-03
+
+### 新增
+- Hunyuan-OCR 模型支持
+
+## [0.1.2] - 2025-11-23
+
+### 新增
+- DeepSeek-OCR 模型支持
+
+## [0.1.1] - 2025-11-12
+
+### 新增
+- Qwen3-VL 系列模型 (2B, 4B, 8B, 32B)
+
+### 修复
+- 为 Qwen3VL 的 tie_word_embeddings 添加 serde 默认值
+
+## [0.1.0] - 2025-10-10
+
+### 新增
+- 初始版本发布
+- Qwen2.5-VL 模型支持
+- MiniCPM4 模型支持
+- VoxCPM 语音生成模型
+- 兼容 OpenAI 的 REST API
+- 所有模型类型的 CLI 界面
@@ -0,0 +1,300 @@
+# CLI Reference
+
+Complete command-line interface reference for aha.
+
+AHA is a high-performance model inference library based on the Candle framework, supporting various multimodal models including vision, language, and audio models.
+
+```bash
+aha [COMMAND] [OPTIONS]
+```
+
+## Global Options
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `-a, --address <ADDRESS>` | Service listen address | 127.0.0.1 |
+| `-p, --port <PORT>` | Service listen port | 10100 |
+| `-m, --model <MODEL>` | Model type (required) | - |
+| `--weight-path <WEIGHT_PATH>` | Local model weight path | - |
+| `--save-dir <SAVE_DIR>` | Model download save directory | ~/.aha/ |
+| `--download-retries <DOWNLOAD_RETRIES>` | Download retry count | 3 |
+| `-h, --help` | Display help information | - |
+| `-V, --version` | Display version number | - |
+
+## Commands
+
+### cli - Download model and start service (default)
+
+Download the specified model and start an HTTP service. This command is used by default when no subcommand is specified.
+
+**Syntax:**
+```bash
+aha cli [OPTIONS] --model <MODEL>
+```
+
+**Options:**
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `-a, --address <ADDRESS>` | Service listen address | 127.0.0.1 |
+| `-p, --port <PORT>` | Service listen port | 10100 |
+| `-m, --model <MODEL>` | Model type (required) | - |
+| `--weight-path <WEIGHT_PATH>` | Local model weight path (skip download if specified) | - |
+| `--save-dir <SAVE_DIR>` | Model download save directory | ~/.aha/ |
+| `--download-retries <DOWNLOAD_RETRIES>` | Download retry count | 3 |
+
+**Examples:**
+
+```bash
+# Download model and start service (default port 10100)
+aha cli -m qwen3vl-2b
+
+# Specify port and save directory
+aha cli -m qwen3vl-2b -p 8080 --save-dir /data/models
+
+# Use local model (skip download)
+aha cli -m qwen3vl-2b --weight-path /path/to/model
+
+# Backward compatible way (equivalent to cli subcommand)
+aha -m qwen3vl-2b
+```
+
+### run - Direct model inference
+
+Run model inference directly without starting an HTTP service. Suitable for one-time inference tasks or batch processing.
+
+**Syntax:**
+```bash
+aha run [OPTIONS] --model <MODEL> --input <INPUT> [--input <INPUT2>] --weight-path <WEIGHT_PATH>
+```
+
+**Options:**
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `-m, --model <MODEL>` | Model type (required) | - |
+| `-i, --input <INPUT>` | Input text or file path (model-specific interpretation, supports 1-2 parameters: input1: prompt text, input2: file path) | - |
+| `-o, --output <OUTPUT>` | Output file path (optional, auto-generated if not specified) | - |
+| `--weight-path <WEIGHT_PATH>` | Local model weight path (required) | - |
+
+**Examples:**
+
+```bash
+# VoxCPM1.5 text-to-speech (single input)
+aha run -m voxcpm1.5 -i "太阳当空照" -o output.wav --weight-path /path/to/model
+
+# VoxCPM1.5 read input from file (single input)
+aha run -m voxcpm1.5 -i "file://./input.txt" --weight-path /path/to/model
+
+# MiniCPM4 text generation (single input)
+aha run -m minicpm4-0.5b -i "你好" --weight-path /path/to/model
+
+# DeepSeek OCR image recognition (single input)
+aha run -m deepseek-ocr -i "image.jpg" --weight-path /path/to/model
+
+# RMBG2.0 background removal (single input)
+aha run -m RMBG2.0 -i "photo.png" -o "no_bg.png" --weight-path /path/to/model
+
+# GLM-ASR speech recognition (two inputs: prompt text + audio file)
+aha run -m glm-asr-nano-2512 -i "请转写这段音频" -i "audio.wav" --weight-path /path/to/model
+
+# Fun-ASR speech recognition (two inputs: prompt text + audio file)
+aha run -m fun-asr-nano-2512 -i "语音转写：" -i "audio.wav" --weight-path /path/to/model
+
+# qwen3 text generation (single input)
+aha run -m qwen3-0.6b -i "你好" --weight-path /path/to/model
+
+# qwen2.5vl image understanding (two inputs: prompt text + image file)
+aha run -m qwen2.5vl-3b -i "请分析图片并提取所有可见文本内容，按从左到右、从上到下的布局，返回纯文本" -i "image.jpg" --weight-path /path/to/model
+
+# Qwen3-ASR speech recognition (single input: audio file)
+aha run -m qwen3asr-0.6b -i "audio.wav" --weight-path /path/to/model
+```
+
+### serv - Start service
+
+Start HTTP service only, without downloading models. Must specify local model path via `--weight-path`.
+
+**Syntax:**
+```bash
+aha serv [OPTIONS] --model <MODEL> --weight-path <WEIGHT_PATH>
+```
+
+**Options:**
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `-a, --address <ADDRESS>` | Service listen address | 127.0.0.1 |
+| `-p, --port <PORT>` | Service listen port | 10100 |
+| `-m, --model <MODEL>` | Model type (required) | - |
+| `--weight-path <WEIGHT_PATH>` | Local model weight path (required) | - |
+
+**Examples:**
+
+```bash
+# Start service with local model
+aha serv -m qwen3vl-2b --weight-path /path/to/model
+
+# Start with specified port
+aha serv -m qwen3vl-2b --weight-path /path/to/model -p 8080
+
+# Specify listen address
+aha serv -m qwen3vl-2b --weight-path /path/to/model -a 0.0.0.0
+```
+
+### download - Download model
+
+Download the specified model only, without starting the service.
+
+**Syntax:**
+```bash
+aha download [OPTIONS] --model <MODEL>
+```
+
+**Options:**
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `-m, --model <MODEL>` | Model type (required) | - |
+| `-s, --save-dir <SAVE_DIR>` | Model download save directory | ~/.aha/ |
+| `--download-retries <DOWNLOAD_RETRIES>` | Download retry count | 3 |
+
+**Examples:**
+
+```bash
+# Download model to default directory
+aha download -m qwen3vl-2b
+
+# Specify save directory
+aha download -m qwen3vl-2b -s /data/models
+
+# Specify download retry count
+aha download -m qwen3vl-2b --download-retries 5
+
+# Download MiniCPM4-0.5B model
+aha download -m minicpm4-0.5b -s models
+```
+
+## Supported Models
+
+| Model ID | Model Name | Description |
+|----------|------------|-------------|
+| `minicpm4-0.5b` | OpenBMB/MiniCPM4-0.5B | OpenBMB MiniCPM4 0.5B model |
+| `qwen2.5vl-3b` | Qwen/Qwen2.5-VL-3B-Instruct | Qwen 2.5 VL 3B model |
+| `qwen2.5vl-7b` | Qwen/Qwen2.5-VL-7B-Instruct | Qwen 2.5 VL 7B model |
+| `qwen3-0.6b` | Qwen/Qwen3-0.6B | Qwen 3 0.6B model |
+| `qwen3vl-2b` | Qwen/Qwen3-VL-2B-Instruct | Qwen 3 VL 2B model |
+| `qwen3vl-4b` | Qwen/Qwen3-VL-4B-Instruct | Qwen 3 VL 4B model |
+| `qwen3vl-8b` | Qwen/Qwen3-VL-8B-Instruct | Qwen 3 VL 8B model |
+| `qwen3vl-32b` | Qwen/Qwen3-VL-32B-Instruct | Qwen 3 VL 32B model |
+| `deepseek-ocr` | deepseek-ai/DeepSeek-OCR | DeepSeek OCR model |
+| `hunyuan-ocr` | Tencent-Hunyuan/HunyuanOCR | Tencent Hunyuan OCR model |
+| `paddleocr-vl` | PaddlePaddle/PaddleOCR-VL | Baidu PaddleOCR VL model |
+| `RMBG2.0` | AI-ModelScope/RMBG-2.0 | RMBG 2.0 background removal model |
+| `voxcpm` | OpenBMB/VoxCPM-0.5B | OpenBMB VoxCPM 0.5B speech synthesis model |
+| `voxcpm1.5` | OpenBMB/VoxCPM1.5 | OpenBMB VoxCPM 1.5 speech synthesis model |
+| `glm-asr-nano-2512` | ZhipuAI/GLM-ASR-Nano-2512 | Zhipu AI ASR Nano 2512 speech recognition model |
+| `fun-asr-nano-2512` | FunAudioLLM/Fun-ASR-Nano-2512 | FunAudioLLM ASR Nano 2512 speech recognition model |
+
+## Common Use Cases
+
+### Scenario 1: Quick start inference service
+
+```bash
+# One command to download and start service
+aha -m qwen3vl-2b
+```
+
+### Scenario 2: Start service with existing model
+
+```bash
+# Assuming model is downloaded to /data/models/Qwen/Qwen3-VL-2B-Instruct
+aha serv -m qwen3vl-2b --weight-path /data/models/Qwen/Qwen3-VL-2B-Instruct
+```
+
+### Scenario 3: Pre-download model
+
+```bash
+# Download model to specified directory for later use
+aha download -m qwen3vl-2b -s /data/models
+
+# Later start with local model
+aha serv -m qwen3vl-2b --weight-path /data/models/Qwen/Qwen3-VL-2B-Instruct
+```
+
+### Scenario 4: Custom service port and address
+
+```bash
+# Start service on 0.0.0.0:8080, allow external access
+aha -m qwen3vl-2b -a 0.0.0.0 -p 8080
+```
+
+## API Endpoints
+
+After the service starts, the following API endpoints are available:
+
+### Chat Completion Endpoint
+- **Endpoint**: `POST /chat/completions`
+- **Function**: Multimodal chat and text generation
+- **Supported Models**: Qwen2.5VL, Qwen3, Qwen3VL, DeepSeekOCR, GLM-ASR-Nano-2512, Fun-ASR-Nano-2512, etc.
+- **Format**: OpenAI Chat Completion format
+- **Streaming Support**: Yes
+
+### Image Processing Endpoint
+- **Endpoint**: `POST /images/remove_background`
+- **Function**: Image background removal
+- **Supported Models**: RMBG-2.0
+- **Format**: OpenAI Chat Completion format
+- **Streaming Support**: No
+
+### Audio Generation Endpoint
+- **Endpoint**: `POST /audio/speech`
+- **Function**: Speech synthesis and generation
+- **Supported Models**: VoxCPM, VoxCPM1.5
+- **Format**: OpenAI Chat Completion format
+- **Streaming Support**: No
+
+## Backward Compatibility
+
+To maintain compatibility with older versions, the following two usage methods are equivalent:
+
+```bash
+# New way (recommended)
+aha cli -m qwen3vl-2b
+
+# Old way (backward compatible)
+aha -m qwen3vl-2b
+```
+
+## Notes
+
+1. **serv subcommand requires `--weight-path`**: Since the `serv` subcommand does not download models, you must specify the path to an already downloaded model via `--weight-path`.
+
+2. **Download retry mechanism**: By default, retries 3 times, waiting 2 seconds after each failure before retrying. You can adjust the retry count with `--download-retries`.
+
+3. **Default save directory**: Models are saved to `~/.aha/` directory by default, which can be customized via `--save-dir` or `-s` parameter.
+
+4. **Port occupation**: Ensure the specified port is not occupied before starting the service. The default port is 10100.
+
+5. **Permission issues**: If saving to a system directory (such as `/data/models`), ensure you have the corresponding write permissions.
+
+## Getting Help
+
+```bash
+# View main help
+aha --help
+
+# View subcommand help
+aha cli --help
+aha serv --help
+aha download --help
+
+# View version information
+aha --version
+```
+
+## See Also
+
+- [Getting Started](./getting-started.md) - Quick start guide
+- [API Documentation](./api.md) - REST API reference
+- [Supported Models](./supported-tools.md) - Available models
@@ -1,6 +1,6 @@
-# AHA 命令行使用说明
+# CLI 参考

-## 概述
+aha 的完整命令行界面参考。

 AHA 是一个基于 Candle 框架的高性能模型推理库，支持多种多模态模型，包括视觉、语言和语音模型。

@@ -272,7 +272,7 @@ aha -m qwen3vl-2b

 2. **下载重试机制**：默认重试 3 次，每次失败后等待 2 秒再重试。可通过 `--download-retries` 调整重试次数。

-3. **默认保存目录**：模型默认保存到 `~/.aha/` 目录下，可通过 `--save-dir` 或 `-d` 参数自定义。
+3. **默认保存目录**：模型默认保存到 `~/.aha/` 目录下，可通过 `--save-dir` 或 `-s` 参数自定义。

 4. **端口占用**：启动服务前确保指定的端口未被占用，默认端口为 10100。

@@ -292,3 +292,9 @@ aha download --help
 # 查看版本信息
 aha --version
 ```
+
+## 另见
+
+- [快速入门](./getting-started.zh-CN.md) - 快速入门指南
+- [API 文档](./api.zh-CN.md) - REST API 参考
+- [支持的模型](./supported-tools.zh-CN.md) - 可用模型
@@ -0,0 +1,411 @@
+# Architecture & Design
+
+This document provides an in-depth look at the architecture and design principles behind AHA.
+
+## Overview
+
+AHA (High-performance AI inference engine) is a Rust-based library built on the [Candle](https://github.com/huggingface/candle) framework. It provides a unified interface for running multiple state-of-the-art AI models locally, without requiring API keys or cloud services.
+
+### Key Characteristics
+
+- **Local-First**: All inference runs on your machine
+- **Multi-Modal**: Support for text, vision, audio, OCR, and ASR models
+- **Cross-Platform**: Linux, macOS, and Windows support
+- **GPU-Accelerated**: Optional CUDA and Metal support
+- **Memory-Safe**: Built with Rust for safety and performance
+- **OpenAI-Compatible**: Easy integration with existing tools
+
+## Architecture Principles
+
+### 1. Local-First Design
+
+AHA is designed to run entirely on your local machine:
+
+- **No cloud dependencies**: All models are downloaded and run locally
+- **Privacy-preserving**: Your data never leaves your machine
+- **No API keys required**: Once downloaded, models work indefinitely
+- **Offline capable**: Models work without internet connection after download
+
+### 2. Unified Model Interface
+
+All models implement a common `GenerateModel` trait, providing:
+
+- Consistent API across different model types
+- Easy model switching without code changes
+- Streaming response support for real-time outputs
+- Standardized error handling
+
+### 3. Cross-Platform Support
+
+AHA abstracts platform differences:
+
+- **Device abstraction**: Automatic CPU/GPU detection and selection
+- **Precision handling**: Dynamic F32/F16/BF16 selection based on hardware
+- **Path management**: Consistent model storage across platforms
+
+## Core Components
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                         CLI Layer                           │
+│  (main.rs - Command parsing, model download, service mgmt)  │
+└─────────────────────────────────────────────────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                        HTTP API Layer                        │
+│  (api.rs - OpenAI-compatible endpoints, streaming, auth)     │
+└─────────────────────────────────────────────────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                     Model Abstraction Layer                  │
+│           (GenerateModel trait - unified interface)          │
+└─────────────────────────────────────────────────────────────┘
+                              │
+        ┌─────────────────────┼─────────────────────┐
+        │                     │                     │
+┌───────▼────────┐  ┌────────▼─────────┐  ┌───────▼────────┐
+│  Text Models   │  │  Vision Models   │  │  Audio Models  │
+│  - Qwen3       │  │  - Qwen2.5VL     │  │  - VoxCPM      │
+│  - MiniCPM4    │  │  - Qwen3VL       │  │  - VoxCPM1.5   │
+└────────────────┘  └──────────────────┘  └────────────────┘
+        │                     │                     │
+┌───────▼────────┐  ┌────────▼─────────┐  ┌───────▼────────┐
+│   OCR Models   │  │   ASR Models     │  │  Image Models  │
+│  - DeepSeek    │  │  - GLM-ASR       │  │  - RMBG2.0     │
+│  - Hunyuan     │  │  - Fun-ASR       │  │                │
+│  - PaddleOCR   │  │  - Qwen3-ASR     │  │                │
+└────────────────┘  └──────────────────┘  └────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                      Utility Modules                         │
+│  - tokenizer: Tokenization utilities                         │
+│  - chat_template: Chat format handling                       │
+│  - position_embed: Positional embeddings                     │
+│  - utils: Common utilities (audio, image, download)          │
+└─────────────────────────────────────────────────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                   Candle ML Framework                        │
+│  (Tensor operations, model loading, device management)       │
+└─────────────────────────────────────────────────────────────┘
+```
+
+### CLI Layer (`src/main.rs`)
+
+The CLI layer provides command-line interface functionality:
+
+- **Command parsing**: Uses `clap` for argument parsing
+- **Model management**: Automatic download and caching
+- **Service control**: Start/stop HTTP server
+- **Direct inference**: Run models without server
+
+**Available Commands**:
+- `cli` - Download model and start service (default)
+- `serv` - Start service with existing model
+- `download` - Download model only
+- `run` - Direct model inference
+- `list` - List supported models
+
+### HTTP API Layer (`src/api.rs`)
+
+The HTTP API layer provides REST endpoints:
+
+- **OpenAI-compatible**: Matches OpenAI API format
+- **Streaming support**: Real-time response generation
+- **Multi-modal**: Handles text, images, and audio
+- **Thread-safe**: Uses RwLock for concurrent requests
+
+**Endpoints**:
+- `POST /chat/completions` - Chat and text generation
+- `POST /images/remove_background` - Image background removal
+- `POST /audio/speech` - Text-to-speech synthesis
+
+### Model Abstraction Layer
+
+All models implement the `GenerateModel` trait:
+
+```rust
+pub trait GenerateModel {
+    // Initialize model with weights
+    fn init(model_path: &str, config: Option<GenerationConfig>) -> Result<Self>;
+
+    // Generate response
+    fn generate(&mut self, prompt: &str, params: GenerationParams) -> Result<String>;
+
+    // Generate with streaming
+    fn generate_stream(&mut self, prompt: &str, params: GenerationParams)
+        -> Result<Box<dyn Iterator<Item = Result<String>>>>;
+}
+```
+
+This provides:
+- **Polymorphism**: Treat different models uniformly
+- **Extensibility**: Easy to add new models
+- **Type safety**: Compile-time guarantees
+
+### Utility Modules
+
+#### Tokenizer (`src/tokenizer/`)
+
+- Loads tokenizers from model configurations
+- Handles special tokens
+- Manages vocabulary
+
+#### Chat Template (`src/chat_template/`)
+
+- Formats chat messages into model prompts
+- Supports multiple chat formats (ChatML, etc.)
+- Handles system messages and role tags
+
+#### Position Embeddings (`src/position_embed/`)
+
+- Implements positional encoding for transformers
+- Supports RoPE (Rotary Position Embedding)
+- Handles M-RoPE for multimodal models
+
+#### Utils (`src/utils/`)
+
+- `audio_utils.rs` - Audio processing (WAV, MP3)
+- `image_utils.rs` - Image processing (resize, encode/decode)
+- `download.rs` - Model downloading with retry logic
+- `common.rs` - Common utilities and constants
+
+## Design Patterns
+
+### 1. Trait-Based Abstraction
+
+The `GenerateModel` trait provides a unified interface:
+
+```rust
+// All models implement this trait
+impl GenerateModel for Qwen3VL { /* ... */ }
+impl GenerateModel for VoxCPM { /* ... */ }
+impl GenerateModel for DeepSeekOCR { /* ... */ }
+
+// Usage is model-agnostic
+let mut model: Box<dyn GenerateModel> = load_model(model_type)?;
+let result = model.generate(prompt, params)?;
+```
+
+### 2. Factory Pattern
+
+Model loading uses a factory function:
+
+```rust
+pub fn load_model(
+    model_type: &str,
+    model_path: &str,
+    device: &Device,
+) -> Result<Box<dyn GenerateModel>> {
+    match model_type {
+        "qwen3vl-2b" => Ok(Box::new(qwen3vl::generate::Qwen3VLGenerate::init(...)?)),
+        "voxcpm1.5" => Ok(Box::new(voxcpm::generate::VoxCPMGenerate::init(...)?)),
+        // ... other models
+        _ => Err(anyhow!("Unsupported model: {}", model_type)),
+    }
+}
+```
+
+### 3. Command Pattern
+
+CLI subcommands encapsulate different operations:
+
+```rust
+match command {
+    Commands::Cli { model, port, address } => { /* download and serve */ }
+    Commands::Serv { model, weight_path, port } => { /* serve only */ }
+    Commands::Download { model, save_dir } => { /* download only */ }
+    Commands::Run { model, input, weight_path } => { /* direct inference */ }
+    Commands::List => { /* list models */ }
+}
+```
+
+## Model Organization
+
+Each model follows a consistent structure:
+
+```
+src/models/{model_name}/
+├── config.rs       # Model configuration and generation parameters
+├── model.rs        # Core model architecture (layers, attention)
+├── generate.rs     # Inference logic (implements GenerateModel trait)
+├── processor.rs    # Model-specific processing (for complex models)
+└── mod.rs          # Module declaration and exports
+```
+
+### Example: Qwen3VL
+
+```
+src/models/qwen3vl/
+├── config.rs       # Qwen3VLConfig, GenerationConfig
+├── model.rs        # Qwen3VL transformer layers, attention mechanisms
+├── generate.rs     # Qwen3VLGenerate implementation
+├── processor.rs    # Image and text processing for multimodal input
+└── mod.rs          # Exports public API
+```
+
+## Performance Optimizations
+
+### GPU Acceleration
+
+AHA supports GPU acceleration through:
+
+- **CUDA**: For NVIDIA GPUs (Linux, Windows)
+- **Metal**: For Apple Silicon (macOS)
+
+Enable with:
+```bash
+cargo build --features cuda    # NVIDIA GPUs
+cargo build --features metal   # Apple Silicon
+```
+
+### Flash Attention
+
+Flash Attention optimizes long-sequence processing:
+
+- Reduces memory usage
+- Improves inference speed
+- Especially beneficial for vision models
+
+Enable with:
+```bash
+cargo build --features "cuda,flash-attn"
+```
+
+### Memory-Mapped Tensors
+
+Models use memory-mapped files for:
+
+- Faster loading times
+- Reduced memory footprint
+- Concurrent model loading
+
+### Precision Optimization
+
+Dynamic precision selection based on hardware:
+
+- **F32**: Maximum accuracy (CPU-only)
+- **F16**: Balanced performance (GPU)
+- **BF16**: Best for modern GPUs
+
+## Security Considerations
+
+### Local-Only Processing
+
+- No external API calls after model download
+- No telemetry or data collection
+- Full control over your data
+
+### Memory Safety
+
+- Rust's ownership system prevents memory leaks
+- No buffer overflows or use-after-free bugs
+- Thread-safe concurrent operations
+
+### Input Validation
+
+- File size limits (5MB strings, 100MB files)
+- Path validation to prevent directory traversal
+- Type-safe request handling
+
+## Data Flow
+
+### Request Flow
+
+```
+┌─────────┐
+│ Client  │
+└────┬────┘
+     │ HTTP Request
+     ▼
+┌──────────────────────────────────────────────────────────┐
+│  Rocket HTTP Server                                      │
+│  - Route request to endpoint                             │
+│  - Parse request body                                    │
+│  - Extract parameters                                    │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  API Handler (api.rs)                                    │
+│  - Acquire model lock                                    │
+│  - Prepare input (tokenize, process images/audio)        │
+│  - Call model.generate() or generate_stream()           │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  Model Implementation (models/{model}/generate.rs)       │
+│  - Load weights from memory-mapped files                 │
+│  - Run forward pass through Candle tensors              │
+│  - Decode output tokens                                  │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  Candle Framework                                        │
+│  - Execute on CPU or GPU device                          │
+│  - Manage tensor operations                              │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  Response Generation                                     │
+│  - Format response (JSON / streaming)                    │
+│  - Return to client                                      │
+└──────────────────────────────────────────────────────────┘
+```
+
+### Model Loading Flow
+
+```
+User specifies model
+        │
+        ▼
+Check if --weight-path provided
+        │
+    ┌───┴───┐
+    │       │
+   Yes      No
+    │       │
+    ▼       ▼
+Use local    Download from ModelScope
+path         │
+    │        ▼
+    │    Save to ~/.aha/{model}/
+    │        │
+    └───┬────┘
+        ▼
+Load model weights into memory
+        │
+        ▼
+Initialize model (init())
+        │
+        ▼
+Ready for inference
+```
+
+## Extension Points
+
+### Adding a New Model
+
+1. Create model directory under `src/models/`
+2. Implement `GenerateModel` trait
+3. Add model to factory function in `mod.rs`
+4. Add CLI mapping in `main.rs`
+5. Add test case in `tests/`
+
+### Custom Processing
+
+Models can override default processing:
+
+- Custom tokenization
+- Special input/output formats
+- Model-specific optimizations
+
+## See Also
+
+- [Installation Guide](./installation.md) - Setup and installation
+- [Getting Started](./getting-started.md) - Quick start guide
+- [API Reference](./api.md) - REST API documentation
+- [Development](./development.md) - Contributing guide
@@ -0,0 +1,411 @@
+# 架构与设计
+
+本文档深入介绍 AHA 的架构和设计原则。
+
+## 概述
+
+AHA（高性能 AI 推理引擎）是一个基于 Rust 的库，构建在 [Candle](https://github.com/huggingface/candle) 框架之上。它提供了统一的接口，可以在本地运行多个最先进的 AI 模型，无需 API 密钥或云服务。
+
+### 核心特性
+
+- **本地优先**：所有推理都在您的机器上运行
+- **多模态支持**：支持文本、视觉、音频、OCR 和 ASR 模型
+- **跨平台**：支持 Linux、macOS 和 Windows
+- **GPU 加速**：可选的 CUDA 和 Metal 支持
+- **内存安全**：使用 Rust 构建，确保安全性和性能
+- **OpenAI 兼容**：易于与现有工具集成
+
+## 架构原则
+
+### 1. 本地优先设计
+
+AHA 设计为完全在本地运行：
+
+- **无云依赖**：所有模型都在本地下载和运行
+- **隐私保护**：您的数据永远不会离开您的机器
+- **无需 API 密钥**：下载后模型可永久使用
+- **离线可用**：下载后模型无需互联网连接
+
+### 2. 统一模型接口
+
+所有模型都实现通用的 `GenerateModel` trait，提供：
+
+- 不同模型类型之间的一致 API
+- 无需更改代码即可轻松切换模型
+- 支持实时输出的流式响应
+- 标准化的错误处理
+
+### 3. 跨平台支持
+
+AHA 抽象了平台差异：
+
+- **设备抽象**：自动 CPU/GPU 检测和选择
+- **精度处理**：基于硬件动态选择 F32/F16/BF16
+- **路径管理**：跨平台一致的模型存储
+
+## 核心组件
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                         CLI 层                              │
+│  (main.rs - 命令解析、模型下载、服务管理)                     │
+└─────────────────────────────────────────────────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                        HTTP API 层                          │
+│  (api.rs - OpenAI 兼容端点、流式传输、认证)                   │
+└─────────────────────────────────────────────────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                     模型抽象层                              │
+│           (GenerateModel trait - 统一接口)                   │
+└─────────────────────────────────────────────────────────────┘
+                              │
+        ┌─────────────────────┼─────────────────────┐
+        │                     │                     │
+┌───────▼────────┐  ┌────────▼─────────┐  ┌───────▼────────┐
+│   文本模型     │  │   视觉模型       │  │   音频模型     │
+│  - Qwen3       │  │  - Qwen2.5VL     │  │  - VoxCPM      │
+│  - MiniCPM4    │  │  - Qwen3VL       │  │  - VoxCPM1.5   │
+└────────────────┘  └──────────────────┘  └────────────────┘
+        │                     │                     │
+┌───────▼────────┐  ┌────────▼─────────┐  ┌───────▼────────┐
+│   OCR 模型     │  │   ASR 模型       │  │   图像模型     │
+│  - DeepSeek    │  │  - GLM-ASR       │  │  - RMBG2.0     │
+│  - Hunyuan     │  │  - Fun-ASR       │  │                │
+│  - PaddleOCR   │  │  - Qwen3-ASR     │  │                │
+└────────────────┘  └──────────────────┘  └────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                       工具模块                              │
+│  - tokenizer: 分词工具                                       │
+│  - chat_template: 聊天格式处理                               │
+│  - position_embed: 位置编码                                  │
+│  - utils: 通用工具（音频、图像、下载）                        │
+└─────────────────────────────────────────────────────────────┘
+                              │
+┌─────────────────────────────────────────────────────────────┐
+│                    Candle ML 框架                           │
+│  (张量操作、模型加载、设备管理)                               │
+└─────────────────────────────────────────────────────────────┘
+```
+
+### CLI 层 (`src/main.rs`)
+
+CLI 层提供命令行界面功能：
+
+- **命令解析**：使用 `clap` 进行参数解析
+- **模型管理**：自动下载和缓存
+- **服务控制**：启动/停止 HTTP 服务器
+- **直接推理**：无需服务器即可运行模型
+
+**可用命令**：
+- `cli` - 下载模型并启动服务（默认）
+- `serv` - 使用现有模型启动服务
+- `download` - 仅下载模型
+- `run` - 直接模型推理
+- `list` - 列出支持的模型
+
+### HTTP API 层 (`src/api.rs`)
+
+HTTP API 层提供 REST 端点：
+
+- **OpenAI 兼容**：匹配 OpenAI API 格式
+- **流式支持**：实时响应生成
+- **多模态**：处理文本、图像和音频
+- **线程安全**：使用 RwLock 处理并发请求
+
+**端点**：
+- `POST /chat/completions` - 聊天和文本生成
+- `POST /images/remove_background` - 图像背景移除
+- `POST /audio/speech` - 文本转语音合成
+
+### 模型抽象层
+
+所有模型都实现 `GenerateModel` trait：
+
+```rust
+pub trait GenerateModel {
+    // 使用权重初始化模型
+    fn init(model_path: &str, config: Option<GenerationConfig>) -> Result<Self>;
+
+    // 生成响应
+    fn generate(&mut self, prompt: &str, params: GenerationParams) -> Result<String>;
+
+    // 流式生成
+    fn generate_stream(&mut self, prompt: &str, params: GenerationParams)
+        -> Result<Box<dyn Iterator<Item = Result<String>>>>;
+}
+```
+
+这提供了：
+- **多态性**：统一处理不同模型
+- **可扩展性**：易于添加新模型
+- **类型安全**：编译时保证
+
+### 工具模块
+
+#### 分词器 (`src/tokenizer/`)
+
+- 从模型配置加载分词器
+- 处理特殊标记
+- 管理词汇表
+
+#### 聊天模板 (`src/chat_template/`)
+
+- 将聊天消息格式化为模型提示
+- 支持多种聊天格式（ChatML 等）
+- 处理系统消息和角色标签
+
+#### 位置编码 (`src/position_embed/`)
+
+- 为 transformer 实现位置编码
+- 支持 RoPE（旋转位置编码）
+- 处理多模态模型的 M-RoPE
+
+#### 工具 (`src/utils/`)
+
+- `audio_utils.rs` - 音频处理（WAV、MP3）
+- `image_utils.rs` - 图像处理（调整大小、编码/解码）
+- `download.rs` - 带重试逻辑的模型下载
+- `common.rs` - 通用工具和常量
+
+## 设计模式
+
+### 1. 基于 Trait 的抽象
+
+`GenerateModel` trait 提供统一接口：
+
+```rust
+// 所有模型都实现此 trait
+impl GenerateModel for Qwen3VL { /* ... */ }
+impl GenerateModel for VoxCPM { /* ... */ }
+impl GenerateModel for DeepSeekOCR { /* ... */ }
+
+// 使用方式与模型无关
+let mut model: Box<dyn GenerateModel> = load_model(model_type)?;
+let result = model.generate(prompt, params)?;
+```
+
+### 2. 工厂模式
+
+模型加载使用工厂函数：
+
+```rust
+pub fn load_model(
+    model_type: &str,
+    model_path: &str,
+    device: &Device,
+) -> Result<Box<dyn GenerateModel>> {
+    match model_type {
+        "qwen3vl-2b" => Ok(Box::new(qwen3vl::generate::Qwen3VLGenerate::init(...)?)),
+        "voxcpm1.5" => Ok(Box::new(voxcpm::generate::VoxCPMGenerate::init(...)?)),
+        // ... 其他模型
+        _ => Err(anyhow!("不支持的模型: {}", model_type)),
+    }
+}
+```
+
+### 3. 命令模式
+
+CLI 子命令封装不同的操作：
+
+```rust
+match command {
+    Commands::Cli { model, port, address } => { /* 下载并启动服务 */ }
+    Commands::Serv { model, weight_path, port } => { /* 仅启动服务 */ }
+    Commands::Download { model, save_dir } => { /* 仅下载 */ }
+    Commands::Run { model, input, weight_path } => { /* 直接推理 */ }
+    Commands::List => { /* 列出模型 */ }
+}
+```
+
+## 模型组织结构
+
+每个模型都遵循一致的结构：
+
+```
+src/models/{model_name}/
+├── config.rs       # 模型配置和生成参数
+├── model.rs        # 核心模型架构（层、注意力）
+├── generate.rs     # 推理逻辑（实现 GenerateModel trait）
+├── processor.rs    # 模型特定处理（用于复杂模型）
+└── mod.rs          # 模块声明和导出
+```
+
+### 示例：Qwen3VL
+
+```
+src/models/qwen3vl/
+├── config.rs       # Qwen3VLConfig、GenerationConfig
+├── model.rs        # Qwen3VL transformer 层、注意力机制
+├── generate.rs     # Qwen3VLGenerate 实现
+├── processor.rs    # 多模态输入的图像和文本处理
+└── mod.rs          # 导出公共 API
+```
+
+## 性能优化
+
+### GPU 加速
+
+AHA 通过以下方式支持 GPU 加速：
+
+- **CUDA**：用于 NVIDIA GPU（Linux、Windows）
+- **Metal**：用于 Apple Silicon（macOS）
+
+启用方式：
+```bash
+cargo build --features cuda    # NVIDIA GPU
+cargo build --features metal   # Apple Silicon
+```
+
+### Flash Attention
+
+Flash Attention 优化长序列处理：
+
+- 减少内存使用
+- 提高推理速度
+- 对视觉模型特别有益
+
+启用方式：
+```bash
+cargo build --features "cuda,flash-attn"
+```
+
+### 内存映射张量
+
+模型使用内存映射文件：
+
+- 更快的加载时间
+- 减少内存占用
+- 支持并发模型加载
+
+### 精度优化
+
+基于硬件的动态精度选择：
+
+- **F32**：最高精度（仅 CPU）
+- **F16**：平衡性能（GPU）
+- **BF16**：最适合现代 GPU
+
+## 安全考虑
+
+### 本地处理
+
+- 模型下载后无外部 API 调用
+- 无遥测或数据收集
+- 完全控制您的数据
+
+### 内存安全
+
+- Rust 所有权系统防止内存泄漏
+- 无缓冲区溢出或使用后释放错误
+- 线程安全的并发操作
+
+### 输入验证
+
+- 文件大小限制（字符串 5MB，文件 100MB）
+- 路径验证防止目录遍历
+- 类型安全的请求处理
+
+## 数据流
+
+### 请求流程
+
+```
+┌─────────┐
+│ 客户端  │
+└────┬────┘
+     │ HTTP 请求
+     ▼
+┌──────────────────────────────────────────────────────────┐
+│  Rocket HTTP 服务器                                      │
+│  - 将请求路由到端点                                      │
+│  - 解析请求体                                            │
+│  - 提取参数                                              │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  API 处理器 (api.rs)                                    │
+│  - 获取模型锁                                            │
+│  - 准备输入（分词、处理图像/音频）                        │
+│  - 调用 model.generate() 或 generate_stream()           │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  模型实现 (models/{model}/generate.rs)                   │
+│  - 从内存映射文件加载权重                                │
+│  - 通过 Candle 张量运行前向传播                          │
+│  - 解码输出标记                                          │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  Candle 框架                                            │
+│  - 在 CPU 或 GPU 设备上执行                              │
+│  - 管理张量操作                                          │
+└────────────┬─────────────────────────────────────────────┘
+             │
+             ▼
+┌──────────────────────────────────────────────────────────┐
+│  响应生成                                               │
+│  - 格式化响应（JSON / 流式）                             │
+│  - 返回给客户端                                          │
+└──────────────────────────────────────────────────────────┘
+```
+
+### 模型加载流程
+
+```
+用户指定模型
+        │
+        ▼
+检查是否提供了 --weight-path
+        │
+    ┌───┴───┐
+    │       │
+   是       否
+    │       │
+    ▼       ▼
+使用本地    从 ModelScope 下载
+路径         │
+    │        ▼
+    │    保存到 ~/.aha/{model}/
+    │        │
+    └───┬────┘
+        ▼
+将模型权重加载到内存
+        │
+        ▼
+初始化模型 (init())
+        │
+        ▼
+准备就绪，可以进行推理
+```
+
+## 扩展点
+
+### 添加新模型
+
+1. 在 `src/models/` 下创建模型目录
+2. 实现 `GenerateModel` trait
+3. 在 `mod.rs` 的工厂函数中添加模型
+4. 在 `main.rs` 中添加 CLI 映射
+5. 在 `tests/` 中添加测试用例
+
+### 自定义处理
+
+模型可以覆盖默认处理：
+
+- 自定义分词
+- 特殊的输入/输出格式
+- 模型特定的优化
+
+## 另见
+
+- [安装指南](./installation.zh-CN.md) - 设置和安装
+- [快速入门](./getting-started.zh-CN.md) - 快速入门指南
+- [API 参考](./api.zh-CN.md) - REST API 文档
+- [开发指南](./development.zh-CN.md) - 贡献指南
@@ -0,0 +1,601 @@
+# Development Guide
+
+This guide covers contributing to AHA, including development setup, adding new models, and submitting contributions.
+
+## Table of Contents
+
+- [Development Setup](#development-setup)
+- [Project Structure](#project-structure)
+- [Adding a New Model](#adding-a-new-model)
+- [Testing](#testing)
+- [Code Style](#code-style)
+- [Submitting Contributions](#submitting-contributions)
+- [Release Process](#release-process)
+
+## Development Setup
+
+### Prerequisites
+
+- Rust 1.85+ (edition 2024)
+- Git
+- Optional: CUDA Toolkit, Metal, FFmpeg (for feature development)
+
+### Clone and Build
+
+```bash
+# Clone the repository
+git clone https://github.com/jhqxxx/aha.git
+cd aha
+
+# Build in debug mode (faster compilation)
+cargo build
+
+# Build in release mode (better performance)
+cargo build --release
+
+# Build with features
+cargo build --features "cuda,flash-attn"
+cargo build --features "metal"
+cargo build --features "ffmpeg"
+```
+
+### Development Workflow
+
+```bash
+# Run the CLI
+cargo run -- -m qwen3-0.6b
+
+# Run tests
+cargo test
+
+# Run specific test
+cargo test test_qwen3vl_generate
+
+# Run with logging
+RUST_LOG=debug cargo run -- -m qwen3-0.6b
+
+# Check code without building
+cargo check
+
+# Format code
+cargo fmt
+
+# Run linter
+cargo clippy
+```
+
+## Project Structure
+
+```
+aha/
+├── Cargo.toml              # Build configuration
+├── src/
+│   ├── main.rs             # CLI entry point
+│   ├── lib.rs              # Library exports
+│   ├── api.rs              # HTTP API handlers
+│   ├── exec/               # CLI command implementations
+│   │   ├── mod.rs
+│   │   ├── cli.rs
+│   │   ├── serv.rs
+│   │   ├── download.rs
+│   │   └── run.rs
+│   ├── models/             # Model implementations
+│   │   ├── mod.rs          # Model factory and exports
+│   │   ├── common/         # Shared model utilities
+│   │   │   ├── config.rs
+│   │   │   └── mod.rs
+│   │   ├── qwen3vl/        # Qwen3-VL model
+│   │   │   ├── config.rs
+│   │   │   ├── model.rs
+│   │   │   ├── generate.rs
+│   │   │   ├── processor.rs
+│   │   │   └── mod.rs
+│   │   ├── voxcpm/         # VoxCPM model
+│   │   └── ...             # Other models
+│   ├── tokenizer/          # Tokenization utilities
+│   ├── chat_template/      # Chat template handling
+│   ├── position_embed/     # Positional embeddings
+│   └── utils/              # Utility functions
+│       ├── audio_utils.rs
+│       ├── image_utils.rs
+│       ├── download.rs
+│       └── common.rs
+├── tests/                  # Integration tests
+│   ├── test_qwen2_5vl.rs
+│   ├── test_qwen3vl.rs
+│   └── ...
+├── examples/               # Example code
+└── docs/                   # Documentation
+```
+
+## Adding a New Model
+
+This section provides a step-by-step guide for adding a new model to AHA.
+
+### Step 1: Create Model Directory
+
+Create a new directory under `src/models/`:
+
+```bash
+mkdir -p src/models/newmodel
+```
+
+### Step 2: Implement Model Files
+
+Create the following files in `src/models/newmodel/`:
+
+#### config.rs
+
+Define model configuration:
+
+```rust
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct NewModelConfig {
+    pub vocab_size: usize,
+    pub hidden_size: usize,
+    pub num_hidden_layers: usize,
+    pub num_attention_heads: usize,
+    // ... other config fields
+}
+
+impl Default for NewModelConfig {
+    fn default() -> Self {
+        Self {
+            vocab_size: 32000,
+            hidden_size: 2048,
+            num_hidden_layers: 24,
+            num_attention_heads: 32,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct GenerationConfig {
+    pub temperature: f64,
+    pub top_p: f64,
+    pub max_tokens: usize,
+}
+
+impl Default for GenerationConfig {
+    fn default() -> Self {
+        Self {
+            temperature: 1.0,
+            top_p: 1.0,
+            max_tokens: 100,
+        }
+    }
+}
+```
+
+#### model.rs
+
+Implement the model architecture:
+
+```rust
+use candle::{Result, Tensor};
+use candle_nn::{Module, VarBuilder};
+
+use super::config::NewModelConfig;
+
+pub struct NewModel {
+    // Define model components
+    config: NewModelConfig,
+}
+
+impl NewModel {
+    pub fn load(vb: VarBuilder, config: &NewModelConfig) -> Result<Self> {
+        // Load model weights
+        Ok(Self {
+            config: config.clone(),
+        })
+    }
+
+    pub fn forward(&self, input_ids: &Tensor) -> Result<Tensor> {
+        // Implement forward pass
+        Ok(input_ids.clone())
+    }
+}
+```
+
+#### generate.rs
+
+Implement the GenerateModel trait:
+
+```rust
+use std::path::Path;
+use anyhow::Result;
+use candle::Device;
+
+use super::config::{GenerationConfig, NewModelConfig};
+use aha::models::common::GenerateModel;
+
+pub struct NewModelGenerate {
+    // Define generate state
+    model: super::model::NewModel,
+    config: GenerationConfig,
+    device: Device,
+}
+
+impl GenerateModel for NewModelGenerate {
+    type Config = NewModelConfig;
+    type GenerateConfig = GenerationConfig;
+
+    fn init(
+        model_path: &str,
+        config: Option<Self::Config>,
+        generate_config: Option<Self::GenerateConfig>,
+    ) -> Result<Self> {
+        // Load model from disk
+        let device = Device::Cpu;
+        let config = config.unwrap_or_default();
+        let generate_config = generate_config.unwrap_or_default();
+
+        // Load weights
+        let model_path = Path::new(model_path);
+        // ... load model implementation
+
+        Ok(Self {
+            model: todo!(),
+            config: generate_config,
+            device,
+        })
+    }
+
+    fn generate(&mut self, prompt: &str) -> Result<String> {
+        // Tokenize prompt
+        // Run inference
+        // Decode output
+        Ok(prompt.to_string())
+    }
+}
+```
+
+#### processor.rs (optional)
+
+For complex models with multimodal input:
+
+```rust
+use anyhow::Result;
+
+pub struct NewModelProcessor {
+    // Processing state
+}
+
+impl NewModelProcessor {
+    pub fn new() -> Self {
+        Self {}
+    }
+
+    pub fn process_text(&self, text: &str) -> Result<Vec<String>> {
+        // Process text input
+        Ok(vec![text.to_string()])
+    }
+
+    pub fn process_image(&self, image_path: &str) -> Result<Tensor> {
+        // Process image input
+        todo!()
+    }
+}
+```
+
+#### mod.rs
+
+Export the model:
+
+```rust
+mod config;
+mod model;
+mod generate;
+pub mod processor;
+
+pub use generate::NewModelGenerate;
+pub use config::{GenerationConfig, NewModelConfig};
+```
+
+### Step 3: Register Model
+
+Update `src/models/mod.rs`:
+
+```rust
+// Add to imports
+pub mod newmodel;
+
+// Add to WhichModel enum
+#[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
+pub enum WhichModel {
+    // ... existing models
+    NewModel,
+}
+
+// Add to model loading
+pub fn load_model(
+    model_type: &WhichModel,
+    model_path: &str,
+    device: &Device,
+) -> Result<Box<dyn common::GenerateModel>> {
+    match model_type {
+        // ... existing models
+        WhichModel::NewModel => {
+            Ok(Box::new(newmodel::NewModelGenerate::init(model_path, None, None)?))
+        }
+    }
+}
+```
+
+### Step 4: Update CLI
+
+Update `src/main.rs` to include the new model in help text and CLI options if needed.
+
+### Step 5: Add Tests
+
+Create `tests/test_newmodel.rs`:
+
+```rust
+use anyhow::Result;
+
+#[test]
+fn test_newmodel_generate() -> Result<()> {
+    let model_path = "path/to/test/model";
+    let mut model = aha::models::newmodel::NewModelGenerate::init(model_path, None, None)?;
+    
+    let result = model.generate("Test prompt")?;
+    assert!(!result.is_empty());
+    
+    Ok(())
+}
+```
+
+### Step 6: Update Documentation
+
+Update the following files:
+- `README.md` - Add model to supported models list
+- `docs/cli.md` - Add model ID to model list
+- `docs/api.md` - Add model to supported models section
+- `CHANGELOG.md` - Add entry for new model
+
+## Testing
+
+### Unit Tests
+
+```bash
+# Run all tests
+cargo test
+
+# Run specific test
+cargo test test_newmodel_generate
+
+# Run with output
+cargo test -- --nocapture
+
+# Run tests in parallel
+cargo test -- --test-threads=4
+```
+
+### Integration Tests
+
+Integration tests are located in `tests/`:
+
+```bash
+# Run specific integration test
+cargo test --test test_qwen3vl
+```
+
+### Manual Testing
+
+Test the CLI and API manually:
+
+```bash
+# Test CLI
+cargo run -- -m newmodel
+
+# Test API
+cargo run -- -m newmodel -p 8080
+curl http://127.0.0.1:8080/chat/completions -d '{...}'
+```
+
+### Model-Specific Tests
+
+Each model should have tests covering:
+- Model loading
+- Basic inference
+- Edge cases
+- Error handling
+
+## Code Style
+
+### Formatting
+
+```bash
+# Format all code
+cargo fmt
+
+# Check formatting without making changes
+cargo fmt --check
+```
+
+### Linting
+
+```bash
+# Run clippy
+cargo clippy
+
+# Fix clippy warnings
+cargo clippy --fix
+```
+
+### Conventions
+
+- Use 4 spaces for indentation
+- Prefer `Result<T>` over `Option<T>` for errors
+- Use `anyhow::Result` for application errors
+- Document public APIs with rustdoc comments
+- Keep functions focused and small
+- Use meaningful variable names
+
+### Documentation
+
+```rust
+/// Generates text using the NewModel.
+///
+/// # Arguments
+///
+/// * `prompt` - The input prompt text
+/// * `max_tokens` - Maximum tokens to generate
+///
+/// # Returns
+///
+/// Generated text as a String
+///
+/// # Examples
+///
+/// ```no_run
+/// let result = model.generate("Hello")?;
+/// ```
+pub fn generate(&mut self, prompt: &str) -> Result<String> {
+    // Implementation
+}
+```
+
+## Submitting Contributions
+
+### Pull Request Process
+
+1. Fork the repository
+2. Create a feature branch:
+   ```bash
+   git checkout -b feature/new-model
+   ```
+3. Make your changes
+4. Add tests
+5. Update documentation
+6. Ensure all tests pass:
+   ```bash
+   cargo test
+   cargo clippy
+   cargo fmt --check
+   ```
+7. Commit and push:
+   ```bash
+   git commit -m "Add NewModel support"
+   git push origin feature/new-model
+   ```
+8. Create a pull request on GitHub
+
+### Pull Request Checklist
+
+- [ ] Tests added/updated
+- [ ] Documentation updated
+- [ ] CHANGELOG.md updated
+- [ ] Code formatted (`cargo fmt`)
+- [ ] Clippy warnings fixed (`cargo clippy`)
+- [ ] All tests pass (`cargo test`)
+- [ ] Commit messages follow conventions
+
+### Commit Message Conventions
+
+Follow conventional commits:
+
+```
+feat: add NewModel support
+fix: correct tensor dimensions in Qwen3VL
+docs: update installation guide
+test: add integration test for VoxCPM
+refactor: simplify model loading logic
+perf: improve inference speed by 20%
+```
+
+## Release Process
+
+### Version Bump
+
+Update `Cargo.toml`:
+
+```toml
+[package]
+version = "0.1.9"  # Bump version
+```
+
+### Update Changelog
+
+Add entry to `CHANGELOG.md`:
+
+```markdown
+## [0.1.9] - 2025-02-06
+
+### Added
+- NewModel support
+
+### Fixed
+- Memory leak in image processing
+
+### Changed
+- Improved error messages
+```
+
+### Create Release
+
+```bash
+# Tag the release
+git tag -a v0.1.9 -m "Release v0.1.9"
+git push origin v0.1.9
+
+# Build release artifacts
+cargo build --release
+
+# Publish to crates.io (if applicable)
+cargo publish
+```
+
+## Debugging
+
+### Logging
+
+Enable debug logging:
+
+```bash
+RUST_LOG=debug cargo run -- -m qwen3-0.6b
+```
+
+Set specific module logging:
+
+```bash
+RUST_LOG=aha::models::qwen3vl=debug cargo run -- -m qwen3-0.6b
+```
+
+### Debugging Tests
+
+```bash
+# Print test output
+cargo test -- --nocapture
+
+# Show backtrace
+RUST_BACKTRACE=1 cargo test
+```
+
+### Common Issues
+
+#### Build Errors
+
+- **Linking errors**: Install required system dependencies
+- **CUDA errors**: Ensure CUDA toolkit is installed
+- **Metal errors**: Check you're on Apple Silicon
+
+#### Runtime Errors
+
+- **Model not found**: Check model path and download
+- **Out of memory**: Use smaller model or enable GPU
+- **Slow inference**: Enable GPU acceleration
+
+## Resources
+
+- [Candle Documentation](https://github.com/huggingface/candle)
+- [Rust Guidelines](https://rust-lang.github.io/api-guidelines/)
+- [Conventional Commits](https://www.conventionalcommits.org/)
+
+## See Also
+
+- [Architecture & Design](./concepts.md) - How AHA works
+- [Installation](./installation.md) - Setup guide
+- [API Reference](./api.md) - API documentation
@@ -0,0 +1,601 @@
+# 开发指南
+
+本指南涵盖为 AHA 做贡献，包括开发设置、添加新模型和提交贡献。
+
+## 目录
+
+- [开发设置](#开发设置)
+- [项目结构](#项目结构)
+- [添加新模型](#添加新模型)
+- [测试](#测试)
+- [代码风格](#代码风格)
+- [提交贡献](#提交贡献)
+- [发布流程](#发布流程)
+
+## 开发设置
+
+### 前置要求
+
+- Rust 1.85+ (edition 2024)
+- Git
+- 可选：CUDA Toolkit、Metal、FFmpeg（用于功能开发）
+
+### 克隆和构建
+
+```bash
+# 克隆仓库
+git clone https://github.com/jhqxxx/aha.git
+cd aha
+
+# 调试模式构建（编译更快）
+cargo build
+
+# 发布模式构建（性能更好）
+cargo build --release
+
+# 使用功能构建
+cargo build --features "cuda,flash-attn"
+cargo build --features "metal"
+cargo build --features "ffmpeg"
+```
+
+### 开发工作流程
+
+```bash
+# 运行 CLI
+cargo run -- -m qwen3-0.6b
+
+# 运行测试
+cargo test
+
+# 运行特定测试
+cargo test test_qwen3vl_generate
+
+# 启用日志运行
+RUST_LOG=debug cargo run -- -m qwen3-0.6b
+
+# 检查代码而不构建
+cargo check
+
+# 格式化代码
+cargo fmt
+
+# 运行 linter
+cargo clippy
+```
+
+## 项目结构
+
+```
+aha/
+├── Cargo.toml              # 构建配置
+├── src/
+│   ├── main.rs             # CLI 入口点
+│   ├── lib.rs              # 库导出
+│   ├── api.rs              # HTTP API 处理程序
+│   ├── exec/               # CLI 命令实现
+│   │   ├── mod.rs
+│   │   ├── cli.rs
+│   │   ├── serv.rs
+│   │   ├── download.rs
+│   │   └── run.rs
+│   ├── models/             # 模型实现
+│   │   ├── mod.rs          # 模型工厂和导出
+│   │   ├── common/         # 共享模型工具
+│   │   │   ├── config.rs
+│   │   │   └── mod.rs
+│   │   ├── qwen3vl/        # Qwen3-VL 模型
+│   │   │   ├── config.rs
+│   │   │   ├── model.rs
+│   │   │   ├── generate.rs
+│   │   │   ├── processor.rs
+│   │   │   └── mod.rs
+│   │   ├── voxcpm/         # VoxCPM 模型
+│   │   └── ...             # 其他模型
+│   ├── tokenizer/          # 分词工具
+│   ├── chat_template/      # 聊天模板处理
+│   ├── position_embed/     # 位置编码
+│   └── utils/              # 工具函数
+│       ├── audio_utils.rs
+│       ├── image_utils.rs
+│       ├── download.rs
+│       └── common.rs
+├── tests/                  # 集成测试
+│   ├── test_qwen2_5vl.rs
+│   ├── test_qwen3vl.rs
+│   └── ...
+├── examples/               # 示例代码
+└── docs/                   # 文档
+```
+
+## 添加新模型
+
+本节提供了向 AHA 添加新模型的分步指南。
+
+### 步骤 1：创建模型目录
+
+在 `src/models/` 下创建新目录：
+
+```bash
+mkdir -p src/models/newmodel
+```
+
+### 步骤 2：实现模型文件
+
+在 `src/models/newmodel/` 中创建以下文件：
+
+#### config.rs
+
+定义模型配置：
+
+```rust
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct NewModelConfig {
+    pub vocab_size: usize,
+    pub hidden_size: usize,
+    pub num_hidden_layers: usize,
+    pub num_attention_heads: usize,
+    // ... 其他配置字段
+}
+
+impl Default for NewModelConfig {
+    fn default() -> Self {
+        Self {
+            vocab_size: 32000,
+            hidden_size: 2048,
+            num_hidden_layers: 24,
+            num_attention_heads: 32,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct GenerationConfig {
+    pub temperature: f64,
+    pub top_p: f64,
+    pub max_tokens: usize,
+}
+
+impl Default for GenerationConfig {
+    fn default() -> Self {
+        Self {
+            temperature: 1.0,
+            top_p: 1.0,
+            max_tokens: 100,
+        }
+    }
+}
+```
+
+#### model.rs
+
+实现模型架构：
+
+```rust
+use candle::{Result, Tensor};
+use candle_nn::{Module, VarBuilder};
+
+use super::config::NewModelConfig;
+
+pub struct NewModel {
+    // 定义模型组件
+    config: NewModelConfig,
+}
+
+impl NewModel {
+    pub fn load(vb: VarBuilder, config: &NewModelConfig) -> Result<Self> {
+        // 加载模型权重
+        Ok(Self {
+            config: config.clone(),
+        })
+    }
+
+    pub fn forward(&self, input_ids: &Tensor) -> Result<Tensor> {
+        // 实现前向传播
+        Ok(input_ids.clone())
+    }
+}
+```
+
+#### generate.rs
+
+实现 GenerateModel trait：
+
+```rust
+use std::path::Path;
+use anyhow::Result;
+use candle::Device;
+
+use super::config::{GenerationConfig, NewModelConfig};
+use aha::models::common::GenerateModel;
+
+pub struct NewModelGenerate {
+    // 定义生成状态
+    model: super::model::NewModel,
+    config: GenerationConfig,
+    device: Device,
+}
+
+impl GenerateModel for NewModelGenerate {
+    type Config = NewModelConfig;
+    type GenerateConfig = GenerationConfig;
+
+    fn init(
+        model_path: &str,
+        config: Option<Self::Config>,
+        generate_config: Option<Self::GenerateConfig>,
+    ) -> Result<Self> {
+        // 从磁盘加载模型
+        let device = Device::Cpu;
+        let config = config.unwrap_or_default();
+        let generate_config = generate_config.unwrap_or_default();
+
+        // 加载权重
+        let model_path = Path::new(model_path);
+        // ... 加载模型实现
+
+        Ok(Self {
+            model: todo!(),
+            config: generate_config,
+            device,
+        })
+    }
+
+    fn generate(&mut self, prompt: &str) -> Result<String> {
+        // 分词提示
+        // 运行推理
+        // 解码输出
+        Ok(prompt.to_string())
+    }
+}
+```
+
+#### processor.rs（可选）
+
+对于具有多模态输入的复杂模型：
+
+```rust
+use anyhow::Result;
+
+pub struct NewModelProcessor {
+    // 处理状态
+}
+
+impl NewModelProcessor {
+    pub fn new() -> Self {
+        Self {}
+    }
+
+    pub fn process_text(&self, text: &str) -> Result<Vec<String>> {
+        // 处理文本输入
+        Ok(vec![text.to_string()])
+    }
+
+    pub fn process_image(&self, image_path: &str) -> Result<Tensor> {
+        // 处理图像输入
+        todo!()
+    }
+}
+```
+
+#### mod.rs
+
+导出模型：
+
+```rust
+mod config;
+mod model;
+mod generate;
+pub mod processor;
+
+pub use generate::NewModelGenerate;
+pub use config::{GenerationConfig, NewModelConfig};
+```
+
+### 步骤 3：注册模型
+
+更新 `src/models/mod.rs`：
+
+```rust
+// 添加到导入
+pub mod newmodel;
+
+// 添加到 WhichModel 枚举
+#[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
+pub enum WhichModel {
+    // ... 现有模型
+    NewModel,
+}
+
+// 添加到模型加载
+pub fn load_model(
+    model_type: &WhichModel,
+    model_path: &str,
+    device: &Device,
+) -> Result<Box<dyn common::GenerateModel>> {
+    match model_type {
+        // ... 现有模型
+        WhichModel::NewModel => {
+            Ok(Box::new(newmodel::NewModelGenerate::init(model_path, None, None)?))
+        }
+    }
+}
+```
+
+### 步骤 4：更新 CLI
+
+更新 `src/main.rs` 以在帮助文本和 CLI 选项中包含新模型（如果需要）。
+
+### 步骤 5：添加测试
+
+创建 `tests/test_newmodel.rs`：
+
+```rust
+use anyhow::Result;
+
+#[test]
+fn test_newmodel_generate() -> Result<()> {
+    let model_path = "path/to/test/model";
+    let mut model = aha::models::newmodel::NewModelGenerate::init(model_path, None, None)?;
+    
+    let result = model.generate("测试提示")?;
+    assert!(!result.is_empty());
+    
+    Ok(())
+}
+```
+
+### 步骤 6：更新文档
+
+更新以下文件：
+- `README.md` - 将模型添加到支持的模型列表
+- `docs/cli.md` - 将模型 ID 添加到模型列表
+- `docs/api.md` - 将模型添加到支持的模型部分
+- `CHANGELOG.md` - 为新模型添加条目
+
+## 测试
+
+### 单元测试
+
+```bash
+# 运行所有测试
+cargo test
+
+# 运行特定测试
+cargo test test_newmodel_generate
+
+# 启用输出运行
+cargo test -- --nocapture
+
+# 并行运行测试
+cargo test -- --test-threads=4
+```
+
+### 集成测试
+
+集成测试位于 `tests/` 中：
+
+```bash
+# 运行特定集成测试
+cargo test --test test_qwen3vl
+```
+
+### 手动测试
+
+手动测试 CLI 和 API：
+
+```bash
+# 测试 CLI
+cargo run -- -m newmodel
+
+# 测试 API
+cargo run -- -m newmodel -p 8080
+curl http://127.0.0.1:8080/chat/completions -d '{...}'
+```
+
+### 模型特定测试
+
+每个模型都应该有涵盖以下内容的测试：
+- 模型加载
+- 基本推理
+- 边缘情况
+- 错误处理
+
+## 代码风格
+
+### 格式化
+
+```bash
+# 格式化所有代码
+cargo fmt
+
+# 检查格式化而不进行更改
+cargo fmt --check
+```
+
+### Linting
+
+```bash
+# 运行 clippy
+cargo clippy
+
+# 修复 clippy 警告
+cargo clippy --fix
+```
+
+### 约定
+
+- 使用 4 个空格缩进
+- 对于错误，优先使用 `Result<T>` 而不是 `Option<T>`
+- 对于应用程序错误使用 `anyhow::Result`
+- 使用 rustdoc 注释记录公共 API
+- 保持函数专注和小型
+- 使用有意义的变量名
+
+### 文档
+
+```rust
+/// 使用 NewModel 生成文本。
+///
+/// # 参数
+///
+/// * `prompt` - 输入提示文本
+/// * `max_tokens` - 要生成的最大令牌数
+///
+/// # 返回
+///
+/// 生成的文本作为字符串
+///
+/// # 示例
+///
+/// ```no_run
+/// let result = model.generate("你好")?;
+/// ```
+pub fn generate(&mut self, prompt: &str) -> Result<String> {
+    // 实现
+}
+```
+
+## 提交贡献
+
+### Pull Request 流程
+
+1. Fork 仓库
+2. 创建功能分支：
+   ```bash
+   git checkout -b feature/new-model
+   ```
+3. 进行更改
+4. 添加测试
+5. 更新文档
+6. 确保所有测试通过：
+   ```bash
+   cargo test
+   cargo clippy
+   cargo fmt --check
+   ```
+7. 提交并推送：
+   ```bash
+   git commit -m "添加 NewModel 支持"
+   git push origin feature/new-model
+   ```
+8. 在 GitHub 上创建 pull request
+
+### Pull Request 清单
+
+- [ ] 已添加/更新测试
+- [ ] 已更新文档
+- [ ] 已更新 CHANGELOG.md
+- [ ] 代码已格式化（`cargo fmt`）
+- [ ] 已修复 Clippy 警告（`cargo clippy`）
+- [ ] 所有测试通过（`cargo test`）
+- [ ] 提交消息遵循约定
+
+### 提交消息约定
+
+遵循约定式提交：
+
+```
+feat: 添加 NewModel 支持
+fix: 修正 Qwen3VL 中的张量维度
+docs: 更新安装指南
+test: 添加 VoxCPM 集成测试
+refactor: 简化模型加载逻辑
+perf: 将推理速度提高 20%
+```
+
+## 发布流程
+
+### 版本升级
+
+更新 `Cargo.toml`：
+
+```toml
+[package]
+version = "0.1.9"  # 升级版本
+```
+
+### 更新变更日志
+
+将条目添加到 `CHANGELOG.md`：
+
+```markdown
+## [0.1.9] - 2025-02-06
+
+### 新增
+- NewModel 支持
+
+### 修复
+- 图像处理中的内存泄漏
+
+### 更改
+- 改进错误消息
+```
+
+### 创建发布
+
+```bash
+# 标记发布
+git tag -a v0.1.9 -m "Release v0.1.9"
+git push origin v0.1.9
+
+# 构建发布产物
+cargo build --release
+
+# 发布到 crates.io（如果适用）
+cargo publish
+```
+
+## 调试
+
+### 日志
+
+启用调试日志：
+
+```bash
+RUST_LOG=debug cargo run -- -m qwen3-0.6b
+```
+
+设置特定模块日志：
+
+```bash
+RUST_LOG=aha::models::qwen3vl=debug cargo run -- -m qwen3-0.6b
+```
+
+### 调试测试
+
+```bash
+# 打印测试输出
+cargo test -- --nocapture
+
+# 显示回溯
+RUST_BACKTRACE=1 cargo test
+```
+
+### 常见问题
+
+#### 构建错误
+
+- **链接错误**：安装所需的系统依赖
+- **CUDA 错误**：确保已安装 CUDA toolkit
+- **Metal 错误**：检查您是否在 Apple Silicon 上
+
+#### 运行时错误
+
+- **未找到模型**：检查模型路径和下载
+- **内存不足**：使用更小的模型或启用 GPU
+- **推理速度慢**：启用 GPU 加速
+
+## 资源
+
+- [Candle 文档](https://github.com/huggingface/candle)
+- [Rust 指南](https://rust-lang.github.io/api-guidelines/)
+- [约定式提交](https://www.conventionalcommits.org/)
+
+## 另见
+
+- [架构与设计](./concepts.zh-CN.md) - AHA 的工作原理
+- [安装](./installation.zh-CN.md) - 设置指南
+- [API 参考](./api.zh-CN.md) - API 文档
@@ -0,0 +1,376 @@
+# Getting Started
+
+Welcome to AHA! This guide will help you get up and running quickly.
+
+## Quick Start (5 Minutes)
+
+### 1. Check Available Models
+
+```bash
+aha list
+```
+
+### 2. Download Your First Model
+
+```bash
+# Download a small text model to start
+aha download -m qwen3-0.6b
+```
+
+### 3. Start the Service
+
+```bash
+# Start the HTTP API server
+aha cli -m qwen3-0.6b
+```
+
+The service will start on `http://127.0.0.1:10100`
+
+### 4. Make Your First API Call
+
+In a new terminal:
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "Hello, AHA!"}
+    ]
+  }'
+```
+
+## Basic Concepts
+
+### What is AHA?
+
+AHA is a local AI inference engine that:
+- Runs models on your machine (no cloud API)
+- Supports multiple model types (text, vision, audio, OCR, ASR)
+- Provides an OpenAI-compatible API
+- Works offline once models are downloaded
+
+### Model Categories
+
+| Category | Description | Example Models |
+|----------|-------------|----------------|
+| **Text** | Text generation and chat | Qwen3, MiniCPM4 |
+| **Vision** | Image understanding | Qwen2.5VL, Qwen3VL |
+| **OCR** | Text extraction from images | DeepSeek-OCR, Hunyuan-OCR |
+| **ASR** | Speech-to-text | GLM-ASR, Fun-ASR, Qwen3-ASR |
+| **Audio** | Text-to-speech | VoxCPM, VoxCPM1.5 |
+| **Image** | Image processing | RMBG2.0 (background removal) |
+
+### CLI Commands
+
+| Command | Purpose |
+|---------|---------|
+| `aha cli` | Download model and start service |
+| `aha serv` | Start service with existing model |
+| `aha download` | Download model only |
+| `aha run` | Direct inference without server |
+| `aha list` | List available models |
+
+## Common Workflows
+
+### Text Generation
+
+```bash
+# Start the service
+aha cli -m qwen3-0.6b
+
+# In another terminal, make a request
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "system", "content": "You are a helpful assistant."},
+      {"role": "user", "content": "Explain quantum computing in simple terms."}
+    ],
+    "max_tokens": 200,
+    "temperature": 0.7
+  }'
+```
+
+### Vision Understanding
+
+```bash
+# Start a vision model
+aha cli -m qwen3vl-2b
+
+# Analyze an image
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3vl-2b",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "Describe this image in detail."},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/image.jpg"}}
+        ]
+      }
+    ]
+  }'
+```
+
+### OCR (Text Extraction)
+
+```bash
+# Start an OCR model
+aha cli -m deepseek-ocr
+
+# Extract text from an image
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "deepseek-ocr",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "Extract all text from this image."},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/document.jpg"}}
+        ]
+      }
+    ]
+  }'
+```
+
+### Speech Recognition (ASR)
+
+```bash
+# Start an ASR model
+aha cli -m glm-asr-nano-2512
+
+# Transcribe audio
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "glm-asr-nano-2512",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "Transcribe this audio."},
+          {"type": "audio_url", "audio_url": {"url": "file:///path/to/audio.wav"}}
+        ]
+      }
+    ]
+  }'
+```
+
+### Text-to-Speech
+
+```bash
+# Start a TTS model
+aha cli -m voxcpm1.5
+
+# Generate speech
+curl http://127.0.0.1:10100/audio/speech \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "voxcpm1.5",
+    "input": "Hello, this is AHA speaking.",
+    "voice": "default"
+  }' \
+  --output speech.wav
+```
+
+### Background Removal
+
+```bash
+# Start RMBG2.0 model
+aha cli -m rmbg2.0
+
+# Remove background from image
+curl http://127.0.0.1:10100/images/remove_background \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "rmbg2.0",
+    "image": "file:///path/to/photo.png"
+  }' \
+  --output no-background.png
+```
+
+### Direct Inference (Without Server)
+
+```bash
+# Run inference directly without starting HTTP server
+aha run -m qwen3-0.6b \
+  -i "Write a haiku about AI" \
+  --weight-path ~/.aha/qwen3-0.6b
+```
+
+## Configuration Options
+
+### Change Port
+
+```bash
+# Use port 8080 instead of default 10100
+aha cli -m qwen3-0.6b -p 8080
+```
+
+### Bind to All Interfaces
+
+```bash
+# Allow external access (use with caution)
+aha cli -m qwen3-0.6b -a 0.0.0.0 -p 8080
+```
+
+### Use Local Model
+
+```bash
+# Skip download, use existing model
+aha serv -m qwen3-0.6b \
+  --weight-path /path/to/model \
+  -p 8080
+```
+
+### Custom Save Directory
+
+```bash
+# Download model to specific directory
+aha download -m qwen3vl-2b -s /data/models
+```
+
+## Streaming Responses
+
+For real-time responses, use streaming:
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "Tell me a story"}
+    ],
+    "stream": true
+  }'
+```
+
+## Model Selection Guide
+
+### For Text Generation
+- **qwen3-0.6b**: Fast, lightweight (~1.2 GB)
+- **minicpm4-0.5b**: Small, efficient (~1 GB)
+
+### For Vision Tasks
+- **qwen3vl-2b**: Balanced performance (~4 GB)
+- **qwen3vl-8b**: Better quality (~16 GB)
+
+### For OCR
+- **deepseek-ocr**: General purpose
+- **hunyuan-ocr**: Good for Chinese text
+- **paddleocr-vl**: Lightweight option
+
+### For Speech Recognition
+- **glm-asr-nano-2512**: Fast, accurate
+- **fun-asr-nano-2512**: Good for Chinese
+- **qwen3asr-0.6b**: Lightweight
+
+### For Text-to-Speech
+- **voxcpm1.5**: High quality Chinese
+
+### For Background Removal
+- **rmbg2.0**: State-of-the-art results
+
+## Tips & Best Practices
+
+### 1. Start Small
+
+Begin with smaller models to understand the workflow:
+```bash
+aha download -m qwen3-0.6b
+```
+
+### 2. Use GPU Acceleration
+
+Build with GPU support for better performance:
+```bash
+# NVIDIA GPUs
+cargo build --release --features cuda
+
+# Apple Silicon
+cargo build --release --features metal
+```
+
+### 3. Pre-download Models
+
+Download models when you have good internet:
+```bash
+aha download -m qwen3vl-2b
+```
+
+Then use them later without internet:
+```bash
+aha serv -m qwen3vl-2b --weight-path ~/.aha/qwen3vl-2b
+```
+
+### 4. Manage Disk Space
+
+Models are stored in `~/.aha/` by default. Clean up if needed:
+```bash
+# Check disk usage
+du -sh ~/.aha/*
+
+# Remove old models
+rm -rf ~/.aha/old-model-name
+```
+
+### 5. Monitor Resources
+
+For large models, monitor your resources:
+```bash
+# Linux
+htop
+nvidia-smi  # For NVIDIA GPUs
+
+# macOS
+Activity Monitor
+```
+
+## Troubleshooting
+
+### Port Already in Use
+
+```bash
+# Use a different port
+aha cli -m qwen3-0.6b -p 8080
+```
+
+### Model Download Failed
+
+```bash
+# Retry with more attempts
+aha download -m qwen3vl-2b --download-retries 5
+```
+
+### Out of Memory
+
+```bash
+# Use a smaller model
+aha cli -m qwen3-0.6b
+```
+
+## Next Steps
+
+1. Explore the [API Reference](./api.md) for detailed endpoint documentation
+2. Read the [CLI Reference](./cli.md) for all command options
+3. Check [Architecture & Design](./concepts.md) to understand how AHA works
+4. See [Development](./development.md) if you want to contribute
+
+## Examples Repository
+
+For more examples, check out the [examples](../examples/) directory in the repository.
+
+## See Also
+
+- [API Reference](./api.md) - Complete API documentation
+- [CLI Reference](./cli.md) - Command-line reference
+- [Installation Guide](./installation.md) - Installation instructions
+- [Development Guide](./development.md) - Contributing guide
@@ -0,0 +1,376 @@
+# 快速入门
+
+欢迎使用 AHA！本指南将帮助您快速上手。
+
+## 快速开始（5 分钟）
+
+### 1. 查看可用模型
+
+```bash
+aha list
+```
+
+### 2. 下载第一个模型
+
+```bash
+# 下载一个小型文本模型开始
+aha download -m qwen3-0.6b
+```
+
+### 3. 启动服务
+
+```bash
+# 启动 HTTP API 服务器
+aha cli -m qwen3-0.6b
+```
+
+服务将在 `http://127.0.0.1:10100` 上启动
+
+### 4. 发起第一个 API 调用
+
+在新终端中：
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "你好，AHA！"}
+    ]
+  }'
+```
+
+## 基本概念
+
+### 什么是 AHA？
+
+AHA 是一个本地 AI 推理引擎，具有以下特点：
+- 在您的机器上运行模型（无需云 API）
+- 支持多种模型类型（文本、视觉、音频、OCR、ASR）
+- 提供 OpenAI 兼容的 API
+- 模型下载后可离线工作
+
+### 模型类别
+
+| 类别 | 描述 | 示例模型 |
+|------|------|----------|
+| **文本** | 文本生成和对话 | Qwen3、MiniCPM4 |
+| **视觉** | 图像理解 | Qwen2.5VL、Qwen3VL |
+| **OCR** | 从图像中提取文本 | DeepSeek-OCR、Hunyuan-OCR |
+| **ASR** | 语音转文本 | GLM-ASR、Fun-ASR、Qwen3-ASR |
+| **音频** | 文本转语音 | VoxCPM、VoxCPM1.5 |
+| **图像** | 图像处理 | RMBG2.0（背景移除） |
+
+### CLI 命令
+
+| 命令 | 用途 |
+|------|------|
+| `aha cli` | 下载模型并启动服务 |
+| `aha serv` | 使用现有模型启动服务 |
+| `aha download` | 仅下载模型 |
+| `aha run` | 直接推理，无需服务器 |
+| `aha list` | 列出可用模型 |
+
+## 常见工作流程
+
+### 文本生成
+
+```bash
+# 启动服务
+aha cli -m qwen3-0.6b
+
+# 在另一个终端中，发起请求
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "system", "content": "你是一个有用的助手。"},
+      {"role": "user", "content": "用简单的术语解释量子计算。"}
+    ],
+    "max_tokens": 200,
+    "temperature": 0.7
+  }'
+```
+
+### 视觉理解
+
+```bash
+# 启动视觉模型
+aha cli -m qwen3vl-2b
+
+# 分析图像
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3vl-2b",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "详细描述这张图片。"},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/image.jpg"}}
+        ]
+      }
+    ]
+  }'
+```
+
+### OCR（文本提取）
+
+```bash
+# 启动 OCR 模型
+aha cli -m deepseek-ocr
+
+# 从图像中提取文本
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "deepseek-ocr",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "提取这张图片中的所有文本。"},
+          {"type": "image_url", "image_url": {"url": "file:///path/to/document.jpg"}}
+        ]
+      }
+    ]
+  }'
+```
+
+### 语音识别（ASR）
+
+```bash
+# 启动 ASR 模型
+aha cli -m glm-asr-nano-2512
+
+# 转写音频
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "glm-asr-nano-2512",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {"type": "text", "text": "转写这段音频。"},
+          {"type": "audio_url", "audio_url": {"url": "file:///path/to/audio.wav"}}
+        ]
+      }
+    ]
+  }'
+```
+
+### 文本转语音
+
+```bash
+# 启动 TTS 模型
+aha cli -m voxcpm1.5
+
+# 生成语音
+curl http://127.0.0.1:10100/audio/speech \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "voxcpm1.5",
+    "input": "你好，这是 AHA 在说话。",
+    "voice": "default"
+  }' \
+  --output speech.wav
+```
+
+### 背景移除
+
+```bash
+# 启动 RMBG2.0 模型
+aha cli -m rmbg2.0
+
+# 移除图像背景
+curl http://127.0.0.1:10100/images/remove_background \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "rmbg2.0",
+    "image": "file:///path/to/photo.png"
+  }' \
+  --output no-background.png
+```
+
+### 直接推理（无需服务器）
+
+```bash
+# 直接运行推理，无需启动 HTTP 服务器
+aha run -m qwen3-0.6b \
+  -i "写一首关于 AI 的俳句" \
+  --weight-path ~/.aha/qwen3-0.6b
+```
+
+## 配置选项
+
+### 更改端口
+
+```bash
+# 使用端口 8080 而不是默认的 10100
+aha cli -m qwen3-0.6b -p 8080
+```
+
+### 绑定到所有接口
+
+```bash
+# 允许外部访问（请谨慎使用）
+aha cli -m qwen3-0.6b -a 0.0.0.0 -p 8080
+```
+
+### 使用本地模型
+
+```bash
+# 跳过下载，使用现有模型
+aha serv -m qwen3-0.6b \
+  --weight-path /path/to/model \
+  -p 8080
+```
+
+### 自定义保存目录
+
+```bash
+# 将模型下载到特定目录
+aha download -m qwen3vl-2b -s /data/models
+```
+
+## 流式响应
+
+对于实时响应，使用流式传输：
+
+```bash
+curl http://127.0.0.1:10100/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-0.6b",
+    "messages": [
+      {"role": "user", "content": "给我讲个故事"}
+    ],
+    "stream": true
+  }'
+```
+
+## 模型选择指南
+
+### 文本生成
+- **qwen3-0.6b**：快速、轻量级（~1.2 GB）
+- **minicpm4-0.5b**：小型、高效（~1 GB）
+
+### 视觉任务
+- **qwen3vl-2b**：平衡性能（~4 GB）
+- **qwen3vl-8b**：更高质量（~16 GB）
+
+### OCR
+- **deepseek-ocr**：通用
+- **hunyuan-ocr**：适合中文文本
+- **paddleocr-vl**：轻量级选项
+
+### 语音识别
+- **glm-asr-nano-2512**：快速、准确
+- **fun-asr-nano-2512**：适合中文
+- **qwen3asr-0.6b**：轻量级
+
+### 文本转语音
+- **voxcpm1.5**：高质量中文
+
+### 背景移除
+- **rmbg2.0**：最先进的结果
+
+## 提示与最佳实践
+
+### 1. 从小开始
+
+从小型模型开始了解工作流程：
+```bash
+aha download -m qwen3-0.6b
+```
+
+### 2. 使用 GPU 加速
+
+使用 GPU 支持构建以获得更好的性能：
+```bash
+# NVIDIA GPU
+cargo build --release --features cuda
+
+# Apple Silicon
+cargo build --release --features metal
+```
+
+### 3. 预先下载模型
+
+在网络良好时下载模型：
+```bash
+aha download -m qwen3vl-2b
+```
+
+稍后在没有网络的情况下使用：
+```bash
+aha serv -m qwen3vl-2b --weight-path ~/.aha/qwen3vl-2b
+```
+
+### 4. 管理磁盘空间
+
+模型默认存储在 `~/.aha/` 中。如需要，清理：
+```bash
+# 检查磁盘使用情况
+du -sh ~/.aha/*
+
+# 删除旧模型
+rm -rf ~/.aha/old-model-name
+```
+
+### 5. 监控资源
+
+对于大型模型，监控您的资源：
+```bash
+# Linux
+htop
+nvidia-smi  # 对于 NVIDIA GPU
+
+# macOS
+活动监视器
+```
+
+## 故障排除
+
+### 端口已被占用
+
+```bash
+# 使用不同的端口
+aha cli -m qwen3-0.6b -p 8080
+```
+
+### 模型下载失败
+
+```bash
+# 重试更多次数
+aha download -m qwen3vl-2b --download-retries 5
+```
+
+### 内存不足
+
+```bash
+# 使用更小的模型
+aha cli -m qwen3-0.6b
+```
+
+## 后续步骤
+
+1. 探索 [API 参考](./api.zh-CN.md) 了解详细的端点文档
+2. 阅读 [CLI 参考](./cli.zh-CN.md) 了解所有命令选项
+3. 查看 [架构与设计](./concepts.zh-CN.md) 了解 AHA 的工作原理
+4. 如果您想贡献，请参阅 [开发指南](./development.zh-CN.md)
+
+## 示例仓库
+
+更多示例，请查看仓库中的 [examples](../examples/) 目录。
+
+## 另见
+
+- [API 参考](./api.zh-CN.md) - 完整的 API 文档
+- [CLI 参考](./cli.zh-CN.md) - 命令行参考
+- [安装指南](./installation.zh-CN.md) - 安装说明
+- [开发指南](./development.zh-CN.md) - 贡献指南
@@ -0,0 +1,420 @@
+# Installation Guide
+
+This guide covers installing and setting up AHA on your system.
+
+## Table of Contents
+
+- [Prerequisites](#prerequisites)
+- [Installation Methods](#installation-methods)
+- [Platform-Specific Instructions](#platform-specific-instructions)
+- [Feature Flags](#feature-flags)
+- [Verification](#verification)
+- [Troubleshooting](#troubleshooting)
+
+## Prerequisites
+
+### Required
+
+- **Rust toolchain**: Rust 1.85 or later (edition 2024)
+  ```bash
+  curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+  ```
+
+- **Git**: For cloning the repository
+  ```bash
+  # Ubuntu/Debian
+  sudo apt-get install git
+  
+  # macOS
+  brew install git
+  
+  # Windows
+  # Download from https://git-scm.com/download/win
+  ```
+
+### Optional (for FFmpeg feature)
+
+- **FFmpeg development libraries**: Required for audio/video processing
+
+## Installation Methods
+
+### Method 1: Build from Source
+
+Clone the repository and build:
+
+```bash
+git clone https://github.com/jhqxxx/aha.git
+cd aha
+
+# Build release version
+cargo build --release
+
+# The binary will be at target/release/aha
+```
+
+### Method 2: Install from Crates.io (when available)
+
+```bash
+cargo install aha
+```
+
+### Method 3: Install with Features
+
+Build with specific features enabled:
+
+```bash
+# With CUDA support (NVIDIA GPUs)
+cargo build --release --features cuda
+
+# With Metal support (Apple Silicon)
+cargo build --release --features metal
+
+# With Flash Attention
+cargo build --release --features "cuda,flash-attn"
+
+# With FFmpeg support
+cargo build --release --features ffmpeg
+```
+
+## Platform-Specific Instructions
+
+### Linux
+
+#### Ubuntu/Debian
+
+```bash
+# Install build dependencies
+sudo apt-get update
+sudo apt-get install -y build-essential pkg-config git clang
+
+# For FFmpeg feature
+sudo apt-get install -y ffmpeg libavutil-dev libavcodec-dev \
+    libavformat-dev libavfilter-dev libavdevice-dev \
+    libswresample-dev libswscale-dev
+
+# For CUDA support, install CUDA toolkit
+# See https://developer.nvidia.com/cuda-downloads
+```
+
+#### Fedora/RHEL
+
+```bash
+# Install build dependencies
+sudo dnf install gcc gcc-c++ make git clang pkg-config
+
+# For FFmpeg feature
+sudo dnf install ffmpeg-devel
+
+# For CUDA support
+sudo dnf install cuda-devel
+```
+
+### macOS
+
+#### Apple Silicon (M1/M2/M3/M4)
+
+```bash
+# Install Rust (if not already installed)
+curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+
+# Install command line tools
+xcode-select --install
+
+# For FFmpeg feature
+brew install ffmpeg
+
+# Build with Metal support for GPU acceleration
+cargo build --release --features metal
+```
+
+#### Intel Mac
+
+```bash
+# Install Rust (if not already installed)
+curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+
+# Install command line tools
+xcode-select --install
+
+# For FFmpeg feature
+brew install ffmpeg
+
+# For CUDA support (if you have NVIDIA GPU)
+# Install CUDA from https://developer.nvidia.com/cuda-downloads
+cargo build --release --features cuda
+```
+
+### Windows
+
+#### Using MSVC
+
+```bash
+# Install Rust from https://rustup.rs/
+# Install Visual Studio Build Tools from https://visualstudio.microsoft.com/downloads/
+
+# For FFmpeg feature
+# Download FFmpeg from https://ffmpeg.org/download.html
+# Set FFMPEG_DIR environment variable to your FFmpeg installation
+
+# Build
+cargo build --release
+```
+
+#### Using WSL2 (Recommended)
+
+```bash
+# Follow Linux instructions inside WSL2
+wsl
+sudo apt-get update
+sudo apt-get install -y build-essential pkg-config git clang
+```
+
+## Feature Flags
+
+aha supports several optional features:
+
+### cuda
+
+Enables CUDA support for NVIDIA GPU acceleration.
+
+```bash
+cargo build --release --features cuda
+```
+
+**Requirements**:
+- NVIDIA GPU
+- CUDA Toolkit 11.0 or later
+- cuDNN library
+
+**Benefits**:
+- 10-50x faster inference
+- Support for larger models
+- Lower CPU usage
+
+### metal
+
+Enables Metal support for Apple Silicon GPU acceleration.
+
+```bash
+cargo build --release --features metal
+```
+
+**Requirements**:
+- Apple Silicon (M1/M2/M3/M4)
+- macOS 11.0 or later
+
+**Benefits**:
+- 5-20x faster inference
+- Lower power consumption
+- Support for larger models
+
+### flash-attn
+
+Enables Flash Attention for optimized long-sequence processing.
+
+```bash
+cargo build --release --features "cuda,flash-attn"
+```
+
+**Requirements**:
+- CUDA feature enabled
+- Supported GPU architecture (compute capability 7.0+)
+
+**Benefits**:
+- Reduced memory usage
+- Faster inference for long sequences
+- Especially beneficial for vision models
+
+**Note**: Must be used with `cuda` feature.
+
+### ffmpeg
+
+Enables FFmpeg support for audio/video processing.
+
+```bash
+cargo build --release --features ffmpeg
+```
+
+**Requirements**:
+- FFmpeg development libraries
+- Platform-specific (see above)
+
+**Benefits**:
+- Extended audio format support (MP3, AAC, etc.)
+- Video processing capabilities
+- Better audio resampling
+
+### Combining Features
+
+You can combine multiple features:
+
+```bash
+# Maximum performance on NVIDIA GPU
+cargo build --release --features "cuda,flash-attn"
+
+# Apple Silicon with audio support
+cargo build --release --features "metal,ffmpeg"
+
+# Everything enabled
+cargo build --release --features "cuda,flash-attn,ffmpeg"
+```
+
+## Verification
+
+After installation, verify that AHA is working:
+
+```bash
+# Check version
+./target/release/aha --version
+
+# List supported models
+./target/release/aha list
+
+# (Or if installed to PATH)
+aha --version
+aha list
+```
+
+Expected output for `aha list`:
+
+```
+#Supported models:
+ 
+ Available models:
+
+Model Name                     ModelScope ID
+-----------------------------------------------------------
+minicpm4-0.5b                  OpenBMB/MiniCPM4-0.5B
+qwen2.5vl-3b                   Qwen/Qwen2.5-VL-3B-Instruct
+qwen2.5vl-7b                   Qwen/Qwen2.5-VL-7B-Instruct
+qwen3-0.6b                     Qwen/Qwen3-0.6B
+qwen3asr-0.6b                  Qwen/Qwen3-ASR-0.6B
+qwen3asr-1.7b                  Qwen/Qwen3-ASR-1.7B
+qwen3vl-4b                     Qwen/Qwen3-VL-2B-Instruct
+qwen3vl-4b                     Qwen/Qwen3-VL-4B-Instruct
+qwen3vl-8b                     Qwen/Qwen3-VL-8B-Instruct
+qwen3vl-32b                    Qwen/Qwen3-VL-32B-Instruct
+deepseek-ocr                   deepseek-ai/DeepSeek-OCR
+hunyuan-ocr                    Tencent-Hunyuan/HunyuanOCR
+paddleocr-vl                   PaddlePaddle/PaddleOCR-VL
+rmbg2.0                        AI-ModelScope/RMBG-2.0
+voxcpm                         OpenBMB/VoxCPM-0.5B
+voxcpm1.5                      OpenBMB/VoxCPM1.5
+glm-asr-nano-2512              ZhipuAI/GLM-ASR-Nano-2512
+fun-asr-nano-2512              FunAudioLLM/Fun-ASR-Nano-2512
+
+```
+
+## Troubleshooting
+
+### Build Errors
+
+#### "error: linking with cc failed"
+
+This usually indicates missing system dependencies.
+
+**Solution**: Install required build tools for your platform (see Platform-Specific Instructions).
+
+#### "error: CUDA not found"
+
+CUDA feature is enabled but CUDA toolkit is not installed.
+
+**Solution**: 
+- Install CUDA toolkit from https://developer.nvidia.com/cuda-downloads
+- Or build without CUDA: `cargo build --release`
+
+#### "error: Metal not available"
+
+Metal feature is enabled but not on supported hardware.
+
+**Solution**:
+- Ensure you're on Apple Silicon
+- Or build without Metal: `cargo build --release`
+
+### Runtime Errors
+
+#### "error while loading shared libraries"
+
+Missing runtime libraries.
+
+**Solution**: Install required libraries (see Platform-Specific Instructions).
+
+#### "Out of memory"
+
+Model is too large for available RAM/VRAM.
+
+**Solution**:
+- Use a smaller model
+- Close other applications
+- Enable GPU acceleration for better memory efficiency
+
+#### "Model download failed"
+
+Network issue or insufficient disk space.
+
+**Solution**:
+- Check internet connection
+- Ensure sufficient disk space in `~/.aha/`
+- Try again: download will resume if interrupted
+
+### Performance Issues
+
+#### Slow inference
+
+**Solutions**:
+1. Enable GPU acceleration: `--features cuda` or `--features metal`
+2. Enable Flash Attention: `--features "cuda,flash-attn"`
+3. Use a smaller model
+4. Check if GPU is being used (should see GPU usage in monitoring tools)
+
+#### High CPU usage
+
+**Solutions**:
+1. Enable GPU acceleration
+2. Reduce batch size
+3. Use model with lower precision
+
+## System Requirements
+*Different models require different hardware and software, for reference.*
+
+### Minimum Requirements
+
+- **CPU**: x86_64 or ARM64
+- **RAM**: 8 GB (16 GB recommended)
+- **Disk**: 10 GB for models (varies by model)
+- **OS**: Linux, macOS, or Windows
+
+### Recommended Requirements
+
+- **CPU**: Modern multi-core processor
+- **RAM**: 32 GB or more
+- **GPU**: NVIDIA GPU (with CUDA) or Apple Silicon
+- **Disk**: SSD with 50+ GB free space
+- **OS**: Linux (Ubuntu 22.04+) or macOS (Monterey+)
+
+## Model Sizes
+
+Approximate download sizes for popular models:
+
+| Model | Size | RAM Usage |
+|-------|------|-----------|
+| qwen3-0.6b | ~1.2 GB | ~2 GB |
+| qwen3vl-2b | ~4 GB | ~6 GB |
+| qwen3vl-8b | ~16 GB | ~20 GB |
+| qwen3vl-32b | ~64 GB | ~70 GB |
+
+## Next Steps
+
+After successful installation:
+
+1. Read the [Getting Started Guide](./getting-started.md)
+2. Download your first model: `aha download -m qwen3-0.6b`
+3. Start the service: `aha cli -m qwen3-0.6b`
+4. Explore the [API Reference](./api.md)
+
+## See Also
+
+- [Getting Started](./getting-started.md) - Quick start guide
+- [CLI Reference](./cli.md) - Command-line usage
+- [API Reference](./api.md) - REST API documentation
+- [Development](./development.md) - Contributing guide
@@ -0,0 +1,419 @@
+# 安装指南
+
+本指南介绍如何在您的系统上安装和设置 AHA。
+
+## 目录
+
+- [前置要求](#前置要求)
+- [安装方法](#安装方法)
+- [平台特定说明](#平台特定说明)
+- [功能特性](#功能特性)
+- [验证安装](#验证安装)
+- [故障排除](#故障排除)
+
+## 前置要求
+
+### 必需
+
+- **Rust 工具链**：Rust 1.85 或更高版本（edition 2024）
+  ```bash
+  curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+  ```
+
+- **Git**：用于克隆仓库
+  ```bash
+  # Ubuntu/Debian
+  sudo apt-get install git
+  
+  # macOS
+  brew install git
+  
+  # Windows
+  # 从 https://git-scm.com/download/win 下载
+  ```
+
+### 可选（用于 FFmpeg 功能）
+
+- **FFmpeg 开发库**：音频/视频处理所需
+
+## 安装方法
+
+### 方法 1：从源码构建
+
+克隆仓库并构建：
+
+```bash
+git clone https://github.com/jhqxxx/aha.git
+cd aha
+
+# 构建发布版本
+cargo build --release
+
+# 二进制文件位于 target/release/aha
+```
+
+### 方法 2：从 Crates.io 安装（当可用时）
+
+```bash
+cargo install aha
+```
+
+### 方法 3：启用特定功能构建
+
+使用特定功能构建：
+
+```bash
+# 启用 CUDA 支持（NVIDIA GPU）
+cargo build --release --features cuda
+
+# 启用 Metal 支持（Apple Silicon）
+cargo build --release --features metal
+
+# 启用 Flash Attention
+cargo build --release --features "cuda,flash-attn"
+
+# 启用 FFmpeg 支持
+cargo build --release --features ffmpeg
+```
+
+## 平台特定说明
+
+### Linux
+
+#### Ubuntu/Debian
+
+```bash
+# 安装构建依赖
+sudo apt-get update
+sudo apt-get install -y build-essential pkg-config git clang
+
+# FFmpeg 功能所需
+sudo apt-get install -y ffmpeg libavutil-dev libavcodec-dev \
+    libavformat-dev libavfilter-dev libavdevice-dev \
+    libswresample-dev libswscale-dev
+
+# CUDA 支持，从 https://developer.nvidia.com/cuda-downloads 安装 CUDA toolkit
+```
+
+#### Fedora/RHEL
+
+```bash
+# 安装构建依赖
+sudo dnf install gcc gcc-c++ make git clang pkg-config
+
+# FFmpeg 功能所需
+sudo dnf install ffmpeg-devel
+
+# CUDA 支持
+sudo dnf install cuda-devel
+```
+
+### macOS
+
+#### Apple Silicon (M1/M2/M3/M4)
+
+```bash
+# 安装 Rust（如果尚未安装）
+curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+
+# 安装命令行工具
+xcode-select --install
+
+# FFmpeg 功能所需
+brew install ffmpeg
+
+# 启用 Metal 支持以进行 GPU 加速
+cargo build --release --features metal
+```
+
+#### Intel Mac
+
+```bash
+# 安装 Rust（如果尚未安装）
+curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+
+# 安装命令行工具
+xcode-select --install
+
+# FFmpeg 功能所需
+brew install ffmpeg
+
+# CUDA 支持（如果您有 NVIDIA GPU）
+# 从 https://developer.nvidia.com/cuda-downloads 安装 CUDA
+cargo build --release --features cuda
+```
+
+### Windows
+
+#### 使用 MSVC
+
+```bash
+# 从 https://rustup.rs/ 安装 Rust
+# 从 https://visualstudio.microsoft.com/downloads/ 安装 Visual Studio Build Tools
+
+# FFmpeg 功能
+# 从 https://ffmpeg.org/download.html 下载 FFmpeg
+# 设置 FFMPEG_DIR 环境变量指向您的 FFmpeg 安装目录
+
+# 构建
+cargo build --release
+```
+
+#### 使用 WSL2（推荐）
+
+```bash
+# 在 WSL2 中按照 Linux 说明操作
+wsl
+sudo apt-get update
+sudo apt-get install -y build-essential pkg-config git clang
+```
+
+## 功能特性
+
+AHA 支持多个可选功能：
+
+### cuda
+
+启用 NVIDIA GPU 加速的 CUDA 支持。
+
+```bash
+cargo build --release --features cuda
+```
+
+**要求**：
+- NVIDIA GPU
+- CUDA Toolkit 11.0 或更高版本
+- cuDNN 库
+
+**优势**：
+- 推理速度提升 10-50 倍
+- 支持更大的模型
+- 降低 CPU 使用率
+
+### metal
+
+启用 Apple Silicon GPU 加速的 Metal 支持。
+
+```bash
+cargo build --release --features metal
+```
+
+**要求**：
+- Apple Silicon (M1/M2/M3/M4)
+- macOS 11.0 或更高版本
+
+**优势**：
+- 推理速度提升 5-20 倍
+- 更低的功耗
+- 支持更大的模型
+
+### flash-attn
+
+启用 Flash Attention 以优化长序列处理。
+
+```bash
+cargo build --release --features "cuda,flash-attn"
+```
+
+**要求**：
+- 启用 CUDA 功能
+- 支持的 GPU 架构（计算能力 7.0+）
+
+**优势**：
+- 减少内存使用
+- 长序列推理更快
+- 对视觉模型特别有益
+
+**注意**：必须与 `cuda` 功能一起使用。
+
+### ffmpeg
+
+启用 FFmpeg 支持以进行音频/视频处理。
+
+```bash
+cargo build --release --features ffmpeg
+```
+
+**要求**：
+- FFmpeg 开发库
+- 特定平台（见上文）
+
+**优势**：
+- 扩展的音频格式支持（MP3、AAC 等）
+- 视频处理能力
+- 更好的音频重采样
+
+### 组合功能
+
+您可以组合多个功能：
+
+```bash
+# NVIDIA GPU 上的最佳性能
+cargo build --release --features "cuda,flash-attn"
+
+# 带音频支持的 Apple Silicon
+cargo build --release --features "metal,ffmpeg"
+
+# 启用所有功能
+cargo build --release --features "cuda,flash-attn,ffmpeg"
+```
+
+## 验证安装
+
+安装后，验证 AHA 是否正常工作：
+
+```bash
+# 检查版本
+./target/release/aha --version
+
+# 列出支持的模型
+./target/release/aha list
+
+# （如果已安装到 PATH）
+aha --version
+aha list
+```
+
+`aha list` 的预期输出：
+
+```shell
+#Supported models:
+ 
+ Available models:
+
+Model Name                     ModelScope ID
+-----------------------------------------------------------
+minicpm4-0.5b                  OpenBMB/MiniCPM4-0.5B
+qwen2.5vl-3b                   Qwen/Qwen2.5-VL-3B-Instruct
+qwen2.5vl-7b                   Qwen/Qwen2.5-VL-7B-Instruct
+qwen3-0.6b                     Qwen/Qwen3-0.6B
+qwen3asr-0.6b                  Qwen/Qwen3-ASR-0.6B
+qwen3asr-1.7b                  Qwen/Qwen3-ASR-1.7B
+qwen3vl-4b                     Qwen/Qwen3-VL-2B-Instruct
+qwen3vl-4b                     Qwen/Qwen3-VL-4B-Instruct
+qwen3vl-8b                     Qwen/Qwen3-VL-8B-Instruct
+qwen3vl-32b                    Qwen/Qwen3-VL-32B-Instruct
+deepseek-ocr                   deepseek-ai/DeepSeek-OCR
+hunyuan-ocr                    Tencent-Hunyuan/HunyuanOCR
+paddleocr-vl                   PaddlePaddle/PaddleOCR-VL
+rmbg2.0                        AI-ModelScope/RMBG-2.0
+voxcpm                         OpenBMB/VoxCPM-0.5B
+voxcpm1.5                      OpenBMB/VoxCPM1.5
+glm-asr-nano-2512              ZhipuAI/GLM-ASR-Nano-2512
+fun-asr-nano-2512              FunAudioLLM/Fun-ASR-Nano-2512
+
+```
+
+## 故障排除
+
+### 构建错误
+
+#### "error: linking with cc failed"
+
+通常表示缺少系统依赖。
+
+**解决方案**：为您的平台安装所需的构建工具（见平台特定说明）。
+
+#### "error: CUDA not found"
+
+启用了 CUDA 功能但未安装 CUDA toolkit。
+
+**解决方案**：
+- 从 https://developer.nvidia.com/cuda-downloads 安装 CUDA toolkit
+- 或不使用 CUDA 构建：`cargo build --release`
+
+#### "error: Metal not available"
+
+启用了 Metal 功能但硬件不支持。
+
+**解决方案**：
+- 确保您在 Apple Silicon 上
+- 或不使用 Metal 构建：`cargo build --release`
+
+### 运行时错误
+
+#### "error while loading shared libraries"
+
+缺少运行时库。
+
+**解决方案**：安装所需的库（见平台特定说明）。
+
+#### "Out of memory"
+
+模型对于可用 RAM/VRAM 太大。
+
+**解决方案**：
+- 使用更小的模型
+- 关闭其他应用程序
+- 启用 GPU 加速以获得更好的内存效率
+
+#### "Model download failed"
+
+网络问题或磁盘空间不足。
+
+**解决方案**：
+- 检查互联网连接
+- 确保 `~/.aha/` 中有足够的磁盘空间
+- 重试：如果中断，下载将恢复
+
+### 性能问题
+
+#### 推理速度慢
+
+**解决方案**：
+1. 启用 GPU 加速：`--features cuda` 或 `--features metal`
+2. 启用 Flash Attention：`--features "cuda,flash-attn"`
+3. 使用更小的模型
+4. 检查是否正在使用 GPU（应在监控工具中看到 GPU 使用率）
+
+#### CPU 使用率高
+
+**解决方案**：
+1. 启用 GPU 加速
+2. 减少批处理大小
+3. 使用低精度模型
+
+## 系统要求
+
+*注：模型不同需求不同的硬件和软件，供参考*
+### 最低要求
+
+- **CPU**：x86_64 或 ARM64
+- **RAM**：8 GB（推荐 16 GB）
+- **磁盘**：10 GB 用于模型（因模型而异）
+- **OS**：Linux、macOS 或 Windows
+
+### 推荐要求
+
+- **CPU**：现代多核处理器
+- **RAM**：24 GB 或更多
+- **GPU**：NVIDIA GPU（带 CUDA）或 Apple Silicon
+- **磁盘**：具有 50+ GB 可用空间的 SSD
+- **OS**：Linux (Ubuntu 22.04+) 或 macOS (Monterey+)
+
+## 模型大小
+
+流行模型的**大致**下载大小：
+
+| 模型 | 大小 | RAM 使用 |
+|------|------|----------|
+| qwen3-0.6b | ~1.2 GB | ~2 GB |
+| qwen3vl-2b | ~4 GB | ~6 GB |
+| qwen3vl-8b | ~16 GB | ~20 GB |
+| qwen3vl-32b | ~64 GB | ~70 GB |
+
+## 后续步骤
+
+成功安装后：
+
+1. 阅读[快速入门指南](./getting-started.zh-CN.md)
+2. 下载您的第一个模型：`aha download -m qwen3-0.6b`
+3. 启动服务：`aha cli -m qwen3-0.6b`
+4. 探索 [API 参考](./api.zh-CN.md)
+
+## 另见
+
+- [快速入门](./getting-started.zh-CN.md) - 快速入门指南
+- [CLI 参考](./cli.zh-CN.md) - 命令行使用
+- [API 参考](./api.zh-CN.md) - REST API 文档
+- [开发指南](./development.zh-CN.md) - 贡献指南
@@ -0,0 +1,92 @@
+# Supported Models
+
+aha supports a growing collection of state-of-the-art AI models across multiple domains.
+
+## Text Generation
+
+| Model | Parameters | Description | Use Case |
+|-------|-----------|-------------|----------|
+| **Qwen2.5-7B** | 7B | General-purpose LLM | Chat, reasoning, code |
+| **Qwen3** | Various | Latest generation | Advanced reasoning |
+| **MiniCPM4** | 4B | Efficient lightweight | Edge deployment |
+
+## Vision & Multimodal
+
+| Model | Type | Description | Resolution |
+|-------|------|-------------|------------|
+| **Qwen2.5-VL** | Vision-Language | Image understanding | Up to 1024x1024 |
+| **Qwen3-VL** | Vision-Language | Enhanced multimodal | Up to 1536x1536 |
+| **MiniCPM-V** | Vision-Language | Lightweight vision | Up to 768x768 |
+
+## Speech Recognition (ASR)
+
+| Model | Language | Real-time | Speed |
+|-------|----------|-----------|-------|
+| **FunASR-Nano** | Chinese/English | Yes | 16x realtime |
+| **GLM-ASR-Nano** | Chinese/English | Yes | 32x realtime |
+
+## OCR
+
+| Model | Languages | Type | Strength |
+|-------|-----------|------|----------|
+| **PaddleOCR-VL** | 80+ | Lightweight | General documents |
+| **Hunyuan-OCR** | Chinese | Deep learning | Complex layouts |
+| **DeepSeek-OCR** | Multi | Scene text | Natural images |
+
+## Audio Processing
+
+| Model | Type | Description |
+|-------|------|-------------|
+| **VoxCPM** | Voice Codec | Neural audio codec |
+| **RMBG-2.0** | Background Removal | Voice isolation |
+
+## Model Formats
+
+All models are served in optimized ONNX format for:
+
+- **Cross-platform compatibility** - Windows, macOS, Linux
+- **CPU acceleration** - AVX2, NEON, SIMD
+- **Edge deployment** - No GPU required
+- **Fast inference** - Optimized runtime
+
+## Model Selection
+
+aha automatically selects the best model for each task. To override:
+
+```bash
+aha chat "Hello" --model qwen2.5-7b
+aha vision --model qwen2.5-vl "Describe this" --image img.jpg
+aha asr --model fun-asr-nano audio.wav
+```
+
+## Model Sources
+
+Models are sourced from:
+
+- [Hugging Face](https://huggingface.co) - Primary model hub
+- [ModelScope](https://modelscope.cn) - Chinese model hub
+- [GitHub Releases](https://github.com) - Backup releases
+
+## Adding New Models
+
+See [Development Guide](./development.md) for instructions on adding new model integrations.
+
+## Model Updates
+
+Models are regularly updated. Check the [releases](https://github.com/yourusername/aha/releases) for the latest versions.
+
+## License
+
+Each model has its own license. Please review the model's license before use in production.
+
+## Performance Benchmarks
+
+Approximate inference speeds on CPU (M1 Pro):
+
+| Model | Task | Tokens/sec |
+|-------|------|------------|
+| Qwen2.5-7B | Text | 25-35 |
+| Qwen2.5-VL | Vision | 20-30 |
+| FunASR-Nano | ASR | 200-500x |
+
+*Benchmarks vary by hardware and input size.*
@@ -0,0 +1,92 @@
+# 支持的模型
+
+aha 支持多个领域的最先进 AI 模型集合。
+
+## 文本生成
+
+| 模型 | 参数量 | 描述 | 使用场景 |
+|------|--------|------|----------|
+| **Qwen2.5-7B** | 7B | 通用大语言模型 | 对话、推理、代码 |
+| **Qwen3** | 多种 | 最新一代 | 高级推理 |
+| **MiniCPM4** | 4B | 高效轻量级 | 边缘部署 |
+
+## 视觉与多模态
+
+| 模型 | 类型 | 描述 | 分辨率 |
+|------|------|------|--------|
+| **Qwen2.5-VL** | 视觉语言 | 图像理解 | 最高 1024x1024 |
+| **Qwen3-VL** | 视觉语言 | 增强多模态 | 最高 1536x1536 |
+| **MiniCPM-V** | 视觉语言 | 轻量视觉 | 最高 768x768 |
+
+## 语音识别 (ASR)
+
+| 模型 | 语言 | 实时 | 速度 |
+|------|------|------|------|
+| **FunASR-Nano** | 中/英 | 是 | 16x 实时 |
+| **GLM-ASR-Nano** | 中/英 | 是 | 32x 实时 |
+
+## OCR
+
+| 模型 | 语言 | 类型 | 优势 |
+|------|------|------|------|
+| **PaddleOCR-VL** | 80+ | 轻量级 | 通用文档 |
+| **Hunyuan-OCR** | 中文 | 深度学习 | 复杂布局 |
+| **DeepSeek-OCR** | 多语言 | 场景文字 | 自然图像 |
+
+## 音频处理
+
+| 模型 | 类型 | 描述 |
+|------|------|------|
+| **VoxCPM** | 语音编解码器 | 神经音频编解码 |
+| **RMBG-2.0** | 背景移除 | 语音隔离 |
+
+## 模型格式
+
+所有模型均采用优化的 ONNX 格式，提供：
+
+- **跨平台兼容** - Windows、macOS、Linux
+- **CPU 加速** - AVX2、NEON、SIMD
+- **边缘部署** - 无需 GPU
+- **快速推理** - 优化运行时
+
+## 模型选择
+
+aha 会自动为每个任务选择最佳模型。如需覆盖：
+
+```bash
+aha chat "你好" --model qwen2.5-7b
+aha vision --model qwen2.5-vl "描述这个" --image img.jpg
+aha asr --model fun-asr-nano audio.wav
+```
+
+## 模型来源
+
+模型来源：
+
+- [Hugging Face](https://huggingface.co) - 主模型中心
+- [ModelScope](https://modelscope.cn) - 中文模型中心
+- [GitHub Releases](https://github.com) - 备份发布
+
+## 添加新模型
+
+参见 [开发指南](./development.zh-CN.md) 了解添加新模型集成的说明。
+
+## 模型更新
+
+模型定期更新。查看 [releases](https://github.com/jhqxxx/aha/releases) 获取最新版本。
+
+## 许可证
+
+每个模型都有自己的许可证。在生产中使用前请查看模型许可证。
+
+## 性能基准
+
+CPU (M1 Pro) 上的近似推理速度：
+
+| 模型 | 任务 | Tokens/秒 |
+|------|------|-----------|
+| Qwen2.5-7B | 文本 | 25-35 |
+| Qwen2.5-VL | 视觉 | 20-30 |
+| FunASR-Nano | ASR | 200-500x |
+
+*基准测试因硬件和输入大小而异。*