Tutti Documentation

Tutti is a high-performance multi-accelerator runtime system designed to enable efficient heterogeneous computing across diverse GPU architectures. It provides unified memory management, kernel portability, and seamless integration with popular AI frameworks.

Contents:

Contents:

• Tutti 入门指南 / Getting Started with Tutti
  • 0. 一分钟看懂 Tutti / What Tutti is in one minute
  • 1. 需要什么机器 / Hardware & topology requirements
  • 2. 安装依赖 / Install dependencies
  • 3. 编译 / Build
  • 4. 严格顺序的启动 / Strict bring-up order
  • 5. 运行示例 / Run the example
  • 6. 用 nsys 看预期结果（可选）/ Profile with nsys (optional)
  • 7. 常见坑 / Common pitfalls
  • 8. 脚本现状 / Script status（截至本机核对）
  • 9. 更多资料 / Further reading
• Architecture
  • Tutti System Architecture
  • Key Designs — Performance & Low Overhead
  • snvme 设计依据：为什么需要一个定制的 NVMe 驱动
• Design Documents
  • Tutti 多加速器、多盘与 NVMe daemon 设计
  • Tutti Runtime 配置、资源与装配流程
  • Kernel Portability — the snvme module across Linux versions and GPU vendors
  • Backend SPI — the DataPath / Resolver / Binding contracts
• Tutorials and Integration Guides
  • Tutti 接入 vLLM 设计文档
  • Tutti × vLLM 操作手册
  • AI 编排协作协议 — 主 Session 完整指令文档
  • 第一部分：你的角色
  • 第二部分：共享目录协议
  • 第三部分：你的工作循环
  • 第四部分：任务卡模板
  • 第五部分：回执模板（子 session 填写，你验收时读）
  • 第六部分：子 Session Prompt（你维护 ai-orch/prompts/sub-session.md，用户贴给各子 session）
  • 第七部分：看板模板（你维护 ai-orch/BOARD.md）
• Build & SNVMe Testing Guide
• Part 1 — Build
  • 1.1 Prepare the environment
  • 1.2 Build Tutti
  • 1.3 Install the snvme kernel module
  • 1.4 Build the test code
• Part 2 — SNVMe Testing
  • Mental model
  • Step 0 — Pre-flight
  • Step 1 — (Re)build the module (only after editing driver code)
  • Step 2 — Rebuild / reload the module (only after editing driver code)
  • Step 3 — Build the test binaries
  • Step 4 — Pick a test device
  • Step 5 — SAFE tests (no bind, won’t disturb any disk)
  • Step 6 — DESTRUCTIVE tests (bind required)
  • Step 7 — GPU tests (only if the NVIDIA driver is loaded)
  • Step 8 — Cleanup & recovery
• Extending Tutti: Adding a Resolver + Binding + DataPath
  • What you need to create
  • What you must NOT modify
  • A second example: striped (multi-device, fused-kernel submission)
  • Step-by-step (memfs example)
  • How the Runtime wires it together
  • Checklist
• GPU Vendor Porting Guide (Tutti cuda_like Framework)
  • Audience
  • The three layers
  • Layer A — Vendor selection framework
  • Layer B — Kernel primitives macro layer
  • Layer C — Direct includes
  • Kernel module P2P adaptation (Metax)
  • Verification
  • Reference: host.h as the canonical shim
  • Quick start for Metax (MUSA / MACA)
• Tutti local-NVMe hardware contract tests
  • 1. 运行前提
  • 2. Test
  • 3. 阶段 3 daemon allocation 补充
• tutti_daemon 启动与部署指南
  • 1. 启动流程概览
  • 2. 环境准备
  • 3. 查找目标 NVMe 及其 PCI 地址
  • 4. 检查或创建 ext4 文件系统
  • 5. cuda-module 构建和 module 前置条件
  • 6. 创建本机 YAML 配置
  • 7. 使用指定 YAML 启动 daemon
  • 8. 启动后生成的设备和目录
  • 9. 优雅停止
  • 10. 阶段 3 canonical 配置与资源分配补充

Key Features

• Multi-Accelerator Support: Run workloads across NVIDIA, AMD, and other GPU vendors

• Unified Memory Management: Intelligent memory allocation and data movement

• Kernel Portability: Write once, run on multiple GPU architectures

• Framework Integration: Seamless integration with vLLM, PyTorch, and other frameworks

• High Performance: Optimized for real-world inference and training workloads

Links

• GitHub Repository: https://github.com/xpu-io/Tutti

• Paper: Tutti: A Multi-Accelerator Runtime System

• License: See LICENSE file in the repository

Version: 0.1 (0.1.1)

Indices and tables

• Index

• Module Index

• Search Page