Documentation Site Map

[GregoryComer]

With the ongoing work on documentation updates (#8178), we are overhauling the navigation for the docs. I'm focusing primarily on the getting started flow, mobile developer experience, and backends initially, with an intent to follow up on LLM-related material, quantization, and advanced reference material in the near future.

Proposed Layout

• Introduction
  • Describe what ExecuTorch is and why users would want to use it. Provide high-level context on user flow and concepts. Can largely be left as-is. We can do a cleanup pass on it, but I think it’s mostly what it should be, content-wise.
• Usage
  • Getting Started with ExecuTorch
    • Intended to provide the user with all of the steps needed to lower and run a model for 90% of common use cases, including inline code snippets. Use a common model (maybe image classifier) in examples. Should cover lowering, Android, iOS, and C++ APIs. Links users to top-level export + platform integration docs.
• Model Export and Lowering
  • Top-level export documentation. Cover end to end export flow, including backend selection, dynamic shapes, serializing to a .pte file, and using pybinding APIs to validate the model.
• Using ExecuTorch on Android
  • Top-level platform docs for runtime integration on Android. Focus on Java APIs and acquiring the ET libraries though Maven/Gradle. Cover library installation and runtime APIs for loading + running a model using Java runtime APIs. Briefly cover relevant delegates (XNN + Vulkan + NPU).
• Using ExecuTorch on iOS
  • Top-level platform docs for runtime integration on iOS. Focus on Objective-C/Swift APIs and acquiring the ET libraries though Swift PM. Cover library installation and runtime APIs for loading + running a model using runtime APIs. Briefly cover relevant delegates (XNN + Core ML + MPS).
• Using ExecuTorch with C++
  • Describe high-level and low-level runtime APIs, and when to use each. Cover CMake build integration and relevant targets.
• Runtime Integration
  • Cover selective build and logging / PAL integration. Briefly cover kernel and backend registration.
• Profiling and Debugging
  • Introduction to devtools and general troubleshooting flow. Provide guidance and links to deeper docs on profiling and model optimization.
• Frequently Asked Questions
  • Cover common issues with install, export, and lowering.
• Building From Source
  • Cover CMake build process and integration, build options, and cross compilation.
• Backends
  • [Top-Level Page]
    • Overview of what ET backends are and how they work (high-level)
      • How to choose a backed
      • List of backends with a brief description of each
  • [Backend-Specific Documentation]
    • Dedicated page per backend. Should include the following:
      • High-level description of the backend
      • Supported hardware / target system requirements
      • Host / development system requirements
      • Partitioner info
      • Partitioner options (for example, per-op mode or AOT compile)
      • Code snippet, including imports
      • Build info, including CMake targets
      • Devtools/profiling integration
• Runtime
• Developer Tools
• API Reference
• Python EXIR
• Python Runtime
• C++ Runtime
• Java Runtime
• Objective-C / Swift Runtime

To Place:

• Working with LLMs
• Quantization
• IR Spec
• Compiler Entry Points
• Contributing

[GregoryComer]

Excalidraw link for proposed flow: https://excalidraw.com/#json=HnX6rjMi2ZUMVrji7HrfG,_ZB0HjZZ7XdQQclWFR1SIA