Revamp iOS documentations

[mergennachin]

📚 The doc issue

Currently there are a few different places for iOS that we wrote over the lifecycle of ExecuTorch development phase (MVP, alpha, beta etc) with different purposes. Some were to showcase functionality (e.g., MV3 and Llama can actually run iOS device), some were to showcase delegate capabilities (XNNPACK, CoreML, MPS integration)... It's time to step back and see from a developer's perspective, and make sure the sum of all the parts make sense and coherent.

• demo app tutorial (https://pytorch.org/executorch/stable/demo-apps-ios.html)

• cross compilation iOS section (https://pytorch.org/executorch/stable/runtime-build-and-cross-compilation.html#ios)

• Llama iOS demo (https://pytorch.org/executorch/stable/llm/llama-demo-ios.html)

  • XNNPACK (https://github.com/pytorch/executorch/blob/main/examples/demo-apps/apple_ios/LLaMA/docs/delegates/xnnpack_README.md)
  • MPS (https://github.com/pytorch/executorch/blob/main/examples/demo-apps/apple_ios/LLaMA/docs/delegates/mps_README.md)

• Integrating and Running ExecuTorch on Apple Platforms (https://pytorch.org/executorch/stable/apple-runtime.html)

• Running with CoreML (https://pytorch.org/executorch/stable/build-run-coreml.html)

---

User journey:

Here's a typical developer user journey from an engineer who's just looking at ExecuTorch for the first time. Their ultimate goal is to build an AI powered iOS app which is performant enough.

• Let's assume modeling is done by different researchers, but still have relatively close, constant communication

Part 1:

• Set up
  -- Clone executorch repo and do minimal setup/installation (./install_executorch.sh --pybind coreml xnnpack)
  -- Ideally should just be able to do pip install executorch in the near future.
• Get a model from somewhere
  -- Download from HuggingFace
  -- Or existing example models e.g., MV3 and Llama3
• Export a model (use XNNPACK as default)
  -- Either export file
  -- Download existing pte file from somewhere like HuggingFace
• Run the model desktop in python via pybind

Part 2:

• Build and run an app. Use swift prebuilts. Not compiling from source
  -- Build and run custom app out-of-source tree.
  -- Use existing demo apps (MV3 and Llama3/Llava) as an example usage
• Further optimizations (CoreML and MPS)

Suggest a potential alternative/fix

No response

cc @byjlw @shoumikhin

[bluejack]

Can confirm. Note that while building from stable release is surely going to be best advice, and you take your lumps when building from the head of the repo, the current instructions don't fully work in several key areas for ios builds when building from current state of the repo:

1. Instructions that specify ./install_requirements.sh do not build executorch. ./install_executorch.sh is also required.

2. the ./build/build_apple_frameworks.sh does not appear to work as documented

3. this recommended guideline does not take a --clean flag. Not sure if that did anything other than rm -rf on cmake-out, but it doesn't do whatever it did.

# Clean cmake cache directory (cmake-out). It's a good practice to do this
# whenever cloning or pulling the upstream repo.
bash install_requirements.sh --clean

[swolchok]

Instructions that specify ./install_requirements.sh do not build executorch. ./install_executorch.sh is also required.

FYI, docs that are supposed to reflect the current state of the repo are available at https://pytorch.org/executorch/main/ ; you can get there by clicking the chevron next to 0.4 in the left sidebar. Bug reports against the current state will be easier for us to take action on.

[bluejack]

I did not notice that chevron! Thanks for the tip. (1) && (3) are resolved by this; and (2) was an oddity in my own environment!

Still support a "developer journey" approach to documentation!