Skip to content

Latest commit

 

History

History
19 lines (15 loc) · 2.04 KB

EXAMPLES.md

File metadata and controls

19 lines (15 loc) · 2.04 KB

Examples Directory

For environment setup and package build instructions, read INSTALL.md. The sections below summarise each supported example and link to the detailed README in that example's folder.

Example Summary Location
Flux + Quantized Attention End-to-end image generation benchmarks (INT8/INT4) using the Flux diffusion model and the PyTorch custom op. examples/flux/README.md
Objective‑C Bridge Minimal host app demonstrating how to call the C FFI directly from Objective‑C. examples/objc/README.md
Python FFI ctypes bindings, unit tests, and CLI utilities for invoking the Metal kernels from Python. examples/python-ffi/README.md
PyTorch Custom Op Drop-in scaled_dot_product_attention replacement for PyTorch with quantized + sparse support. examples/pytorch-custom-op-ffi/README.md
Rust FFI Rust crate that links against the C API and exercises forward attention with benchmarks. examples/rust-ffi/README.md
Swift Reference Snippets Standalone Swift programs illustrating pure-Swift quantized attention and bridge benchmarks. examples/swift/README.md

Usage Notes

  • Examples rely on the Swift package being built (swift build) and, where applicable, the quantized Metal shaders compiled. Always run through the steps in INSTALL.md first.
  • Some examples (Flux, PyTorch) build a native extension that expects DYLD_LIBRARY_PATH to include .build/…/release. The README for each example calls this out explicitly.
  • When experimenting with quantized attention, pay attention to the expected scale/zero-point metadata. The Python and Flux examples show how to set mfa_set_scale_arrays or pass per-tensor parameters.
  • Additional examples or language ports should add a README.md and extend this table so users can discover them.