| --- |
| license: apache-2.0 |
| tags: |
| - ardy |
| - motion-generation |
| - macos |
| - apple-silicon |
| - mps |
| - pytorch |
| --- |
| |
| # ARDY on Apple Silicon |
|
|
| Open-source macOS compatibility tools for running NVIDIA ARDY locally through |
| PyTorch Metal (MPS). This repository does not contain or redistribute model |
| weights. |
|
|
| [Watch the Mac/MPS demo](demo/ardy-macos-mps-demo.mp4) |
|
|
|  |
|
|
| ## How it works |
|
|
| 1. Download the official ARDY source. |
| 2. Run the `ardy-macos-tools` installer on an Apple Silicon Mac. |
| 3. Accept Meta Llama 3 access and authenticate with Hugging Face. |
| 4. ARDY downloads its official checkpoint and text-encoder weights directly. |
| 5. The compatibility service runs generation through Apple's MPS backend. |
|
|
| The installer, source code, checksums, and complete instructions live in the |
| [GitHub project](https://github.com/trev222/ardy-macos-tools). Model files, |
| caches, tokens, and the PocketAI application are not included. |
|
|
| ## Measured Mac performance |
|
|
| In two desktop sessions on an Apple M5 Max with 128 GB unified memory, a |
| four-second, 80-frame motion generated in 1.90–3.62 seconds after the models |
| were resident. Peak MPS driver allocation was 15.23 GiB. Cached cold loading |
| took 145.6–260.3 seconds and varied with desktop conditions. |
|
|
| See the [full reproducible benchmark](https://github.com/trev222/ardy-macos-tools/blob/main/benchmarks/APPLE_M5_MAX_128GB.md) |
| for individual runs, CPU measurements, methodology, and caveats. |
|
|
| Built with Meta Llama 3. ARDY checkpoints remain subject to the NVIDIA Open |
| Model Agreement. See the project's third-party notices before use. |
|
|
