Request: Apple Silicon / Mac Metal or MLX support for JoyCaption

[Phlypper]

Hi JoyCaption developers,
I’m using JoyCaption locally on an Apple Silicon MacBook Pro with an M4 Max GPU. My use case is accessibility: I am building a local video-to-audio-description workflow for blind users, where JoyCaption captions sampled video frames and the app turns those captions into spoken narration.
The current Hugging Face / LLaVA-style JoyCaption model can run on Mac in some setups, but Apple Silicon support is still difficult compared with CUDA. PyTorch MPS / Metal compatibility, memory use, dtype handling, and speed can be challenging, especially with newer or larger JoyCaption versions.
I wanted to ask whether you would consider one of the following for future JoyCaption releases:
Better tested Apple Silicon support through PyTorch MPS / Metal
An MLX-compatible version for Apple Silicon
A Core ML export or conversion guide
A smaller or quantized Mac-friendly model
Clear Mac setup notes for Transformers users
This would be very helpful for accessibility-focused local apps. Many blind users and Mac users would benefit from a strong local image/video captioning model that does not require NVIDIA CUDA hardware.
My current system:
Apple Silicon MacBook Pro, M4 Max
macOS
PyTorch MPS / Metal
Hugging Face Transformers
Local JoyCaption model folder
Thank you for building JoyCaption and making it available to the community.