Support MLX / Apple Silicon (MPS)

[catch6]

Hi FireRedTeam team!

Thanks for open-sourcing FireRedTTS3. I noticed the current requirements.txt pins flash_attn==2.8.3, which is CUDA-only and blocks running on Apple Silicon (Mac M-series).

Feature request: Add MLX or MPS support so FireRedTTS3 can run locally on Apple Silicon, similar to how the community has ported FireRedTTS / other TTS models to mlx-audio.

A few options that would help:

1. Make flash_attn optional and fall back to a standard attention implementation on non-NVIDIA devices.
2. Provide MLX conversion scripts for the checkpoints.
3. Document the minimal steps to run on an MPS backend.

Happy to contribute if you can point me to where the attention is wired in. Thanks!

[catch6]

Clarifying priority: MLX is the preferred target over MPS on Apple Silicon.

Why MLX:

• It is Apple\s purpose-built framework for the unified-memory M-series chips, with Metal-optimized ops that generally run notably faster than PyTorch\s MPS backend.
• PyTorch MPS still silently falls back to CPU for many under-optimized ops, so real-world throughput is often much lower.

So if resources allow, MLX conversion would be the most valuable option. Thanks!

[Talpik]

#6 here