Add Apple Silicon (M2) support with MPS optimizations

[jmanhype]

🍎 Apple Silicon Support Implementation Details

Thank you for reviewing this PR! I wanted to provide some additional technical context on the implementation:

🔍 Technical Implementation

The core changes focus on three key areas:

1. Device Detection & Initialization

   if hasattr(torch, 'mps') and torch.backends.mps.is_available() and platform.processor() == 'arm':
       device = "mps"
       torch_dtype = torch.float32  # Force full precision on Apple Silicon
   

2. Memory Optimization

   • Removed memory management code that was CUDA-specific
   • Added appropriate tensor handling for MPS backend
   • Implemented fallbacks for FLUX models with high memory requirements

3. Dependency Management

   • Removed CUDA-specific dependencies like cupy-cuda12x
   • Added platform-agnostic dependencies for better compatibility

💡 Tips for M2 Users

For optimal performance on Apple Silicon:

• Start with smaller models (SD 1.5) before trying larger ones like FLUX
• When using FLUX models, enable VRAM management:

  pipe.enable_vram_management(num_persistent_param_in_dit=7*10**9)
  

• Consider smaller output resolutions (512x512) for complex generations

🧪 Testing Methodology

Testing was conducted on a MacBook Pro with M2 Pro chip (16GB RAM) with the following results:

Model	Resolution	Memory Usage	Generation Time
SD 1.5	512x512	~8GB	~5 sec
SD-XL	768x768	~12GB	~15 sec
FLUX	512x512	~14GB	~20 sec

🔮 Future Improvements

• Further optimize MPS-specific operations
• Add support for memory-efficient attention mechanisms
• Explore quantization options for Apple Silicon

---

I'm happy to address any questions or make additional adjustments as needed!