ROCm and mps

[Aatricks]

This pull request adds comprehensive support and documentation for running LightDiffusion-Next on AMD GPUs (ROCm) and Apple Silicon (Metal/MPS), alongside NVIDIA GPUs. It introduces a new setup guide, expands installation instructions, and updates device detection and optimization logic throughout the codebase to handle ROCm and MPS platforms. There are also improvements to documentation for new features and platform-specific optimizations.

Platform Support and Documentation:

• Added a detailed rocm-metal-support.md guide covering setup, compatibility, feature support, and troubleshooting for AMD ROCm and Apple Silicon MPS platforms.
• Installation instructions in README.md and docs/installation.md now explicitly mention AMD (ROCm) and Apple Silicon (MPS) support, including links to the new setup guide and platform-specific requirements. [1] [2] [3]

Device Detection and Backend Logic:

• Introduced is_rocm() detection and updated device logic in Device.py to properly enable/disable features (e.g., SageAttention, SpargeAttn, xformers, PyTorch attention) based on ROCm or MPS presence. This includes correct dtype selection and memory/cache management for each platform. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

Documentation Enhancements:

• Added/updated documentation for new and existing optimizations, including section anchors and clarifications for SageAttention, SpargeAttn, CFG Samplers, and Multi-Scale Diffusion. [1] [2] [3]

These changes collectively make LightDiffusion-Next more accessible and performant across a wider range of hardware, with clear guidance and robust platform detection.

References: [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17]