# ROCm (AMD GPU) Support Plan
## Goals
- Make AMD/ROCm support explicit and robust on Linux.
- Avoid CUDA-only optimizations that may break on ROCm.
- Provide version-agnostic installation/verification guidance and compatibility notes.
## Scope
- Device detection and reporting in backend.
- Safe defaults for dtype, attention slicing, and memory/offload behavior.
- Optional UI labels for ROCm.
- Documentation updates (README + troubleshooting).
## Plan
1) Research and confirm ROCm detection strategy
- Use torch.version.hip to detect ROCm.
- Validate that torch.cuda.is_available() is true on ROCm builds.
2) Add explicit ROCm device detection
- Update detect_device() to return "rocm" when torch.version.hip is set.
- Keep CUDA/MPS/CPU order but make ROCm distinct from CUDA.
3) Update hardware heuristics for ROCm
- Treat "rocm" like "cuda" for VRAM checks in get_vram_gb().
- Extend attention slicing logic to include ROCm VRAM thresholds.
4) Guard CUDA-only optimizations
- Gate torch.compile + SDNQ quantized matmul to NVIDIA-only unless ROCm is known good.
- Gate enable_model_cpu_offload() to NVIDIA-only unless ROCm is confirmed.
- Add log warnings when skipping these paths on ROCm.
5) Update UI device labels (optional)
- Add model_desc_*_rocm strings in static i18n JSON files.
- Ensure fallback labels remain correct if not present.
6) Documentation
- Expand README with version-agnostic ROCm install/verify guidance and links.
- Add a short compatibility matrix by ROCm major/minor (best-effort, user-reported).
- Add a troubleshooting section (common ROCm pitfalls + diagnostics).
7) Validation
- Smoke test on ROCm environment (or ask a user to confirm).
- Confirm device reporting and model selection behavior.
## Risks / Open Questions
- ROCm support varies by GPU architecture, kernel version, and ROCm version.
- torch.compile stability on ROCm depends on Triton/Inductor versions.
- Some diffusers kernels may not be ROCm-friendly yet.
- Driver/runtime/PyTorch version skew is the most common source of issues.
## Deliverables
- Code changes in hardware and engine modules.
- Updated README and any ROCm troubleshooting notes.
- Optional i18n label updates for ROCm device display.
