- Add Dockerfile.rocm based on rocm/pytorch:rocm7.2.3
- Add docker-compose-rocm.yaml with /dev/kfd + /dev/dri passthrough
- Add ROCm Docker section to README with quick start and notes
- Use onnxruntime_migraphx from repo.radeon.com for GPU-accelerated ONNX
- Pin starlette<1.0.0 to work around #2762
- Set ROCBLAS_USE_HIPBLASLT=0 for RDNA4 stability
Tested: GPT-SoVITS v2 inference on Radeon AI PRO R9700 (gfx1201)
with ROCm 7.2.3, PyTorch 2.9.1, fp16. WebUI and API both working.
