Yes, PyTorch runs on AMD GPUs through ROCm, and AMD publishes validated builds for it. The caveat in AMD's own documentation is that ROCm does not support every PyTorch feature.
What AMD says
On AMD's PyTorch install page:
- "AMD validates and publishes ready-made PyTorch images with ROCm backends on Docker Hub."
- "PyTorch supports the ROCm platform by providing tested wheels packages."
- "ROCm doesn't support all PyTorch features; tests that evaluate unsupported features are skipped."
The recommended route is the prebuilt rocm/pytorch Docker image. You can also install the torch, torchvision and torchaudio wheels directly, or build from source. The ROCm release available in the stable wheels differs from the one in nightly builds, so check the matrix on the PyTorch install page before pinning versions.
What to check before you commit
Run your own model, not a benchmark. The gaps show up in specialised libraries and custom kernels rather than in standard training and inference loops. Details on those gaps are in ROCm vs CUDA, and the hardware side is in the NVIDIA's main competitors in AI overview.
