I'm excited to announce that the full ROCm stack moved from [testing] to [community] today. To celebrate this, PyTorch packages with ROCm support are now in [community]! A simple code to check if the ROCm backend works on your machine is
if torch.cuda.is_available() and torch.version.hip: d = torch.device('cuda') a = torch.rand(1, 2).to(d) print(a + 0) print("TESTS PASSED!") else: print("ROCm not supported with PyTorch installation")
After you installed python-pytorch-rocm or the optimized python-pytorch-opt-rocm from [community], save the code snippet as rocm.py and run
The output should be similar to
tensor([[0.9949, 0.5596]], device='cuda:0') TESTS PASSED!