I am trying to install PyTorch from https://github.com/ROCmSoftwarePlatform/pytorch, following the instructions there.
I have a server with 2 MD EPYC 7302 CPUs and 2 AMD Instinct MI100 GPUs, running Ubuntu 20.04.3. I have installed ROCm 4.3 and can compile and run sample code from https://github.com/ROCm-Developer-Tools/HIP.
I follow the instructions to compile PyTorch from source and install and I end up running into errors for the "gfx90a" target. This is where I get the idea that maybe I should not just be compiling for the default targets. I can glean from various examples that I can set this with the PYTORCH_ROCM_ARCH environment variable. rocminfo tells me that the "Name" of the GPUs is "gfx908". Is this what I should use for target architecture?
The more general question: where do I look up which possible target architectures exist? I have been browsing around various AMD compiler documentation with no luck. Where can I see which architectures specific GPU product names correspond to?
