We're trying to run some @PyTorch benchmarks for our open-source inference server NOS (https://github.com/autonomi-ai/nos) and wanted to get a deeper understanding of the state of ROCm compatibility vs. CUDA.