new here, what's your stack: early impressions

[rocmrunner40]

That matches what I was seeing. I ended up in about the same place using PyTorch. The biggest change for me was simplifying the stack instead of adding one more optimization layer.