09-26-2026, 01:01 AM
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.
still benchmarking everything
