odd little inference appliances: early impressions

[bytepilot]

That matches what I was seeing. I ended up in about the same place using PyTorch. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.