[freeqaz]

Is it possible to offload some layers to CPU and still train in a reasonable amount of time?

[generalizations]

There’s also that pruning tool that was on hn in the last couple weeks. It seemed to work really well on the larger models, and could reduce size by 30-50%