AFAIK the model can’t be quantized during backprop, so right there you’d need a ton of RAM.

Backprop is faster bc it can be parallelized, but IIRC you need to hold an entire copy of the model for each backprop process.