Co-composition uses more memory than normal sampling about jukebox HOT 2 OPEN

I decided to go ahead and try upsampling from there, and CUDA then failed to allocate enough memory unless I deleted top_prior as suggested for a hosted runtime in this block

Unfortunately, there's no workaround. These priors are large and unused levels have to be unloaded

This worked for a while, but the continuation failed after I had reached 64 seconds of audio

Decreasing the sampling batch size could help

from jukebox.

I decided to go ahead and try upsampling from there, and CUDA then failed to allocate enough memory unless I deleted top_prior as suggested for a hosted runtime in this block

Unfortunately, there's no workaround. These priors are large and unused levels have to be unloaded

The curious bit is that this unloading is not necessary when I sample even larger amounts of audio, so long as I am running normal sampling via sample.py and not co-composition as implemented in the jupyter notebook. In both cases I am running on local hardware.

So it seems that for some reason the priors are oversized specifically when running co-composition.

from jukebox.