Pytorch gpu parallelism vram for nvidia smi for implementation with jupyter notebook and miniconda for deforum

Question

I am wondering if it is possible to span GPU VRAM and compute power using pytorch and nvidia smi. The spanning would work for the deforum collab as a local instance using jupyter notebook and miniconda. How would I implement pytorch to use all gpu and VRam available and patch that in to miniconda controlling the local collab of deforum so that their collab used all gpu resources? But I haven't found too many threads dealing with parallelism and nvidia smi. Any guidance here of possibilities would be awesome.