Each cell gets a GPU only while it runs, and only if it needs one. Between cells your notebook holds no GPU; its variables are kept and come back for the next cell, on whichever machine it runs.

LooseLeaf reads your notebook as you type and sends a cell that uses CUDA to a GPU, and anything else to a CPU. To choose yourself, right-click a cell, then Run on → GPU or CPU; the cell then shows “set by you”, and Run on → Automatically gives the choice back. A shell command (!nvidia-smi, !python train.py) runs on a CPU unless you choose GPU.

While a cell runs, a line under it says where: “Waiting for a GPU · #2” when every GPU is busy, then “Running on GPU”.

A notebook's variables live as long as its kernel. Closing the tab keeps them for 30 min with no page open and no cell running; restarting or shutting down the kernel, or a platform restart, ends them. Files in your home are permanent. Packages you pip install stay until the kernel shuts down; a restart keeps them.

A cell that defines something that cannot be kept says so. It works while the same worker stays warm; a later cell on another worker fails with the reason, and running the defining cell again fixes it. So open files in the cell that uses them, start a server and use it in the cells right after, and train in chunks across cells. If a cell's worker crashes or goes past its memory limit, the cell says so; what earlier cells saved is kept.