Storage, checkpoints and data transfer

A fine-tune stores three things: the base model, the dataset and the checkpoints. The base model and dataset are fixed in size. The checkpoints grow with every save.

A full fine-tune saves the whole model at each checkpoint. A LoRA run saves only the added weights, which are far smaller. So the method changes the storage line as well as the GPU line.

Storage is billed by the gigabyte over time, whether or not a GPU is attached. A volume left behind after the run is billed until you delete it. This is the most common surprise on a fine-tune bill.

Some storage belongs to the machine and disappears with it. Some persists on its own. Know which one your checkpoints are on before you stop the machine.

Moving the dataset in is usually free. Moving the finished model out is sometimes billed by the gigabyte. The rate is rarely on the GPU pricing page. Look under network or storage pricing, and add the size of your final model to the estimate.

Keep the last few checkpoints and the best one, and delete the rest as the run goes. Download the final model when the run ends. Then delete the volume, or keep only what the next run needs.

Loading the base model and dataset takes time on every fresh machine, and that time is GPU time. A volume that persists saves it on every rerun. Weigh the storage cost against the GPU minutes saved.

On Runcrate, volumes keep files between instances, so the base model, the dataset and the checkpoints are there when the next GPU starts. One prepaid balance pays for GPUs, models and storage, so both lines draw on the same balance.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.