One can keep all tensors in the RAM, just push whatever needed to GPU VRAM, basically limited by PCIe speed. Or some intelligent strategy with read-ahead from SSD if one's RAM is limited. There are even GPUs with their own SSDs.