Question: memory management

got a 2080 in front of me and about to get a 4090. I noticed comfy loads/unloads model as needed, probably to save ram. This can be infuenced by careful placement of nodes in the 'correct' (reverse) order.

Is there information about how comfy handles vRam? i had multiple situation where I got OOMs on 8gb vram doing a resize through ergan in pixel space, but on a second run it worked fine.

Understanding how comfy handles loaders/models/lora/etc priority would help us create more efficient workflows, in other words ensure we are operating on the correct' model for a branch , then switch accross a different one to not have this constant swapping, especially while batching.

Finally, is it fair to say that comfy will attempt to 'cache' models in vram if we have enough of it? I noticed this leads to confusion when it comes to nodes like efficiency nodes, which has a 'cache model' option.

Thank you.