Memory management / Load and Unload the model

[Hamzajani153]

How does the Automatic1111 backend handle memory deallocation in VRAM and RAM when unloading a model before loading a new one? In cases where there isn’t enough VRAM to load a new model after unloading the current one, how does the system handle this situation?Tell me example code if it is present in github repo for this memory issue.

[CyrilCommando]

In my experience, it doesn't. It loads a model into RAM when it starts, keeps it there, then blows past all 20gb of your RAM loading it a second time when you click generate. Whenever you select a different model it doesn't free what's consumed by the one loaded upon launch, and proceeds to blow past 20+GB ram and starts bleeding into swapspace.