Idea: Keep model in the GPU memory

[linonetwo]

Seems stable difussion webui will keep model (checkpoint) in the GPU memory, so you don't need to send it again, is this correct?

And I found ComfyUI seems keep it in normal memory, and send it to GPU on each workflow execution.

Will SD webui's approach faster?

[comfyanonymous]

You can use: --highvram in ComfyUI to keep the model in GPU memory.

[linonetwo]

Basically the same

when drawing, the usage is high

when finished, it drop to 3.1GB, compare to 1-2GB when without --highvram

[linonetwo]

Anyway, I'm not sure if this affects performance, so never mind.

[WASasquatch]

I don't believe the model is stored on the "GPU". It's stored in RAM, for quick access by the GPU when computing. Least it was common practice to move to CPU (RAM) when not actually computing.