Open murphylanga opened 3 months ago
Is there a way to clear the memory (VRAM) after a workflow run?
There's an API route for it (post /free
with { "unload_models": true, "free_memory": true }
, and Swarm has a button in the Server tab for that. I don't think comfy itself currently has a button for it
Or possibly even during the run?
that'd require a node or CLI arg. There are relevant CLI args listed here: https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/cli_args.py#L103-L109
Generally by default comfy is designed to try to avoid running out of memory and will automatically unload on the fly as needed. This might not work as intended with some custom nodes if they load models and don't use that system.
Thank you very much for your answer
Is there a way to clear the memory (VRAM) after a workflow run? Or possibly even during the run? I am using a laptop RTX3070 graphics card. And I use it to run very large workflows. I have optimized the workflows to such an extent that I can switch all sorts of things on and off. For some things (e.g. FaceSwap or Controlnet) ComfyUi then switches to lowvram mode. That wouldn't be too bad, because I have to be patient here, with less VRAM. But ComfyUI does not find its way back, even if the memory-hungry nodes are switched off. For example, upscaling could also be performed without Controlnet. Unfortunately, I then always have to reload Comfy, but then I can also upscale the image without any problems. Maybe these CustomNodes are a first approach: https://github.com/ntdviet/comfyui-ext I have installed the LatentGarbageCollector node from there and it looks like there is enough free memory to work without lowvram. If it is not possible during the workflow run, should it somehow be possible to switch comfy back to normalvrm before a new run? Or when you reload a workflow? So far, unfortunately, only restart works for me. I would be very happy if someone knows or finds a solution.