RAM OOM

[twitset]

First of all, thank you very much for the wan_ksampler node - it runs much faster than the standard sampler.

I noticed that over 40-50 generations RAM fills up and my debian terminates comfy. I have isolated the issue to the wan_ksampler node and I suspect there is a memory leak somewhere. I tried to find out where but I failed as I am not very familiar with the code. I am running comfy headless without UI and with --preview-method none.

What I suspect is one or more of these things:

• previewer objects inside prepare_callback (potentially large decoders or image objects)
• callback closures (may capture tensors and references)
• latent_image, noise, and noise_mask (multi-MB tensors)

This would be my first candidate
callback = latent_preview.prepare_callback(model_high_noise, steps)

It might be necessary to delete the callback using del

del callback

I would appreciate if you could have a look as you are much more familiar. If you don't have time, let me know and I will dig in.

[stduhpf]

Thanks for looking into this so thoroughly. I assumed python was properly garbage collected, I will take a look at this later.