WAN 2.2 + MOE: OOM with 6gb card (Like its loading high noise then low noise without unload)

[ivanlr-cpu]

For WAN 2.2, I use two ksamplers (the classic way). Yes, it works with a 6gb card because I create a 480x480 video with 6fps. Comfyui can clean the used memory between HIGH and LOW phases.
I wanted to use MOE with the Last image (LF), and in fact it works well, by loading two safetensors (HIGH and LOW noise). It runs okay for the first steps (HIGH noise phase), but then, before starting last 3 steps (LOW noise phase), it gives an OOM.

Perhaps, as a possible feature, have a way to use 'two' MOEs, with "Clean VRAM used' or other specialized node?

[axymeus]

This. This is unfortunately necessary since Comfy isn't smart enough to manage memory on its own most of the time. A simple parameter to force unload the low noise model would be great.

[stduhpf]

I might not have the time to look into that, but if you want to submit a PR with that change, then I could review and merge it.

Until then, you can use the SplitSigmasAtT node that is included in this repo (see the "custom" workflows to figure out how it works). This will probably allow you to unload the models just like in the classic workflow, while having the optimal switching step.