Boundary and/or sigma_shift setting is ignored.

[GalaxyTimeMachine]

I am using WAN2.2 and have these settings:

The high/low noise split is always at 50% of the steps, but, with these settings, it should be 2 steps high and 8 steps low.
So, I end up with a 50/50 split with however many steps I use:

Running high noise model...
Requested to load WAN21
loaded partially 14900.072977840311 14898.994262695312 247
Patching comfy attention to use sageattn
100%|████████████████████████████████████████████████████████████████████████████████████| 5/5 [02:30<00:00, 30.16s/it]
Restoring initial comfy attention
Running low noise model...
Requested to load WAN21
loaded completely 14896.072977840311 13627.511375427246 True
Patching comfy attention to use sageattn
100%|████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:46<00:00,  9.23s/it]

[GalaxyTimeMachine]

As this repo doesn't seem to be maintained; I've fixed this in my own fork, here: https://github.com/GalaxyTimeMachine/ComfyUI-WanMoeKSampler

[stduhpf]

I beleive the high/low noise split is always at 50% of the steps with these setting because you're using the default sigma_shift=8 with boundary=0.875, which will always make the split happen near 50% of the denoising steps.

Either lowering the sigma_shift value to around 2 or increasing the boundary to approximately 0.95 (or something in between for both) would make the split happen at step 2 with 10 total steps.

Maybe I'm missing something, but at least according to the image you attached, the behavior you're describing seems like the intended behavior?

[GalaxyTimeMachine]

I was moving the sigma up and down, and the split was always at 50% of steps. Now, with my changes, the split increases along with the sigma_shift. That was probably just a bad example.

[darksidewalker]

@GalaxyTimeMachine thank you for your effort. Did your optimizations get committed into this repo or do I need to clone yours?

[stduhpf]

I didn't commit these changes (yet) because I fundamentally disagree with some of the changes implemented (for example the removal of the SplitSigmasAtT node), and also because I can't reproduce this particular issue, nor do I understand how it can happen to other people.

[darksidewalker]

Okay, I see. Maybe I need more testing if shift really is not happening.

[GalaxyTimeMachine]

@darksidewalker You can clone mine and give it a try. It has additional features for CFG control, so you can start with a higher CFG and control it to gradually decrease to 1, and how far into the gen that happens.