Going OOM on every workflow I try.
Thanks for making this int8 checkpoint but I'm having a problem where every time it reaches the first sampler It spike my memory ram+pagefile and oom's me. rtx 3090 and 64gb of system ram. Any ideas? trying to work through a solution, just thought I'd ask.
Thanks again for all your effort on these int8's
edit: also the vaes in the checkpoint are not recognized. Not sure if that's something I'm doing but just letting you know.
It is a unet only checkpoint so you will need to load the text encoders and vaes separately. It could be a windows issue with the RAM, for me it works fine on 64gb, and I honestly have no clue how to unfuck the shit comfy aimdo pulls on windows...
It works fine for my at rtx3060 12gb + 64gb ram.
The problem is only when I try implement distilled loras from orginal 10eros workflow with Ltx2 Lora Loader Advanced then I got very high ram demand but this nodes probably don't work with int8 node.
So far I mostly use int8 Sulphur Distilled and this don't have any issues exept some loras not works like it should. But model itself is very capable.
Edit:
Win10.
Thanks for the input guys, I'll keep trying to figure it out.
I've tried adding more proper model caching (for like the 3rd time by now) here https://github.com/BobJohnson24/ComfyUI-INT8-Fast/tree/FileLoader
you can give it a try if you'd like. I've checked and at least on my machine it matches the memory usage of FP8 models perfectly.
I've tried adding more proper model caching (for like the 3rd time by now) here https://github.com/BobJohnson24/ComfyUI-INT8-Fast/tree/FileLoader
you can give it a try if you'd like. I've checked and at least on my machine it matches the memory usage of FP8 models perfectly.
Didn't help me anything for loras - anything I try (any node including int8 group) make ram use gigantic (and I got 64Gb ram). And I need Distilled Lora to work with 10eros model otherwise model isn't useble.