FP8mixed vs INT8 convrot agressive speedboost
Hi there,
After using your INT8_convrot_agressive quant on complex edit (such as small writings or several objects transfer), I didn't notice any speed increase, however the precision is much better compared to FP8Mixed.
Just used the native node on ComfyUI, didn't tried with https://github.com/BobJohnson24/ComfyUI-INT8-Fast/tree/main (maybe that one could be faster though?)
RTX 5080
Ram 96GB DDR5
Ubuntu 24.04
ComfyUI: v0.27.1
I used a custom workflow found on CivitAI with 6 reference images. I'm around 720s for the generation both with FP8Mixed and Int8, sageattention 2.2, res 1408x1408.
30steps
Cfg1
Sampler res_2s
Scheduler normal
In the end that's just a message to say thank you! If you have advices to speed up please don't hesitate!
various newer models suddenly get int8 convrot versions. even set as default in comfy templates. And speed boost is good, even on a 5090 Flux.2 dev needs all the help it can get.
according to gpt5.6:
" I would expect:
INT8 ConvRot and FP8-scaled to have approximately the same model VRAM footprint.
INT8 ConvRot to be somewhere between slightly slower and noticeably faster, depending on the architecture and kernel.
Quality differences to usually be subtle, but ConvRot should have fewer quantization-induced deviations."