One thing is not obvious to me is how ConvRot can be applicable beyond diffusion models. Especially for LLM decoding, as each ConvRot would be more expensive for a given decoding vector, and it is required now, so you cannot easily get the benefit for prefill only, while maintaining the same decoding performance.

Alternative title: "Our invention makes INT8 better on older consumer cards without FP8 support"

>"FP8, which has been used as the standard until now, was only supported by the RTX 40/50 series. In contrast,

INT8 ConvRot

is attracting attention because it can benefit from hardware support even on the RTX 20/30 series."

Related:

https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui