Comfy-Org

/

Boogu-Image

like 124

Follow

Comfy Org 10.8k

nvfp4 and fp8 for boogu_image_turbo_hotfix, please! Thanks so much

fp8 hotfixes?

I won't be doing fp8 because int8-convrot is just much better, faster and better quality on all Nvidia GPUs. I've uploaded that now, but currently it needs very latest ComfyUI and comfy-kitchen versions.

I won't be doing fp8 because int8-convrot is just much better, faster and better quality on all Nvidia GPUs. I've uploaded that now, but currently it needs very latest ComfyUI and comfy-kitchen versions.

Is this a brand-new quantization format, and is FP8 about to be deprecated? Does this apply only to bboogu, or will it extend to future models as well?

I won't be doing fp8 because int8-convrot is just much better, faster and better quality on all Nvidia GPUs. I've uploaded that now, but currently it needs very latest ComfyUI and comfy-kitchen versions.

Is this a brand-new quantization format, and is FP8 about to be deprecated? Does this apply only to bboogu, or will it extend to future models as well?

Community has been using it for a while through Triton and custom nodes, it is now getting native support through cuda, and we'll monitor how well it performs with different models/GPUs before fully deciding that.

It does already look like fp8 is unnecessary for some models since int8-convrot is better quality and thus also allows quantizing more layers, ending up also faster on all Nvidia GPUs.

will int4 be natively support too? nunchaku int4 is as fast as nvfp4, and can work on 40s

I won't be doing fp8 because int8-convrot is just much better, faster and better quality on all Nvidia GPUs. I've uploaded that now, but currently it needs very latest ComfyUI and comfy-kitchen versions.

Makes sense. Thank you!

I won't be doing fp8 because int8-convrot is just much better, faster and better quality on all Nvidia GPUs. I've uploaded that now, but currently it needs very latest ComfyUI and comfy-kitchen versions.

Is this a brand-new quantization format, and is FP8 about to be deprecated? Does this apply only to bboogu, or will it extend to future models as well?

Community has been using it for a while through Triton and custom nodes, it is now getting native support through cuda, and we'll monitor how well it performs with different models/GPUs before fully deciding that.

It does already look like fp8 is unnecessary for some models since int8-convrot is better quality and thus also allows quantizing more layers, ending up also faster on all Nvidia GPUs.

Hope this int8-convrot could be supported on intel XPU which also have int8_mm accelerate ability. Maybe a triton backend is needed in comfy-kitchen.

I won't be doing fp8 because int8-convrot is just much better, faster and better quality on all Nvidia GPUs. I've uploaded that now, but currently it needs very latest ComfyUI and comfy-kitchen versions.

Is this a brand-new quantization format, and is FP8 about to be deprecated? Does this apply only to bboogu, or will it extend to future models as well?

Community has been using it for a while through Triton and custom nodes, it is now getting native support through cuda, and we'll monitor how well it performs with different models/GPUs before fully deciding that.

It does already look like fp8 is unnecessary for some models since int8-convrot is better quality and thus also allows quantizing more layers, ending up also faster on all Nvidia GPUs.

Will there be an INT8 ConvRot version of Qwen3-VL-8B?

Please add boogu_image_edit_turbo_int8 nvfp4

Please add boogu_image_edit_turbo_hotfix. Thanks!

哈哈哈,new version a lot better. kj is in vacation i think, we have to wait until he returns 哈哈

They use scripts from https://github.com/Comfy-Org/comfy-quants. Maybe you can convert yourself

The hotfix is performing much better. I converted the model to mxfp8 and it seems to work well. I uploaded it here: boogu_image_edit_turbo_hotfix_mxfp8.safetensors

int8 is faster and better than mxfp8

Please add boogu_image_edit_turbo_hotfix_int8_convrot.safetensors，Thanks a lot!

and boogu_image_base int8convrot please :)