fp8 gguf versions of FastHunyuan

#4
by boyetosekuji - opened

FastHunyuan is an accelerated HunyuanVideo model. It can sample high quality videos with 6 diffusion steps. That brings around 8X speed up compared to the original HunyuanVideo with 50 steps.
https://huggingface.co./FastVideo/FastHunyuan
https://x.com/TXhunyuan/status/1869282002786292097

this needs to happen! keep up the good work

would be great hopefully city96 sees the request

Please please please, GGUF of FastHunyuan!

Please please please, GGUF of FastHunyuan!

there is a fp8 version and i am using it hunyuan_video_FastVideo_720_fp8_e4m3fn.safetensors here https://huggingface.co./Kijai/HunyuanVideo_comfy/tree/main, i have not checked how fast it is vs the distill version.

i prefer GGUF Q8. Usually better than FP8

Please please please, GGUF of FastHunyuan!

there is a fp8 version and i am using it hunyuan_video_FastVideo_720_fp8_e4m3fn.safetensors here https://huggingface.co./Kijai/HunyuanVideo_comfy/tree/main, i have not checked how fast it is vs the distill version.

It's 12GB against Q4 which requires only 8GB card. I also tried a Hunyuan's block swap node in ComfyUI to handle my 8GB limit, but got a crash on sampler stage.

can we please get the GGUF for fasthunyuan, fp8 doesnt work on macs

can we please get the GGUF for fasthunyuan, fp8 doesnt work on macs

dude please do gguf version of fasthunyuan, c'mon mahn can't wait anymore.

Sorry, been super busy with real life stuff. Anyway, uploaded them here: https://huggingface.co./city96/FastHunyuan-gguf

Only tested on single frame generation and probably with the wrong settings but it seems to work.

Sorry, been super busy with real life stuff. Anyway, uploaded them here: https://huggingface.co./city96/FastHunyuan-gguf

Only tested on single frame generation and probably with the wrong settings but it seems to work.

It works, THANK YOU! Tested on the same workflow on my laptop 4060 8GB (704x400, 61 frames) FastHunyuan Q4_K_S 6 steps ~2 minutes vs Hunyuan Q4_K_S 20 steps ~6 mins

Sign up or log in to comment