weight dtype "default" very slow
First of all i have a 4090 and 64 GB of RAM in my machine
i am using comfy ui
now i am wondering what´s the problem with the weight dtype
when i set the weight dtype to "default" it takes up to 10 minuten for a simple 1024x1024 image to generate
but when i set the weight dtype to "fp8_e4m3fn" it is generating the images super fast - like 14 seconds
so what did i wrong? :D
it is btw not using a lot of RAM - only 12 GB of RAM used (not meaning VRAM)
default is bfloat16 and is ~23GB, so your 4090's vram is still not enough.
however, I can run default unet with 4070 12G and about 2.5s/it (uses a lot of ram).
so I think your case can be faster, maybe some wrong settings
For me it's the other way round if I use fp8_e4 or fp8_e5... it's really slow but on default I get 5s/it with my 3060 12G. I did the following. Updated Nvidia drivers and in Nvidia settings disable Sysmem Fallback policy so it will only use your VRAM and not your system ram.
Make sure comfyui is install on a main drive and not on an external USB hard drive. This will effect load times. I can craft in seconds with my 4090