float8_e4m3fn quantized model?

#4
by adm1223 - opened

Could you provide float8_e4m3fn quantized model?
Could you in the future avoid int8 quantization for audio, image and video models they are inferior to either FP8 or FP8_mixed.

adm1223 changed discussion title from Why the mixed model has lower sizer than int8? to Why the mixed model has lower size than int8?
adm1223 changed discussion title from Why the mixed model has lower size than int8? to float8_e4m3fn quantized model?

I cannot avoid INT8 because people use them. But I do have an fp8 dev model:

https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-dev

Thanks a lot @joeygambino I have seen the "dev" version and I thought that I have to download the distilled lora for it which is another 8 GB . I was hoping you might consider doing fp8_mixed (preferred) or fp8 for the distilled model which I'm currently commenting on. It should be around the same size as int8 but with superior quality output. If it's too much I'll gladly stick to my original plan dev model+distilled lora.
I don't really use int8 models even for text_encoder they are inferior in compare to fp8_mixed which I usually use or fp8.

The fp8 checkpoint is also in this repo, went live this morning. Sorry, just saw your message.

https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-comfy-native/blob/main/LTX25dist-echoVid-070T30-v2-DiT-comfy-fp8.safetensors

adm1223 changed discussion status to closed

Sign up or log in to comment