t5xxl_fp8_e4m3fn_scaled.safetensors and t5-v1_1-xxl-encoder-Q8_0.gguf

[Amazon90]

What are the differences between these two models? How should I make my choice?In terms of model precision, which one is higher?

https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/blob/main/t5-v1_1-xxl-encoder-Q8_0.gguf

https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp8_e4m3fn_scaled.safetensors

[Kwisss]

I tried both but could not really see a difference in prompt following, but the other to clip encoders do the most heavy lifting I think

[city96]

I think the new scaled FP8 ones should be on-par with Q8_0 for T5. The logic is fairly similar from what I can tell but I haven't tested them myself.

The main benefit of FP8 is that it's faster (even more so on 40XX cards with the launch arg) + natively supported in ComfyUI.

I guess the main benefit for GGUF would be lower initial system memory usage on windows due to the way the model is loaded.

[Amazon90]

Thank you. Although I have a 40XX graphics card, I chose GGUF for its lower RAM usage since I only have 16GB of RAM.