What's the difference between all these t5XXL models?
What are the differences between your T5XXL fp32 to Q8 gguf model to these other two? Which would be the most optimal to use in a strictly ComfyUI gguf use case? Thanks for your work!
calcuis/pig-encoder/t5xxl_fp32-q8_0.gguf (This one.)
chatpig/encoder/t5xxl-encoder-q8_0.gguf
chatpig/t5-v1_1-xxl-encoder-fp32-gguf/t5xxl-encoder-fp32-q8_0.gguf
pig encoder load faster; since the tensor name is 100% matched with torch/safetensors; you could make it yourself using convertor zero and it can be reversed back to safetensors using convertor reverse; the other two are converted using llama.cpp's convert_hf_to_gguf.py; this one will rename all the tensors with specific config and the config is different by models (this due to the tensor names are no standards, but they want to standardize it and shorten it for low end complier's 64-char length limitation), need extra effort and step to convert it to torch tensors; each time the llama.cpp update, they will add the new model configs, and transformers need these configs to work; if you plan to use it with transformers and diffusers inference, might need to take the later two (for these two, the last one's quality better since it was converted using the original fp32 version instead of bf16 version; you could check the model card for the source); btw, gguf node supports all of them