Disty0

/

Z-Image-Turbo-SDNQ-uint4-svd-r32

like 44

You can quantize the original model on model load. Quantization doesn't really add any extra overhead or memory usage.

I only upload UINT4 SVD versions because svd takes a little bit longer to quantize and uint4 svd has the minimum file size with good quality.

I also don't want to waste a terabyte of disk space just for different quantization types of the same model. Use the original model and quantize on load.

But i can make an exception for Z-Image only for INT8 as it is still small and we can use INT8 MatMul without any bit packing or svd overhead.

https://huggingface.co/Disty0/Z-Image-Turbo-SDNQ-int8

You can quantize the original model on model load. Quantization doesn't really add any extra overhead or memory usage.
https://huggingface.co/Disty0/Z-Image-Turbo-SDNQ-int8

love you~ and offering prebuild files is even better♥️
👄♥️👄♥️👄♥️👄♥️👄♥️👄♥️👄♥️👄♥️👄♥️👄♥️🔥🤗👊