More quantization visualization types (repost)

This is an issue of visualization. MXFP4 will be better than Q4\_0 simply due to the fact that Q4\_0 is integer quantization, while MXFP4 is float quantization. Most of the weights and activations of LLMs are closer to zero and just some of them differ in magnitude. Interestingly, we don't need to have low error for these high magnitude weights/activations -- we only need them to differ from all of the others. Conversely, since most of the w/a are of low magnitude, we need to differentiate between them. FP4 quantization (MXFP4, NVFP4) have relatively low quality for high magnitude values, since there are not a lot of values that can be represented in fp4, but they have much higher amount of representable values for lower magnitude. In contrast, Int4 have more or less uniform distribution of representable values, having bigger gaps around zero. In images, this is just what we need. Int4 quantizes pixel values in high contrast images in aesthetically pleasing way. FP4 clamps higher values, while retaining much higher precision around zero -- but since the images are high contrast with wide histogram of colors, it looks bad. LLMs have a different value distribution, thus, FP4 wins against Int4 (Q4\_0). TLDR: Images and weights/activations of LLMs have different value distributions, thus, FP4 is optimal for quantizing LLMs and images are a bad visualization tool.