Why so few 8 bit capable models?

#13
by ibivibiv - opened

Just genuine curiosity, it seems like things are either full on 32bit, fp16 or 4bit/3bit quantized for the most part. Is there something special about 8bit quantized that makes it undesirable? For instance I can easily fit the large models at 4bit but an fp16 of them stretches beyond my vram. 8bit would fit and use up in most cases a larger amount of my VRAM. Is it a performance thing that there isn't much difference between 8 and 4 or is it more that there aren't many people that could do 8 and not do 16 so there just isn't a demand? It just seemed really odd to me that 8bit just isn't very prevalent at all in the community.

Just genuine curiosity, it seems like things are either full on 32bit, fp16 or 4bit/3bit quantized for the most part. Is there something special about 8bit quantized that makes it undesirable? For instance I can easily fit the large models at 4bit but an fp16 of them stretches beyond my vram. 8bit would fit and use up in most cases a larger amount of my VRAM. Is it a performance thing that there isn't much difference between 8 and 4 or is it more that there aren't many people that could do 8 and not do 16 so there just isn't a demand? It just seemed really odd to me that 8bit just isn't very prevalent at all in the community.

Very simple. 8-bit is slower than 4-bit because of memory bandwidth. There's just more gigabytes to copy around.

On CPU however (ggml llama.cpp), TheBloke does often provide an 8-bit option.

I agree that 4-bit quantization is no good. 5-bit should be a minimum.

Sign up or log in to comment