zlacker

> Are there a lot of options how "how far" do you quantize?

So many: https://www.reddit.com/r/LocalLLaMA/comments/1ba55rj/overvie...

> How much VRAM does it take to get the 92-95% you are speaking of?

For inference, it's heavily dependent on the size of the weights (plus context). Quantizing an f32 or f16 model to q4/mxfp4 won't necessarily use 92-95% less VRAM, but it's pretty close for smaller contexts.

>>bigyab+(OP)
Thank you. Could you give a tl;dr on "the full model needs ____ this much VRAM and if you do _____ the most common quantization method it will run in ____ this much VRAM" rough estimate please?

>>Muffin+z3
It’s a trivial calculation to make (+/- 10%).

Number of params == “variables” in memory

VRAM footprint ~= number of params * size of a param

A 4B model at 8 bits will result in 4GB vram give or take, same as params. At 4 bits ~= 2GB and so on. Kimi is about 512GB at 4 bits.