> You can shrink the model to a fraction of its "full" size and get 92-95% same performance, with less VRAM use.
replyAre there a lot of options how "how far" do you quantize? How much VRAM does it take to get the 92-95% you are speaking of?
> Are there a lot of options how "how far" do you quantize?
replySo 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.
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?
reply