3 ms·
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
by MuffinFlavored 8mo ago
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?
- omneity 8mo agoIt’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.