4 ms·
I've been told the 4 bit quantization slows it down, but don't quote me on this since I was unable to benchmark at 8 bit locally In any case, you're right it m
by enlyth 4y ago
I've been told the 4 bit quantization slows it down, but don't quote me on this since I was unable to benchmark at 8 bit locally
In any case, you're right it might not be as significant, however, the quality of the output increases with 8/16bit, and running 65B is completely impossible on 24GB
- riku_iki 4y agoIt's not impossible, there are several projects which load model layer by layer for execution from the disk or ram, but it will be much slower.