Comment by dabockster

You don’t need it if you use llamacpp on Windows, or if you compile it on Linux with CUDA 13 and the correct kernel HMM support, and you’re only using MoE models (which, tbh, you should be doing anyways).