Fast embedding, sampling, and freq_cis computation of decoding on CPU

[kaizizzzzzz]

Hi, I have compared the latency of embedding and freq_cis (which are the former steps before computing the transformer layers) on CPU and GPU in "gpt-fast" repo. The latency of CPU is still much bigger than doing on GPU even on the decoding stage (do not need to access multiple token_id, but only one). So I'm curious about the extra latency and wondering if LLAMA.CPP has some optimizations to speedup these computations on CPU, because I think this non-computation step should also be fast on CPU.

Thanks!