This change upstreams llamafile's vectorized expf() functions. This lets us compute softmax and silu more accurately than the short[65536] lookup table that GGML previously used to make this operation go faster. We can support aarch64 and sse2+ with the worst case rounding error of 2 ulp. I wrote avx2 and avx512 implementations as well but they didn't offer much advantage compared to sse2+fma to be worth the code complexity.

Hacker News is a community-driven platform for sharing and discussing technology news, startups, and programming-related topics. Through user submissions and comments, Hacker News offers insights into emerging technology trends, industry developments, and entrepreneurial ventures. Readers can participate in discussions, share their insights, and stay informed about the latest advancements in technology and innovation.

Hacker News

A pull request to rewrite silu and softmax for CPU in ggml code, resulting in significant performance improvements with support for aarch64 and sse2+.

ggml : rewrite silu and softmax for cpu by jart · Pull Request #7154 · ggerganov/llama.cpp