I recently published a paper on arXiv/ePrint about accelerating TFHE with ternary secrets. This repo contains the core optimized kernel—2-bit encoding, sparse AVX-512 FMA. It's dependency-free C. Benchmarks show 2.25x SIMD and 23x sparse speedups. I'm open-sourcing it to advance work in efficient FHE and 1.58-bit LLMs. Feedback welcome.
HyperFoldUK•1h ago