CYBER WATCH / EQUINOXE INTELLIGENCE
← All articles
b12284: CUDA: Add OCP FP8 support (#28413)
llama.cpp Releases · 2026-11-11 12:01:57
PR b12284 adds FP8 support to the CUDA build of llama.cpp. It introduces FP8 intrinsics, optimizes the GEMV kernel for FP8, handles conversions and strided copies, and provides a cuBLAS fallback for Hopper and other SKUs. It also includes precision handling, workspace management, and weight‑scale fusion, along with performance tests on an RTX PRO 4500.Read original source ↗