Local Ai
b9820
Release b9820 of llama.cpp introduces scheduler optimizations to reduce synchronizations during split compute and improves CUDA performance with fewer synchronizations between tokens. The update inclu
Release b9820 of llama.cpp introduces scheduler optimizations to reduce synchronizations during split compute and improves CUDA performance with fewer synchronizations between tokens. The update includes CPU-to-CUDA copy capability for async tensor operations and relaxes synchronization requirements between input copies on supported backends like CUDA, with other backends such as Vulkan potentially adopting similar changes.
Source: llama.cpp Releases | 2026-06-26