CYBER WATCH / EQUINOXE INTELLIGENCE

← All articles

b12051: vulkan: add tensor-parallel allreduce function (#25051)

llama.cpp Releases · 2026-11-02 20:00:30

This update enhances Vulkan support by adding a tensor‑parallel all‑reduce function, requiring shared semaphores for communication, and fixing ring all‑reduce races on three or more GPUs. It aligns host memory regions, folds the all‑reduce into compute submissions to cut vkQueueSubmit calls, and optimizes synchronization with timeline semaphores, removing external dependencies and improving communication buffer handling.

Read original source ↗