nasone32/llama.cpp-RDNA3-7900xtx-opt
This repository carries a patched llama.cpp build aimed at RDNA3 cards such as the RX 7900 XTX, combining unreleased pull requests, AMD kernel patches and the RDNA Boost series. It adds an internal HIP all-reduce for two GPUs, adaptive MTP draft depth, direct lazy loading and an experimental MoE expert cache. Builds target ROCm and are tested on Linux.
Maintenance activity