hikkian/shura.cpp is a C++ inference engine forked from llama.cpp. The README describes running MoE models on graphics cards shared between gaming and desktop use while managing VRAM. The latest release version is not given in the provided record, and the license is MIT. The core point conveyed by the repository description and the README is a VRAM governor inside llama-server: when the desktop demands memory, it shrinks the MoE expert cache and gives graphics memory back.
According to the README, it covers a configuration that runs a 35B MoE on a 16GB gaming graphics card. When the desktop's free VRAM shrinks, it reduces the expert cache in one step. Prefill, which handles long inputs, is aborted and rolled back, then resumed. This goes beyond merely shrinking the cache and is designed so that long-input processing also continues. The README publishes measurements from an AMD graphics card, and the provided research record states that it was verified on AMD gfx1101 (ROCm). The use case this repository therefore presents is an environment where a single GPU is shared between the desktop and inference work.
However, the README notes that there are no measurement results yet for the CUDA build. The published description and verification evidence are limited to AMD graphics cards, so there is insufficient basis to expect the same behavior and performance on other GPUs. In particular, no CUDA results are presented, so it is too early to judge this as a general-purpose VRAM management feature usable across multiple GPUs.