a775828b-dot/xtier is a Python plugin that runs mixture-of-experts models (MoE) split across CPU and GPU in SGLang. According to the repository description, it follows the upstream project's main branch. FACTS does not list a latest release, so no specific version or change released this time can be confirmed. What can be reviewed here is the project structure and description published on the GitHub repository.
This plugin is designed to place each expert in GPU memory or host memory depending on how frequently it is used, and to compute on whichever side is needed. Large embedding tables keep some rows on storage and read them from there. The README explains that it wraps the standard mixture-of-experts module, so multiple models can run in hybrid mode without writing model-specific code. The repository contains backends, a cache, a scheduler and a set of model-specific optimizations, and also includes test directories, benchmark scripts and results, and run materials for different configurations.
However, the README states that the project is at an early development stage, and explains that supporting some models requires a separate SGLang branch. The published benchmark materials alone do not reproduce the performance figures, so it is difficult to judge the actual throughput or memory savings. For now, it is close to an early-stage plugin that lets you look at the implementation structure for hybrid inference in SGLang.