wwoosshh/geulpuv2 provides a deterministic engine and debugger for running GGUF language models on local GPUs. The repository's latest release is v0.9.0. The provided material lists nothing as newly changed in this version, so release-specific changes cannot be identified. The core of the project is a precision policy documented so that the same request reproduces the same tokens and probability values. The engine, server, and app are implemented in Glolang, a Korean keyword programming language.
This repository focuses on reproducing model execution results and looking at the differences. Runs can be recorded and replayed, and the outcome of changing one element at a point where the run branches can be compared with it. It also provides a per-token analysis screen. According to the README, it supports CUDA and Vulkan, and it includes a browser chat screen and an OpenAI-compatible interface. The repository language is listed as Python and the license is MIT. Here, determinism refers to a design that aims to pin down results even under execution conditions including batching or slots, prompt caching, and speculative decoding.
However, the provided material presents no performance comparisons or external validation results, so it is hard to judge the extent to which the precision policy holds in real environments. Therefore, rather than the specific changes in v0.9.0 or the engine's performance advantage, it is more accurate to see it as a repository whose debugging structure for recording and comparing runs has been confirmed.