Pennyroyal
An optimized SGLang-derived inference runtime tuned for running Qwen3.8 models on a single NVIDIA RTX PRO 6000 Blackwell GPU, with prebuilt launch configurations, a long context window, and disk-backed caching.
Pennyroyal is a patched, SGLang-derived runtime with launch configurations qualified for one NVIDIA RTX PRO 6000 Blackwell Workstation GPU (96 GB, SM120, tensor parallelism 1). It ships two qualified model profiles built from the same patched source: Qwen3.8-27B FP8 with DFlash2, and Qwen3.8 Flash-Next NVFP4 with native next-token prediction and speculative decoding.
Both profiles support a prebuilt container for the fastest setup, or a native build and install; an optional terminal setup utility walks through model, GPU and cache settings. The Flash-Next profile keeps a large cache in RAM or can trade RAM for SSD input/output through a disk-backed option, while the 27B/DFlash2 profile uses a large in-memory cache; a disk budget bounds cache growth on disk.
A prebuilt Docker image bundles the runtime, CUDA build tools and a data-transfer library for both profiles; the host still supplies the NVIDIA driver, model files and writable cache folders.