Running DiffusionGemma Locally on an RTX 5090 with vLLM: The One-Time Pioneer Cost of Bleeding-Edge Model Bleeding-Edge GPU

I stood up Google's diffusion language model DiffusionGemma as a vLLM OpenAI-compatible server on my home RTX 5090 (Blackwell / 32GB), so Open WebUI could use it as just another model. Declarative, on-demand on NixOS. This is the write-up of the traps I hit (a VRAM precheck that kills startup instantly, a container that wedges in a crash loop, the 32GB constraint that leaves NVFP4 as the only quantization that fits) and the lesson I drew from it: mature models are a one-liner, bleeding-edge ones you pioneer by hand exactly once.