High-performance uncensored Gemma-4-26B inference on NVIDIA DGX Spark using vLLM - 45+ tok/s - View it on GitHub