This is a reusable SparkRun v4 adaptation of Tony D's two-node DGX Spark deployment for GLM-5.3-Flash NVFP4 with DFlash2 speculative decoding.
Upstream credit: The container image, SM121 patches, launch settings, debugging work, and performance findings come from Tony D (
tonyd2wild) and contributors to GLM-5.3-Flash-NVFP4-DFlash2-2x-DGX-Spark. This Gist only translates that work into a portable SparkRun recipe. It is