Self-hosted alternatives to spark-environment-setup

S

spark-environment-setup

Set up a working ML training/inference environment on NVIDIA DGX Spark

spark-environment-setup is an agent skill from the wshobson/agents collection for setting up a working machine learning training and inference environment on NVIDIA DGX Spark. The DGX Spark uses a GB10 Grace Blackwell chip, with an aarch64 CPU, an SM121 GPU, 128 GB of unified memory and CUDA 13. The skill points out that this is a narrower and newer platform than a standard x86 CUDA 12 machine, so package selection and binary compatibility matter more, and the ecosystem of wheels for aarch64 with CUDA 13 is still filling in. It covers installing PyTorch, Unsloth, TRL, vLLM and xformers on the platform, along with checking that a framework is really using the GPU and not silently falling back to the CPU. It also covers the choice between using an NGC container and installing with pip, and restoring a working setup after an operating system reinstall or a base image update, with a verification from scratch. The skill applies when setting up a fresh DGX Spark for training or inference, when an import error mentions a missing CUDA runtime or symbol, when a wheel installs but will not load, when a framework install fails or hangs, and when a setup needs to be re-verified. Installing frameworks changes the machine, so the steps should be run in a planned order and checked after each one. The skill is installed with the skills CLI from the wshobson/agents repository. Its skills.sh page showed about 2.3K installs, with Gen Agent Trust Hub and Socket passing and Snyk showing a warning at the time of the scan.