Skip to content
Get packs

DGX Spark Environment Setup

Set up PyTorch and ML training tools correctly on NVIDIA DGX Spark (aarch64, CUDA 13)

Source Download skill
Repository
wshobson/agents
Stars
42,095
Licence
MIT
Files
3 files

What it does

This skill gives an AI agent a checklist for building a working machine-learning environment on an NVIDIA DGX Spark, a small workstation with an ARM CPU and a GB10 GPU. It tells the agent to prefer pre-validated NGC Docker containers over ad hoc pip installs, to pin exact package versions, and to check for CUDA version mismatches that cause confusing import errors. The agent uses these rules to diagnose failures and suggest the right install path for PyTorch, Unsloth, TRL, vLLM and similar tools.

Great for

  • Engineers setting up a fresh DGX Spark for fine-tuning or inference
  • People hitting libcudart or wheel-ABI import errors on an ARM/CUDA 13 machine
  • Teams restoring a Spark environment after an OS reinstall

Skip it if

  • Standard x86 CUDA 12 machines, where the ARM and CUDA 13 specifics do not apply and versions in the file may be wrong
  • Anyone expecting it to be current: the version pins are dated and depend on the stack-matrix file's 'Last verified' date, so they can go stale and cause failures if followed blindly
  • Non-technical users, since it assumes Docker, pip and command-line comfort

Try asking

  • “My PyTorch import on my DGX Spark says undefined symbol libcudart, what should I check?”
  • “Set up an Unsloth fine-tuning environment on DGX Spark using the right container.”
  • “Should I use an NGC container or bare pip to install vLLM on my Spark?”
  • “Verify my DGX Spark can see the GPU before I start a training run.”
Step-by-step guide: how to use DGX Spark Environment SetupInstall, first run, examples, tips and FAQs

Install in 3 steps

  1. 1

    Download the skill

    The ZIP contains the 3 original files, the MIT licence and an attribution note.

    spark-environment-setup.zip
  2. 2

    Add it to your agent

    Claude Code (all projects)

    mkdir -p ~/.claude/skills && unzip spark-environment-setup.zip -d ~/.claude/skills/

    Claude Code / Agent SDK (this project only)

    mkdir -p .claude/skills && unzip spark-environment-setup.zip -d .claude/skills/

    Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.

  3. 3

    Ask for the task

    No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.

Get better results

  • Check the 'Last verified' date in references/stack-matrix.md before using pinned versions, and update pins if they are old.
  • Resolve and pin the digest of moving tags like unsloth/unsloth:dgxspark-latest before any reproducible run.
  • Review the commands the agent proposes before running them, since they pull images and install packages from the internet.

What's inside

  • references/container-workflow.md
  • references/stack-matrix.md
  • SKILL.md

Redistributed unmodified from wshobson/agents under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.