Skip to content
Get packs

Guide · Agent skill

How to use the DGX Spark Environment Setup agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The spark-environment-setup 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 is for engineers who are setting up a fresh Spark for fine-tuning or inference, or who are chasing import errors on this ARM platform. After reading this guide you will know what the skill does, how to install it, how to ask for it in plain language, and what it needs before you rely on it.

DGX Spark Environment Setupwshobson/agents · MIT

Quick start

  1. 1Download and unzip the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or the project's .claude/skills/ folder. Alternatively, upload it under Settings → Capabilities → Skills in the Claude apps.
  2. 2Confirm your hardware and Docker setup. Check that you are on a DGX Spark, that Docker has the NVIDIA container runtime, and that you can reach nvcr.io.
  3. 3Ask for the environment in plain language. Describe what you want to install or the error you see, including the tool name and any error text.
  4. 4Run the GPU visibility check. Run the Python check that prints whether CUDA is available and which CUDA version torch was built against, before installing project packages.
  5. 5Choose a container or bare pip. Use the NGC PyTorch or Unsloth container by default. Use bare pip only when neither container fits, and follow the install order the agent gives.
  6. 6Check the pins and known issues. Confirm the Last verified date in the reference files and look at recent upstream issues before a long run.

What is an agent skill, and what does this one make the agent do?

An agent skill is a folder with a SKILL.md file and optional reference documents. When your AI agent sees a request that matches the skill's description, it reads those instructions and follows them. This skill is written for the DGX Spark platform, which uses an aarch64 CPU, a GB10 GPU and the CUDA version the platform ships with.

The skill 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. It covers PyTorch, Unsloth, TRL, vLLM and similar tools, and gives the agent a short verification step to run before anything expensive.

  • Recommends an NGC PyTorch container for general training and inference work.
  • Recommends an Unsloth container for Unsloth-centred fine-tuning.
  • Falls back to bare pip only when neither container fits, and then follows a fixed install order.
  • Diagnoses ABI mismatches, where a wheel built for one CUDA major version is loaded on a system with another.

How do you install the spark-environment-setup skill?

Download the ZIP of the skill from this page. Then install it in one of two ways. For a personal setup, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder so it applies only there.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills. Either way, the skill folder should contain SKILL.md along with the references/ folder, because the skill points the agent to references/container-workflow.md and references/stack-matrix.md for the details.

  • Personal install: unzip into ~/.claude/skills/.
  • Project install: unzip into the project's .claude/skills/.
  • Claude apps: Settings → Capabilities → Skills, then upload.

How do you trigger the skill with natural requests?

You do not need to name the skill. The agent matches your request against the skill's description, which covers installing PyTorch, Unsloth, TRL or vLLM on DGX Spark, libcudart or wheel-ABI errors on aarch64, and choosing between NGC containers and bare pip installs. Describe the problem in your own words and include the machine details the agent needs.

Requests that work well mention the Spark, the tool and the error text. Pasting the exact error message helps the agent pick the right branch of the checklist.

  • "I'm setting up a fresh DGX Spark for LoRA fine-tuning. Should I use an NGC container or pip?"
  • "My torch import on the Spark fails with a libcudart error. Walk me through the fix."
  • "Restoring the Spark after an OS reinstall. Verify the GPU environment from scratch."
  • "vLLM installed but the model runs on CPU. What do I check first?"

What does a typical check look like?

The first check the skill asks for is whether PyTorch can see the GPU. Run this inside your container or Python environment before installing project packages.

If the output starts with False, do not reinstall a wheel straight away. The skill has the agent check, in order, whether nvidia-smi works, whether CUDA_VISIBLE_DEVICES is set correctly, the permissions on /dev/nvidia*, and whether the CUDA version reported by torch matches the system. A mismatch is one cause among several.

  • If nvidia-smi does not show the GPU, the problem is usually runtime, visibility or permissions rather than the wheel.
  • Confirm the wheel's CUDA build before reinstalling anything.
python3 -c "import torch; print(torch.cuda.is_available(), torch.version.cuda)" # expected output on a working setup: True followed by a 13.x CUDA version

What do you need before the skill is useful?

The skill assumes a specific stack. It needs NVIDIA DGX Spark hardware, Docker with the NVIDIA container runtime, access to nvcr.io (NVIDIA's NGC registry), and Python with pip. The packages it references, such as Hugging Face tools, are installed at pinned versions.

If your machine is not a DGX Spark, the skill's ARM and CUDA-specific advice does not apply. Its version pins may also be wrong for you.

  • NVIDIA DGX Spark hardware (GB10, aarch64).
  • Docker with the NVIDIA container runtime.
  • Access to nvcr.io and Python with pip.
  • Comfort with the command line, Docker and pip.

What are the safety and network notes?

This skill does not run scripts bundled with it. It does use the network, because the workflows pull container images from nvcr.io and install packages from package indexes, and the agent will recommend those steps. Review each command before the agent runs it, especially pulls and installs on a machine you care about.

The version pins are dated. The reference files carry a Last verified date that governs how stale they are. Do not follow the pins blindly on a later date. Check the NVIDIA playbook issues and the stack-matrix reference before trusting a recipe for a long training run.

  • Scripts: none are bundled to run.
  • Network: yes, for container pulls and package installs.
  • Pin digests for moving container tags before reuse.
  • Use --no-deps on the Unsloth install as the skill describes, to avoid pulling incompatible torch or triton builds.
docker run --runtime=nvidia --gpus all -it --rm \ nvcr.io/nvidia/pytorch:25.09-py3

What tips make the skill work better?

Start from a container rather than a bare install. The skill's reasoning is that a container locks Triton, xformers and transformers together against a combination already validated on this hardware. Bare pip leaves that resolution to you.

Expect some tools to be fragile. The skill says flash-attn should be skipped on this hardware, that xformers needs a source build, that vLLM uses nightly wheels, and that LLaMA-Factory and NeMo are fragile on Spark. Check upstream issues first for those.

  • Verify the GPU before installing project packages.
  • Mention the exact error text in your request.
  • If Triton kernel compilation fails during training, the skill points to setting TRITON_PTXAS_PATH and checking the stack-matrix reference.
  • Ask the agent to restate the Last verified date it is relying on.

What people use it for

  • Setting up a fresh DGX Spark for LoRA or full fine-tuning with Unsloth or TRL.
  • Diagnosing a libcudart import error after a wheel installs cleanly.
  • Deciding between an NGC container and a bare pip install for a new project.
  • Re-verifying a Spark environment after an OS reinstall.

Frequently asked questions

How do I install the DGX Spark Environment Setup skill?

Download the skill ZIP from this page and unzip it into ~/.claude/skills/ for a personal install, or into a project's .claude/skills/ folder for a single project. You can also upload it in the Claude apps under Settings → Capabilities → Skills.

What are the requirements for spark-environment-setup?

You need NVIDIA DGX Spark hardware with a GB10 GPU and an aarch64 CPU, Docker with the NVIDIA container runtime, access to nvcr.io, and Python with pip. The skill assumes command-line comfort.

Is the spark-environment-setup skill free?

The skill is published under the MIT licence. The source does not state any price, so check the directory page for any terms beyond the licence.

Does this skill work on a normal x86 CUDA machine?

No. The skill is written for the DGX Spark's ARM CPU and GB10 GPU. On a standard x86 CUDA machine its ARM and CUDA-specific advice does not apply, and its version pins may be wrong.

Why does my torch import fail with a libcudart error on DGX Spark?

This is usually a CUDA ABI mismatch, where a wheel was built against a different libcudart major version from the one on the system. Check torch.version.cuda first. Using a matched NGC container or a cu130 aarch64 wheel is the usual fix.

Is spark-environment-setup an alternative to the NGC containers?

No. The skill recommends the NGC PyTorch and Unsloth containers as the default and uses them to guide your setup. Bare pip is the fallback when a container does not fit your project.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you are setting up a DGX Spark for fine-tuning or inference and want an agent to walk you through container choice, the ABI check and the install order. Pick something else if you are on an x86 CUDA machine, if you are not comfortable with Docker and pip, or if you need the version pins to be current without checking them yourself.

Download the skill