Quick Start

Install

# Install uv if you don't have it
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install sparkrun and set up your cluster
uvx sparkrun setup

The setup wizard walks you through cluster creation, SSH mesh, and CX7 networking interactively. See the Setup Wizard Walkthrough for a detailed step-by-step guide.

Run a model (single node)

No cluster setup needed — just point sparkrun at a host and run:

# Run on localhost (or your default cluster's first host)
sparkrun run qwen3-1.7b-vllm

# Run on a specific remote host
sparkrun run qwen3-1.7b-vllm --hosts 10.24.11.13

sparkrun launches containers in the background and follows logs. Ctrl+C detaches from logs — it never kills your inference job. Your model keeps serving.

Run multi-node (tensor parallelism)

With a cluster configured, scale across multiple DGX Sparks:

sparkrun run qwen3-1.7b-vllm --tp 2

Each DGX Spark has one GPU with 128 GB unified memory, so --tp 2 splits the model across 2 nodes. See Multi-Node Tensor Parallelism for a full walkthrough.

Inspect a recipe

sparkrun show nemotron3-nano-30b-nvfp4-vllm

This displays recipe details including a VRAM estimation that tells you whether the model fits in memory before you launch.

Status, logs, and stop

# Check what's running
$ sparkrun status

Job: @sparkrun-transitional/qwen3-1.7b-llama-cpp  [e0ed758410a2]  (1 container(s))
  solo       10.24.11.13 (ib: 192.168.11.13)          Up 45 minutes             ghcr.io/spark-arena/dgx-llama-cpp:latest
  logs: sparkrun logs e0ed758410a2
  stop: sparkrun stop e0ed758410a2

# Re-attach to logs (Ctrl+C is always safe) -- you can reference recipe or job ID
sparkrun logs @sparkrun-transitional/qwen3-1.7b-llama-cpp

# Stop a workload (shown here using job ID for unambiguous reference)
sparkrun stop e0ed758410a2

Note

If you launched with --tp (modifying the recipe default), pass the same --tp flag to stop and logs so they resolve the same cluster ID. Or, use the cluster ID from sparkrun status directly:

Spark Arena

Spark Arena is the community hub for DGX Spark recipe benchmarks — browse benchmark results, then run them directly with sparkrun.

Official Recipes

Official Recipes are maintained by the Spark Arena team and hosted on GitHub. They are tested and optimized for NVIDIA DGX Spark systems.

Community Recipes

Community Recipes are contributed by the community and hosted on GitHub.

Next steps

• Your First Model — step-by-step tutorial
• Recipe Format — write your own recipes
• Registries — add community or private recipe registries
• Networking — CX7 configuration for multi-node clusters