Community-contributed inference recipes for NVIDIA DGX Spark

Share your optimized model configs, discover what others are running, and benchmark on Spark Arena.

---

This is the community recipe registry for sparkrun. Anyone can submit a recipe via pull request and make it available to every sparkrun user.

Community recipes are run with the @community prefix:

sparkrun run @community/my-awesome-recipe

Submit a Recipe

1. Fork this repository
2. Create your recipe YAML in recipes/:

   recipes/
     my-model-vllm/
       recipe.yaml
   

3. Open a pull request — describe what the recipe runs, which runtime it targets, and expected VRAM usage
4. Once merged, your recipe is live for everyone

Recipe format

Recipes follow the standard sparkrun recipe schema. See the recipe authoring docs for the full specification.

Run Community Recipes

# List available community recipes
sparkrun list @community

# Run a community recipe
sparkrun run @community/my-awesome-recipe

# Check VRAM requirements before launching
sparkrun show @community/my-awesome-recipe

Benchmark Your Recipes

Got a recipe that screams? Prove it.

Submit your recipe benchmark to Spark Arena to benchmark your recipes and see how they stack up. Publish your numbers and show the community what DGX Spark can do.

# Login to Spark Arena (if not already logged in) -- see https://spark-arena.com for details
sparkrun arena login

# Benchmark your recipes (this will run the recipe, benchmark it, stop it, and upload the result to Spark Arena in one step)
sparkrun arena benchmark @community/my-awesome-recipe

Then you can go to https://spark-arena.com/leaderboard to see how you stack up!

Links

• sparkrun — the tool that runs recipes
• sparkrun docs — full documentation
• Spark Arena — community hub and benchmarking