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/awesome-recipe-username

You should follow the following directory structure:

recipes/{model}/{user}/{model}-{quant}-{mtp}-{runtime}-{user}.yaml

• You can add a README.md and supporting files under recipes/{model}/{user}; however, note that only the recipe yaml file will be considered part of the recipe.
• The model name should come first for tab completion at the command line.
• The recipe filename is the canonical lookup for the recipe, so it cannot overlap with other recipes and should provide sufficient data to understand the key inputs.
• If quantization is not applicable, still provide the native dtype (e.g. bf16 for bfloat16, fp8, etc.) for clarity. (re: {quant})
• If MTP/dflash/etc. not applicable, then leave it out. (re: {mtp})
• All community recipes should include your username at the end of the recipe filename for differentiation.

NOTE: Integration with Spark Arena Benchmarks coming soon!

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

   recipes/
     model-name/
       user-name/
         {model}-{quant}-{mtp}-{runtime}-{user}.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

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

# 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

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!

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