Recipe Commands

List recipes

sparkrun recipe list [query]
sparkrun list [query]          # alias

List available recipes from all configured registries. Optionally filter by a search query.

Options

Option	Description
--all / -a	Include recipes from hidden registries
--runtime	Filter by runtime (e.g., vllm, sglang)
--registry	Filter by registry name
--json	Emit machine-readable JSON instead of the table

The output table includes:

Column	Description
Name	Recipe display name
Runtime	Inference runtime (vllm, sglang, llama-cpp, trtllm)
TP	Default tensor parallelism setting (- if not set)
Nodes	Minimum node count (min_nodes)
GPU Mem	Default gpu_memory_utilization setting (- if not set)
Registry	Source registry name or local
File	Recipe filename stem

Scoping to a registry

The query itself accepts the same @registry syntax recipe names use:

sparkrun list @community        # same as --registry community
sparkrun list @community/       # same
sparkrun list @community/qwen   # --registry community, query "qwen"

Naming a registry implies --all, so its recipes appear even if the registry is hidden by default. An unknown or disabled registry — whether named via the shorthand or --registry — is reported as an error listing the available registries, rather than silently yielding “No recipes found”. Combining a conflicting @scope with --registry is also an error.

list vs search

Both scan the same sources; they differ in how duplicates are treated.

• list shows one row per unqualified name, preferring working-directory recipes.
• search shows every copy.

That distinction matters because a registry’s recipe directory is scanned recursively, so 3x-spark-cluster/foo.yaml and 4x-spark-cluster/foo.yaml are genuinely different recipes that share a name. Only literal repeats of the same file are ever dropped.

Show recipe details

sparkrun recipe show <recipe>
sparkrun show <recipe>         # alias

Display detailed recipe information including defaults, metadata, and VRAM estimation. The VRAM estimator auto-detects model architecture from HuggingFace.

Recipes can be local names, file paths, URLs, or Spark Arena shortcuts:

# Show a local/registry recipe
sparkrun show qwen3-1.7b-vllm

# Show a Spark Arena recipe
sparkrun show @spark-arena/<recipe-id>

To save a copy of a recipe for customization, use sparkrun export recipe:

sparkrun export recipe @spark-arena/<recipe-id> --save my-recipe.yaml

Options

Option	Description
--no-vram	Skip VRAM estimation display
--tp / --tensor-parallel	Override tensor parallelism for VRAM estimate
--gpu-mem	Override GPU memory utilization (0.0-1.0) for VRAM estimate

Search recipes

sparkrun recipe search <query>
sparkrun search <query>        # alias

Search for recipes by name, model, or description across all registries.

Options

Option	Description
--all / -a	Include recipes from hidden registries
--runtime	Filter by runtime (e.g., vllm, sglang)
--registry	Filter by registry name
--json	Emit machine-readable JSON instead of the table

The output table includes:

Column	Description
Name	Recipe display name
Runtime	Inference runtime
TP	Default tensor parallelism setting (- if not set)
Nodes	Minimum node count
GPU Mem	Default gpu_memory_utilization setting (- if not set)
Model	HuggingFace model identifier
Registry	Source registry name

Validate a recipe

sparkrun recipe validate <recipe>

Check a recipe file for issues:

• Missing required fields (model, runtime)
• Invalid mode values
• min_nodes / max_nodes consistency
• Valid metadata.model_params and metadata.model_dtype
• Runtime-specific validation

Update recipes

sparkrun recipe update
sparkrun recipe update --registry community   # just one

Refreshes the configured recipe registries from git, so newly published recipes become resolvable. sparkrun update does this too, alongside upgrading sparkrun itself.

Estimate VRAM usage

sparkrun recipe vram <recipe> [options]

Option	Description
--tp / --tensor-parallel	Override tensor parallelism
--max-model-len	Override max sequence length
--gpu-mem	Override gpu_memory_utilization (0.0-1.0)
--no-auto-detect	Skip HuggingFace model auto-detection