sagemaker.train.recipe_resolver

Recipe resolution with 3-level override precedence for Nova model training.

Functions

flatten_resolved_recipe(resolved)	Flatten a resolved recipe dict into a single-level key-value map.
render_template(template, override_spec)	Render a Hub recipe template by filling {{placeholder}} values.

Classes

RecipeResolver(recipe_template, override_spec)

Resolves a 3-level recipe configuration for Nova model training.

class sagemaker.train.recipe_resolver.RecipeResolver(recipe_template: Dict[str, Any], override_spec: Dict[str, Any], user_recipe_path: str | None = None, overrides: Dict[str, Any] | None = None, protected_keys: Set[str] | None = None, full_recipe_template: Dict[str, Any] | None = None, compute: Compute | HyperPodCompute | None = None)

Bases: object

Resolves a 3-level recipe configuration for Nova model training.

Precedence (highest wins):

1. Programmatic overrides (dict)

2. User recipe (YAML file)

3. Base defaults (rendered from Hub template + override-params spec)

Immutable after construction — all inputs are deep-copied. resolve() is idempotent: second call returns cached result.

get_resolved_recipe() → Dict[str, Any]

Return the resolved recipe as a read-only deep copy.

Callable before or after train()/evaluate(). Triggers resolution on first call if not already resolved.

Returns:

Deep copy of the fully resolved recipe dict.

resolve() → Dict[str, Any]

Perform template render, 3-level merge, and validation.

Returns:

The fully resolved recipe as a plain dict.

Raises:

ValueError – If validation fails for any parameter.

sagemaker.train.recipe_resolver.flatten_resolved_recipe(resolved: Dict[str, Any]) → Dict[str, Any]

Flatten a resolved recipe dict into a single-level key-value map.

Recursively walks all nested dicts and extracts scalar leaf values keyed by their leaf key name. Used by trainers and evaluators to apply resolved recipe values as flat hyperparameters to the SageMaker training API.

For nested structures like:

training_config:

lr_scheduler:

warmup_steps: 15 min_lr: 1e-6

This produces: {“warmup_steps”: 15, “min_lr”: 1e-6}

If duplicate leaf keys exist at different nesting levels, the last one encountered wins (depth-first traversal).

Parameters:

resolved – The resolved recipe dict (nested by section).

Returns:

Flat dict of all scalar leaf key-value pairs across all sections.

sagemaker.train.recipe_resolver.render_template(template: Dict[str, Any], override_spec: Dict[str, Any]) → Tuple[Dict[str, Any], Dict[str, str]]

Render a Hub recipe template by filling {{placeholder}} values.

Parameters:

• template – Hub recipe template dict containing ‘{{key}}’ placeholders.

• override_spec – Flat dict mapping spec keys to their metadata (including ‘default’, ‘type’, ‘min’, ‘max’, ‘enum’).

Returns:

Tuple of (rendered_dict, key_path_map) where key_path_map maps flat spec keys to their dotpath location in the recipe structure. e.g. {“learning_rate”: “training_config.learning_rate”}