sagemaker.core.training.utils

sagemaker.core.training.utils#

Training utilities.

Functions

build_nova_hyperpod_manifest_s3_uri(...)

Build the HyperPod manifest.json S3 URI for a Nova training job.

build_nova_manifest_s3_uri(s3_output_path, ...)

Build the serverless manifest.json S3 URI for a Nova training job.

build_nova_output_tar_gz_s3_uri(...)

Build the output.tar.gz S3 URI for a Nova training job.

convert_unassigned_to_none(instance)

Convert Unassigned values to None for any instance.

read_nova_checkpoint_uri_from_manifest(...)

Read the checkpoint URI from a raw manifest.json object in S3.

resolve_nova_checkpoint_uri(s3_client, ...)

Resolve the Nova checkpoint (escrow) URI from a training job's output.

validate_instance_preferences(compute)

Client-side validation for Compute.instance_preferences (server remains the source of truth).
sagemaker.core.training.utils.build_nova_hyperpod_manifest_s3_uri(s3_output_path: str, training_job_name: str) → str[source]#

Build the HyperPod manifest.json S3 URI for a Nova training job.

HyperPod jobs write the manifest directly under the job directory: <s3_output_path>/<training_job_name>/manifest.json.

Parameters:

  • s3_output_path – The training job’s output_data_config.s3_output_path.

  • training_job_name – The training job name.

Returns:

Fully-qualified S3 URI to the job’s manifest.json.

sagemaker.core.training.utils.build_nova_manifest_s3_uri(s3_output_path: str, training_job_name: str) → str[source]#

Build the serverless manifest.json S3 URI for a Nova training job.

Serverless jobs write the manifest under a nested output directory: <s3_output_path>/<training_job_name>/output/output/manifest.json.

Parameters:

  • s3_output_path – The training job’s output_data_config.s3_output_path.

  • training_job_name – The training job name.

Returns:

Fully-qualified S3 URI to the job’s manifest.json.

sagemaker.core.training.utils.build_nova_output_tar_gz_s3_uri(s3_output_path: str, training_job_name: str) → str[source]#

Build the output.tar.gz S3 URI for a Nova training job.

Parameters:

  • s3_output_path – The training job’s output_data_config.s3_output_path.

  • training_job_name – The training job name.

Returns:

Fully-qualified S3 URI to the job’s output.tar.gz.

sagemaker.core.training.utils.convert_unassigned_to_none(instance) → Any[source]#

Convert Unassigned values to None for any instance.

sagemaker.core.training.utils.read_nova_checkpoint_uri_from_manifest(s3_client, s3_uri: str) → str[source]#

Read the checkpoint URI from a raw manifest.json object in S3.

Parameters:

  • s3_client – A boto3 S3 client.

  • s3_uri – S3 URI of the manifest.json object.

Returns:

The checkpoint_s3_bucket value from the manifest.

Raises:

ValueError – If the object is missing, unparseable, or lacks the key.

sagemaker.core.training.utils.resolve_nova_checkpoint_uri(s3_client, s3_output_path: str, training_job_name: str) → str[source]#

Resolve the Nova checkpoint (escrow) URI from a training job’s output.

Reads checkpoint_s3_bucket from the job’s manifest.json. The manifest is first looked up as a raw object, and if that fails, it falls back to the copy packaged inside output.tar.gz.

Parameters:

  • s3_client – A boto3 S3 client.

  • s3_output_path – The training job’s output_data_config.s3_output_path.

  • training_job_name – The training job name.

Returns:

The checkpoint URI recorded in the manifest.

Raises:

ValueError – If the checkpoint URI cannot be resolved from any known output layout.

sagemaker.core.training.utils.validate_instance_preferences(compute) → None[source]#

Client-side validation for Compute.instance_preferences (server remains the source of truth).

  • instance_preferences is mutually exclusive with the classic single-cluster fields instance_type / instance_groups / instance_placement_config, and with managed spot training.

  • Instance types must not repeat across preferences.

  • Count mode: exactly one of the top-level instance_count (applies to whichever preference wins) with no per-preference instance_count, or an instance_count on EVERY element with the top-level unset. Both-set, partial, and neither are rejected.

  • Training plans: the top-level (whole-job) training_plan_arn is mutually exclusive with per-preference training_plan_arns.

List-size limits (max preferences, max plans per preference) are deliberately NOT enforced client-side: they are server-side configurable, so raising them must not require a new SDK release.

No-op when instance_preferences is not set.

Contents