vertebrae.profiling

Measured inference-resource profiles for local representation benchmarks.

Classes

`AdapterOperationResult`	Outcome of an optional framework-specific profiling operation.
`ResourceAdapterMetadata`	Framework, device, and synchronization context supplied by an adapter.
`DeviceMemoryMeasurement`	Framework allocator memory observed for one profiling session.
`ModelFootprintMeasurement`	In-memory model state exposed by a framework adapter.
`DeploymentArtifact`	Explicit local artifact whose deployment footprint should be measured.
`ArtifactFootprint`	Measured footprint for one declared deployment artifact.
`DistributedResourceProfile`	Aggregation of independent worker-local profiling windows.
`BaseResourceProfileAdapter`	Default unavailable implementations for custom resource adapters.
`ResourceProfiler`	Observe real extractor calls without introducing repeat inference.
`TorchResourceProfileAdapter`	Resource hooks for a Torch-backed extractor without forcing model loading.
`TensorFlowResourceProfileAdapter`	Resource hooks shared by Keras and TensorFlow Hub extractors.
`JAXResourceProfileAdapter`	Synchronization and device metadata for JAX-backed extractors.
`KerasResourceProfileAdapter`	Backend-aware resource hooks for multi-backend Keras models.
`ONNXResourceProfileAdapter`	Default unavailable implementations for custom resource adapters.

Functions

`resource_profile_columns`(profile, *[, prefix])	Flatten stable local or worker-aggregate fields for result tables.
`with_distributed_embedding_footprint`(profile, ...[, ...])	Update variant storage while retaining distributed inference evidence.
`resource_profile_from_dict`(value)	Reconstruct a typed local resource profile from serialized data.
`distributed_resource_profile_from_dict`(value)	Reconstruct a typed distributed resource summary from serialized data.
`resource_profile_like_from_dict`(value)	Reconstruct either supported profiling scope from a serialized payload.
`aggregate_distributed_resource_profiles`(profiles, *, ...)	Aggregate worker observations without presenting them as one local run.

Module Contents

class vertebrae.profiling.AdapterOperationResult[source]: Outcome of an optional framework-specific profiling operation.

class vertebrae.profiling.ResourceAdapterMetadata[source]: Framework, device, and synchronization context supplied by an adapter.

class vertebrae.profiling.DeviceMemoryMeasurement[source]: Framework allocator memory observed for one profiling session.

class vertebrae.profiling.ModelFootprintMeasurement[source]: In-memory model state exposed by a framework adapter.

class vertebrae.profiling.DeploymentArtifact[source]: Explicit local artifact whose deployment footprint should be measured.

class vertebrae.profiling.ArtifactFootprint[source]: Measured footprint for one declared deployment artifact.

class vertebrae.profiling.DistributedResourceProfile[source]: Aggregation of independent worker-local profiling windows.

vertebrae.profiling.resource_profile_columns(profile, *, prefix='')[source]

Flatten stable local or worker-aggregate fields for result tables.

Parameters:

profile (ResourceProfileLike)
prefix (str)

Return type:

Dict[str, Any]

class vertebrae.profiling.BaseResourceProfileAdapter[source]: Default unavailable implementations for custom resource adapters.

class vertebrae.profiling.ResourceProfiler(config, extractor, *, streaming, context=None)[source]

Observe real extractor calls without introducing repeat inference.

Parameters:

config (vertebrae.config.ResourceProfilingConfig)
extractor (Any)
streaming (bool)
context (Optional[Dict[str, Any]])

vertebrae.profiling.with_distributed_embedding_footprint(profile, raw_embeddings, evaluated_embeddings, *, store=None, raw_key=None, evaluated_key=None, raw_stat=None, evaluated_stat=None, persisted_storage=True)[source]

Update variant storage while retaining distributed inference evidence.

Parameters:

profile (Optional[DistributedResourceProfile])
raw_embeddings (Any)
evaluated_embeddings (Any)
store (Any)
raw_key (Optional[str])
evaluated_key (Optional[str])
raw_stat (Any)
evaluated_stat (Any)
persisted_storage (bool)

Return type:

Optional[DistributedResourceProfile]

vertebrae.profiling.resource_profile_from_dict(value)[source]

Reconstruct a typed local resource profile from serialized data.

Parameters:: value (Dict[str, Any])
Return type:: ResourceProfile

vertebrae.profiling.distributed_resource_profile_from_dict(value)[source]

Reconstruct a typed distributed resource summary from serialized data.

Parameters:: value (Dict[str, Any])
Return type:: DistributedResourceProfile

vertebrae.profiling.resource_profile_like_from_dict(value)[source]

Reconstruct either supported profiling scope from a serialized payload.

Parameters:: value (Optional[Dict[str, Any]])
Return type:: Optional[ResourceProfileLike]

vertebrae.profiling.aggregate_distributed_resource_profiles(profiles, *, merged_embeddings, all_shard_keys=None, store=None, merged_key=None, merged_stat=None, persisted_storage=True)[source]

Aggregate worker observations without presenting them as one local run.

Parameters:

profiles (Sequence[Tuple[str, ResourceProfile]])
merged_embeddings (Any)
all_shard_keys (Optional[Sequence[str]])
store (Any)
merged_key (Optional[str])
merged_stat (Any)
persisted_storage (bool)

Return type:

DistributedResourceProfile

class vertebrae.profiling.TorchResourceProfileAdapter(extractor, artifacts=(), *, model_getter=None, device_resolver=None, torch_loader=None)[source]

Bases: BaseResourceProfileAdapter

Resource hooks for a Torch-backed extractor without forcing model loading.

Parameters:

extractor (Any)
artifacts (Sequence[Union[str, DeploymentArtifact]])
model_getter (Optional[Callable[[], Any]])
device_resolver (Optional[Callable[[Any], Any]])
torch_loader (Optional[Callable[[], Any]])

class vertebrae.profiling.TensorFlowResourceProfileAdapter(extractor, artifacts=(), *, model_getter=None, backend='tensorflow', profiling_device=None)[source]

Bases: BaseResourceProfileAdapter

Resource hooks shared by Keras and TensorFlow Hub extractors.

Parameters:

extractor (Any)
artifacts (Sequence[Union[str, DeploymentArtifact]])
model_getter (Optional[Callable[[], Any]])
backend (str)
profiling_device (Optional[str])

class vertebrae.profiling.JAXResourceProfileAdapter(extractor, *, values_getter=None, jax_loader=None, backend='jax', profiling_device=None)[source]

Bases: BaseResourceProfileAdapter

Synchronization and device metadata for JAX-backed extractors.

Parameters:

extractor (Any)
values_getter (Optional[Callable[[], Any]])
jax_loader (Optional[Callable[[], Any]])
backend (str)
profiling_device (Optional[str])

class vertebrae.profiling.KerasResourceProfileAdapter(extractor, artifacts=(), *, profiling_device=None)[source]

Bases: BaseResourceProfileAdapter

Backend-aware resource hooks for multi-backend Keras models.

Parameters:

extractor (Any)
artifacts (Sequence[Union[str, DeploymentArtifact]])
profiling_device (Optional[str])

class vertebrae.profiling.ONNXResourceProfileAdapter(extractor, external_artifacts=())[source]

Bases: BaseResourceProfileAdapter

Default unavailable implementations for custom resource adapters.

Parameters:

extractor (Any)
external_artifacts (Sequence[Union[str, DeploymentArtifact]])