sampling#

Classes

SparseSampleManifest

SparseSampleRecord

SparseSamplingPolicy

Functions

sample_replacement_candidates

Select at most replacement_cap layer candidates independently per width.

sample_subblock_configs

Select layer-independent teacher, single-axis, and pairwise subblocks.
class SparseSampleManifest#

Bases: object

__init__(*, mode, policy, eligible, selected, excluded)#
Parameters:
Return type:

None

eligible: tuple[SparseSampleRecord, ...]#
excluded: tuple[SparseSampleRecord, ...]#
property identity: str#
mode: str#
policy: SparseSamplingPolicy#
selected: tuple[SparseSampleRecord, ...]#
to_dict(*, include_identity=True)#
Parameters:

include_identity (bool)

Return type:

dict[str, Any]

class SparseSampleRecord#

Bases: object

__init__(*, sample_id, candidate_id, layer_idx, hidden_width, subblock_kind, subblock_name, changed_axes, block_config, subblock_config, no_op=False, reason='eligible')#
Parameters:

  • sample_id (str)

  • candidate_id (str)

  • layer_idx (int)

  • hidden_width (int | None)

  • subblock_kind (str)

  • subblock_name (str)

  • changed_axes (tuple[str, ...])

  • block_config (dict[str, Any])

  • subblock_config (dict[str, Any] | None)

  • no_op (bool)

  • reason (str)

Return type:

None

block_config: dict[str, Any]#
candidate_id: str#
changed_axes: tuple[str, ...]#
hidden_width: int | None#
layer_idx: int#
no_op: bool = False#
reason: str = 'eligible'#
sample_id: str#
subblock_config: dict[str, Any] | None#
subblock_kind: str#
subblock_name: str#
to_dict()#
Return type:

dict[str, Any]

class SparseSamplingPolicy#

Bases: object

__init__(*, max_pairwise_per_family=4, replacement_cap=50, seed=42)#
Parameters:

  • max_pairwise_per_family (int)

  • replacement_cap (int)

  • seed (int)

Return type:

None

max_pairwise_per_family: int = 4#
replacement_cap: int = 50#
seed: int = 42#
sample_replacement_candidates(candidates, *, policy=None)#

Select at most replacement_cap layer candidates independently per width.

Parameters:
Return type:

SparseSampleManifest

sample_subblock_configs(candidates, *, policy=None)#

Select layer-independent teacher, single-axis, and pairwise subblocks.

Parameters:
Return type:

SparseSampleManifest