post_training_compression

post_training_compression ¶

Hydra configuration dataclasses for the post-training compression endpoint.

PreparationConfig `dataclass` ¶

PreparationConfig(replace_frozen_batchnorm=True, fuse_conv_batchnorm=True)

Pre-quantization model preparation settings.

Attributes:

Name	Type	Description
`replace_frozen_batchnorm`	`bool`	Whether FrozenBatchNorm layers become plain BatchNorm before fusion.
`fuse_conv_batchnorm`	`bool`	Whether Conv and BatchNorm pairs are fused.

BasePrunerConfig `dataclass` ¶

BasePrunerConfig(amount=MISSING)

Base config for pruning strategies.

Attributes:

Name	Type	Description
`amount`	`float`	Fraction of weights to prune, in (0, 1).

UnstructuredPrunerConfig `dataclass` ¶

UnstructuredPrunerConfig(amount=MISSING, _target_='versatil.post_training_compression.pruning.UnstructuredPruner', layer_types=None)

Bases: BasePrunerConfig

Global L1 unstructured weight pruning.

Note

When layer_types is null, convolution and linear layers are pruned. Use the ${prunable_layer:*} resolver in YAML to constrain the set.

Attributes:

Name	Type	Description
`_target_`	`str`	Import path instantiated by Hydra.
`layer_types`	`list[str] \| None`	PrunableLayerType values to target. Defaults to convolution and linear layers (normalization scales and embedding tables are usually not good pruning targets).

StructuredPrunerConfig `dataclass` ¶

StructuredPrunerConfig(amount=MISSING, _target_='versatil.post_training_compression.pruning.StructuredPruner', norm_order=1, dimension=0, layer_types=None)

Bases: BasePrunerConfig

Per-layer structured weight pruning using Lp-norm channel ranking.

Note

When layer_types is null, defaults to Conv1d, Linear and Conv2d. Use ${prunable_layer:*} resolver in YAML to add types.

Attributes:

Name	Type	Description
`_target_`	`str`	Import path instantiated by Hydra.
`norm_order`	`int`	The p in Lp-norm used to rank channels (e.g., 1 for L1, 2 for L2).
`dimension`	`int`	Weight tensor dimension along which to prune.
`layer_types`	`list[str] \| None`	PrunableLayerType values to target. Defaults to Conv1d, Conv2d, and Linear.

CompressionTargetConfig `dataclass` ¶

CompressionTargetConfig(_target_='versatil.post_training_compression.compression_target.CompressionTarget', module_path=MISSING, preparation='${preparation}', pruning='${pruning}')

Per-module preparation and pruning scheme with inheritance.

Note

Absent fields in YAML inherit from the global config via Hydra interpolation defaults. Explicit null means skip.

Attributes:

Name	Type	Description
`_target_`	`str`	Import path instantiated by Hydra.
`module_path`	`str`	Dotted path to the target submodule, or empty string for the full policy.
`preparation`	`PreparationConfig \| None`	BN replacement and fusion settings.
`pruning`	`list[Any] \| None`	Pruning strategies to apply sequentially.

TorchInductorBackendConfig `dataclass` ¶

TorchInductorBackendConfig(_target_='versatil.post_training_compression.deployment_backends.torch_inductor.TorchInductorBackend')

Torch inductor deployment backend that writes .pt2 artifacts.

ExecutorchXNNPACKBackendConfig `dataclass` ¶

ExecutorchXNNPACKBackendConfig(_target_='versatil.post_training_compression.deployment_backends.executorch_xnnpack.ExecutorchXNNPACKBackend', max_batch_size=32)

ExecuTorch backend that lowers artifacts to XNNPACK .pte files.

Attributes:

Name	Type	Description
`_target_`	`str`	Import path instantiated by Hydra.
`max_batch_size`	`int`	Upper bound for dynamic batch execution in the serialized ExecuTorch program.

PostTrainingCompressorConfig `dataclass` ¶

PostTrainingCompressorConfig(_target_='versatil.post_training_compression.compressor.PostTrainingCompressor', checkpoint_path=MISSING, checkpoint_name=value, output_directory=None, calibration_steps=32, generate_report=False, modules=list(), preparation=PreparationConfig(), pruning=None, quantization=None, deployment_backend=None)

Top-level config for the post-training compression endpoint.

Note

Global fields serve as defaults inherited by per-module configs via Hydra interpolation.

Attributes:

Name	Type	Description
`_target_`	`str`	Import path instantiated by Hydra.
`checkpoint_path`	`str`	Path to the training checkpoint directory.
`checkpoint_name`	`str`	Checkpoint filename inside the directory.
`output_directory`	`str \| None`	Where to save compressed output. Defaults to checkpoint_path/compressed/.
`calibration_steps`	`int`	Number of calibration batches for static quantization.
`generate_report`	`bool`	Whether to generate a quantization report after saving. Disabled by default since it runs additional forward passes for benchmarking.
`modules`	`list[CompressionTargetConfig]`	Per-module compression schemes (empty = global).
`preparation`	`PreparationConfig`	Global preparation settings.
`pruning`	`list[Any] \| None`	Global pruning strategies (inherited by modules).
`quantization`	`Any \| None`	Quantization workflow. `None` exports the float model without quantization.
`deployment_backend`	`Any \| None`	Deployment backend that owns artifact format and lowering. Defaults to torch inductor.

post_training_compression