API Reference

coreai_opt

coreai_opt - A library for PyTorch model compression and optimizations.

coreai_opt.CoreMLExportError(dtype, context)	Raised when a model cannot be exported to the CoreML backend.
coreai_opt.ExportBackend(value, *args, **kwargs)	Enum representing supported model export backends.

coreai_opt.casting

Casting related utilities including FP32 -> FP16 and INT32 -> INT16 passes.

coreai_opt.casting.cast_fp32_to_fp16(...)	Convert a torch exported program from FP32 to FP16 where applicable.
coreai_opt.casting.cast_int32_to_int16(...)	Convert INT32/INT64 tensors to INT16 in a torch exported program.
coreai_opt.casting.cast_to_16_bit_precision(...)	Convert a torch exported program to 16-bit precision: FP32→FP16 and INT32/64→INT16.

coreai_opt.config

Configuration and specification modules for coreai_opt.

coreai_opt.config.CompressionConfig	Top level configuration class for model compression.
coreai_opt.config.CompressionSpec	Base class for compression specifications.
coreai_opt.config.CompressionType(value[, ...])	Enum representing compression techniques applied to the model.
coreai_opt.config.ModuleCompressionConfig	Abstract base configuration class for module-level compression settings.
coreai_opt.config.OpCompressionConfig	Abstract base configuration class for op-level compression settings.
coreai_opt.config.WeightOnlyModuleValidationMixin()	Mixin that adds weight-only validation to ModuleCompressionConfig subclasses.
coreai_opt.config.WeightOnlyOpValidationMixin()	Mixin that adds weight-only validation to OpCompressionConfig subclasses.

coreai_opt.config.spec

Base abstractions for compression specs, simulators, and component factories.

coreai_opt.config.spec.CompressionComponentFactoryBase()	Abstract base class for compression component factories.
coreai_opt.config.spec.CompressionSimulatorBase(...)	Abstract base class for compression simulators.
coreai_opt.config.spec.CompressionTargetTensor(value)	Enum to specify the target tensor for compression.

coreai_opt.coreai_utils

Core AI MLIR-level compression transforms.

coreai_opt.coreai_utils.CompressionGranularity(value)	Enum representing the granularity of quantization for Core AI weight compression.
coreai_opt.coreai_utils.DType(value[, ...])	Enum representing data types for Core AI weight compression.
coreai_opt.coreai_utils.palettize_weights(...)	Palettize weights in a Core AI AIProgram (MLIR<CoreAI> IR) by using Core AI ops.
coreai_opt.coreai_utils.quantize_weights(...)	Quantize weights in a Core AI AIProgram (MLIR<CoreAI> IR) by using Core AI ops.
coreai_opt.coreai_utils.sparsify_weights(...)	Sparsify weights in a Core AI AIProgram (MLIR<CoreAI> IR) by using Core AI ops.

coreai_opt.coreai_utils.common

Common enums and constants for coreai_opt.coreai_utils.

coreai_opt.coreai_utils.common.QScheme(value)

Enum representing the quantization scheme.

coreai_opt.inspection

Utilities for inspecting model operations and compression configuration.

coreai_opt.inspection.ModelInspector(model, ...)	Inspect operations in a PyTorch model for compression configuration.
coreai_opt.inspection.ModelSummary(model, mode)	Complete listing of operations discovered in a model.
coreai_opt.inspection.ModuleContext(...)	One level of the nn.Module nesting hierarchy.
coreai_opt.inspection.ModuleInfo(...)	A node in the nn.Module hierarchy with its directly-owned ops.
coreai_opt.inspection.OpInfo(op_name, ...)	Information about a single operation discovered in a model.
coreai_opt.inspection.SourceFrame(filename, ...)	A single frame in the source call stack leading to an operation.

coreai_opt.palettization

Palettization specification and utilities for weight compression via lookup tables.

coreai_opt.palettization.KMeansPalettizer(model)	K-means palettizer with integrated supported operations strategy.
coreai_opt.palettization.KMeansPalettizerConfig	Top-level configuration class for kmeans palettization.
coreai_opt.palettization.ModuleKMeansPalettizerConfig	Configuration for palettizing a specific module using K-means clustering.
coreai_opt.palettization.PalettizationSpec	Specification for palettization compression of neural network weights.

coreai_opt.palettization.config

Palettization configuration classes.

coreai_opt.palettization.config.OpKMeansPalettizerConfig

Configuration class for palettization at the operation level.

coreai_opt.palettization.spec

Palettization specs, granularity classes, and factory functions.

coreai_opt.palettization.spec.PalettizationGranularity	Base class for palettization granularity specifications.
coreai_opt.palettization.spec.PerGroupedChannelGranularity	Per-grouped-channel palettization granularity.
coreai_opt.palettization.spec.PerTensorGranularity	Per-tensor palettization granularity.
coreai_opt.palettization.spec.default_weight_palettization_spec()

coreai_opt.pruning

Pruning infrastructure for coreai_opt.

coreai_opt.pruning.MagnitudePruner(model[, ...])	Apply magnitude-based pruning to a model.
coreai_opt.pruning.MagnitudePrunerConfig	Top-level configuration for magnitude pruning.
coreai_opt.pruning.ModuleMagnitudePrunerConfig	Module-level pruning configuration.
coreai_opt.pruning.PruningSpec	Specification for pruning tensors.

coreai_opt.pruning.config

Pruning configuration exports.

coreai_opt.pruning.config.ConstantSparsitySchedule	Step function: zero before begin_step, target_sparsity at and after.
coreai_opt.pruning.config.OpMagnitudePrunerConfig	Operation-level pruning configuration.
coreai_opt.pruning.config.PolynomialDecaySchedule	Polynomial schedule from initial_sparsity to target_sparsity.
coreai_opt.pruning.config.SparsityScheduleBase	Abstract base for sparsity schedules used by MagnitudePruner.

coreai_opt.pruning.spec

Pruning spec components: specs, schemes, and parametrizations.

coreai_opt.pruning.spec.ChannelStructured	Channel-structured pruning scheme.
coreai_opt.pruning.spec.PruneImplBase(...)	Abstract base for pruning parametrizations that mask a layer's weight.
coreai_opt.pruning.spec.PruningScheme	Base class for pruning scheme specifications.
coreai_opt.pruning.spec.Unstructured	Unstructured pruning scheme.
coreai_opt.pruning.spec.default_weight_pruning_spec()	Return the default pruning spec for weight tensors.

coreai_opt.quantization

Quantization compressor, configuration, specs, and granularity classes.

coreai_opt.quantization.ExecutionMode(value, ...)	Enum representing quantization execution modes.
coreai_opt.quantization.ModuleQuantizerConfig	Configuration class for quantization at the module level.
coreai_opt.quantization.QuantizationSpec	Specification for quantizing tensors in neural networks.
coreai_opt.quantization.Quantizer(model[, ...])	Unified quantizer API that provides a single entry point for various quantization workflows, including:
coreai_opt.quantization.QuantizerConfig	Top-level configuration class for quantization.

coreai_opt.quantization.config

Quantization configuration classes and execution mode.

coreai_opt.quantization.config.OpQuantizerConfig	Configuration class for quantization at the operation level.
coreai_opt.quantization.config.QATSchedule	Schedule for controlling observer and fake quantization state in QAT.

coreai_opt.quantization.spec

Quantization specs, schemes, granularity classes, and parameter calculators.

coreai_opt.quantization.spec.GlobalMinMaxQParamsCalculator(...)	Computes scale and zero point by tracking the running min/max.
coreai_opt.quantization.spec.MinMaxRangeCalculator(...)	Range calculator that computes the range of a given tensor as the min and max values of the tensor.
coreai_opt.quantization.spec.MovingAverageQParamsCalculator(...)	Computes the scale and zero point using a moving average of the range.
coreai_opt.quantization.spec.PerBlockGranularity	Per-block quantization granularity.
coreai_opt.quantization.spec.PerChannelGranularity	Per-channel quantization granularity.
coreai_opt.quantization.spec.PerTensorGranularity	Per-tensor quantization granularity.
coreai_opt.quantization.spec.QParamsCalculatorBase(...)	Base class for implementing logic to calculate quantization parameters (scale, zero_point, minval) given min/max values.
coreai_opt.quantization.spec.QuantizationComponentFactory()	Factory class for creating quantization components from QuantizationSpec.
coreai_opt.quantization.spec.QuantizationFormulation(value)	Formula used to map between quantized integers and dequantized values.
coreai_opt.quantization.spec.QuantizationGranularity	Base class for quantization granularity specifications.
coreai_opt.quantization.spec.QuantizationScheme(value)
coreai_opt.quantization.spec.RangeCalculatorBase(...)	Base class and registry for classes used to compute the range of a given tensor.
coreai_opt.quantization.spec.RunningRangeMixin(...)	Mixin for calculators that maintain running min/max range buffers.
coreai_opt.quantization.spec.StaticQParamsCalculator(...)	Computes scale and zero point using min/max values from the current tensor.
coreai_opt.quantization.spec.default_activation_quantization_spec()
coreai_opt.quantization.spec.default_weight_quantization_spec()

coreai_opt.quantization.spec.fake_quantize

Fake quantization implementation base class and default implementation.

coreai_opt.quantization.spec.fake_quantize.FakeQuantizeImplBase(...)

Base class for implementing fake quantization