dmx.compressor.layer_reconstruction

Classes

`ApproximationFunctionTuner`(module, search_space)
`LayerReconstructionMixin`(args, *kwargs)	This mixin equips DmxModule with layer-reconstruction functionalities.
`OptimalBrainCompressor`(module)

class dmx.compressor.layer_reconstruction.ApproximationFunctionTuner(module, search_space)

Bases: object

optimize(input, *args, **kwargs)

class dmx.compressor.layer_reconstruction.LayerReconstructionMixin(*args, **kwargs)

Bases: object

This mixin equips DmxModule with layer-reconstruction functionalities. Layer-reconstruction is any post-training process by which certain module parameters are fitted to optimize a local objective, usually by passing data (input activations) through the module. Examples are traditional static activation calibration, static SmoothQuant calibration, Optimal Brain Compression, etc.

calibrating_quantizers(hyperparams) → None

calibrating_smoothquant(hyperparams) → None

enable_approximation_function_tuning(state: bool, hyperparams) → None

enable_optimal_brain_compression(state: bool, hyperparams) → None

enable_quantizer_calib(state: bool, hyperparams) → None

enable_smoothquant_calib(state: bool, hyperparams) → None

optimal_brain_compressing(hyperparams) → None

slanc_tuning(hyperparams) → None

tuning_approximation_function(hyperparams) → None

update_smoothquant_scale(input)

class dmx.compressor.layer_reconstruction.OptimalBrainCompressor(module)

Bases: object

H: Tensor | None = None

apply(microblock_size=1, block_size=128, percdamp=0.01)

measure_hessian(inp)