Module quantize 

Precision conversion passes for NPU quantization.

Rewrites array<f32> global variable types to lower-precision element types suitable for specific NPU backends.

Supports both naive type rewriting and calibration-based quantization with proper scale/zero_point computation.

Structs

CalibrationData

Calibration data for the entire module.

F32ToBf16

Converts array<f32> global variables to array<bf16>.

F32ToF16

Converts array<f32> global variables to array<f16>.

F32ToInt8

Converts array<f32> global variables to array<i8>.

MixedPrecisionPass

Mixed-precision quantization pass.

MixedPrecisionPolicy

Policy for choosing per-layer precision in mixed-precision quantization.

PerChannelQuantParams

Per-channel quantization parameters.

QuantizationParams

Parameters describing how floating-point values were quantized to integers.

SensitivityScore

Sensitivity score for a single layer (lower = more sensitive).

TensorCalibration

Calibration data for a single tensor, identified by binding.

Functions

compute_calibrated_params

Compute per-tensor quantization parameters from calibration data and module.

estimate_sensitivity

Compute sensitivity scores for all storage globals in a module.

policy_from_sensitivity

Build a mixed-precision policy from sensitivity scores.

Type Aliases

LayerPrecision

Target precision for a single layer/variable in mixed-precision mode.