Accuracy Improvement Methods

Layer-wise Percentile

Improve quantization accuracy by calibrating per-layer clipping ranges using percentile statistics.

Quark ONNX Quantization Tutorial For Layerwise Percentile

Cross Layer Equalization (CLE)

Equalize weight ranges across adjacent layers to reduce quantization error before calibration.

Quark ONNX Quantization Tutorial For Cross Layer Equalization (CLE)

ADAQuant

Apply adaptive rounding of quantized weights to minimize layer-wise reconstruction error.

Quark ONNX Quantization Tutorial For AdaQuant

ADARound

Learn task-aware rounding decisions for quantized weights to recover post-quantization accuracy.

Quark ONNX Quantization Tutorial For AdaRound

Mixed Precision

Assign different bit-widths to different layers to balance accuracy and performance.

Quark ONNX Quantization Tutorial For Mixed Precision

Smooth Quant

Migrate quantization difficulty from activations to weights via mathematically equivalent scaling.

Quark ONNX Quantization Tutorial For Smooth Quant