Interactive Quiz

Symmetric quantization uses an offset zero-point while asymmetric quantization maps zero to zero.

Symmetric quantization maps the floating zero to the quantized zero, whereas asymmetric quantization maps the floating minimum and maximum range to the quantized range with an offset zero-point.

Asymmetric quantization is always faster than symmetric quantization.

Symmetric quantization is used only for activations, asymmetric quantization only for weights.

QAT allows quantization without precision loss even at 1 bit.

QAT integrates the quantization process during training, allowing the model to adapt to quantization-induced errors.

PTQ requires fewer computations than QAT, making QAT unnecessary for large models.

QAT is only compatible with symmetric quantization.

Adjust the model weights to optimize accuracy.

Find the range of values for activations or weights that minimizes quantization error.

Reduce the model size by decreasing the number of bits used.

Modify the model structure to improve inference speed.

To increase the precision of quantized weights by reducing the dynamic range.

To reduce the effect of extreme values (outliers) that distort the mapping and increase quantization error.

To ensure that quantization is always symmetric around zero.

To speed up the calculation of the scale factor and zero-point.

Dynamic quantization calculates the scale factor and zero-point before inference, while static quantization does so during inference.

Static quantization requires a calibration dataset to compute the scale factor and zero-point before inference, whereas dynamic quantization calculates them during inference for each layer.

Dynamic quantization is always faster than static quantization.

Static quantization can only be used for weights, not for activations.