Future

[aksheyd]

• Move symmetric block quantization to chapter 4
• Add asymmetric, precision aware block quantization as new algorithm (this is where the range is no longer -127...127 -> you transpose scale onto min...max (instead of just [-max, max] and then add a zero-point to shift the effective scale onto zero-point as zero)
• Move to chapter 5
• Add adaptive / mixed-precision quantization using a calibrated threshold (this is where you change the precision based on the blocks itself, you can also manipulate block size programmaticlaly)
• Move to Chapter 6
• Train a simple model to use scale and zero point as weight and bias to learn effective block quantization
• Move to Chapter 7