3 items tagged with "quantization"
Quantization stores model weights at lower bit-widths to cut memory and speed inference; full precision preserves maximum accuracy. The trade-off is size and speed versus fidelity.
Quantization reduces the numerical precision of a model's weights and activations, for example from 32-bit floating point to 8-bit or 4-bit integers, ...
Quantization reduces the numeric precision of a model's weights and activations, shrinking memory use and speeding inference with limited accuracy loss.