Quantization-Aware Training
Simulating low-precision rounding during the forward pass while training, so a network's weights adapt to quantization error before the model ever ships.
Every forward pass rounds weights (and often activations) onto a low-precision grid, called fake quantization, so the network experiences the same rounding error during training that it will face at deployment. Since rounding has a gradient of zero almost everywhere, the backward pass uses a straight-through estimator that treats the rounding step as the identity function, letting the underlying full-precision weight keep receiving a usable gradient.
This consistently outperforms quantizing an already-trained network afterward, because the optimizer gets many chances to find weights that are robust to rounding rather than surprised by it.