Quantization

Reducing the numeric precision used to store a model's weights, such as from 16-bit to 8-bit or 4-bit, to shrink memory footprint and speed up inference. It is usually applied after training, though quantization-aware training folds it into the training loop instead, and it can also be combined with LoRA: quantizing a frozen base model, then training small adapters on top of it.

Why exams ask this

Tested as a deployment tradeoff rather than a training technique. The correct answer trades a small, usually acceptable accuracy loss for a large cut in memory and inference cost.

Relevant to

Related concepts

Resources

No resources linked to this concept yet.