Study
LLMs · Quantisation
Post-training quantisation for inference: absmax and zero-point int8, per-tensor vs per-channel scales, how one outlier ruins a whole tensor's scale, and PyTorch dynamic quantisation on a small model.
- Quantisation maps floats to small integers with a scale (and sometimes a zero point), trading a little accuracy for memory and speed.
- One large outlier stretches the scale for everyone; per-channel or blockwise scales limit the damage.