Squeeze the Weights

Round a row of weights to 8, 4, 3 or 2 bits, add one outlier, and see why a shared scale fails and per-group scales rescue it.

From Chapter 11: Inside Modern LLMs