Why it matters

Adafactor cuts optimizer memory. Understanding shapes large-model training.

Advertisement

The architecture

Instead of full v for each param, store row + column stats.

Reconstruct approximation.

Adafactor structureFull v (Adam)per-paramRow + col statsAdafactorReconstructouter productSaves optimizer memory ~half; used for T5, PaLM, Gemini training
Adafactor.
Advertisement

How it works end to end

Row + col sum of squares.

v_ij ~= row_i * col_j / total.

Optional momentum.

Relative step: scale-invariant.