Why it matters

bnb int8 makes big models loadable. Understanding shapes quant choice.

Advertisement

The architecture

Row-wise scaling per weight.

Outlier features kept fp16.

Two matmuls combined.

bitsandbytes flowRow-wise scaleint8 weightsOutlier featureskept fp16Combined matmulint8 + fp16Dettmers 2022; also 8-bit optimizers in library
bnb int8.
Advertisement

How it works end to end

load_in_8bit=True in HF.

Quality: negligible drop.

Memory: ~50%.

Speed: often slower than fp16.