Why it matters

SqueezeLLM pushes low bits. Understanding shapes quant frontier.

Advertisement

The architecture

Sensitivity-based clustering (k-means).

Sparse outlier storage.

Dense + sparse combined.

SqueezeLLM structureSensitivity analysisper weightCluster + sparse outliersnon-uniformCombineddense + sparseKim et al. 2023; better quality than uniform quant at same bits
SqueezeLLM.
Advertisement

How it works end to end

Non-uniform quantization.

Outliers stored sparse.

Custom kernels needed.

3-4 bit target.