Skip to content
#

int3

Here are 4 public repositories matching this topic...

Language: All
Filter by language

Asymmetric NSN: training-free KV cache quantization by separating vector norm and direction. Covers 3-bit/5-bit evaluations across 8 models, bit-packed 3-bit caches for looped Transformers (Ouro, Nanbeige; 64–68% smaller), and a fused Triton V kernel.

  • Updated Nov 10, 2026
  • Python

Add this topic to your repo

To associate your repository with the int3 topic, visit your repo's landing page and select "manage topics."

Learn more