ExTernD: Overcoming the Accuracy Limits of Ternary Quantization

·2026.07.16 22:31

Key point

A new PTQ methodology is proposed that achieves efficient Ternary Quantization while minimizing accuracy loss through matrix decomposition.

Details

Noting that the existing Ternary PTQ approach, which uses a fixed matrix size, has limitations in maintaining accuracy, we propose ExTernD, a method that decomposes a matrix into two ternary matrices and an internal diagonal scaling matrix.

This approach allows the Inner Rank to be arbitrarily expanded, making it possible to maintain high accuracy regardless of the quantization level. It also has the advantage of maximizing the benefits of ternary operations while incurring very little increase in VRAM usage compared to existing quantization methods.

llm-quantization ptq ternary-quantization matrix-decomposition externd

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.