DCT-Based Structured Transforms for Efficient 2-Bit LLM Quantization
September 11, 2026
A new quantization algorithm replaces dense random orthogonal matrices with sign-randomized Discrete Cosine Transforms (DCT), reducing per-iteration complexity from O(N^2) to O(N log N). The method enables stable 2-bit clustering by using a greedy algorithm with closed-form initialization, bypassing traditional k-means bottlenecks.
HOW THIS AFFECTS YOU
●
builderYou can achieve lower-latency inference through more efficient weight quantization.
●
researcherYou can leverage improved convergence properties in low-bitweight quantization schemes.