The Best 2-Bit Quantization on Paper Loses 2.27x to a Trellis Kernel Once You Actually Serve It

arXiv 2609.02652·medium signalXBlueskyLinkedInCopy link

Leech-lattice vector quantization holds the strongest reported 2-bit quality under its own protocol, but no implementation of the multi-shell decoder the rate requires existed; this paper supplies one for the full 301-class codebook with a fused dequantize-plus-matvec kernel and measures batch-1 decode-phase GEMV cost. The key distinction is that the in-VRAM rate is a separate design axis from the on-disk rate: four bit-exact layouts timed in one process show binary bit planes beating one-hot masks at 4.80 bits per weight, 2.15x FP16. Run in the same process against deployed AWQ (4-bit) and QTIP (2-bit) kernels, the trellis kernel reads 2.40x fewer bytes and runs 2.27x faster, with the time gap tracking the traffic gap — the price of unfolding a codebook too large for a lookup table.