Qualcomm's NeurIPS paper: Leech lattice vector quantization compresses LLMs without codebooks
HowDevelop · x · 2026-11-07
Qualcomm Research published "Leech Lattice Vector Quantization for Efficient LLM Compression" at NeurIPS 2026. Scalar quantization is fundamentally limited by information-theoretic bounds; vector quantization overcomes this by encoding parameter blocks jointly but usually requires expensive lookups or explicit codebooks. The paper leverages the 24-dimensional Leech lattice—the highest-dimensional lattice with optimal sphere packing and kissing configurations. Key contributions:
- Extends a search algorithm based on the extended Golay code to support indexing (bitstring conversion without materializing the codebook),
- Enables angular search over unions of Leech lattice shells,
- Proposes a fully parallelizable dequantization kernel,
- Provides a geometric reinterpretation of shape–gain quantization combined with GPTQ-style Hessian corrections as a retraction onto a product of spheres (Spherical GPTQ),
- Shows low-angular-distortion LLVQ reduces sensitivity to Hadamard/rotation preprocessing, enabling a strong Hadamard-free PTQ for on-device inference.
More from Infra
- Terrarium open-sources local VM sandbox giving AI agents explicit, auditable permissions — v_0009 · 2026-11-07
- DeepSeek outages draw attention; analyst says it needs hundreds of thousands of inference GPUs — teortaxesTex · 2026-11-07
- BUI decouples Blender's GPU UI kernel into a standalone runtime: 18MB memory, 12ms cold start — Obijuan_cube · 2026-11-07
- TSMC unveils BEOL oxide-semiconductor memory: 0.75V operation, sub-8ns R/W cycles — Terminator857 · 2026-11-07
- PJM data: data-center demand drove a minority of 2026 power cost hikes; electricity itself rose 47.7% — kimmonismus · 2026-11-07
- Box CEO Aaron Levie: long-running AI agents let inference run anywhere, reshaping data center economics — inductionheads · 2026-11-07