FibQuant: Universal Vector Quantization for Random-Access KV-Cache Compression
FibQuant: универсальное векторное квантование для сжатия KV-кеша с произвольным доступом
2026-05-12
SCID: 54.1/wzs76nyb
Discuss with AI
Beta-quantile radii and Fibonacci/Roberts–Kronecker directionskey–value (KV) cache compressionnormalize–rotate–store interfaceradial–angular codebookrandom-access universal vector quantization
Figures from the paper
Abstract (AI)
Long-context inference is increasingly a memory-traffic problem. The culprit is the key--value (KV) cache: it grows with context length, batch size, layers, and heads, and it is read at every decoding step. Rotation-based scalar codecs meet this systems constraint by storing a norm, applying a shared random rotation, and quantizing one coordinate at a time. They are universal and random-access, but they discard the geometry created by the normalization step. After a Haar rotation, a block of $k$ consecutive coordinates is not a product source; it is a spherical-Beta source on the unit ball. We introduce \textsc{FibQuant}, a universal fixed-rate vector quantizer that keeps the same normalize--rotate--store interface while replacing scalar tables by a shared radial--angular codebook matched to this canonical source. The codebook combines Beta-quantile radii, Fibonacci\,/\,Roberts--Kronecker quasi-uniform directions, and multi-restart Lloyd--Max refinement. We prove that the resulting vector code strictly improves on its scalar product specialization at matched rate, with a high-rate gain that separates into a cell-shaping factor and a density-matching factor. The same construction gives a dense rate axis, including fractional-bit and sub-one-bit operating points, without calibration or variable-length addresses. On GPT-2 small KV caches, \textsc{FibQuant} traces a memory--fidelity frontier from $5\times$ compression at $0.99$ attention cosine similarity to $34\times$ at $0.95$. End-to-end on TinyLlama-1.1B, it is within $0.10$ perplexity of fp16 at $4\times$ compression and has $3.6\times$ lower perplexity than scalar \textsc{TurboQuant} at $b = 2$ ($8\times$ compression), where scalar random-access quantization begins to fail.
Key Findings
1
FibQuant is a universal fixed-rate vector quantizer compatible with the normalize–rotate–store interface used for random-access KV-cache compression.
2
FibQuant provides a dense rate axis including fractional-bit and sub-one-bit operating points without calibration or variable-length addresses.
3
FibQuant replaces scalar tables with a shared radial–angular codebook matched to the spherical-Beta source induced by Haar rotation, using Beta-quantile radii and quasi-uniform (Fibonacci/Roberts–Kronecker) directions refined by multi-restart Lloyd–Max.
4
On GPT-2 small KV caches, FibQuant achieves a memory–fidelity frontier from 5× compression at 0.99 attention cosine similarity to 34× at 0.95; on TinyLlama-1.1B it is within 0.10 perplexity of fp16 at 4× compression and has 3.6× lower perplexity than scalar TurboQuant at b=2 (8× compression).
5
The vector code provably strictly improves on its scalar product specialization at matched rate, with high-rate gains decomposing into a cell-shaping factor and a density-matching factor.
Research Object
Key–value (KV) cache for transformer models (random-access KV-cache used during long-context inference)
Research Subject
Universal fixed-rate vector quantization (FibQuant) for random-access compression of the KV-cache, specifically a radial–angular codebook replacing scalar codecs to improve compression–fidelity trade-offs, enable fractional-bit rates, and reduce memory traffic during decoding
Publication Details
Publication Date
2026-05-12
Journal
Publisher
ISSN
Cited by
0
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest