Query-Whitened Low-Rank Snapshots of Gated DeltaNet States for Hybrid Prefix Caching
Hybrid language models that interleave full attention with Gated DeltaNet (GDN) layers cannot reuse a cached prefix from its key-value (KV) cache alone: they also need the recurrent state at the reuse point. Engines therefore store state snapshots, which cost tens of megabytes each and limit how many restore points fit in memory. We compress these snapshots without training. Each GDN head state is truncated to low rank under a single global rank budget shared by all heads, after whitening by the second moment of the head's queries, so that the truncation minimises the expected readout error rather than the Frobenius error; we show that pooling whitened eigenvalues across heads is the optimal allocation for this summed error. On Qwen3.5-2B and 9B, query whitening cuts next-token NLL damage at 16x state compression by 34-46% relative to unweighted truncation, and the compressed snapshot (state plus convolution cache, int8 factors) is 1.4-2.0 MB instead of 52 MB for 9B. On state-sensitive RULER common-word extraction and on LongBench QA, where discarding the state costs up to 55 points, the compressed snapshot's point estimates stay within 3 points of the exact one. Against our reimplementation of tail replay, which rebuilds the state from cached per-token hidden states, compression is more accurate on 2B (significant in three of four comparisons, +3.0 to +6.8 points) and matches it on 9B, while storing tens to hundreds of times less data and restoring in 0.2 ms. Code and per-item results will be released.
Authors
- Nihal Krishna Shrivastava
Institutions
- Birla Institute of Technology and Science, Pilani - Goa Campus (IN)
- Birla Institute of Technology and Science, Pilani (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24160128
- Primary Topic
- Topic Modeling
- Type
- preprint