Skip to content

SVD INT8 KV Cache Paging

SVD INT8 KV Cache Paging solves the memory bottleneck of ultra-long context sequences (32K–128K) by decomposing attention keys and values into low-rank singular subspaces.


Algorithmic Overview

For a Key/Value projection tensor \(\mathbf{K} \in \mathbb{R}^{B \times H \times S \times D}\), we compute a calibrated spectral truncated SVD:

\[\mathbf{K} \approx \mathbf{U}_r \mathbf{\Sigma}_r \mathbf{V}_r^T\]

Where rank \(r = 64 \ll D\).

During serving, Turing Engine stores: 1. Low-Rank Projection Basis: \(\mathbf{V}_r \in \mathbb{R}^{D \times r}\) (cached per layer). 2. Symmetric INT8 Quantized Singular Vectors: \(\mathbf{Z} = \text{quant8}(\mathbf{K} \mathbf{V}_r) \in \mathbb{Z}_8^{B \times H \times S \times r}\).


Memory Compression Impact

Sequence Context FP16 KV Memory Turing SVD INT8 Compression Ratio Top-1 Needle Retrieval
8K Context 2.56 GB 0.64 GB -75.0% 100.0%
32K Context 10.24 GB 2.56 GB -75.0% 100.0%
64K Context 20.48 GB 5.12 GB -75.0% 100.0%
128K Context 40.96 GB 10.24 GB -75.0% 100.0%