arXiv:2609.28029v1 Announce Type: cross Abstract: The key-value (KV) cache of autoregressive transformers can be viewed as a fourth-order tensor spanning attention heads, tokens, features, and groupe...
arXiv:2607.12550v3 Announce Type: replace-cross Abstract: The key-value (KV) cache has become the dominant memory cost of transformer inference: it grows with batch size, context length, and depth, a...
arXiv:2607. 12550v1 Announce Type: new Abstract: The key-value (KV) cache has become the dominant memory cost of transformer inference.
Activation-Aware Weight Tensorization (AWT) is a training‑free calibration wrapper that preconditions each weight matrix with a diagonal activation‑derived scale before applying standard Tensor Train (TT) or Tree Tensor Network (TTN) decompositions. By adding only an input‑side elementwise rescaling, AWT improves functional fidelity of post‑training tensor‑network compression across several large language models (Llama 3.1 8B, Ministral 8B, Qwen2.5 7B) at 2–6× compression, closing a significant portion of the perplexity gap to dense baselines and enhancing downstream task performance. The method demonstrates that diagonal preconditioning offers a robustness‑modularity trade‑off, outperforming full‑covariance oracles in many cases while remaining modular and solver‑agnostic.
The paper introduces JoLT, a training‑free compressor that jointly allocates rank and precision for key‑value (KV) cache compression in long‑context language models. JoLT treats grouped prefill caches as fourth‑order tensors, applies partial Tucker decomposition along token and feature modes, and uses a rotated low‑bit quantizer for residuals, all governed by a single Lagrangian dual under a global byte constraint. Across five models from four architecture families, JoLT achieves 2–3× compression with less than 0.2% perplexity loss, and near‑lossless retrieval accuracy on LLaMA‑3.1‑8B at 64K context up to 3× compression.
arXiv:2606. 03465v1 Announce Type: cross Abstract: Post-training compression is essential for deploying large language models (LLMs) under tight resource constraints.
arXiv:2411. 09816v5 Announce Type: replace Abstract: Large neural networks achieve state-of-the-art performance on many tasks, yet their sheer size hinders deployment on resource-constrained devices.
Kronecker Product Attention (KPA) is introduced as a memory‑efficient method for compressing the key‑value cache in autoregressive models. By representing key and value activations as sums of Kronecker products of compact, token‑dependent matrix factors, KPA generalizes Tensor Product Attention and can capture richer block‑structured patterns. Empirical results show that KPA matches or surpasses existing techniques such as MHA, MQA, GQA, and MLA across multiple model scales and benchmarks, achieving up to 10× compression while preserving relative‑position information through a structured rotary embedding.
Post-training compression is essential for deploying large language models (LLMs) under tight resource constraints. Tensor decompositions have emerged as a promising direction, offering compact parameterizations well suited to Transformer weight structures.
arXiv:2609.15838v1 Announce Type: cross Abstract: Per-matrix singular value decomposition (SVD) truncation is Eckart-Young optimal in the whitened Frobenius norm, but errors from independently compre...
The paper introduces GaugeLasso, a method that applies symmetric group‑lasso penalties to transformer channels during training, enabling entire tensor slices to be zeroed out while maintaining dense tensors for GPU efficiency. By calibrating channel penalties based on inference utility per compute, the network self‑organizes into depth‑dependent structural profiles that can be dramatically smaller than the original architecture, achieving up to 255‑fold compression on a polynomial division task and outperforming hand‑designed baselines on language modeling and autoencoding benchmarks. The approach also accelerates training and reveals over‑provisioned axes that guide subsequent design iterations.
Modern transformers pair impressive capabilities with substantial memory and compute demands. Low-rank weight factorization reduces both while keeping the matrices dense, and thus efficient on standar...