Transformer-FSQ: Discrete Latents for Extreme KV-Cache Compression

Long-context language modeling is increasingly constrained by the Key-Value (KV) cache, whose memory and decode-time access costs scale linearly with the prefix length. This bottleneck has motivated a range of context-compression methods, from token-level summarization to recent optimization-based KV compression methods. These post-hoc methods operate on the KV cache of a fixed pretrained model, so their effectiveness is fundamentally limited by how well the model's internal representations can be compressed. In this work, we formalize the notion of KV compressibility and show that it is a property of the learned representations, rather than of the context alone. We prove that almost any sequence-to-vector function admits both highly compressible and inherently non-compressible transformer implementations, highlighting the need to guide transformers toward compressible representations during training. Motivated by this, we propose KV-Compression Aware Training (KV-CAT), a continued pretraining procedure that incentivizes the emergence of compressible representations. We introduce a train-time KV sparsification policy that masks KV slots during training. This forces the model to use fewer KV slots and encourages it to learn representations amenable to post-hoc compression. Empirically, we show that KV-CAT improves the quality-budget tradeoff of downstream compression methods across retrieval, long-context question answering, and perplexity-based evaluation of compressed-prefix continuation.

Recent work on KV cache quantization, culminating in TurboQuant, has approached the Shannon entropy limit for per-vector compression of transformer key-value caches. We observe that this limit applies to a strictly weaker problem than the one that actually matters: compressing the KV cache as a sequence. The tokens stored in a KV cache are not arbitrary floating-point data -- they are samples from the exact formal language the model was trained on, and the model is by construction a near-optimal predictor of that language. We introduce sequential KV compression, a two-layer architecture that exploits this structure. The first layer, probabilistic prefix deduplication, identifies semantically equivalent shared prefixes across sessions using the trie metric d_T(s, s') = -log_2 P_M(s ^ s') from Probabilistic Language Tries (PLTs). The second layer, predictive delta coding, stores only the residual of each new KV vector from the model's own prediction of it, achieving a per-token entropy bound of H(KV_{i+1} | KV_{<=i}) <= H(token_{i+1} | token_{<=i}). We prove that at typical language model perplexity -- approximately 10-20 for fluent English text -- this bound is 3.3-4.3 bits on average per token position, compared to TurboQuant's 3 bits per vector component (with typical attention heads having 64-128 components). The theoretical compression ratio over TurboQuant is approximately 914,000x at the Shannon limit. Even at 1000x above the entropy floor -- a deliberately pessimistic worst-case overhead, two orders of magnitude above the 2-5x typical of practical source coders -- the ratio remains approximately 914x over TurboQuant, with compression improving rather than degrading as context length grows. The two layers are orthogonal and compose with existing per-vector quantization methods including TurboQuant.

Standard Transformer attention uses identical dimensionality for queries, keys, and values, yet these components serve different roles: queries and keys produce scalar attention weights (selection), while values carry rich representations (value transfer). We show that selection requires only $O(\log N)$ dimensions to distinguish among $N$ relevant token categories (e.g., syntactic roles, semantic clusters, positional patterns) -- far fewer than value transfer needs. We introduce factored keys, which exploit this asymmetry to physically shrink the KV cache of any pretrained model without retraining from scratch -- unlike Grouped-Query Attention (GQA) and Multi-Head Latent Attention (MLA), which must be designed into the architecture before pretraining. We factorize each key projection $W_K \approx A_{d \times r} B_{r \times d}$ via truncated singular value decomposition (SVD) (where $r$ is the chosen compression dimension), set $W_K' = A$ as the new key projection producing compact $r$-dimensional keys for the cache, and absorb $B^\top$ into the query projection ($W_Q' = W_Q B^\top$) at zero cost -- since queries are never cached. At the 7B scale, training from scratch with $r = d/4$ (where $d$ is the model dimension) matches full-attention perplexity ($9.24$ vs $9.25$ PPL after 20B tokens, mean over two seeds) while using 12% fewer parameters and training 8% faster. For existing models, SVD followed by QK fine-tuning (3 epochs, less than 1% of pretraining data) achieves 75% key cache savings at roughly 2% quality cost on both GPT-2 and Mistral-7B. The approach composes with GQA and quantization for up to $16\times$ combined key cache compression. For a 7B model serving a 128K context, factored keys save 25 GB of KV cache per user, enabling roughly 60% more concurrent users on identical hardware.