Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu · 2026 · arXiv (Cornell University)
DOI: 10.48550/arxiv.2610.31633 · cited by 0 works (OpenAlex) · open-access PDF
Be Looped finplex là giá trị kết hợp quy mô cao cấp trong khi giữ số tham số được cố định bằng cách lặp đi lặp lại các kích hoạt thông qua cùng một khối các lớp. Trong suy luận, mỗi lặp lại tạo ra KV-cache độc lập thông qua, gây ra bộ nhớ để phát triển tuyến tính. Để giải quyết thách thức này, công việc trước đã đề xuất KV-shareability cơ chế cơ bản (tức là, chia sẻ KV-cache qua chiều sâu), thấy rằng, trong một số trường hợp, KV-cache của sự lặp lại cuối cùng có thể được chia sẻ với các lặp lại trước đó mà không mất chất lượng, một thuộc tính mà chúng tôi chỉ định là KV-shareability. Tuy nhiên, các cơ chế cơ bản của KV-shareability vẫn còn kém hiểu, với các kiến trúc được đào tạo cùng nhau không thể hiện nó. Trong bài báo
Machine translation (Meta M2M100); the English original follows.
Looped architectures scale quality while keeping parameter count fixed by repeatedly passing activations through the same block of layers. During inference, each recurrence generates independent KV-caches, causing memory to grow linearly. To address this challenge, prior work has proposed KV-sharing (i.e., sharing KV-caches across depth), finding that, in some cases, the KV-cache of the final recurrence can be shared with the earlier recurrences without quality loss, a property we denote as KV-shareability. However, the mechanisms underlying KV-shareability remain poorly understood, with even identically trained looped architectures not consistently exhibiting it. In this paper, we study how KV-shareability can be enabled in looped models. We propose a two-stage method: in the first stage, we use stochastic depth during pretraining, which induces contractive residual dynamics that drive representations toward a common limit across recurrences. In the second stage, we propose BeNice!, a post-training method that broadcasts the final recurrence's KV-cache into the first recurrence, training it to produce compatible residuals while preserving the original model's quality. As a result, BeNice! enables KV-shareability while matching standard finetuning's unshared validation perplexity. Using BeNice!, we show that KV-sharing is complementary to alternative KV-compression methods, composing with memory-efficient sequence mixers (e.g., multi-latent attention, Kimi delta attention) and low-precision KV-cache quantization (e.g., fp8, mxfp8, nvfp4). Finally, in simulated low-memory, long-context inference, KV-sharing improves cache hit rates by up to 31.4%, increases throughput by up to 23.6%, and reduces p90 time-to-first-token by up to 83.6%.
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
Topics: Quantization (signal processing), Computer science, Residual, Property (philosophy), Block (permutation group theory), Algorithm
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.31633, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange