Looped CED 循环架构让 KV 缓存零增长，推理吞吐提升 3.6 倍
Looping Deeper Without Growing the KV Cache: Recurrent Depth for Causal Encoder-Decoders
循环加深模型本来会把 KV 缓存撑爆，这篇论文找到放编码器侧循环的办法。在 552B 的 DeepSeek-V4.1-Flash 上实测，KV 少 36.8 倍、吞吐高 3.6 倍。做推理服务和长上下文部署的工程师值得看。
论文提出 Looped CED（因果编码器-解码器）架构，通过循环加深计算而 KV 缓存零逐轮增长。相比普通 looped Transformer，它以 1/6 的全局 KV 恢复了 86-93% 的循环收益。机制上，循环被放在编码器侧以迭代精炼全局 KV 表示，训练时用随机循环深度诱导不动点均衡，使每轮可无损复用上一轮的 KV 缓存。
实验在 552B 参数的 DeepSeek-V4.1-Flash 上进行，在一个 4xGB200 节点上循环 8 轮：普通 per-layer/per-round KV 方案连一个 128K token 请求都放不下，而 Looped CED 每 token KV 减少 36.8 倍，可同时容纳 38 个 128K 请求（与未循环模型持平），8K prompt 上吞吐提升 3.6 倍。这为长上下文推理服务在显存受限场景下的批处理扩展提供了具体路径。
Looping Deeper Without Growing the KV Cache: Recurrent Depth for Causal Encoder-Decoders
Looped language models deepen computation without adding parameters, yet caching full-context key-value (KV) states across layers and rounds severely bounds serving batch size. We design a looped causal encoder-decoder (CED) that achieves recurrent depth with zero round-wise KV growth, addressing where and how to loop an asymmetric stack in which only sparse encoder layers generate global KV. We find that recurrence must be placed in the encoder to iteratively refine global KV representations rather than re-reading static decoder projections, which both drives language-modeling gains and preserves long-context scaling. Furthermore, training the encoder loop under stochastic recurrence depth induces a fixed-point equilibrium where every round losslessly shares the last round's KV cache. Compounding cross-layer and cross-round KV sharing, the resulting Looped CED matches an unlooped CED's minimal memory footprint and recovers 86-93% of a looped Transformer's gain with 1/6 of its global KV. To test serving at production scale, we loop the 552B-parameter DeepSeek-V4.1-Flash for 8 rounds on one 4xGB200 node: with per-layer and per-round KV the loop cannot fit a single 128K-token request, whereas Looped CED holds 36.8x less KV per token, fits 38 such requests (as many as the unlooped model), and serves 3.6x higher throughput on 8K-token prompts.
本条情报归属 AITOP「论文」核心栏目，重点聚焦中国 AI 产业与国产大模型生态。在跨区域产业观察坐标中，该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：DeepSeek、KV Cache、推理优化、开源模型，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始信息发布自一手权威信源「arXiv: DeepSeek」（Tier-1 一手发布），具备高确定性参考价值。
