DeepSeek's CED applies YOCO-style shared KV caching only to global CSA2 branches
stochasticchasm · x · 2026-09-11
- In DeepSeek's CED, the YOCO-like trick is applied only to the global branch of each CSA2 layer, uniquely per layer.
- Each decoder CSA2 layer has a global branch plus a local sliding-window branch: the SWA branch is computed from current-layer hidden states, while the global branch's KV cache is just a linear projection of encoder hidden states — so no work is needed at prefill, and decode KVs come from a quick linear projection.
- YOCO itself is simple: the first half computes normally, hidden states are projected once more to K/Vs, and that single KV cache is shared across every layer in the second half.
Related event: Inside KV-Cache Sharing: How DeepSeek CED and GLM 5.2 Differ(4 posts)→
More from Research
- Stanford's PaperFormalizer turns research papers into Lean-verified versions with author in the loop — StanfordUniversity · 2026-10-27
- JockeyBench uses Japanese horse racing to test LLM agents — all betting strategies lose money — Zhongwei Xie · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- Hugging Face TRL makes score centering + masked importance sampling default in async GRPO — _lewtun · 2026-10-27
- AutoResearch agent's weird training curve gets dug into instead of discarded — ZimingLiu11 · 2026-10-27
- EMNLP 2026: sub-billion-parameter LMs show superhuman entity tracking, human-like fleeting memory — GretaTuckute · 2026-10-27