DeepSeek's CED vs GLM's KV reuse: a developer unpacks how the cache-sharing modes actually differ
stochasticchasm · x · 2026-09-11
- stochasticchasm finds the emerging naming taxonomy for KV-cache reuse modes a bit funny but useful, and wonders what you'd call a mode that reuses indices but not KV, as in GLM 5.2.
- In a follow-up they detail DeepSeek's CED: the same idea, but applied only to the global branch of each CSA2 layer, uniquely per layer. In the decoder, every CSA2 layer has a global branch and a local SWA branch — the SWA branch computes from current-layer hidden states, while the global branch's KV cache is just a linear projection of encoder hidden states, and it's skipped during prefill.
- Useful architectural detail for anyone following long-context inference and KV-cache economics.
Related event: Inside KV-Cache Sharing: How DeepSeek CED and GLM 5.2 Differ(4 posts)→
More from Research
- FLASH: GPU-native deformable sim learns cloth folding in minutes on one RTX 5090 — ChongZzZhang · 2026-10-26
- LEMON: zero-shot LLM misbehavior monitor flags hallucinations from attention maps — yoavartzi · 2026-10-26
- Behind the EMNLP keynote: LMs as simulated learners to generate child language acquisition hypotheses — najoungkim · 2026-10-26
- The paper behind the EMNLP 2026 keynote on bigram 'simulated learners' — najoungkim · 2026-10-26
- Laying out DeepLIFT as a standalone computation graph: from sequence to attributions directly — jmschreiber91 · 2026-10-26
- Speculative Diffusion Paper Maps the Speed-Quality Pareto Frontier via Two Tradeoffs — UCLA · 2026-10-26