Inside KV-Cache Sharing: How DeepSeek CED and GLM 5.2 Differ
Developers break down KV-cache sharing designs: YOCO lets all layers in the model's second half share one KV cache; DeepSeek's CED applies a similar YOCO-style global branch per CSA2 layer, caching only encoder KV; GLM 5.2 scores per token while sharing block indices for flexibility and performance.
2026-09-11 ~ 2026-09-11 · 4 related posts
- Episode 1: Inside DeepSeek V4.1 Flash: YOCO at its core and KV cache reuse(2026-09-10, 7 posts)
- Episode 2: DeepSeek V4.1 Flash Architecture Breaks Down Asymmetric Design(2026-09-11, 2 posts)
- Episode 3: Inside KV-Cache Sharing: How DeepSeek CED and GLM 5.2 Differ(2026-09-11, 4 posts)
- YOCO explained: one shared KV cache reused across the model's second half — stochasticchasm · 2026-09-11
- DeepSeek's CED applies YOCO-style shared KV caching only to global CSA2 branches — stochasticchasm · 2026-09-11
- DeepSeek's CED vs GLM's KV reuse: a developer unpacks how the cache-sharing modes actually differ — stochasticchasm · 2026-09-11
- GLM 5.2's KV cache sharing modes explained: blockwise to tokenwise scoring — stochasticchasm · 2026-09-11