以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

DeepSeek V4 CSA

4 atoms · 跨 1 天 · 首见 2026-05-10 · 最近 2026-05-10

三色: 🟦 fact 0 · 🟥 take 4 · stance ▲0/▼0/◆2

来源: X 4

时态: aging:4

标签: 好观点:3 · 好思考:2

2026-05-10 [老化中] · 架构组成
My interpretation of DeepSeek V4 CSA: it’s mostly a composition of several strong recent ideas: • SWA + sink blocks (StreamingLLM) • But "sink" tokens are chosen via top-k selection • KV cache compression (many inference papers explored variants of this) • top-k key selection from query (reminds me of cross-attention from image gen world) • shared key-value attention across layers (shazeer goat) • sink bias is also included (gpt-OSS)
tags: 好观点·好思考 · → daily
2026-05-10 [老化中] · 架构简化概念
Whole architecture can be made simplier conceptually: 1. Compress KV cache 32x across all layers 2. Just use SWA+GA with sink blocks.
tags: 好思考 · → daily
2026-05-10 [老化中] · 与sink概念差异
imo it's very different from sink conceptually
tags: 好观点 · → daily
2026-05-10 [老化中] · 与sink概念差异观点同意
i agree
tags: 好观点 · → daily