KDA

AI overview

Headlines · 2

3d
RossiLH demos a lightweight KIMI K3 serving module
RossiLH presented a CPU-based interactive demonstration of parts of the KIMI K3 serving module across 14 tabs. It covers KDA speculative replay, replay-consistency checks, ReplaySSM ring buffering, unified paging for MLA and KDA storage, two-stage prefix-cache lookup, engine comparisons, and Fleet scheduling with consistent-hash routing and budget isolation. The author describes the components as reduced-scale versions or approximations and presents the project as part of his learning process.
掘金 · 人工智能 · 🔥 0
6d
How KDA Changes Caching for Kimi K3
A Juejin AI source-code analysis examines how Kimi K3 combines token-addressable MLA history with KDA convolution and recurrent state. It argues that prefix caching must restore a coherent set of MLA, convolution, and recurrent objects at the same boundary rather than merely reuse KV blocks. The article compares SGLang’s KDA-aware checkpointing and unified-memory approach with vLLM’s state blocks, metadata, fine-grained partial hits, and retention strategies.
掘金 · 人工智能 · 🔥 0

Experience and discussion from the community

Parallel Agent Teams: Small-Team Collaboration, Multi-Team Exploration

Optimizing a GPU kernel or completing a set of formal proofs often requires repeated trial and error: try a different implementation, run an evaluation, and keep iterating in whichever direction proves effective. When giving a model more test-time compute, the most direct approach is to let more agents try simultaneously; if the task only needs one best result, the value of the other attempts lies…

Agent PAT 测试时计算多智能体 Benchmark