Skip to content
View original post on X: vLLMOfficial
Published
40/100

vLLM explains K3.1 attention and serving architecture

AIvLLM says K3.1 combines DSA sparse attention with recurrent Kimi Delta Attention, storing MLA latents, indexer keys and KDA state in one packed cache. The serving work adds token-sharded prefill, exact top-k merging for decode context parallelism, and a fused low-latency decode path.
vLLMVerified on XOfficial
@vllm_project
View on X

Part of a thread · earlier post

K3.1 combines DSA sparse attention with recurrent Kimi Delta Attention.

One packed cache holds MLA latents, indexer keys and KDA state. The serving work adds token-sharded prefill, exact top-k merging for decode context parallelism, and a fused low-latency decode path.

Source details

vLLM · x.comPublished · added here

Original title: K3.1 combines DSA sparse attention with recurrent Kimi Delta Attention.

Topics

Deployment & engineeringChina AITutorials & how-to