Why it matters

Prefix sharing shapes multi-tenant efficiency. Understanding shapes serving.

Advertisement

The architecture

Detect common prefix.

Cache KV for prefix.

Requests point to shared.

Shared prefix cachingRequestswith common prefixRadix triedetect sharedShare KVone copyRadixAttention (SGLang) generalizes; vLLM prefix caching is similar
Shared prefix.
Advertisement

How it works end to end

Radix trie for prefix detection.

Copy-on-write when diverging.

Enables multi-turn optimization.