Found 1 document(s)
Page 1 of 1
-
Distributed LLM serving relies on a global index to locate reusable key-value cache (KVCache) blocks through prefix matching. Losing the index after a crash or restart forces affected requests to recompute existing KVCac...