MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641125334 A) filed by Vellore Institute Of Technology on October 21, 2026, for System And Method For Epoch-Fenced, Address-Preserving Recovery Of A Language Model Key-Value Cache Held In Compute Express Link Attached Memory.
Inventors include Jagadeesan S; and Kishore E.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: The present invention discloses a system and method for recovering a large language model inference session without copying its key-value (KV) cache. The cache is held in an address range (116) of a memory pool (114) attached through a Compute Express Link fabric to several computenodes (106,108, 110).A fabricmanager (112) issues acomputenodeonly arevocable execution grant over that range, bound to an ownership epoch. Each decoding step writes and verifies the KV entries and then atomically publishes a recovery marker of KV version, token index, block version and decoder-state version. On failure, the fabric manager revokes the grant, increments the epoch and grants the same address range to a replacement node, which resumes from the last valid marker. Requests carrying a superseded epoch are rejected. Hot-set-first activation, predictive migration and failure-domain-aware node selection shorten recovery.
Disclaimer: Curated by HT Syndication.
