LiveLoRA: Exact KV Cache Migration Across LoRA Updates for LLM Agents

Updating an LLM adapter should not require rebuilding the inference state of every active session from its entire interaction history. Because affected key-value (KV) cache entries were computed under the previous model version, reusing them mixes versions, whereas restoring coherence through full-history re-prefill incurs work that grows with context length and fleet size. We introduce LiveLora, a model-runtime co-design that makes final-block K/V migration exact for a specific LoRA adapter family under explicit model and arithmetic assumptions. LiveLora freezes the lower decoder blocks and fixes the down-projection matrices of the final-block K/V adapters, leaving each token's base projections and low-rank coordinates invariant across adapter versions. To publish an update, the runtime reconstructs final-block K/V tensors for the new version, recomputes only the latest token through the final block, and commits the adapter, cache, and next-token logits as one coherent state. On Qwen2.5-3B-Instruct, the within-run median paired speedup over full-history re-prefill averages $68.98\times$ across five runs for 32 active sessions with 8K tokens each. Across all seven evaluated settings, migrated final-block K/V tensors match independently captured references with an observed maximum absolute error of zero; we also quantify the resulting storage and cached-decoding overheads. LiveLora turns online adapter publication into a version-coherent model-runtime state transition, allowing active sessions to adopt new adapter versions without replaying their interaction histories.