MESI-KV — Cross-Layer Directory Coherence & Attention-Taint-Bounded KV Invalidation

Zero-Training Plug-and-Play Architecture

When external files, configs, or database rows mutate mid-trajectory, unmodified CausalLM caches fail silently—either keeping stale memoized Value (V) vectors (Leyline / PatchKV) or flushing 100% of downstream KV blocks (Concord). MESI-KV tracks Python AST symbol, config key, and SQLite row-version granules in a 4-state (S, M, T, I) Lease Directory, accumulates Order-Statistic Calibrated Attention Taint in FP32 pre-scaled chunks, and surgically recomputes only mutated + tainted downstream blocks while rotating clean suffix blocks via closed-form $SO(2)$ relative RoPE shifts $R_\Delta$.

Accuracy and KV Savings

Quick Install

pip install git+https://github.com/batistado/mesikv.git

Citation

@article{syed2026mesikv,
  title   = {{MESI-KV}: Cross-Layer Directory Coherence and Attention-Taint-Bounded Invalidation Across Mutable Workspaces, Semantic Tool Caches, and Disaggregated {KV} Pools},
  author  = {Syed, Mohammed Kamran and Rentala, Anudeep and Rahul, Somesh and Pai, Raghuram Malpe and Dara, Krishna Suman},
  journal = {IEEE Transactions on Parallel and Distributed Systems (TPDS)},
  year    = {2026},
  url     = {https://github.com/batistado/mesikv}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for batistado/MESI-KV

Base model

Qwen/Qwen2.5-3B
Finetuned
(1,680)
this model

Dataset used to train batistado/MESI-KV