R23: MPCache: MPC-Friendly KV Cache Eviction for Efficient Private Large Language Model Inference

Status: PLLM reimplementation
Year: 2025
Authors: Wenxuan Zeng, Ye Dong, Jinjin Zhou, Junming Ma, Jin Tan, Runsheng Wang, Meng Li
External source: MPCache: MPC-Friendly KV Cache Eviction for Efficient Private Large Language Model Inference

Source contribution

MPCache evaluates MPC-friendly static and dynamic KV-cache eviction in a three-party SecretFlow/SPU setting; its artifact focuses state selection and operator costs.

PLLM perspective

Reproduce cached generation and use its explicit KV state as a comparison oracle for PLLM decoder state without inferring complete protected boundary coverage.

No MPCache reproduction has been executed. Current PLLM work implements a deterministic structural DecoderPlan adaptation for dense Qwen2/Qwen3 fixed-shape caches; protected execution, fidelity, assurance, and matched benchmarks remain unchecked.

PLLM implements only a structural DecoderPlan adaptation. It does not reproduce MPCache's three-party protected runtime. Fidelity, protected execution, assurance, and matched benchmark status remain unchecked.

The source's boundary remains binding: Three online semi-honest parties and incomplete protected-boundary evidence are incompatible with the one-evaluator complete-model target.

PLLM modules and documentation

Target module identifiers: protocol.mpcache, state.secret_shared_kv, gate.mpcache_attention. These identifiers describe scoped work unless this page explicitly links an implementation source. Dependencies: None.

• Research backlog
• Reimplementation backlog
• Clean-room workflow
• PLLM method status
• Decoder plan usage
• MPCache structural adaptation source

Backlog

This source is assigned to reimplementation priority 5, Protected KV-cache policy. Its source-registry order is 23. Follow current gates and unfinished work in the research backlog.