RFC R-State — Resumable State Transfer Format for Heterogeneous LLM Pipeline Architectures
Abstract
R-State is a proposed serialization format for transferring the internal state of large language models between stages of a heterogeneous inference pipeline without detokenization and re-tokenization. It is designed to reduce latency, preserve intermediate activations and KV-cache, and enable more modular multi-model workflows through a structured packet containing metadata and tensor payloads
Fields of Study & Subjects
Machine Learning and Algorithms
Natural Language Processing Techniques
Machine Learning and Data Classification
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 1 May 2026
- Language
- EN