Reproduce “Beyond Compression Ratio: High-Throughput KV Compression for Disaggregated LLM Serving” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junhong Wu, Zedong Liu, Xinyang Ma, Yuhan Chen, Guangming Tan, Dingwen Tao

FastKVComp accelerates KV-cache compression in disaggregated LLM serving. It preserves KVServe’s policy, fusing distribution transformation with low-bit quantization to reduce intermediate GPU-memory traffic and using offline-selected lossless coding. On the host-staged NCCL shared-memory path, Qwen2.5-7B-Instruct on LongBench’s 2WikiMQA achieves 2.04× Raw Transfer output throughput. This finding is specific to that workload and path.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.