Reproduce “Scaling LLM Inference via Parallelism: Throughput-Latency Trade-off” 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
Javad Mirzaei, Jeebak Mitra
The authors compare tensor- and pipeline-parallel LLM serving configurations across workloads, TTFT targets, and context lengths. They evaluate mean time to first token and inter-token latency, then measure goodput as throughput multiplied by the fraction of requests meeting a TTFT threshold. Across four configurations on eight GPUs, preferred setups shift from PP-heavy to TP-heavy as targets relax, within tested workloads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.