Article

Don't Let a Few Network Failures Slow the Entire AllReduce

OptCC gives an information-theoretic lower bound and a four-stage pipelined AllReduce algorithm for bandwidth-asymmetric failures, staying within 2%–6% of NCCL's fault-free ring …

peiqing-chen

• Jun 1, 2026 • 1 min read

Systems for ML

Enabling Performant and Flexible Model-Internal Observability for LLM Inference

A performant, flexible deep model inspector that turns internal LLM observability into a first-class systems primitive — only 0.4%–6.8% offline / ~6% online overhead, 2×–15× lower …

Nengneng Yu

• May 11, 2026 • 1 min read

ML Algorithms

TabSyM: A Generative Pipeline for Small Multi-Cohort Omics Tabular Data

Diffusion-based generative pipeline for small, high-dimensional, cross-cohort omics tabular data.

Nengneng Yu

• Jul 14, 2025 • 1 min read