UCCL-EP: Portable Expert-Parallel Communication — Full Results

Full evaluation of UCCL-EP across NVIDIA and AMD GPUs, AWS EFA, InfiniBand, and Broadcom NICs — with application-level results on SGLang, vLLM inference and Megatron-LM training.

Previewing UCCL-EP: Flexible and Efficient Expert Parallelism for Cloud and Beyond

GPU-driven communication (e.g., DeepEP) is the key to efficient and large-scale EP, but it cannot run on heterogeneous platforms in the public cloud due to tight coupling between GPU and NIC.

Everything You Want to Know about KV Cache Transfer Engine

There have been many KV cache transfer engines for PD disaggregation, but nearly no benchmarks on their performance. This blog serves for this purpose---benchmarking and analyzing the performance of v...

How to Debug NCCL Performance Issues for ML Workloads?

NCCL is notoriously hard to debug. In this post, we will go through our journey of debugging NCCL performance issues and how UCCL can help this process.

UCCL-Tran: An Extensible Software Transport Layer for GPU Networking

UCCL-Tran is designed to be fast and extensible to meet the challenging requirements of modern ML/LLM workloads

About Us

About UCCL team