UCCL-EP: An expert parallel communications kernel without owning the NIC

▤ Summary

This article explains how UCCL-EP adapts DeepEP-style expert parallel communication to work across different GPU and NIC combinations. It replaces GPU-owned NIC control with a proxy-thread design that moves commands through pinned host memory while keeping payloads in GPU memory. The approach preserves the key send, signal, and quiet semantics needed by expert parallel kernels without requiring NVIDIA-specific hardware. The article also notes that UCCL-EP already supports EFA and generic RDMA verbs backends and runs on AMD GPUs through ROCm. The main takeaway is that this design broadens high-performance EP communication to heterogeneous infrastructure.

▦ Linked Companies

Doubleword

$1M to $5M