Contribute to NIXL’s EFA/Libfabric backend — one-sided GPU-to-GPU and GPU-to-storage transfers for disaggregated inference, used to move KV-cache and model state between prefill and decode workers at scale.
Back to projects
Jul 31, 2026
1 min read
NIXL
An open-source, high-performance point-to-point data transfer library for GPU inference clusters, part of the AI-Dynamo project.
