fabric-lib: RDMA Point-to-Point Communication for LLM Systems

TL;DR

fabric-lib provides a hardware-agnostic, high-throughput point-to-point communication library for large language model systems, enabling flexible, portable, and efficient data transfer across diverse NICs.

Contribution

It introduces fabric-lib, a novel abstraction layer that unifies NIC functionality for LLM systems, improving portability and performance.

Findings

Achieves 400 Gbps peak throughput on NVIDIA ConnectX-7 and AWS EFA.

Enables disaggregated inference with dynamic scaling in production systems.

Reduces latency for trillion-parameter RL weight updates to 1.3 seconds.

Abstract

Emerging Large Language Model (LLM) system patterns, such as disaggregated inference, Mixture-of-Experts (MoE) routing, and asynchronous reinforcement fine-tuning, require flexible point-to-point communication beyond simple collectives. Existing implementations are locked to specific Network Interface Controllers (NICs), hindering integration into inference engines and portability across hardware providers. We present fabric-lib, which bridges the functionality of common NICs to expose a uniform interface. fabric-lib exposes one-sided WriteImm operations with a ImmCounter primitive for completion notification, without ordering assumptions of network transport, transparently managing multiple NICs per GPU. We demonstrate peak throughput of 400 Gbps on both NVIDIA ConnectX-7 and AWS Elastic Fabric Adapter (EFA). We showcase fabric-lib through three production systems: (1) KvCache transfer for disaggregated inference with dynamic scaling, (2) RL weight updates achieving 1.3 seconds for trillion-parameter models, and (3) MoE dispatch/combine implementation exceeding DeepEP decode latency on ConnectX-7, with the first viable latencies on EFA. We demonstrate that our portable point-to-point communication complements collectives while avoiding lock-in. fabric-lib is open-sourced at https://github.com/perplexityai/pplx-garden/