Joint Training on AMD and NVIDIA GPUs

TL;DR

This paper introduces a novel heterogeneous training method for AMD and NVIDIA GPUs, enabling efficient cross-vendor communication and achieving near-native throughput for large language model training.

Contribution

It proposes a Device-Direct Communication approach with CPU-offloading P2P, significantly improving cross-vendor GPU data transfer performance.

Findings

Achieves up to 98% of NVIDIA homogeneous system throughput

Maintains training stability and correctness

Demonstrates effectiveness on LLaMA-8B and Qwen2-7B models

Abstract

As large language models continue to scale, training demands on compute and system capacity grow rapidly, making single-vendor homogeneous clusters insufficient. This paper presents a technical solution for heterogeneous mixed training in AMD-NVIDIA environments. We first adopt a compatibility-oriented approach based on CPU-Forwarding Communication, with differentiated communication back-end selection across parallel groups and multi-NIC parallel data transfer. To achieve higher performance, we further propose another Device-Direct Communication approach, integrating a CPU-offloading P2P mechanism to enable direct cross-vendor GPU data transfer without host-memory staging. Experiments on LLaMA-8B and Qwen2-7B demonstrate that the proposed Device-Direct Communication approach achieves up to 98% of the throughput of an NVIDIA homogeneous system, while preserving training stability and correctness.