HyperParallel: A Supernode-Affinity AI Framework

TL;DR

HyperParallel is a novel AI framework designed to efficiently utilize supernode architectures by integrating hardware-aware orchestration, hierarchical memory management, and declarative parallel strategies, significantly improving training and inference performance.

Contribution

It introduces a supernode-affinity AI framework with hardware-aware features and novel modules, addressing inefficiencies in existing frameworks for large-scale supernode architectures.

Findings

Enhanced training and inference efficiency

Reduced programming complexity and tuning overhead

Effective utilization of supernode architectures

Abstract

The emergence of large-scale, sparse, multimodal, and agentic AI models has coincided with a shift in hardware toward supernode architectures that integrate hundreds to thousands of accelerators with ultra-low-latency interconnects and unified memory pools. However, existing AI frameworks are not designed to exploit these architectures efficiently, leading to high programming complexity, load imbalance, and poor memory utilization. In this paper, we propose a supernode-affinity AI framework that treats the supernode as a single logical computer and embeds hardware-aware orchestration into the framework. Implemented in MindSpore, our HyperParallel architecture comprises HyperOffload for automated hierarchical memory management, HyperMPMD for fine-grained MPMD parallelism across heterogeneous workloads, and HyperShard for declarative parallel strategy specification. Together, these techniques significantly improve training and inference efficiency while reducing parallel programming and system tuning overhead, demonstrating the necessity of supernode affinity for next-generation AI frameworks.