MegaScale-Omni: A Hyper-Scale, Workload-Resilient System for MultiModal LLM Training in Production

TL;DR

MegaScale-Omni is a scalable, resilient system for training multimodal large language models that adapts dynamically to workloads, achieving significant throughput improvements over existing systems.

Contribution

It introduces a novel parallelism scheme, unified representations, and workload balancing techniques for efficient, flexible MLLM training at hyper-scale.

Findings

Achieved up to 7.57x throughput improvement under dynamic workloads.

Implemented long-short sequence parallelism for variable-length samples.

Deployed on thousands of GPUs for large-scale MLLM training.

Abstract

As the foundational component of versatile AI applications, training an multimodal large language model (MLLM) relies on multimodal datasets with dynamic modality mixture proportions and sample length distributions. However, existing MLLM systems remain inefficient under dynamic workloads, due to statically coupled decisions of resource allocation and model parallelization between encoders and the LLM backbone. This paper presents MegaScale-Omni, an industrial-grade MLLM training system tailored for dynamic workload adaption and hyper-scale deployment. MegaScale-Omni is built upon the training scheme of encoder-LLM multiplexing with three key innovations: (1) Decoupled parallelism strategies with long-short sequence parallelism for encoders to process variable-length samples, and full-fledged 5D parallelism for the LLM backbone, both organized under a communication-efficient parallelization layout. (2) Unified encoder-LLM representations for flexible, extensible colocation, and a new paradigm of encoder-LLM joint pipeline with workload resilience. (3) Workload balancing techniques via decentralized grouped reordering in data loaders and adaptive resharding from encoder to LLM ranks. MegaScale-Omni is deployed as the foundation of our in-house large-scale MLLM training tasks with thousands of GPUs. Our experimental results demonstrate $1.27\times$-$7.57\times$ throughput improvement under production-grade dynamic workloads, as compared to four state-of-the-art systems.