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

2026-05-09

AI Summary

Existing multimodal large language model (MLLM) training systems suffer from inefficiency under dynamic workloads due to the static coupling between vision encoders and LLM backbones. This work proposes an efficient training system that decouples short- and long-sequence parallelism, integrates a unified encoder-LLM representation, and introduces a novel joint pipelining paradigm within a 5D model parallelism layout. To further enhance scalability, the system incorporates decentralized grouped reordering and adaptive resharding mechanisms, enabling dynamic load balancing and communication optimization. Experiments on a thousand-GPU cluster demonstrate that, under realistic dynamic workloads, the proposed system achieves 1.27× to 7.57× higher throughput compared to four state-of-the-art baselines.

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.

Problem

• multimodal large language model
• dynamic workload
• model parallelism
• resource allocation
• training efficiency

Innovation

• Decoupled Parallelism
• Encoder-LLM Joint Pipeline
• Workload Resilience
• Adaptive Resharding
• Multimodal LLM Training