Scalable Training of Mixture-of-Experts Models with Megatron Core

• URL: http://arxiv.org/abs/2603.07685v2
• Date: Tue, 10 Mar 2026 06:23:58 GMT
• Title: Scalable Training of Mixture-of-Experts Models with Megatron Core
• Abstract summary: Scaling Mixture-of-Experts (MoE) training introduces systems challenges absent in dense models.<n>Because each token activates only a subset of experts, this sparsity allows total parameters to grow much faster than per-token computation.<n>We address these challenges for MoE training through integrated optimizations spanning memory (fine-dimensional recomputation, offloading), communication, and computation.
• Score: 26.9162079065285
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Scaling Mixture-of-Experts (MoE) training introduces systems challenges absent in dense models. Because each token activates only a subset of experts, this sparsity allows total parameters to grow much faster than per-token computation, creating coupled constraints across memory, communication, and computation. Optimizing one dimension often shifts pressure to another, demanding co-design across the full system stack. We address these challenges for MoE training through integrated optimizations spanning memory (fine-grained recomputation, offloading, etc.), communication (optimized dispatchers, overlapping, etc.), and computation (Grouped GEMM, fusions, CUDA Graphs, etc.). The framework also provides Parallel Folding for flexible multi-dimensional parallelism, low-precision training support for FP8 and NVFP4, and efficient long-context training. On NVIDIA GB300 and GB200, it achieves 1,233/1,048 TFLOPS/GPU for DeepSeek-V3-685B and 974/919 TFLOPS/GPU for Qwen3-235B. As a performant, scalable, and production-ready open-source solution, it has been used across academia and industry for training MoE models ranging from billions to trillions of parameters on clusters scaling up to thousands of GPUs. This report explains how these techniques work, their trade-offs, and their interactions at the systems level, providing practical guidance for scaling MoE models with Megatron Core.

Related papers

• Mixture-of-Parallelisms: Towards Memory-Efficient Training Stack for Mixture-of-Experts Models [67.17268530365189]
  This paper showcases a memory-efficient training stack for Mixture-of-Experts (MoE) models.<n>It combines and specializes various existing and novel parallelism techniques at different layers and stages of the MoE model training pipeline.
  arXiv  Detail & Related papers  (2026-07-02T08:06:57Z)
• Characterizing the Efficiency of Distributed Training: A Power, Performance, and Thermal Perspective [6.51239603014107]
  Large Language Models (LLMs) have pushed training workloads beyond the limits of single-node analysis.<n>We present a comprehensive characterization of LLM training across diverse real-world workloads and hardware platforms.
  arXiv  Detail & Related papers  (2025-09-12T16:05:07Z)
• Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs [111.69640966866059]
  Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models.<n>In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs.<n>The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware.
  arXiv  Detail & Related papers  (2025-05-07T15:46:36Z)
• MoE Parallel Folding: Heterogeneous Parallelism Mappings for Efficient Large-Scale MoE Model Training with Megatron Core [11.40633051522406]
  We introduce an end-to-end training framework for large-scale MoE models.<n>MoE Parallel Folding is a novel strategy that decouples the parallelization of attention and MoE in Transformer models.<n>A flexible token-level dispatcher supports both token-dropping and token-dropless MoE training.
  arXiv  Detail & Related papers  (2025-04-21T08:39:47Z)
• AutoHete: An Automatic and Efficient Heterogeneous Training System for LLMs [68.99086112477565]
  Transformer-based large language models (LLMs) have demonstrated exceptional capabilities in sequence modeling and text generation.<n>Existing heterogeneous training methods significantly expand the scale of trainable models but introduce substantial communication overheads and CPU workloads.<n>We propose AutoHete, an automatic and efficient heterogeneous training system compatible with both single- GPU and multi- GPU environments.
  arXiv  Detail & Related papers  (2025-02-27T14:46:22Z)
• Democratizing AI: Open-source Scalable LLM Training on GPU-based Supercomputers [65.35142508909892]
  We present a novel four-dimensional hybrid parallel algorithm implemented in a highly scalable, portable, open-source framework called AxoNN.<n>We demonstrate fine-tuning of a 405-billion parameter LLM using AxoNN on Frontier.
  arXiv  Detail & Related papers  (2025-02-12T06:05:52Z)
• EPS-MoE: Expert Pipeline Scheduler for Cost-Efficient MoE Inference [49.94169109038806]
  This paper introduces EPS-MoE, a novel expert pipeline scheduler for MoE that surpasses the existing parallelism schemes.<n>Our results demonstrate at most 52.4% improvement in prefill throughput compared to existing parallel inference methods.
  arXiv  Detail & Related papers  (2024-10-16T05:17:49Z)
• TorchTitan: One-stop PyTorch native solution for production ready LLM pre-training [17.157552816494427]
  This paper introduces TorchTitan, an open-source, PyTorch-native distributed training system.<n>It unifies state-of-the-art techniques, streamlining integration and reducing overhead.<n>We evaluate TorchTitan on the Llama 3.1 family of large language models (LLMs)
  arXiv  Detail & Related papers  (2024-10-09T03:26:11Z)
• OmniBal: Towards Fast Instruction-Tuning for Vision-Language Models via Omniverse Computation Balance [67.37017498784748]
  Large-scale 3D parallel training on vision-language instruction-tuning models leads to an imbalanced computation load across different devices.<n>We rebalance the computational load from data, model, and memory perspectives, achieving more balanced computation across devices.<n>Our method's efficacy and generalizability are further validated across various models and datasets.
  arXiv  Detail & Related papers  (2024-07-30T12:02:58Z)
• MoESys: A Distributed and Efficient Mixture-of-Experts Training and Inference System for Internet Services [32.278096820269816]
  We present a novel MoESys that boosts efficiency in both large-scale training and inference. Specifically, in the training procedure, the proposed MoESys adopts an Elastic MoE training strategy with 2D prefetch and Fusion communication over Hierarchical storage. For scalable inference in a single node, MoESys builds the CPU-GPU memory jointly into a ring of sections to load the model, and executes the computation tasks across the memory sections in a round-robin manner for efficient inference.
  arXiv  Detail & Related papers  (2022-05-20T09:09:27Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.