Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs

TL;DR

This paper presents a comprehensive approach to efficiently train a 718-billion-parameter sparse language model on Ascend NPUs, optimizing hardware utilization and demonstrating scalable training of large MoE models.

Contribution

The paper introduces Pangu Ultra MoE, a large-scale sparse language model trained efficiently on Ascend NPUs, with optimized system strategies and simulation-based hyperparameter selection.

Findings

Achieved 30.0% MFU during training on 6K NPUs.

Optimized Expert Parallelism reduces communication overhead.

Demonstrated scalable training of a 718-billion-parameter MoE model.

Abstract

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. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. 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. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.