Saved in:
Bibliographic Details
Main Authors: Zhang, Wuyue, Huang, Chongdong, You, Chunbo, Gu, Cheng, Wang, Fengjuan, Sun, Mou
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.02731
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912939419631616
author Zhang, Wuyue
Huang, Chongdong
You, Chunbo
Gu, Cheng
Wang, Fengjuan
Sun, Mou
author_facet Zhang, Wuyue
Huang, Chongdong
You, Chunbo
Gu, Cheng
Wang, Fengjuan
Sun, Mou
contents Training large-scale Mixture-of-Experts (MoE) models is bottlenecked by activation memory and expert-parallel communication, yet FP4 training remains impractical on Hopper-class GPUs without native MXFP4 or NVFP4 support. In this work, we present a training recipe that enables MXFP4 efficiency for MoE models on Hopper architectures without native 4-bit computation support. A central challenge is to integrate FP4 into an existing BF16/FP8 hybrid training pipeline without incurring costly precision round-trips (e.g., FP4 $\leftrightarrow$ BF16 $\leftrightarrow$ FP8). We address this challenge by introducing direct FP8-to-FP4 quantization and de-quantization, together with scaling-aware FP4 row-wise to column-wise conversion, enabling FP4 activations and expert-parallel communication with minimal overhead. Core MoE computations are executed in FP8, while activations and expert-parallel communication are compressed using MXFP4, achieving substantial memory and bandwidth savings without degrading convergence. At the 671B parameter scale, our method achieves end-to-end training performance comparable to strong FP8 baselines, while reducing peak activation memory by 14.8\% (11.8 GB) and improving training throughput by 12.5\%, from 1157 to 1302 tokens per GPU per second. These results show that FP4 efficiency can be practically realized for large-scale MoE training through careful software-hardware co-design, even without native FP4 Tensor Core support.
format Preprint
id arxiv_https___arxiv_org_abs_2603_02731
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Practical FP4 Training for Large-Scale MoE Models on Hopper GPUs
Zhang, Wuyue
Huang, Chongdong
You, Chunbo
Gu, Cheng
Wang, Fengjuan
Sun, Mou
Machine Learning
Artificial Intelligence
Training large-scale Mixture-of-Experts (MoE) models is bottlenecked by activation memory and expert-parallel communication, yet FP4 training remains impractical on Hopper-class GPUs without native MXFP4 or NVFP4 support. In this work, we present a training recipe that enables MXFP4 efficiency for MoE models on Hopper architectures without native 4-bit computation support. A central challenge is to integrate FP4 into an existing BF16/FP8 hybrid training pipeline without incurring costly precision round-trips (e.g., FP4 $\leftrightarrow$ BF16 $\leftrightarrow$ FP8). We address this challenge by introducing direct FP8-to-FP4 quantization and de-quantization, together with scaling-aware FP4 row-wise to column-wise conversion, enabling FP4 activations and expert-parallel communication with minimal overhead. Core MoE computations are executed in FP8, while activations and expert-parallel communication are compressed using MXFP4, achieving substantial memory and bandwidth savings without degrading convergence. At the 671B parameter scale, our method achieves end-to-end training performance comparable to strong FP8 baselines, while reducing peak activation memory by 14.8\% (11.8 GB) and improving training throughput by 12.5\%, from 1157 to 1302 tokens per GPU per second. These results show that FP4 efficiency can be practically realized for large-scale MoE training through careful software-hardware co-design, even without native FP4 Tensor Core support.
title Practical FP4 Training for Large-Scale MoE Models on Hopper GPUs
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2603.02731