FP8-Flow-MoE: A Casting-Free FP8 Recipe without Double Quantization Error

TL;DR

The paper introduces FP8-Flow-MoE, a novel FP8 training method for large MoE models that reduces memory and increases throughput by eliminating redundant casts and ensuring quantization consistency, while maintaining stability.

Contribution

It presents a new FP8-centric dataflow with scaling-aware operations that avoids double quantization errors and improves efficiency in large-scale MoE training.

Findings

Up to 21% higher throughput compared to BF16

16.5 GB lower memory usage per GPU

Stable convergence maintained in large models

Abstract

Training large Mixture-of-Experts (MoE) models remains computationally prohibitive due to their extreme compute and memory demands. Although low-precision training promises to accelerate computation and reduce memory footprint, existing implementations still rely on BF16-dominated dataflows with frequent quantize-dequantize (Q/DQ) conversions. These redundant casts erode much of FP8's theoretical efficiency. However, naively removing these casts by keeping dataflows entirely in FP8 introduces double quantization error: tensors quantized along different dimensions accumulate inconsistent scaling factors, degrading numerical stability. We propose FP8-Flow-MoE, an FP8 training recipe featuring a quantization-consistent FP8-centric dataflow with a scaling-aware transpose and fused FP8 operators that streamline computation and eliminate explicit cast operations from 12 to 2. Evaluations on a 671B-parameter MoE model demonstrate up to 21\% higher throughput and 16.5 GB lower memory usage per GPU compared to BF16 and na\"ive FP8 baselines, while maintaining stable convergence. We provide a plug-and-play FP8 recipe compatible with TransformerEngine and Megatron-LM, which will be open-sourced soon.