FP8-Flow-MoE: A Casting-Free FP8 Recipe without Double Quantization Error
This work addresses the high compute and memory demands for researchers and practitioners training large MoE models, offering a plug-and-play solution that is incremental in optimizing low-precision training.
The paper tackles the computational inefficiency of training large Mixture-of-Experts (MoE) models by proposing FP8-Flow-MoE, an FP8 training recipe that eliminates redundant casts to reduce double quantization error, achieving up to 21% higher throughput and 16.5 GB lower memory usage per GPU while maintaining stable convergence.
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 naïve 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.