Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation
This work addresses the challenge of efficient and accurate low-precision training for large-scale AI models, offering a significant improvement over existing methods but is incremental as it builds on prior quantized training techniques.
The paper tackles the problem of accuracy loss in fully-quantized pre-training of large language models using the NVFP4 format by introducing MS-EDEN, a novel unbiased quantization routine with over 2x lower quantization error than stochastic rounding, and Quartet II, a fully-NVFP4 scheme for linear layers, achieving up to 4.2x speedup over BF16 on NVIDIA Blackwell GPUs with validation on models up to 1.9B parameters.
The NVFP4 lower-precision format, supported in hardware by NVIDIA Blackwell GPUs, promises to allow, for the first time, end-to-end fully-quantized pre-training of massive models such as LLMs. Yet, existing quantized training methods still sacrifice some of the representation capacity of this format in favor of more accurate unbiased quantized gradient estimation by stochastic rounding (SR), losing noticeable accuracy relative to standard FP16 and FP8 training. In this paper, improve the state of the art for quantized training in NVFP4 via a novel unbiased quantization routine for micro-scaled formats, called MS-EDEN, that has more than 2x lower quantization error than SR. We integrate it into a novel fully-NVFP4 quantization scheme for linear layers, called Quartet II. We show analytically that Quartet II achieves consistently better gradient estimation across all major matrix multiplications, both on the forward and on the backward passes. In addition, our proposal synergizes well with recent training improvements aimed specifically at NVFP4. We further validate Quartet II on end-to-end LLM training with up to 1.9B parameters on 38B tokens. We provide kernels for execution on NVIDIA Blackwell GPUs with up to 4.2x speedup over BF16. Our code is available at https://github.com/IST-DASLab/Quartet-II .