As transformer models continue to grow, developers need practical ways to improve post-training and fine-tuning throughput without sacrificing model quality or adding fragile precision-conversion workflows. In this session, we will show how to fine-tune models with NVIDIA NVFP4, a 4-bit low-precision training format, using the PyTorch-native NVIDIA NeMo Megatron Bridge stack.

We will explain how NVFP4 reduces memory usage and increases throughput compared with BF16, while preserving downstream accuracy when applied with the right mixed-precision recipe. Attendees will learn how Megatron Bridge enables developers to switch between BF16, FP8, MXFP8, and NVFP4 through configuration changes rather than rewriting model or optimizer code while keep the PyTorch functionalities intact. The session will cover practical considerations for stable NVFP4 fine-tuning, including calibration, hierarchical scaling, and selectively retaining BF16 precision in sensitive transformer layers.