Defeating the Training-Inference Mismatch via FP16

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Qi, Penghui, Liu, Zichen, Zhou, Xiangxin, Pang, Tianyu, Du, Chao, Lee, Wee Sun, Lin, Min
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914124592578560
author Qi, Penghui
Liu, Zichen
Zhou, Xiangxin
Pang, Tianyu
Du, Chao
Lee, Wee Sun
Lin, Min
author_facet Qi, Penghui
Liu, Zichen
Zhou, Xiangxin
Pang, Tianyu
Du, Chao
Lee, Wee Sun
Lin, Min
contents Reinforcement learning (RL) fine-tuning of large language models (LLMs) often suffers from instability due to the numerical mismatch between the training and inference policies. While prior work has attempted to mitigate this issue through algorithmic corrections or engineering alignments, we show that its root cause lies in the floating point precision itself. The widely adopted BF16, despite its large dynamic range, introduces large rounding errors that breaks the consistency between training and inference. In this work, we demonstrate that simply reverting to \textbf{FP16} effectively eliminates this mismatch. The change is simple, fully supported by modern frameworks with only a few lines of code change, and requires no modification to the model architecture or learning algorithm. Our results suggest that using FP16 uniformly yields more stable optimization, faster convergence, and stronger performance across diverse tasks, algorithms and frameworks. We hope these findings motivate a broader reconsideration of precision trade-offs in RL fine-tuning.
format Preprint
id arxiv_https___arxiv_org_abs_2510_26788
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Defeating the Training-Inference Mismatch via FP16
Qi, Penghui
Liu, Zichen
Zhou, Xiangxin
Pang, Tianyu
Du, Chao
Lee, Wee Sun
Lin, Min
Machine Learning
Artificial Intelligence
Computation and Language
Reinforcement learning (RL) fine-tuning of large language models (LLMs) often suffers from instability due to the numerical mismatch between the training and inference policies. While prior work has attempted to mitigate this issue through algorithmic corrections or engineering alignments, we show that its root cause lies in the floating point precision itself. The widely adopted BF16, despite its large dynamic range, introduces large rounding errors that breaks the consistency between training and inference. In this work, we demonstrate that simply reverting to \textbf{FP16} effectively eliminates this mismatch. The change is simple, fully supported by modern frameworks with only a few lines of code change, and requires no modification to the model architecture or learning algorithm. Our results suggest that using FP16 uniformly yields more stable optimization, faster convergence, and stronger performance across diverse tasks, algorithms and frameworks. We hope these findings motivate a broader reconsideration of precision trade-offs in RL fine-tuning.
title Defeating the Training-Inference Mismatch via FP16
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2510.26788