CMPS: Constrained Mixed Precision Search

ICLR 2026 Conference Submission9349 Authors

Keywords: Neural Networks, Quantization, Mixed Precision, Post Training, Transformers, LLMs, Large Language Models, Hardware, GPU, Data Formats, Bit Allocation, FP32, MXFP, BFP, MXINT, GPU, TPU, AI Accelerators

Abstract: The increasing complexity of deep neural networks (DNNs) requires effective model compression to reduce their computational and memory footprints for deployment on resource-constrained hardware. Mixed-precision search is a prominent bit allocation method based on neural architecture search (NAS) that has been shown to significantly reduce the DNN footprint while preserving the accuracy of the model by allocating bits to each layers based on their quantization sensitivity. However, mixed-precision search is often defined as a dual optimization problem handled with a single heuristic objective function, which does not provide strong guarantees of the resulting compression rate. We propose a post-training reformulation of mixed precision search as an explicit constrained optimization problem, solved using interior-point methods within a framework based on NAS. Our method requires minimal calibration data, as few as 128 samples, in a post-training setting. We corroborate this approach with experiments that span multiple transformer architectures with up to 4 billion parameters, using the MXFP family of data formats. We show that this constrained formulation provides users with higher resolution over compression rates, and we show that explicitly satisfying hardware budgets while optimizing for accuracy can outperform uniform allocation methods, improving performance by up to several standard deviations over the uniform baselines.

Primary Area: infrastructure, software libraries, hardware, systems, etc.

Submission Number: 9349