IQFormerLite: a hardware-efficient framework for real-time automatic modulation recognition on edge NPUs

Purpose This paper aims to address the deployment gap between recent automatic modulation recognition (AMR) models and real-time edge neural processing unit (NPU) inference, where parameter redundancy, hardware-incompatible operators and costly fixed frequency-domain preprocessing limit practical use. Design/methodology/approach IQFormerLite is designed for Rockchip RK3588 NPU deployment. It replaces serialization-heavy and attention-based components with a parallel-friendly, convolution-centric backbone. A large-kernel global context block captures long-range dependencies, while a learnable Kolmogorov–Arnold network–based filterbank performs adaptive spectral extraction from raw I/Q sequences. Experiments use RadioML2016.10A and RadioML2016.10B with rockchip neural network (RKNN)-based on-device profiling. Findings IQFormerLite remains competitive with IQFormer while reducing parameters by about 63.8%, from approximately 0.35 M to 0.13 M. Under INT8 deployment on RK3588, it achieves a throughput of 6387.84 samples/s and a latency of 0.157 ms, corresponding to a 14.83 × speedup over the baseline, while maintaining 63.11% accuracy. Research limitations/implications The evaluation is limited to RadioML2016.10A, RadioML2016.10B and the RK3588 platform. Further validation on additional datasets, over-the-air signals and channel-mismatch conditions would strengthen generalizability. Practical implications The framework supports real-time AMR deployment on edge NPUs by reducing model size, improving compiler compatibility and avoiding fixed frequency-domain preprocessing. Originality/value This paper provides a hardware-aligned AMR framework that bridges algorithmic design and practical NPU deployment through LKGC and LKF modules.