Global numerical modeling is advancing into the era of kilometer-scale, non-hydrostatic simulations with integrated AI capabilities, while heterogeneous computing emerges as a pivotal trend in high-performance computing (HPC). As a strong candidate for next-generation global kilometer-scale general circulation models (GCMs), the A-grid dynamical core based on the Low Mach number Approximate Riemann Solver (LMARS) must address key challenges: ensuring monotonicity while preserving sharp gradients, mitigating CFL constraints caused by vertically propagating sound waves, and integrating AI-driven computational power. This work presents LMARSpy, a GPU-optimized non-hydrostatic dynamical core with a sharpness-preserving monotonicity limiter and a conservative vertical implicit solver. Designed for GPU efficiency, LMARSpy leverages a Python-based high-performance computing platform to ensure robust compatibility across heterogeneous computing environments. Benchmark tests validate the model‘s innovations: the monotonicity limiter effectively suppresses non-physical oscillations while maintaining high-order accuracy, incurring only a 10.4% increase in GPU computational cost; the vertical implicit solver alleviates CFL limitations, achieving at least an order-of-magnitude improvement in efficiency when horizontal grid spacing significantly exceeds vertical spacing; and the Python-based HPC platform enables seamless operation on both CPU and GPU architectures, with a single GPU delivering performance equivalent to clusters exceeding 325 standard CPU cores. Furthermore, the PyTorch backend provides inherent compatibility with machine learning, positioning LMARSpy as a cutting-edge tool to propel global numerical modeling into the AI era.

Dynamical Core,LMARS,Monotonicity Limiter,Vertical Implicit Solver,GPU