We present GRACE, a new GPU-accelerated numerical-relativity framework designed to run efficiently on heterogeneous high-performance computing platforms. Developed from scratch and built exclusively on open-source libraries, GRACE employs Kokkos for performance portability across CPU and GPU architectures and p4est for adaptive mesh refinement. The code evolves the equations of ideal GRMHD—with divergence-free magnetic fields maintained by constrained transport—self-consistently coupled to the Einstein equations in the Z4c formulation, on fixed or adaptively refined grids. We validate the implementation against a suite of standard tests, ranging from magnetized shock tubes and the magnetic rotor in flat spacetime, through (magnetized) Bondi accretion onto a Schwarzschild black hole and the ringdown of a perturbed spinning puncture, to neutron-star oscillation spectra in fixed and dynamical spacetimes and the merger of binary black holes. As more demanding applications, we evolve two binary neutron-star mergers—an equal-mass, unmagnetized system with an ideal-gas equation of state and an unequal-mass, magnetized system with a finite-temperature tabulated equation of state—finding the inspiral dynamics to agree well with the FIL code. We also report single-device throughput together with strong- and weak-scaling results on multiple GPU and CPU architectures. GRACE is publicly released together with GRACEpy, a basic post-processing and data-analysis environment.