Résumé : X-ray binaries (XRBs) display spectral state transitions that are accompanied by substantial changes in the hardness, luminosity, and structure of the accretion flow. We developed a GPU-accelerated cooling toolkit for general relativistic magnetohydrodynamic (GRMHD) simulations of accreting black holes that uses texture memory for fast retrieval of pre-computed values. The toolkit incorporates bremsstrahlung, synchrotron, inverse Compton radiation and Coulomb collision processes. We implemented our toolkit into a GRMHD code and used it to simulate a magnetically arrested disk in the context of the XRB low/hard state around a Kerr black hole. We explored the mass accretion rate in the $\sim (10^{-6}-0.3) \dot{M}_{\rm Edd}$ range, where $\dot{M}_{\rm Edd}$ is the Eddington accretion rate. Our simulations reveal that for low accretion rates ($\dot{M} \lesssim 0.01 \dot{M}_{\rm Edd}$), the flow settles into a geometrically thick, low-density, two-temperature hot accretion flow. At higher accretion rates, the flow turns into a cold single-temperature thin disk at $r_{\rm in} \gtrsim 50 r_g$. Inside, the disk breaks up into single-temperature thin filaments embedded into a two-temperature hot thick flow. Our GPU texture memory accelerated cooling prescription is $3-5$ times faster than the standard radiation M1 closure methods, and $\sim5$ times faster than storing the lookup table in global memory.