Radiative transfer is often the dominant mode of heat transfer in fires, and solving the governing radiative transfer equation (RTE) in CFD fire simulations is computationally intensive. This work develops a versatile toolkit for training neural surrogates to solve various RTEs across different geometries and boundary conditions. Principal Component Analysis is applied for dimension reduction to enable efficient training of high-dimensional surrogate solvers. The mesh free nature of these surrogates enables them to overcome the ray effect suffered by traditional solvers. Our results demonstrate that neural surrogate can provide fast and accurate radiation predictions for practical problems important to fire safety research.