Sparse code multiple access (SCMA) uses multi-dimensional sparse codewords to transmit user data. The expectation propagation algorithm (EPA) exploiting the sparse property shows linear complexity growth and thus is preferred for multi-user detection. To further reduce the complexity, a convergence-aware based EPA for uplink MIMO SCMA systems is proposed. Techniques including user termination, antenna termination, and codebook reduction are adopted. The user termination must be combined with the iteration constraint to avoid misjudgement. The antenna termination can stop the computations related with certain antennas having strong channel gains. Only possible codewords are considered in the reduced codebook to eliminate unnecessary calculation for posterior probability. From simulation results, we show that these three techniques can strike a balance between performance and complexity and more than 50% complexity can be saved.