Evaluating real-world einsum expressions can lead to hyper-sparse intermediate tensors that are hard to predict, we exploit this dynamic sparsity in a hybrid algorithm.