Creating Conditional Trainable Random PyTorch Tensor

Is it possible to create a random torch tensor with a condition (say that elements add up to 1),which can act as trainable tensor.