This new opt was crashing with the tensor.min() tensor.max() CAReduce. I fix this. I test the opt with many more reduction then just sum.