bugfix. If you used Value, TensorValue, tensor.value(), the value was not considered as an input. So this was making the inplace_opt optimizer optimize it while it should not be doing it.