Created by: willthefrog

cherry pick #24409

test the error message with follow code snippet:

import paddle
from paddle import fluid
import paddle.fluid.core as core
import numpy as np

def test_error_message(skip_grad_var=[], check_sub_result=False):
    x = np.random.uniform(0.1, 0.6, (2, 3, 4)).astype("float32")
    y = np.random.uniform(0.1, 0.6, (2, 3, 4)).astype("float32")
    sub_res = x - y
    sub_res = sub_res.reshape((2, 3 * 4))
    output = sub_res * sub_res
    out_grad = np.random.random_sample((2, 1)).astype(np.float32)
    sub_result = np.random.uniform(0.1, 0.6, (1, 3 * 4)).astype("float32")

    var_dict = {
        'x': x,
        'y': y,
        'sub_result': sub_result if check_sub_result else sub_res,
        'out': output,
        'out@GRAD': out_grad,
    }

    program = fluid.Program()
    with fluid.program_guard(program):
        block = program.global_block()
        for name in var_dict:
            block.create_var(
                name=name,
                dtype='float32',
                shape=var_dict[name].shape)
        l2dist_op = block.append_op(
            type="squared_l2_distance",
            inputs={
                "X": block.var('x'),
                "Y": block.var('y'),
            },
            outputs={
                "Out": block.var('out'),
                "sub_result": block.var('sub_result')})

        # generate backward op_desc
        grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc(
            l2dist_op.desc, set(), [])
        grad_op_desc = grad_op_desc_list[0]
        new_op_desc = block.desc.append_op()
        new_op_desc.copy_from(grad_op_desc)
        grad_vars = [name for name in grad_op_desc.output_arg_names()
                     if name not in skip_grad_var]
        for var_name in grad_vars:
            block.desc.var(var_name.encode("ascii"))
        grad_op_desc.infer_var_type(block.desc)
        grad_op_desc.infer_shape(block.desc)
        for arg in grad_vars:
            grad_var = block.desc.find_var(arg.encode("ascii"))
            grad_var.set_dtype(core.VarDesc.VarType.FP32)

        # remove forward op, only grad is left
        block.ops = []
        block.desc._remove_op(0, 1)
        exe = fluid.Executor(fluid.CPUPlace())
        exe.run(program, feed={name: var_dict[name]
                               for name in ['x', 'y', 'sub_result', 'out@GRAD']})

# NOTE run those one at a time, comment out the others

# InvalidArgumentError: First dimension of gradient must be greater or equal than first dimension of target. But received gradient dimension = 1 and target dimension is 2.
test_error_message(check_sub_result=True)

# NotFoundError: variable(X@GRAD) cannot be found in scope for operator 'squared_l2_distance_grad'.
test_error_message(skip_grad_var=['x@GRAD'])

# NotFoundError: variable(Y@GRAD) cannot be found in scope for operator 'squared_l2_distance_grad'.
test_error_message(skip_grad_var=['y@GRAD'])