How to get a new attribute in backward

I’ve defined a new attribute named man in the grad_output with the backward function. I want to get this attribute when I call this function the second time,but I can`t get it. It is strange that when I use PDB I can get the man value.
code:

import torch
import torch.nn as nn
import pdb
import os
import random
import numpy as np


class LinearFunction(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input, weight):
        print('input id: {}'.format(id(input)))
        ctx.save_for_backward(input, weight)
        output = input.mm(weight.t())
        if not hasattr(input, 'man'):
            output.man = 1
        else:
            output.man = input.man + 1
        print('output.man :{}'.format(output.man))

        return output

    @staticmethod
    def backward(ctx, grad_output):
        input, weight = ctx.saved_tensors
        grad_input = grad_weight = grad_bias = None
        if ctx.needs_input_grad[0]:
            grad_input = grad_output.mm(weight)
        if ctx.needs_input_grad[1]:
            grad_weight = grad_output.t().mm(input)
        # pdb.set_trace()
        print(id(grad_output), id(grad_input))
        if not hasattr(grad_output, 'man'):
            setattr(grad_input, 'man', 1)
        else:
            grad_input.man = grad_output.man + 1
        print('linear {} grad_input.man:{}'.format(id(grad_output), grad_input.man))

        return grad_input, grad_weight, grad_bias


class Linear(nn.Module):
    def __init__(self, input_features, output_features):
        super(Linear, self).__init__()
        self.input_features = input_features
        self.output_features = output_features
        self.weight = nn.Parameter(torch.Tensor(output_features, input_features))
        self.weight.data.uniform_(-0.1, 0.1)

    def forward(self, input):
        return LinearFunction.apply(input, self.weight)


def seed_torch(seed=1029):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True


seed_torch()

x = torch.rand((2, 4), requires_grad=True)
y = Linear(input_features=4, output_features=3)(x)
y = Linear(input_features=3, output_features=3)(y)
z = torch.sum(y)

y.retain_grad()

z.backward()

output without pdb:

input id: 140034962031456
output.man :1
input id: 140034962034408
output.man :2
140034962080032 140034962080248
linear 140034962080032 grad_input.man:1
140034962080032 140034962080176
linear 140034962080032 grad_input.man:1

output with pdb:

input id: 139694370501472
output.man :1
input id: 139694370504424
output.man :2
> /data1/UAL/Codes/UAL2.6/tools/eval/test.py(32)backward()
-> print(id(grad_output), id(grad_input))
(Pdb) c
c
139694370545952 139694370546168
linear 139694370545952 grad_input.man:1
> /data1/UAL/Codes/UAL2.6/tools/eval/test.py(32)backward()
-> print(id(grad_output), id(grad_input))
(Pdb) c
c
139694370546168 139694370547608
linear 139694370546168 grad_input.man:2

What should I do to get the value of the attribute named man