Set negative learning rate in torch.optim. One way fails while another works, why?

I want to define an optimizer in pytorch for my code.

The code needs to have a negative learning rate for a special alogrithm.

When I write code like this:

import torch

class Net(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = torch.nn.Linear(3, 2)

    def forward(self, x):
        return self.linear1(x)

if __name__ == '__main__':
    net = Net()
    main_model_params = [{
        "params": [p for _, p in net.named_parameters()],
        "weight_decay": 0.0,
    }]
    optimizer = torch.optim.AdamW(
                main_model_params,
                lr = -0.1
            )
    print(optimizer)

Then an error will happens:

Traceback (most recent call last):
  File ".\experiment\optim_test.py", line 18, in <module>
    optimizer = torch.optim.AdamW(
  File "C:\ProgramData\Miniconda3\lib\site-packages\torch\optim\adamw.py", line 70, in __init__
    raise ValueError("Invalid learning rate: {}".format(lr))
ValueError: Invalid learning rate: -0.1

But when I change a method to write the code:

import torch
import torch.nn

class Net(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = torch.nn.Linear(3, 2)

    def forward(self, x):
        return self.linear1(x)

if __name__ == '__main__':
    net = Net()
    main_model_params = [{
        "params": [p for _, p in net.named_parameters()],
        "weight_decay": 0.0,
        "lr": -0.1
    }]
    optimizer = torch.optim.AdamW(
                main_model_params,
            )
    print(optimizer)

It can be executed rightly:

AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    eps: 1e-08
    lr: -0.1
    maximize: False
    weight_decay: 0.0
)

Who can tell me why?