In register_full_backward_hook - grad is None for first module

1

Why are gradients not defined in the first module (the last one when backward)?
Thank you.

import torch, torch.nn as nn

model = nn.Sequential(nn.Linear(3, 2),  nn.Linear(2, 1) )

def hook(module, grad_input, grad_output):        
    for k,p in module.named_parameters():
        print(f"{k:20s} {tuple(p.grad.shape) if p.grad is not None else p.grad }  {tuple(p.data.shape)}")    

model[0].register_full_backward_hook(hook)
model[1].register_full_backward_hook(hook)

x = torch.randn(1, 3)
out = model(x)
out.mean().backward()

Output:

weight               (1, 2)  (1, 2)
bias                 (1,)  (1,)
weight               None  (2, 3)
bias                 None  (2,)
2

see: register_full_backward_hook does not consistently fire #56380