Is the operation "torch.max" differentiable？

[tangcongyi]

Thank you for your contribution! There is a question for you: In keep_autoaugment.py, there are some codes:

"images_half.requires_grad = True

    if self.early:
        preds = model(images_half,True)
    else:
        preds = model(images_half)
    
    **score, _ = torch.max(preds, 1)**
    score.mean().backward()
    slc_, _ = torch.max(torch.abs(images_half.grad), dim=1)"

Is the operation "torch.max" differentiable？In my project, it seems to stop the process of backward.

I would appreciate that if you could help me to handle this question!!!!!!!!!!!!!!!!!!