MacBook M4 processor, using 'mps' generates an unexpected sharp rise in loss in some training step

[Jessen-Li]

MacBook Pro, chip: Apple M4 Pro
Section 5.2, page 149,
when using cpu for computation, device = torch.device("cuda" if torch.cuda.is_available() else 'cpu'),
the resulting plot is

when using mps, device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu'),
the resulting plot is

There is a weird sharp rise of train loss and val loss in step 000020

Ep 2 (Step 000010): Train loss 6.747, Val loss 7.042
Ep 2 (Step 000015): Train loss 6.061, Val loss 6.626
every effort moves you.                                                 
Ep 3 (Step 000020): Train loss 9.083, Val loss 10.437
Ep 3 (Step 000025): Train loss 5.652, Val loss 6.434

Does it say that there is some issues with mps, the GPU of M4 chip?

[kulinseth]

Can you provide the details to reproduce and what Torch version you are using ?

[Jessen-Li]

use the same code as that in the book.
python -c "import torch; print(torch.version)"
2.7.0

[casinca]

This sounds expected and related to this issue, Sebastian replied #625 (comment)