MPS support

[aman-17]

Added MPS support.

[dirkgr]

Is this all good to go then?

[aman-17]

Yeah, just removed precision and batch_size from main function as you've mentioned last week in the meeting.

[dirkgr]

Did you try this to make sure it still works on MPS, and also on GPS?

[aman-17]

Did you try this to make sure it still works on MPS, and also on GPS?

It runs fine on both MPS and GPU.
Our .toml file installs Pytorch 2.6.0, but we need to downgrade to 2.5.1 for training.

[dirkgr]

Our .toml file installs Pytorch 2.6.0, but we need to downgrade to 2.5.1 for training.

Why is that?

[aman-17]

Our .toml file installs Pytorch 2.6.0, but we need to downgrade to 2.5.1 for training.

Why is that?

Pytorch released 2.6 and changed the default weights_only=True when loading the models which is restricting PosixPath's.

Our.toml file for your reference:

dependencies = [
    "torch>=2.1",
    ...
    ]

And error with torch==2.6.0:

(1) In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
(2) Alternatively, to load with `weights_only=True` please check the recommended steps in the following error message. WeightsUnpickler error: Unsupported global: GLOBAL pathlib.PosixPath was not an allowed global by default. Please use `torch.serialization.add_safe_globals([PosixPath])` or the `torch.serialization.safe_globals([PosixPath])` context manager to allowlist this global if you trust this class/function.

[dirkgr]

we changed the default value of the weights_only

Can you just change it where we call the function? Pass in the old value?

[aman-17]

we changed the default value of the weights_only

Can you just change it where we call the function? Pass in the old value?

Passed weights_only=False explicitly. Works well with both the torch versions(2.5.x and 2.6.0)

[dirkgr]

One more thing

[dirkgr]

Thank you!