[PT FE] Support torch==2.6.0

[mvafin]

Details:

• Support torch==2.6.0

Tickets:

• CVS-162009

[rkazants]

Here update is also required: https://github.com/openvinotoolkit/openvino/blob/master/src/bindings/python/src/openvino/preprocess/torchvision/requirements.txt

Ideally, that file should removed in favor of the single file

[mvafin]

Here update is also required: https://github.com/openvinotoolkit/openvino/blob/master/src/bindings/python/src/openvino/preprocess/torchvision/requirements.txt

Ideally, that file should removed in favor of the single file

No, the update is not required there. It is a product requirements, not a test requirements, they only limit lower bound and upper bound for python 3.9

[cavusmustafa]

It seems like this torch.arange with torch.tensor inputs fail to trace with both make_fx and aot_autograd. The reason aot_autograd passes the test is that it adds graph break while tracing and falls back before it reaches to the openvino compiler function (fx_openvino in backend.py).

I can also reproduce this issue with torch inductor backend on my side. Below is a short reproducer:

import torch

def arange_fn(x):
    return torch.arange(x)

end = 5
end_tensor = torch.tensor(end)

compiled_fn = torch.compile(arange_fn, fullgraph=True)

# Executes without any issue
out = arange_fn(end)
print("Eager scalar: ", out)

# Executes without any issue
out = arange_fn(end_tensor)
print("Eager torch.tensor: ", out)

# Executes without any issue
out = compiled_fn(end)
print("Inductor scalar: ", out)

# Fails
out = compiled_fn(end_tensor)
print("Inductor torch.tensor: ", out)

I am not sure if this is a bug in pytorch side or simply not supported by aot_autograd and make_fx. Either way, we may not be able to fix it in openvino backend side.

And alternative way to test this op to change the input type to a regular scalar when testing for torch.compile maybe?

[mvafin]

That seems like a problem. I think tensor input should be better supported then scalar input. Makes sense to create an issue to torch.