UP031: ruff 0.4.2 generated invalid fix it code

[Skylion007]

                msg="\n".join(
                    [
                        "alpha = alpha_c + %.2g" % shift,
                        "expected_grad: %.5g" % expected_grad,
                        "actual_grad: %.5g" % actual_grad,
                        "error = %.2g" % torch.abs(expected_grad - actual_grad).max(),
                    ]

became

                msg="\n".join(
                    [
                        f"alpha = alpha_c + {shift:.2g}",
                        f"expected_grad: {expected_grad:.5g}",
                        f"actual_grad: {actual_grad:.5g}",
                        f"error = {torch.abs(expected_grad - actual_grad).max():.2g}",
                    ]

which throws an error. I know it's an unsafe fix, but unsafe fixes should not generate invalid code. Specifically:

Traceback (most recent call last):
  File "/var/lib/jenkins/workspace/test/distributions/test_distributions.py", line 4658, in test_dirichlet_multivariate
    f"expected_grad: {expected_grad:.5g}",
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_tensor.py", line 989, in __format__
    return object.__format__(self, format_spec)
TypeError: unsupported format string passed to Tensor.__format__

from this PR: pytorch/pytorch#125031

[AlexWaygood]

Guessing this is due to the recent change to UP031? Cc. @plredmond

[charliermarsh]

Why can't tensors be passed-in there, but can be formatted with % formatting?

[Skylion007]

@charliermarsh I think it's related to classes overriding object.__format__ magic method. Does the '%' method cast it to a string before calling __format__ on it or something?

[Skylion007]

@charliermarsh Ah, apparently in the first code, it's implicitly converted before: "Here, you're using the old-style string formatting (%-formatting), where shift, a PyTorch tensor, is implicitly converted to a numeric type (float or integer) when it's formatted. This conversion happens before the formatting is applied, so the process doesn't raise any issues."

import torch
x = torch.tensor([1.0])
print("Value: %.5g" % x)  # Works because x is cast as a float before __format__ is called
print(f"Value: {x:.5g}") # fails because Tensor.__format__ is called

[Skylion007]

I think explicitly casting to float in all float specific f-string fixits could solve this issue?

[charliermarsh]

I think that makes sense. E.g., when there's a :.5g specifier?

[Skylion007]

There are probably other specifies that do this casting f (or maybe some to int for d) etc.

[plredmond]

@charliermarsh perhaps we could add this kind of explicit-conversion to the UP031 rule after we can infer the types of the expressions? alternatively, since we parse the format string we could use that information to select an explicit-conversion function?