PyTorch backend warns on every read-only constant: torch.as_tensor on non-writable array

[jessegrabowski]

Rewrites can produce constants whose data is read-only, and pytorch_typify_tensor hands them straight to torch.as_tensor, which shares memory and warns that writes are undefined behavior. Anything running filterwarnings = error fails on it.

import numpy as np
import pytensor
import pytensor.tensor as pt

x = pt.vector("x")
fn = pytensor.function([x], pt.grad((x**2).sum(), x), mode="PYTORCH")
fn(np.ones(3))  # UserWarning: The given NumPy array is not writable

The read-only array comes from TensorConstant.unique_value, so the forward pass is fine and only a rewrite in the gradient trips it.

Potential fix (requires testing):

def pytorch_typify_tensor(data, dtype=None, **kwargs):
    if isinstance(data, np.ndarray) and not data.flags.writeable:
        data = data.copy()
    return torch.as_tensor(data, dtype=dtype)

[ricardoV94]

We don't write on constants so why not just filter the warning on the spot?