Constant folding override

[srikris-sridhar]

Issue around constant folding when working with views that share memory. I have a fix in the local branch. Will put up a PR shortly.

    # Ensure we copy the array to avoid BufferError when numpy operations
    # (like split) create views that share memory with original tensors
    if isinstance(array, np.ndarray) and not array.flags["OWNDATA"]:
        array = array.copy()

[srikris-sridhar]

@justinchuby please assign this to me. I'll put up the fix. I filed the issue so I would remember to do this.

[justinchuby]

Since we don't do in-place operations on np arrays. I was hoping this should not happen? Could you share a repro?

[srikris-sridhar]

That's exactly what I'm doing now. Chasing my repro. It's in one of the 500 models. I'm testing. I have to isolate it so might take me a little time.