Constrain numpy version to be less than 2.x

[mhucka]

Although the latest release of Cirq is compatible with NumPy 2, other
dependencies for TFQ are not, and this leads to failures at run-time.
For now, tell pip not to install NumPy 2.

[MichaelBroughton]

Does this cause a breakage in the pip package too ? If so we might want to put this constraint there as well ?

[mhucka]

Does this cause a breakage in the pip package too ? If so we might want to put this constraint there as well ?

You would think that would make sense, but unfortunately, it doesn't end up working. The reason is the sequence of commands. First,

pip install -r requirements.txt

then followed separately by

pip install -U cirq --pre

and while we could constrain the version of NumPy in the requirements file to be < 2.0, the second command (installing Cirq) will cause NumPy 2 to be brought in. Pip doesn't remember the constraints between invocations.

I think what needs to happen is to make TFQ work with Cirq 1.5, at which point, the version of Cirq defined in requirements.txt can be set to 1.5 (instead of 1.3 as it is now), and then the version of NumPy in requirements.txt can be limited to < 2.0. Having the requirements and packages be installed in one go will make pip satisfy the simultaneous constraints.

For now, changing the compatibility workflow to use the pip constraints feature will let the CI checks pass.