Numpy 1.24 breaks master.

[tanujkhattar]

Description of the issue
Numpy 1.24 was released recently (see release notes) and it breaks master branch.

Using alias for built-in types was deprecated in Numpy 1.20 (See https://numpy.org/doc/stable/release/1.20.0-notes.html#using-the-aliases-of-builtin-types-like-np-int-is-deprecated) and was finally removed in the 1.24 release. As a result, all cirq code that uses aliases for built-in types like np.bool breaks.

Specifically, we'd need to fix the the proper_repr function to use bool instead of np.bool, which would end up changing the repr of a bunch a Cirq classes.

Cirq/cirq-core/cirq/_compat.py

Line 143 in db33aad

return f'np.array({value.tolist()!r}, dtype=np.{value.dtype})'

Other occurrences of np.bool are in repr and repr_inward files, which can be updated directly to use bool instead of np.bool. Specifically, the following two files:

Cirq/cirq-google/cirq_google/json_test_data/cirq.google.EngineResult.repr

Line 1 in db33aad

cirq_google.EngineResult(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), records={'m': np.array([[[True, True, False, True, False]], [[False, True, True, False, False]], [[True, False, True, False, True]]], dtype=np.bool)}, job_id='my_job_id', job_finished_time=datetime.datetime(2022, 4, 1, 8, 23, 45, tzinfo=datetime.timezone.utc))

Cirq/cirq-core/cirq/protocols/json_test_data/TrialResult.repr_inward

Line 1 in db33aad

[cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[True, True, False, True, False], [False, True, True, False, False], [True, False, True, False, True]], dtype=np.bool)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[1, 1, 0, 1, 0], [0, 1, 1, 0, 0], [1, 0, 1, 0, 1]], dtype=np.uint8)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.int8)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.int16)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.int32)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.int64)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.uint8)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.uint16)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.uint32)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.uint64)}), cirq.ResultDict(params=cirq.ParamResolver({sympy.Symbol('a'): 0.5}), measurements={'m': np.array([[1, 1, 0, 1, 0], [0, 1, 1, 0, 0], [1, 0, 1, 0, 1]], dtype=np.uint8), 'n': np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=np.int64)})]

Cirq version

1.1.0.dev

[tanujkhattar]

Another source of conflict is that numba (a Cirq dependency due to cirq-contrib) supports numpy only upto 1.23; so we are blocked on them before we can upgrade to 1.24. See https://numba.readthedocs.io/en/stable/user/installing.html#compatibility

For now, I will limit numpy version to 1.23 and then we can work on the upgrade once numba adds support for 1.24.

[mpharrigan]

can we add a comment to the requirements.txt linking to this issue so it's clear why we're restricting the version

[mpharrigan]

Is anyone assigned to this high priority issue?

[tanujkhattar]

I think this should be unblocked by #5991 and a new release of numba that now supports numpy 1.24 (https://numba.readthedocs.io/en/stable/user/installing.html#compatibility)

[mpharrigan]

sweeeet! Regardless of what numba is doing, we can un-pin it in our dependencies, yes?

[tanujkhattar]

I think so, let's find out - #6149

[ChrisPattison]

Is there a plan to release soon with numpy 1.24 support?