Tests do not pass on Apple M1 running python 3

[mathDR]

Hi! I just git cloned the repo into a new virtual environment. I ran pip install -e . on the repo, then pytest tests. That failed (I needed to pip install pytest).

After running pytest tests I show 71 failed, 11 passed, 2 skipped in 14.64 s.

I am running macOS Monterey on an Apple M1 Max running python 3.9.10.

A particular instance was running pytest test_noise.py:

test_noise.py FF.                                                                                                 [100%]

======================================================= FAILURES ========================================================
_____________________________________________________ test_diagonal _____________________________________________________

    def test_diagonal():
        N = 50
        random = np.random.default_rng(9432)
        diag = random.normal(size=N)
        noise = tinygp.noise.Diagonal(diag=diag)
>       check_noise_model(noise, np.diag(diag))

test_noise.py:42: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

noise = Diagonal(diag=array([ 1.58860105,  0.38873816,  0.43751174, -0.11334617,  0.14765921,
       -1.65553085, -0.62329537,...44718,  1.76214413,  1.19522174, -0.97858096,
       -0.46264887, -1.94102813,  0.30831414, -0.31562414, -1.43893292]))
dense_rep = array([[ 1.58860105,  0.        ,  0.        , ...,  0.        ,
         0.        ,  0.        ],
       [ 0.       ...414,  0.        ],
       [ 0.        ,  0.        ,  0.        , ...,  0.        ,
         0.        , -1.43893292]])

    def check_noise_model(noise, dense_rep):
        random = np.random.default_rng(6675)
    
        np.testing.assert_allclose(noise.diagonal(), jnp.diag(dense_rep))
        np.testing.assert_allclose(noise + np.zeros_like(dense_rep), dense_rep)
    
        y1 = random.normal(size=dense_rep.shape)
>       np.testing.assert_allclose(noise + y1, dense_rep + y1)
E       AssertionError: 
E       Not equal to tolerance rtol=1e-07, atol=0
E       
E       Mismatched elements: 6 / 2500 (0.24%)
E       Max absolute difference: 1.392429e-07
E       Max relative difference: 1.21363102e-06
E        x: array([[ 1.922384, -1.28015 , -0.315337, ..., -0.567188,  0.508422,
E               -0.657338],
E              [-1.391238,  0.80504 ,  0.164051, ...,  0.481563, -1.469477,...
E        y: array([[ 1.922384, -1.28015 , -0.315337, ..., -0.567188,  0.508422,
E               -0.657338],
E              [-1.391238,  0.80504 ,  0.164051, ...,  0.481563, -1.469477,...

test_noise.py:17: AssertionError
______________________________________________________ test_banded ______________________________________________________

    def test_banded():
        N, J = 50, 5
        random = np.random.default_rng(9432)
    
        # Create a random symmetric banded matrix
        R = random.normal(size=(N, N))
        R[np.triu_indices(N, J + 1)] = 0
        R[np.tril_indices(N)] = R.T[np.tril_indices(N)]
    
        # Extract the diagonal and off-diagonal elements
        diag = np.diag(R)
        off_diags = np.zeros((N, J))
        for j in range(J):
            off_diags[: N - j - 1, j] = R[
                (np.arange(0, N - j - 1), np.arange(j + 1, N))
            ]
    
        noise = tinygp.noise.Banded(diag=diag, off_diags=off_diags)
>       check_noise_model(noise, R)

test_noise.py:63: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

noise = Banded(diag=array([ 1.58860105, -2.16543909, -0.52223767,  2.62247777, -0.51765523,
        0.41917606,  1.36994167,  ....00000000e+00],
       [ 0.00000000e+00,  0.00000000e+00,  0.00000000e+00,
         0.00000000e+00,  0.00000000e+00]]))
dense_rep = array([[ 1.58860105,  0.38873816,  0.43751174, ...,  0.        ,
         0.        ,  0.        ],
       [ 0.3887381...348, -1.30796807],
       [ 0.        ,  0.        ,  0.        , ...,  2.26915227,
        -1.30796807, -1.53816113]])

    def check_noise_model(noise, dense_rep):
        random = np.random.default_rng(6675)
    
        np.testing.assert_allclose(noise.diagonal(), jnp.diag(dense_rep))
        np.testing.assert_allclose(noise + np.zeros_like(dense_rep), dense_rep)
    
        y1 = random.normal(size=dense_rep.shape)
>       np.testing.assert_allclose(noise + y1, dense_rep + y1)
E       AssertionError: 
E       Not equal to tolerance rtol=1e-07, atol=0
E       
E       Mismatched elements: 52 / 2500 (2.08%)
E       Max absolute difference: 3.63120751e-07
E       Max relative difference: 1.09484999e-05
E        x: array([[ 1.922384e+00, -8.914119e-01,  1.221748e-01, ..., -5.671883e-01,
E                5.084218e-01, -6.573380e-01],
E              [-1.002500e+00, -1.749138e+00,  2.532022e+00, ...,  4.815633e-01,...
E        y: array([[ 1.922384e+00, -8.914119e-01,  1.221748e-01, ..., -5.671883e-01,
E                5.084218e-01, -6.573380e-01],
E              [-1.002500e+00, -1.749137e+00,  2.532022e+00, ...,  4.815633e-01,...

test_noise.py:17: AssertionError
================================================ short test summary info ================================================
FAILED test_noise.py::test_diagonal - AssertionError: 
FAILED test_noise.py::test_banded - AssertionError: 
============================================== 2 failed, 1 passed in 0.50s ==============================================

[dfm]

This is to be expected - the tests must be run with double precision enabled. The preferred mode for running the tests (as described in the docs) is with tox which will handle the settings appropriately.