set seed before each test (#5468)

**Context:**

We've been seeing a lot of stochastic failures lately. We have random
tests that rely on the global numpy rng. Adding new tests or changing
around the order tests are run in can cause tests to randomly fail,
eating large amounts of CI and developer time.

**Description of the Change:**

Sets a global numpy seed before each test.

**Benefits:**

No more CI flakiness.

**Possible Drawbacks:**

Might still be a "special" seed, with failures on other seed values. We
should periodically update this seed just for testing robustness.

**Related GitHub Issues:**

tests/conftest.py

@@ -39,6 +39,12 @@ class DummyDevice(DefaultGaussian):
  _operation_map["Kerr"] = lambda *x, **y: np.identity(2)
 
 
+@pytest.fixture(autouse=True)
+def set_numpy_seed():
+ np.random.seed(9872653)
+ yield
+
+
 @pytest.fixture(scope="session")
 def tol():
  """Numerical tolerance for equality tests."""

tests/gradients/finite_diff/test_finite_difference_shot_vec.py

@@ -718,7 +718,7 @@ def test_multiple_expectation_values(self, approx_order, strategy, validate):
 
  assert isinstance(res[0], tuple)
  assert len(res[0]) == 2
- assert np.allclose(res[0], [-np.sin(x), 0], atol=0.1, rtol=0)
+ assert np.allclose(res[0], [-np.sin(x), 0], atol=0.15, rtol=0)
  assert isinstance(res[0][0], numpy.ndarray)
  assert isinstance(res[0][1], numpy.ndarray)
 

tests/gradients/parameter_shift/test_parameter_shift_shot_vec.py

@@ -1305,7 +1305,7 @@ def test_non_involutory_variance_single_param(self):
  for gradF in all_gradF:
  assert isinstance(gradF, np.ndarray)
  assert gradF.shape == ()
- assert gradF == pytest.approx(expected, abs=1)
+ assert qml.math.allclose(gradF, expected, atol=2 * _herm_shot_vec_tol)
 
  @flaky(max_runs=5)
  def test_non_involutory_variance_multi_param(self):
@@ -1359,7 +1359,7 @@ def test_non_involutory_variance_multi_param(self):
  # Note: the tolerances here are significantly higher than in usual tests
  # due to the stochasticity of the test case
  assert gradF[0] == pytest.approx(expected, abs=2)
- assert gradF[1] == pytest.approx(expected, abs=1)
+ assert qml.math.allclose(gradF[1], expected, atol=1.5)
 
  @flaky(max_runs=8)
  def test_involutory_and_noninvolutory_variance_single_param(self):
@@ -1486,8 +1486,8 @@ def test_involutory_and_noninvolutory_variance_multi_param(self):
 
  assert np.allclose(shot_vec_result[0][0], expected[0], atol=1)
  assert np.allclose(shot_vec_result[0][1], expected[1], atol=1)
- assert np.allclose(shot_vec_result[1][0], expected[2], atol=1)
- assert np.allclose(shot_vec_result[1][1], expected[3], atol=1)
+ assert np.allclose(shot_vec_result[1][0], expected[2], atol=1.5)
+ assert np.allclose(shot_vec_result[1][1], expected[3], atol=1.5)
 
  @pytest.mark.parametrize("ind", [0, 1])
  def test_var_and_probs_single_param(self, ind):

tests/transforms/test_hamiltonian_expand.py

@@ -345,9 +345,7 @@ def test_processing_function_shot_vectors_broadcasting(self, H, expected, groupi
  """Tests that the processing function works with shot vectors, parameter broadcasting,
  and grouping with different number of coefficients in each group"""
 
- np.random.seed(824)
- dev_with_shot_vector = qml.device("default.qubit", shots=[(8000, 4)])
-
+ dev_with_shot_vector = qml.device("default.qubit", shots=[(10000, 4)])
  if grouping:
  H.compute_grouping()
 

tests/transforms/test_qcut.py

@@ -2506,7 +2506,7 @@ def target_circuit(v):
  qml.RX(2.3, wires=2)
  return qml.expval(qml.PauliZ(wires=0) @ qml.PauliZ(wires=2))
 
- dev = dev_fn(wires=2, shots=10000)
+ dev = dev_fn(wires=2, shots=20000)
 
  @partial(qml.cut_circuit_mc, classical_processing_fn=fn)
  @qml.qnode(dev)