diff --git a/src/qibo/backends/numpy.py b/src/qibo/backends/numpy.py
index 30cf2fd5dc..6bc47097a7 100644
--- a/src/qibo/backends/numpy.py
+++ b/src/qibo/backends/numpy.py
@@ -118,7 +118,8 @@ def asmatrix_fused(self, fgate):
         for gate in fgate.gates:
             # transfer gate matrix to numpy as it is more efficient for
             # small tensor calculations
-            gmatrix = gate.asmatrix(self)
+            # explicit to_numpy see https://github.com/qiboteam/qibo/issues/928
+            gmatrix = self.to_numpy(gate.asmatrix(self))
             # Kronecker product with identity is needed to make the
             # original matrix have shape (2**rank x 2**rank)
             eye = np.eye(2 ** (rank - len(gate.qubits)), dtype=self.dtype)
diff --git a/tests/test_models_variational.py b/tests/test_models_variational.py
index 7777d8b3b9..795a7b57b2 100644
--- a/tests/test_models_variational.py
+++ b/tests/test_models_variational.py
@@ -166,7 +166,7 @@ def test_qaoa_execution(backend, solver, dense, accel=None):
             u = expm(-1j * p * m_matrix)
         else:
             u = expm(-1j * p * h_matrix)
-        target_state = u @ target_state
+        target_state = backend.cast(u) @ target_state
 
     qaoa = models.QAOA(h, mixer=m, solver=solver, accelerators=accel)
     qaoa.set_parameters(params)
@@ -196,7 +196,7 @@ def test_qaoa_callbacks(backend, accelerators):
     h_matrix = backend.to_numpy(h.matrix)
     m_matrix = backend.to_numpy(qaoa.mixer.matrix)
     calc_energy = lambda s: (s.conj() * h_matrix.dot(s)).sum()
-    target_state = np.copy(state)
+    target_state = backend.to_numpy(state)
     target_energy = [calc_energy(target_state)]
     for i, p in enumerate(params):
         if i % 2:
diff --git a/tests/test_quantum_info_superoperator_transformations.py b/tests/test_quantum_info_superoperator_transformations.py
index d32c22cf24..5e8868d2ca 100644
--- a/tests/test_quantum_info_superoperator_transformations.py
+++ b/tests/test_quantum_info_superoperator_transformations.py
@@ -314,7 +314,9 @@ def test_liouville_to_choi(backend, order):
     choi = liouville_to_choi(test_superop, order, backend)
 
     axes = [1, 2] if order == "row" else [0, 3]
-    test_choi = np.reshape(test_superop, [2] * 4).swapaxes(*axes).reshape([4, 4])
+    test_choi = backend.cast(
+        np.reshape(test_superop, [2] * 4).swapaxes(*axes).reshape([4, 4])
+    )
 
     backend.assert_allclose(
         backend.calculate_norm(choi - test_choi) < PRECISION_TOL, True
@@ -324,12 +326,15 @@ def test_liouville_to_choi(backend, order):
 @pytest.mark.parametrize("order", ["row", "column"])
 def test_choi_to_liouville(backend, order):
     axes = [1, 2] if order == "row" else [0, 3]
-    test_choi = np.reshape(test_superop, [2] * 4).swapaxes(*axes).reshape([4, 4])
+    test_choi = backend.cast(
+        np.reshape(test_superop, [2] * 4).swapaxes(*axes).reshape([4, 4])
+    )
 
     liouville = choi_to_liouville(test_choi, order=order, backend=backend)
 
     backend.assert_allclose(
-        backend.calculate_norm(liouville - test_superop) < PRECISION_TOL, True
+        backend.calculate_norm(liouville - backend.cast(test_superop)) < PRECISION_TOL,
+        True,
     )
 
 
@@ -343,7 +348,9 @@ def test_choi_to_kraus(
     backend, order, validate_cp, test_a0, test_a1, test_kraus_left, test_kraus_right
 ):
     axes = [1, 2] if order == "row" else [0, 3]
-    test_choi = np.reshape(test_superop, [2] * 4).swapaxes(*axes).reshape([4, 4])
+    test_choi = backend.cast(
+        np.reshape(test_superop, [2] * 4).swapaxes(*axes).reshape([4, 4])
+    )
 
     with pytest.raises(TypeError):
         choi_to_kraus(test_choi, str(PRECISION_TOL), backend=backend)
@@ -888,7 +895,7 @@ def test_reshuffling(backend, order):
     reshuffled = _reshuffling(reshuffled, order, backend=backend)
 
     backend.assert_allclose(
-        np.linalg.norm(reshuffled - test_superop) < PRECISION_TOL, True
+        np.linalg.norm(reshuffled - backend.cast(test_superop)) < PRECISION_TOL, True
     )
 
     axes = [1, 2] if order == "row" else [0, 3]
@@ -898,5 +905,5 @@ def test_reshuffling(backend, order):
     reshuffled = _reshuffling(reshuffled, order, backend=backend)
 
     backend.assert_allclose(
-        np.linalg.norm(reshuffled - test_choi) < PRECISION_TOL, True
+        np.linalg.norm(reshuffled - backend.cast(test_choi)) < PRECISION_TOL, True
     )