Use higher precision for linspace with double (#3029)

python/src/convert.cpp

@@ -259,6 +259,8 @@ nb::object tolist(mx::array& a) {
       return to_list<float>(a, 0, 0);
     case mx::bfloat16:
       return to_list<mx::bfloat16_t, float>(a, 0, 0);
+    case mx::float64:
+      return to_list<double>(a, 0, 0);
     case mx::complex64:
       return to_list<std::complex<float>>(a, 0, 0);
     default:

python/tests/test_double.py

@@ -292,6 +292,15 @@ class TestDouble(mlx_tests.MLXTestCase):
         b = np.array(a)
         self.assertTrue(np.array_equal(a, b))
 
+        a = mx.array([1.0, 2.0], mx.float64)
+        b = a.tolist()
+        self.assertEqual(b, [1.0, 2.0])
+
+    def test_linspace(self):
+        with mx.stream(mx.cpu):
+            vals = mx.linspace(0, math.pi, 2, mx.float64)
+            self.assertEqual(vals.tolist()[1], math.pi)
+
 
 if __name__ == "__main__":
     mlx_tests.MLXTestRunner()