only float Tensors have gradient [pr] (#10475)

test/test_tensor.py

@@ -494,7 +494,7 @@ class TestTinygrad(unittest.TestCase):
       _a = Tensor([3]) in [Tensor([3]), Tensor([4]), Tensor([5])]
 
   def test_repr_with_grad(self):
-    a = Tensor([1], requires_grad=True)
+    a = Tensor([1.0], requires_grad=True)
     b = Tensor([1])
     c = (a + b).sum().backward()
     print(a)

test/unit/test_gradient.py

@@ -104,6 +104,11 @@ class TestTensorGradient(unittest.TestCase):
     x_casted = x_reshaped.cast(dtypes.float16)
     x_casted.mean().gradient(x_reshaped)
 
+  def test_non_float_tensor_raise(self):
+    x = Tensor([1, 2, 3])
+    with self.assertRaises(RuntimeError): x.sum().gradient(x)
+    with self.assertRaises(RuntimeError): x.float().sum().gradient(x)
+
 class TestRealizeMeansRealize(unittest.TestCase):
   def test_randn_realizes(self):
     x = Tensor.randn(2, 3, 64, 64, requires_grad=True).realize()

tinygrad/tensor.py

@@ -907,6 +907,7 @@ class Tensor(MathTrait):
     ```
     """
     assert gradient is not None or self.shape == tuple(), "when no gradient is provided, backward must be called on a scalar tensor"
+    if not (self.is_floating_point() and all(t.is_floating_point() for t in targets)): raise RuntimeError("only float Tensors have gradient")
     if gradient is None: gradient = Tensor(1.0, dtype=self.dtype, device=self.device, requires_grad=False)
     rets = []
     target_uops = [x.lazydata for x in targets]