fix device of Tensor.arange inside Tensor.one_hot (tinygrad#3199)

it should have the same device as self

test/test_tensor.py

@@ -437,5 +437,12 @@ def test_reduce_default(self):
     np.testing.assert_equal(Tensor([]).sum().numpy(), 0)
     np.testing.assert_equal(Tensor([]).mean().numpy(), 0)
 
+class TestTensorCreationDevice(unittest.TestCase):
+  # test auxiliary tensors are created on the same device
+  def test_one_hot(self):
+    y = Tensor([1, 2, 3]).to("CPU")
+    x = y.one_hot(10)
+    x.realize()
+
 if __name__ == '__main__':
   unittest.main()

tinygrad/tensor.py

@@ -901,7 +901,8 @@ def dropout(self, p=0.5) -> Tensor:
     if not Tensor.training or p == 0: return self
     return self * (Tensor.rand(*self.shape, requires_grad=False, device=self.device) >= p) * (1/(1.0 - p))
 
-  def one_hot(self, num_classes:int, **kwargs) -> Tensor: return Tensor.where(self[..., None] == Tensor.arange(num_classes), 1, 0, **kwargs)
+  def one_hot(self, num_classes:int) -> Tensor:
+    return (self[..., None] == Tensor.arange(num_classes, requires_grad=False, device=self.device)).where(1, 0)
 
   def scaled_dot_product_attention(self, key:Tensor, value:Tensor, attn_mask:Optional[Tensor]=None, dropout_p:float=0.0, is_causal:bool=False) -> Tensor:  # noqa: E501
     # NOTE: it works if key, value have symbolic shape