diff --git a/examples/basics.rs b/examples/basics.rs
index f01f7871..d5d4f72b 100644
--- a/examples/basics.rs
+++ b/examples/basics.rs
@@ -2,7 +2,7 @@ use anyhow::Result;
 use candle::{Device, Tensor};
 
 fn main() -> Result<()> {
-    let x = Tensor::var(&[3f32, 1., 4.], Device::Cpu)?;
+    let x = Tensor::var(&[3f32, 1., 4.], &Device::Cpu)?;
     let y = (((&x * &x)? + &x * 5f64)? + 4f64)?;
     println!("{:?}", y.to_vec1::<f32>()?);
     Ok(())
diff --git a/examples/cuda_basics.rs b/examples/cuda_basics.rs
index e1dca6a9..a9647469 100644
--- a/examples/cuda_basics.rs
+++ b/examples/cuda_basics.rs
@@ -1,9 +1,12 @@
 use anyhow::Result;
-use candle::{DType, Device, Tensor};
+use candle::{Device, Tensor};
 
 fn main() -> Result<()> {
     let device = Device::new_cuda(0)?;
-    let x = Tensor::zeros(4, DType::F32, device)?;
+    let x = Tensor::new(&[3f32, 1., 4., 1., 5.], &device)?;
+    let y = Tensor::new(&[2f32, 7., 1., 8., 2.], &device)?;
     println!("{:?}", x.to_vec1::<f32>()?);
+    let z = (x + y)?;
+    println!("{:?}", z.to_vec1::<f32>()?);
     Ok(())
 }
diff --git a/src/tensor.rs b/src/tensor.rs
index be642329..a1262334 100644
--- a/src/tensor.rs
+++ b/src/tensor.rs
@@ -84,7 +84,7 @@ impl Tensor {
     fn ones_impl<S: Into<Shape>>(
         shape: S,
         dtype: DType,
-        device: Device,
+        device: &Device,
         is_variable: bool,
     ) -> Result<Self> {
         let shape = shape.into();
@@ -101,22 +101,22 @@ impl Tensor {
         Ok(Self(Arc::new(tensor_)))
     }
 
-    pub fn ones<S: Into<Shape>>(shape: S, dtype: DType, device: Device) -> Result<Self> {
+    pub fn ones<S: Into<Shape>>(shape: S, dtype: DType, device: &Device) -> Result<Self> {
         Self::ones_impl(shape, dtype, device, false)
     }
 
-    pub fn ones_var<S: Into<Shape>>(shape: S, dtype: DType, device: Device) -> Result<Self> {
+    pub fn ones_var<S: Into<Shape>>(shape: S, dtype: DType, device: &Device) -> Result<Self> {
         Self::ones_impl(shape, dtype, device, true)
     }
 
     pub fn ones_like(&self) -> Result<Self> {
-        Tensor::ones(self.shape(), self.dtype(), self.device())
+        Tensor::ones(self.shape(), self.dtype(), &self.device())
     }
 
     fn zeros_impl<S: Into<Shape>>(
         shape: S,
         dtype: DType,
-        device: Device,
+        device: &Device,
         is_variable: bool,
     ) -> Result<Self> {
         let shape = shape.into();
@@ -133,21 +133,21 @@ impl Tensor {
         Ok(Self(Arc::new(tensor_)))
     }
 
-    pub fn zeros<S: Into<Shape>>(shape: S, dtype: DType, device: Device) -> Result<Self> {
+    pub fn zeros<S: Into<Shape>>(shape: S, dtype: DType, device: &Device) -> Result<Self> {
         Self::zeros_impl(shape, dtype, device, false)
     }
 
-    pub fn zeros_var<S: Into<Shape>>(shape: S, dtype: DType, device: Device) -> Result<Self> {
+    pub fn zeros_var<S: Into<Shape>>(shape: S, dtype: DType, device: &Device) -> Result<Self> {
         Self::zeros_impl(shape, dtype, device, true)
     }
 
     pub fn zeros_like(&self) -> Result<Self> {
-        Tensor::zeros(self.shape(), self.dtype(), self.device())
+        Tensor::zeros(self.shape(), self.dtype(), &self.device())
     }
 
     pub fn new_impl<A: crate::device::NdArray>(
         array: A,
-        device: Device,
+        device: &Device,
         is_variable: bool,
     ) -> Result<Self> {
         let shape = array.shape()?;
@@ -164,11 +164,11 @@ impl Tensor {
         Ok(Self(Arc::new(tensor_)))
     }
 
-    pub fn new<A: crate::device::NdArray>(array: A, device: Device) -> Result<Self> {
+    pub fn new<A: crate::device::NdArray>(array: A, device: &Device) -> Result<Self> {
         Self::new_impl(array, device, false)
     }
 
-    pub fn var<A: crate::device::NdArray>(array: A, device: Device) -> Result<Self> {
+    pub fn var<A: crate::device::NdArray>(array: A, device: &Device) -> Result<Self> {
         Self::new_impl(array, device, true)
     }
 
diff --git a/tests/grad_tests.rs b/tests/grad_tests.rs
index 432b1520..56186e5d 100644
--- a/tests/grad_tests.rs
+++ b/tests/grad_tests.rs
@@ -3,7 +3,7 @@ use candle::{Device, Tensor};
 
 #[test]
 fn simple_grad() -> Result<()> {
-    let x = Tensor::var(&[3f32, 1., 4.], Device::Cpu)?;
+    let x = Tensor::var(&[3f32, 1., 4.], &Device::Cpu)?;
     let y = (((&x * &x)? + &x * 5f64)? + 4f64)?;
     let grads = y.backward()?;
     let grad_x = grads.get(&x).context("no grad for x")?;
diff --git a/tests/tensor_tests.rs b/tests/tensor_tests.rs
index fb2d84d9..81c2e801 100644
--- a/tests/tensor_tests.rs
+++ b/tests/tensor_tests.rs
@@ -2,7 +2,7 @@ use candle::{DType, Device, Result, Tensor};
 
 #[test]
 fn zeros() -> Result<()> {
-    let tensor = Tensor::zeros((5, 2), DType::F32, Device::Cpu)?;
+    let tensor = Tensor::zeros((5, 2), DType::F32, &Device::Cpu)?;
     let (dim1, dim2) = tensor.shape().r2()?;
     assert_eq!(dim1, 5);
     assert_eq!(dim2, 2);
@@ -11,7 +11,7 @@ fn zeros() -> Result<()> {
 
 #[test]
 fn add_mul() -> Result<()> {
-    let tensor = Tensor::new(&[3f32, 1., 4.], Device::Cpu)?;
+    let tensor = Tensor::new(&[3f32, 1., 4.], &Device::Cpu)?;
     let dim1 = tensor.shape().r1()?;
     assert_eq!(dim1, 3);
     let content: Vec<f32> = tensor.to_vec1()?;
@@ -28,7 +28,7 @@ fn add_mul() -> Result<()> {
 #[test]
 fn tensor_2d() -> Result<()> {
     let data = &[[3f32, 1., 4., 1., 5.], [2., 1., 7., 8., 2.]];
-    let tensor = Tensor::new(data, Device::Cpu)?;
+    let tensor = Tensor::new(data, &Device::Cpu)?;
     let dims = tensor.shape().r2()?;
     assert_eq!(dims, (2, 5));
     let content: Vec<Vec<f32>> = tensor.to_vec2()?;
@@ -39,9 +39,9 @@ fn tensor_2d() -> Result<()> {
 #[test]
 fn binary_op() -> Result<()> {
     let data = &[[3f32, 1., 4., 1., 5.], [2., 1., 7., 8., 2.]];
-    let tensor = Tensor::new(data, Device::Cpu)?;
+    let tensor = Tensor::new(data, &Device::Cpu)?;
     let data2 = &[[5f32, 5., 5., 5., 5.], [2., 1., 7., 8., 2.]];
-    let tensor2 = Tensor::new(data2, Device::Cpu)?;
+    let tensor2 = Tensor::new(data2, &Device::Cpu)?;
     let tensor = (&tensor + (&tensor * &tensor)? / (&tensor + &tensor2))?;
     let dims = tensor.shape().r2()?;
     assert_eq!(dims, (2, 5));