diff --git a/candle-core/src/lib.rs b/candle-core/src/lib.rs
index ddd446ee..fa85f6e0 100644
--- a/candle-core/src/lib.rs
+++ b/candle-core/src/lib.rs
@@ -63,6 +63,7 @@ pub mod shape;
 mod storage;
 mod strided_index;
 mod tensor;
+pub mod test_utils;
 pub mod utils;
 mod variable;
 
diff --git a/candle-core/tests/test_utils.rs b/candle-core/src/test_utils.rs
similarity index 82%
rename from candle-core/tests/test_utils.rs
rename to candle-core/src/test_utils.rs
index 327e88c6..8ff73fc0 100644
--- a/candle-core/tests/test_utils.rs
+++ b/candle-core/src/test_utils.rs
@@ -1,9 +1,4 @@
-#![allow(dead_code)]
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use candle_core::{Result, Tensor};
+use crate::{Result, Tensor};
 
 #[macro_export]
 macro_rules! test_device {
@@ -23,6 +18,12 @@ macro_rules! test_device {
     };
 }
 
+pub fn to_vec0_round(t: &Tensor, digits: i32) -> Result<f32> {
+    let b = 10f32.powi(digits);
+    let t = t.to_vec0::<f32>()?;
+    Ok(f32::round(t * b) / b)
+}
+
 pub fn to_vec1_round(t: &Tensor, digits: i32) -> Result<Vec<f32>> {
     let b = 10f32.powi(digits);
     let t = t.to_vec1::<f32>()?;
@@ -40,7 +41,7 @@ pub fn to_vec2_round(t: &Tensor, digits: i32) -> Result<Vec<Vec<f32>>> {
     Ok(t)
 }
 
-pub fn to_vec3_round(t: Tensor, digits: i32) -> Result<Vec<Vec<Vec<f32>>>> {
+pub fn to_vec3_round(t: &Tensor, digits: i32) -> Result<Vec<Vec<Vec<f32>>>> {
     let b = 10f32.powi(digits);
     let t = t.to_vec3::<f32>()?;
     let t = t
diff --git a/candle-core/tests/conv_tests.rs b/candle-core/tests/conv_tests.rs
index 7a4c2956..2fdd4c74 100644
--- a/candle-core/tests/conv_tests.rs
+++ b/candle-core/tests/conv_tests.rs
@@ -1,6 +1,5 @@
-mod test_utils;
 use anyhow::Result;
-use candle_core::{Device, Tensor};
+use candle_core::{test_device, test_utils, Device, Tensor};
 
 /* This test is based on the following script.
 import torch
diff --git a/candle-core/tests/custom_op_tests.rs b/candle-core/tests/custom_op_tests.rs
index 7ec04c6a..cff0aebe 100644
--- a/candle-core/tests/custom_op_tests.rs
+++ b/candle-core/tests/custom_op_tests.rs
@@ -1,10 +1,8 @@
 use candle_core::backend::BackendStorage;
 use candle_core::cpu_backend;
+use candle_core::test_utils::to_vec1_round;
 use candle_core::{CpuStorage, CustomOp1, DType, Device, Error, Layout, Result, Shape, Tensor};
 
-mod test_utils;
-use test_utils::to_vec1_round;
-
 fn fwd<T: num_traits::Float>(v: T, alpha: f64) -> T {
     if v.is_sign_positive() {
         v
diff --git a/candle-core/tests/grad_tests.rs b/candle-core/tests/grad_tests.rs
index c44a7ea7..81b1338a 100644
--- a/candle-core/tests/grad_tests.rs
+++ b/candle-core/tests/grad_tests.rs
@@ -1,6 +1,5 @@
 use anyhow::{Context, Result};
-use candle_core::{Device, Shape, Tensor, Var};
-mod test_utils;
+use candle_core::{test_device, test_utils, Device, Shape, Tensor, Var};
 
 fn simple_grad(device: &Device) -> Result<()> {
     let x = Var::new(&[3f32, 1., 4.], device)?;
diff --git a/candle-core/tests/indexing_tests.rs b/candle-core/tests/indexing_tests.rs
index 1a77b32d..9c88f319 100644
--- a/candle-core/tests/indexing_tests.rs
+++ b/candle-core/tests/indexing_tests.rs
@@ -1,8 +1,6 @@
 use anyhow::Result;
 use candle_core::{Device, IndexOp, Tensor};
 
-mod test_utils;
-
 #[test]
 fn integer_index() -> Result<()> {
     let dev = Device::Cpu;
diff --git a/candle-core/tests/layout_tests.rs b/candle-core/tests/layout_tests.rs
index d7b89605..1b29476f 100644
--- a/candle-core/tests/layout_tests.rs
+++ b/candle-core/tests/layout_tests.rs
@@ -1,5 +1,4 @@
-mod test_utils;
-use candle::{Device, IndexOp, Result, Tensor};
+use candle::{test_device, Device, IndexOp, Result, Tensor};
 use candle_core as candle;
 
 fn contiguous(device: &Device) -> Result<()> {
diff --git a/candle-core/tests/pool_tests.rs b/candle-core/tests/pool_tests.rs
index bfdbc40d..b8c007b8 100644
--- a/candle-core/tests/pool_tests.rs
+++ b/candle-core/tests/pool_tests.rs
@@ -1,5 +1,4 @@
-mod test_utils;
-use candle_core::{Device, IndexOp, Result, Tensor};
+use candle_core::{test_device, test_utils, Device, IndexOp, Result, Tensor};
 
 // https://github.com/huggingface/candle/issues/364
 fn avg_pool2d(dev: &Device) -> Result<()> {
@@ -56,14 +55,17 @@ fn avg_pool2d_pytorch(dev: &Device) -> Result<()> {
     .reshape((1, 2, 4, 4))?;
     let pool = t.avg_pool2d((2, 2), (2, 2))?.squeeze(0)?;
     assert_eq!(
-        test_utils::to_vec3_round(pool, 4)?,
+        test_utils::to_vec3_round(&pool, 4)?,
         [
             [[-1.1926, -0.0395], [0.2688, 0.1871]],
             [[0.1835, -0.1606], [0.6249, 0.3217]]
         ]
     );
     let pool = t.avg_pool2d((3, 3), (3, 3))?.squeeze(0)?;
-    assert_eq!(test_utils::to_vec3_round(pool, 4)?, [[[0.085]], [[0.0078]]]);
+    assert_eq!(
+        test_utils::to_vec3_round(&pool, 4)?,
+        [[[0.085]], [[0.0078]]]
+    );
 
     let t = t.reshape((1, 1, 4, 8))?;
     let pool = t.avg_pool2d((2, 2), (2, 2))?.squeeze(0)?.squeeze(0)?;
diff --git a/candle-core/tests/quantized_tests.rs b/candle-core/tests/quantized_tests.rs
index d178b89b..4fd5e90c 100644
--- a/candle-core/tests/quantized_tests.rs
+++ b/candle-core/tests/quantized_tests.rs
@@ -1,11 +1,10 @@
 use candle_core::{
     quantized::{self, GgmlDType},
+    test_utils::to_vec2_round,
     Device, Result, Tensor,
 };
 use quantized::{k_quants, GgmlType};
-mod test_utils;
 use rand::prelude::*;
-use test_utils::to_vec2_round;
 
 const GGML_TEST_SIZE: usize = 32 * 128;
 
diff --git a/candle-core/tests/tensor_tests.rs b/candle-core/tests/tensor_tests.rs
index 1038efdd..6af43196 100644
--- a/candle-core/tests/tensor_tests.rs
+++ b/candle-core/tests/tensor_tests.rs
@@ -1,5 +1,4 @@
-mod test_utils;
-use candle_core::{DType, Device, IndexOp, Result, Tensor};
+use candle_core::{test_device, DType, Device, IndexOp, Result, Tensor};
 
 fn zeros(device: &Device) -> Result<()> {
     let tensor = Tensor::zeros((5, 2), DType::F32, device)?;
diff --git a/candle-nn/tests/batch_norm.rs b/candle-nn/tests/batch_norm.rs
index 7a3cfc18..209fc10a 100644
--- a/candle-nn/tests/batch_norm.rs
+++ b/candle-nn/tests/batch_norm.rs
@@ -4,10 +4,8 @@ extern crate intel_mkl_src;
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
 
-mod test_utils;
-
 use anyhow::Result;
-use candle::{DType, Device, Tensor};
+use candle::{test_utils, DType, Device, Tensor};
 use candle_nn::BatchNorm;
 
 /* The test below has been generated using the following PyTorch code:
diff --git a/candle-nn/tests/group_norm.rs b/candle-nn/tests/group_norm.rs
index eff66d17..8145a220 100644
--- a/candle-nn/tests/group_norm.rs
+++ b/candle-nn/tests/group_norm.rs
@@ -25,10 +25,9 @@ extern crate intel_mkl_src;
 extern crate accelerate_src;
 
 use anyhow::Result;
+use candle::test_utils::to_vec3_round;
 use candle::{Device, Tensor};
 use candle_nn::{GroupNorm, Module};
-mod test_utils;
-use test_utils::to_vec3_round;
 
 #[test]
 fn group_norm() -> Result<()> {
@@ -60,7 +59,7 @@ fn group_norm() -> Result<()> {
         device,
     )?;
     assert_eq!(
-        to_vec3_round(gn2.forward(&input)?, 4)?,
+        to_vec3_round(&gn2.forward(&input)?, 4)?,
         &[
             [
                 [-0.1653, 0.3748, -0.7866],
@@ -81,7 +80,7 @@ fn group_norm() -> Result<()> {
         ]
     );
     assert_eq!(
-        to_vec3_round(gn3.forward(&input)?, 4)?,
+        to_vec3_round(&gn3.forward(&input)?, 4)?,
         &[
             [
                 [0.4560, 1.4014, -0.6313],
diff --git a/candle-nn/tests/layer_norm.rs b/candle-nn/tests/layer_norm.rs
index 0f43d804..f81c29bd 100644
--- a/candle-nn/tests/layer_norm.rs
+++ b/candle-nn/tests/layer_norm.rs
@@ -5,11 +5,9 @@ extern crate intel_mkl_src;
 extern crate accelerate_src;
 
 use anyhow::Result;
-use candle::{Device, Tensor};
+use candle::{test_utils, Device, Tensor};
 use candle_nn::{LayerNorm, Module};
 
-mod test_utils;
-
 #[test]
 fn layer_norm() -> Result<()> {
     let device = &Device::Cpu;
@@ -28,7 +26,7 @@ fn layer_norm() -> Result<()> {
     let inp = Tensor::new(&[[[1f32, 2., 3.], [4., 5., 6.], [9., 8., 7.]]], device)?;
     let res = ln.forward(&inp)?;
     assert_eq!(
-        test_utils::to_vec3_round(res.clone(), 4)?,
+        test_utils::to_vec3_round(&res, 4)?,
         [[
             [-3.1742, 0.5, 4.1742],
             [-3.1742, 0.5, 4.1742],
@@ -41,7 +39,7 @@ fn layer_norm() -> Result<()> {
     let std = (res.broadcast_sub(&mean)?.sqr()?.sum_keepdim(2)?.sqrt()? / 3.0)?;
     // The standard deviation should be sqrt(`w`).
     assert_eq!(
-        test_utils::to_vec3_round(std, 4)?,
+        test_utils::to_vec3_round(&std, 4)?,
         [[[1.7321], [1.7321], [1.7321]]]
     );
     Ok(())
diff --git a/candle-nn/tests/loss.rs b/candle-nn/tests/loss.rs
index c075c7fb..d772f176 100644
--- a/candle-nn/tests/loss.rs
+++ b/candle-nn/tests/loss.rs
@@ -4,9 +4,8 @@ extern crate intel_mkl_src;
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
 
+use candle::test_utils::to_vec0_round;
 use candle::{Device, Result, Tensor};
-mod test_utils;
-use test_utils::to_vec0_round;
 
 /* Equivalent python code:
 import torch
diff --git a/candle-nn/tests/ops.rs b/candle-nn/tests/ops.rs
index fcf39fd8..4ba8cfcc 100644
--- a/candle-nn/tests/ops.rs
+++ b/candle-nn/tests/ops.rs
@@ -4,10 +4,7 @@ extern crate intel_mkl_src;
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
 
-mod test_utils;
-use test_utils::to_vec3_round;
-
-use candle::{Device, Result, Tensor};
+use candle::{test_utils::to_vec3_round, Device, Result, Tensor};
 
 #[test]
 fn softmax() -> Result<()> {
@@ -18,7 +15,7 @@ fn softmax() -> Result<()> {
     let t1 = candle_nn::ops::softmax(&tensor.log()?, 1)?;
     let t2 = candle_nn::ops::softmax(&tensor.log()?, 2)?;
     assert_eq!(
-        to_vec3_round(t0, 4)?,
+        to_vec3_round(&t0, 4)?,
         &[
             // 3/5, 1/2, 4/11
             [[0.6, 0.5, 0.3636], [0.1111, 0.7143, 0.5294]],
@@ -27,7 +24,7 @@ fn softmax() -> Result<()> {
         ]
     );
     assert_eq!(
-        to_vec3_round(t1, 4)?,
+        to_vec3_round(&t1, 4)?,
         &[
             // 3/4, 1/6, 4/13
             [[0.75, 0.1667, 0.3077], [0.25, 0.8333, 0.6923]],
@@ -36,7 +33,7 @@ fn softmax() -> Result<()> {
         ]
     );
     assert_eq!(
-        to_vec3_round(t2, 4)?,
+        to_vec3_round(&t2, 4)?,
         &[
             // (3, 1, 4) / 8, (1, 5, 9) / 15
             [[0.375, 0.125, 0.5], [0.0667, 0.3333, 0.6]],
diff --git a/candle-nn/tests/optim.rs b/candle-nn/tests/optim.rs
index f1d3b3f5..673d0455 100644
--- a/candle-nn/tests/optim.rs
+++ b/candle-nn/tests/optim.rs
@@ -4,8 +4,7 @@ extern crate intel_mkl_src;
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
 
-mod test_utils;
-use test_utils::{to_vec0_round, to_vec2_round};
+use candle::test_utils::{to_vec0_round, to_vec2_round};
 
 use anyhow::Result;
 use candle::{Device, Tensor, Var};
diff --git a/candle-nn/tests/test_utils.rs b/candle-nn/tests/test_utils.rs
deleted file mode 100644
index bb422cd9..00000000
--- a/candle-nn/tests/test_utils.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-#![allow(dead_code)]
-use candle::{Result, Tensor};
-
-pub fn to_vec0_round(t: &Tensor, digits: i32) -> Result<f32> {
-    let b = 10f32.powi(digits);
-    let t = t.to_vec0::<f32>()?;
-    Ok(f32::round(t * b) / b)
-}
-
-pub fn to_vec1_round(t: &Tensor, digits: i32) -> Result<Vec<f32>> {
-    let b = 10f32.powi(digits);
-    let t = t.to_vec1::<f32>()?;
-    let t = t.iter().map(|t| f32::round(t * b) / b).collect();
-    Ok(t)
-}
-
-pub fn to_vec2_round(t: &Tensor, digits: i32) -> Result<Vec<Vec<f32>>> {
-    let b = 10f32.powi(digits);
-    let t = t.to_vec2::<f32>()?;
-    let t = t
-        .iter()
-        .map(|t| t.iter().map(|t| f32::round(t * b) / b).collect())
-        .collect();
-    Ok(t)
-}
-
-pub fn to_vec3_round(t: Tensor, digits: i32) -> Result<Vec<Vec<Vec<f32>>>> {
-    let b = 10f32.powi(digits);
-    let t = t.to_vec3::<f32>()?;
-    let t = t
-        .iter()
-        .map(|t| {
-            t.iter()
-                .map(|t| t.iter().map(|t| f32::round(t * b) / b).collect())
-                .collect()
-        })
-        .collect();
-    Ok(t)
-}