tensorflow is not built with CUDA

**Describe the bug**
Even with GPU instance, tensorflow ( from sagemaker-distribution conda environment ) can't recognize the GPU.
`nvidia-smi` show I'm using GPU instance, and pytorch can recognize the GPU.

**To Reproduce**
Steps to reproduce the behavior:
1. Start GPU runtime.
2. launch jupyter notebook ( with sagemaker-distribution kernel )
3. type
```python
import tensorflow as tf
import torch

tf.config.list_physical_devices('GPU') # []
tf.test.is_built_with_cuda() # False
torch.cuda.is_available() # True
```
4. See error
TensorFlow can't recognize GPU, and is not even built with CUDA.

**Expected behavior**
```python
tf.config.list_physical_devices('GPU') # at least one GPU should be in the list
tf.test.is_built_with_cuda() # True
```

**Screenshots**
If applicable, add screenshots to help explain your problem.

**Desktop (please complete the following information):**
- OS: windows 11
- Browser : chrome
- Version : 126.0.6478.127 (Official Build) (64-bit)

The issue names the sagemaker-distribution conda environment and a Jupyter notebook, but no repository files or tests. Start by reproducing the Python checks in a GPU runtime, then inspect how TensorFlow is installed and built in that kernel. Done means tf.test.is_built_with_cuda() returns True and tf.config.list_physical_devices('GPU') lists at least one GPU.

Written by the indexing model from the issue text.