Fix get_device_properties crash when CUDA is installed but no GPU

[Jah-yee]

Good day

Problem

In src/transformers/testing_utils.py, the get_device_properties() function checks IS_CUDA_SYSTEM to determine whether to call torch.cuda.get_device_capability(). However, IS_CUDA_SYSTEM is set to True when torch.version.cuda is not None, regardless of whether a GPU is actually available.

When CUDA is installed (e.g., CUDA toolkit on a headless cloud instance with no GPU), but no GPU device is present, torch.cuda.get_device_capability() raises a RuntimeError, causing the function to crash.

Solution

Added a guard check for torch.cuda.is_available() before calling torch.cuda.get_device_capability():

# Before
if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM:

# After
if (IS_CUDA_SYSTEM or IS_ROCM_SYSTEM) and torch.cuda.is_available():

When CUDA is installed but no GPU is available, the function will now fall through to the else branch and return (torch_device, None, None) instead of crashing.

Testing

• The fix is a minimal, targeted change — only one line modified.
• The reproduction is straightforward: run get_device_properties() on a CUDA-installed but GPU-less environment (e.g., cloud CPU-only instance with CUDA toolkit).
• The existing test test_get_device_properties should continue to pass on GPU-enabled machines.

Fixes #45341

---

Thank you for your attention. If there are any issues or suggestions, please leave a comment and I will address them promptly.

Warmly,
RoomWithOutRoof

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=45509&sha=92806c