CPU-only execution fails with AssertionError: Torch not compiled with CUDA enabled

Bug: CPU-only execution fails with AssertionError: Torch not compiled with CUDA enabled

Description

I am trying to run the benchmark suite on a machine with a CPU-only PyTorch installation. The execution fails with:

AssertionError: Torch not compiled with CUDA enabled

From my investigation, at some point the code sets the device to cuda:0, which eventually triggers the error. However, I could not identify where this device selection happens.

What I checked

• I reviewed the documented CLI arguments and could not find any option to force CPU execution.
• I searched through the configuration and entry points but the codebase is relatively large and the device-selection logic is difficult to follow.
• I added some debug prints and confirmed that cuda:0 is being selected internally before the failure occurs.

Expected behavior

If CUDA is unavailable (or if PyTorch was compiled without CUDA support), the benchmark should either:

1. Automatically fall back to CPU execution, or
2. Expose a documented option/configuration parameter to explicitly select cpu, or
3. Fail early with a clear message explaining that CUDA is currently required.

A common pattern would be:

device = "cuda" if torch.cuda.is_available() else "cpu"

instead of assuming CUDA is always available.

Environment

• OS: macos
• conda env as specified in the readme

Question

Is CPU-only execution currently supported?

If yes, could you point me to the correct way to force CPU execution?

If not, would you consider adding a CPU fallback or a documented device-selection option?

Start by tracing the documented CLI arguments, configuration, and benchmark entry points to find where the device is assigned, then reproduce the failure with a CPU-only PyTorch installation. Done means CPU execution works or device selection is documented, or the benchmark fails early with a clear CUDA requirement.