I triggered gpu tests to confirm this is the issue

Ah, thanks for catching.

FYI, @hipudding

I’m very sorry for the trouble caused by my previous submission. I originally thought that passing CI would ensure all necessary test cases were executed. In future development, I will make sure to run all relevant test cases locally beforehand to prevent similar issues from being merged upstream.

The issue occurred in the deserialize_from_numpy_or_scalar function. Even when with_tensor_transport explicitly specifies CUDA, the function still automatically selects the currently available accelerator. In this particular scenario, since no GPU was available, it fell back to using a CPU tensor. I have fixed this issue locally and confirmed that the test_torch_tensor_transport_gpu test case now passes.

I don’t have the necessary permissions to view the postmerge results. May I ask if there are any other test cases that have failed?

What should I do next? Should I create a new PR after fix this issue? @stephanie-wang

Hi @hipudding , no worries, thanks for following up on this!

You can create a new PR which contains your original changes and the new fix. The GPU tests need to triggered manually in CI, and we can help you with that. But you can also test them locally, e.g., test_torch_tensor_dag.py, test_torch_tensor_transport. Set RAY_PYTEST_USE_GPU to true when you test.