Executing Ray Train with PyTorch

Ray Libraries (Data, Train, Tune, Serve) Ray Train
1

How severe does this issue affect your experience of using Ray?

  • High: It blocks me to complete my task.

I am trying to migrate my code to use Ray Train, etc. However, when tried to execute the FASHION MNIST Example from Ray Docs,
I get following error:


RuntimeError: use_libuv was requested but PyTorch was build without libuv support

The error is stemming from:

site-packages\torch\distributed\rendezvous.py", line 189, in _create_c10d_store
return TCPStore(
^^^^^^^^^
RuntimeError: use_libuv was requested but PyTorch was build without libuv support

Because I’m running it on a Windows CPU machine at the moment, so I switched the GPU flag to False. Following are my library versions with Python 3.11.0:

PyTorch version: 2.5.1+cpu
Ray version: 2.40.0

I have tried several fixes but those include setting the libuv flag to False which I can’t as I’m using the that from within the Ray.

2

Hi.

This issue seems to be an issue with your PyTorch installation. In PyTorch 2.4, libuv was made the default backend for TCPStore initialization: Introduction to Libuv TCPStore Backend — PyTorch Tutorials 2.5.0+cu124 documentation

I’m not too sure of the right way to build on Windows with libuv support, and there even seems to be an open issue for the same problem: PyTorch defaults to using libuv but is built without support for it on Windows · Issue #139990 · pytorch/pytorch · GitHub

As an alternative, you can use the older backend by setting USE_LIBUV=0 in your environment. Make sure to add this environment variable at ray initialization with runtime_env

For the Fashion MNIST example, you can do the following:

if __name__ == "__main__":
+  ray.init(runtime_env={"env_vars": {"USE_LIBUV": "0"}}) 
    train_fashion_mnist(num_workers=4, use_gpu=True)

This way, all the Ray Train workers will have this environment variable set.

1 Like
3

Really appreciate your help!