fix: init_process_group_dummy on non-gpu devices (regression)

dmlcloud/util/distributed.py

@@ -1,6 +1,7 @@
 import os
 from contextlib import contextmanager
 
+import torch
 import torch.distributed as dist
 
 from .tcp import find_free_port, get_local_ips
@@ -87,7 +88,7 @@ def init_process_group_dummy(**kwargs):
     """
     backend = kwargs.get('backend', None)
     if backend is None:
-        backend = 'cpu:gloo,cuda:nccl' if dist.is_nccl_available() else 'gloo'
+        backend = 'cpu:gloo,cuda:nccl' if dist.is_nccl_available() and torch.cuda.is_available() else 'gloo'
     store = dist.HashStore()
     dist.init_process_group(store=store, rank=0, world_size=1, backend=backend, **kwargs)