[fix] ShardedDDP - cpu testfix - remove Gloo/CPU (#350)

* no idea about the root issue, but it proved to be fairly narrowed (gloo+cpu+python3.8+no cuda installed) so I guess that's out of scope for fairscale

fairscale/utils/testing.py

@@ -31,6 +31,7 @@ import logging
 import multiprocessing
 import os
 import random
+import sys
 import tempfile
 from typing import Any, Callable, Dict, List, Optional, Tuple
 
@@ -53,6 +54,10 @@ skip_if_single_gpu = pytest.mark.skipif(
     not torch.cuda.is_available() or torch.cuda.device_count() < 2, reason="multiple GPUs required"
 )
 
+skip_if_py38 = pytest.mark.skipif(
+    sys.version_info.major == 3 and sys.version_info.minor == 8, reason="Python3.8 is skipped"
+)
+
 _, filename_mpi = tempfile.mkstemp()
 
 

tests/nn/data_parallel/test_sharded_ddp.py

@@ -21,7 +21,7 @@ from torch.nn.parallel import DistributedDataParallel as DDP
 
 from fairscale.nn.data_parallel import ShardedDataParallel
 from fairscale.optim import OSS
-from fairscale.utils.testing import GPT2, skip_if_no_cuda, skip_if_single_gpu
+from fairscale.utils.testing import GPT2, skip_if_no_cuda, skip_if_py38, skip_if_single_gpu
 
 
 def run_one_step(rank, world_size, backend, device, temp_file_name):
@@ -112,16 +112,17 @@ def run_test(backend, device, world_size=2):
     mp.spawn(run_one_step, args=(world_size, backend, device, temp_file_name), nprocs=world_size, join=True)
 
 
-def test_step_on_cpu():
-    run_test(backend=dist.Backend.GLOO, device=torch.device("cpu"), world_size=4)
-
-
 @skip_if_no_cuda
 @skip_if_single_gpu
-def test_step_on_gpu():
+def test_step_gpu():
     run_test(backend=dist.Backend.NCCL, device=torch.device("cuda"))
 
 
+@skip_if_py38
+def test_step_cpu():
+    run_test(backend=dist.Backend.GLOO, device=torch.device("cpu"))
+
+
 def run_ddp_parity(rank, world_size, backend, temp_file_name):
     url = "file://" + temp_file_name
     dist.init_process_group(init_method=url, backend=backend, rank=rank, world_size=world_size)