Try forking before asyncio usage

Author: borzunov

tests/test_allreduce.py

@@ -13,8 +13,8 @@
 from hivemind.p2p import P2P
 
 
-@pytest.mark.forked
 @pytest.mark.asyncio
+@pytest.mark.forked
 async def test_partitioning():
     all_tensors = [
         torch.randn(30_000, 128),
@@ -61,8 +61,8 @@ async def write_tensors():
     ],
 )
 @pytest.mark.parametrize("peer_fractions", [(0.33, 0.44, 0.23), (0.5, 0.5), (0.1, 0.0, 0.9), (1.0,), (0.1,) * 9])
-@pytest.mark.forked
 @pytest.mark.asyncio
+@pytest.mark.forked
 async def test_partitioning_edge_cases(tensors: Sequence[torch.Tensor], peer_fractions: Sequence[float]):
     partition = TensorPartContainer(tensors, peer_fractions, part_size_bytes=16)
     for peer_index in range(len(peer_fractions)):
@@ -75,8 +75,8 @@ async def test_partitioning_edge_cases(tensors: Sequence[torch.Tensor], peer_fra
         tensor_index += 1
 
 
-@pytest.mark.forked
 @pytest.mark.asyncio
+@pytest.mark.forked
 async def test_partitioning_asynchronous():
     """ensure that tensor partitioning does not interfere with asynchronous code"""
     tensors = [torch.randn(2048, 2048), torch.randn(1024, 4096), torch.randn(4096, 1024), torch.randn(30_000, 1024)]
@@ -114,8 +114,8 @@ async def wait_synchronously():
 @pytest.mark.parametrize("num_senders", [1, 2, 4, 10])
 @pytest.mark.parametrize("num_parts", [0, 1, 100])
 @pytest.mark.parametrize("synchronize_prob", [1.0, 0.1, 0.0])
-@pytest.mark.forked
 @pytest.mark.asyncio
+@pytest.mark.forked
 async def test_reducer(num_senders: int, num_parts: int, synchronize_prob: float):
     tensor_part_shapes = [torch.Size([i]) for i in range(num_parts)]
     reducer = TensorPartReducer(tensor_part_shapes, num_senders)
@@ -168,8 +168,8 @@ async def send_tensors(sender_index: int):
         ((AUX, AUX, AUX, AUX), (0.0, 0.0, 0.0, 0.0), (1, 2, 3, 4), 2**20),
     ],
 )
-@pytest.mark.forked
 @pytest.mark.asyncio
+@pytest.mark.forked
 async def test_allreduce_protocol(peer_modes, averaging_weights, peer_fractions, part_size_bytes):
     """Run group allreduce protocol manually without grpc, see if the internal logic is working as intended"""
 

tests/test_averaging.py

@@ -17,8 +17,8 @@
 from test_utils.dht_swarms import launch_dht_instances
 
 
-@pytest.mark.forked
 @pytest.mark.asyncio
+@pytest.mark.forked
 async def test_key_manager():
     dht = hivemind.DHT(start=True)
     key_manager = GroupKeyManager(

tests/test_dht.py

@@ -104,8 +104,8 @@ def test_run_coroutine():
     dht.shutdown()
 
 
-@pytest.mark.forked
 @pytest.mark.asyncio
+@pytest.mark.forked
 async def test_dht_get_visible_maddrs():
     # test 1: IPv4 localhost multiaddr is visible by default