Improve the process cleanup logic when running tests

tests/conftest.py

@@ -1,6 +1,5 @@
 import asyncio
 import gc
-from contextlib import suppress
 
 import psutil
 import pytest
@@ -40,13 +39,12 @@ def cleanup_children():
 
     children = psutil.Process().children(recursive=True)
     if children:
-        logger.info(f"Cleaning up {len(children)} leftover child processes")
-        for child in children:
-            with suppress(psutil.NoSuchProcess):
-                child.terminate()
-        psutil.wait_procs(children, timeout=1)
-        for child in children:
-            with suppress(psutil.NoSuchProcess):
-                child.kill()
+        gone, alive = psutil.wait_procs(children, timeout=0.1)
+        logger.debug(f"Cleaning up {len(alive)} leftover child processes")
+        for child in alive:
+            child.terminate()
+        gone, alive = psutil.wait_procs(alive, timeout=1)
+        for child in alive:
+            child.kill()
 
     MPFuture.reset_backend()

tests/test_allreduce_fault_tolerance.py

@@ -209,3 +209,4 @@ def _make_tensors():
 
     for averager in averagers:
         averager.shutdown()
+    dht.shutdown()

tests/test_dht_experts.py

@@ -82,6 +82,10 @@ def test_beam_search(
         assert all(isinstance(e, hivemind.RemoteExpert) for experts in batch_experts for e in experts)
         assert all(len(experts) == beam_size for experts in batch_experts)
 
+    you.shutdown()
+    for dht_instance in dht_instances:
+        dht_instance.shutdown()
+
 
 @pytest.mark.forked
 def test_dht_single_node():
@@ -119,6 +123,8 @@ def test_dht_single_node():
     with pytest.raises(AssertionError):
         beam_search.get_active_successors(["e.1.2.", "e.2", "e.4.5."])
 
+    node.shutdown()
+
 
 def test_uid_patterns():
     valid_experts = [