Add before suite hook to set mp start method.

The default multiprocessing start method is "fork" which is not compatible with
with runtime assertions that it is set to spawn. When running unit tests, it's
possible to call an external library that sets the start method to "fork".
Here we enforce the start method to be "spawn" for all tests before executing.

tests/python/conftest.py

@@ -0,0 +1,14 @@
+import multiprocessing
+import pytest
+
+
+# The default multiprocessing start method is "fork" which is not compatible with
+# with runtime assertions that it is set to spawn. When running unit tests, it's
+# possible to call an external library that sets the start method to "fork".
+# Here we enforce the start method to be "spawn" for all tests before executing.
+@pytest.fixture(scope="session", autouse=True)
+def initialize_data_environment():
+    try:
+        multiprocessing.set_start_method("spawn")
+    except Exception:
+        pass

tests/python/test_code_composition.py

@@ -1,4 +1,3 @@
-import multiprocessing
 from unittest import TestCase
 
 from dolma.core.data_types import Document
@@ -40,11 +39,6 @@ def baz():
 
 class TestDolmaCodeProseCompositionClassifier(TestCase):
     def setUp(self) -> None:
-        try:
-            multiprocessing.set_start_method("spawn")
-        except Exception:
-            pass
-
         self.code_composition_tagger = CodeProseCompositionClassifier()
 
     def test_prose_text(self):