Comment

src/accelerate/test_utils/scripts/test_notebook.py

@@ -0,0 +1,10 @@
+# Test file to ensure that in general certain situational setups for notebooks work.
+
+import torch
+
+
+assert not torch.cuda.is_initialized(), "CUDA was initialized before the test script."
+
+from accelerate import Accelerator  # noqa
+
+assert not torch.cuda.is_initialized(), "CUDA was initialized upon importing the `Accelerator` class."

tests/test_accelerator.py

@@ -1,3 +1,4 @@
+import inspect
 import json
 import os
 import tempfile
@@ -6,11 +7,12 @@ from unittest.mock import patch
 import torch
 from torch.utils.data import DataLoader, TensorDataset
 
+import accelerate
 from accelerate import DistributedType, infer_auto_device_map, init_empty_weights
 from accelerate.accelerator import Accelerator
 from accelerate.state import GradientState, PartialState
 from accelerate.test_utils import require_bnb, require_multi_gpu, slow
-from accelerate.test_utils.testing import AccelerateTestCase, require_cuda
+from accelerate.test_utils.testing import AccelerateTestCase, execute_subprocess_async, require_cuda
 from accelerate.utils import patch_environment
 
 
@@ -331,10 +333,9 @@ class AcceleratorTester(AccelerateTestCase):
         _ = accelerator.prepare(sgd)
 
     @require_cuda
-    def test_a_cuda_initialization_on_import(self):
-        # Everything else is already initialized, now we just need to check that the cuda device is *not* initialized
-        initialized = torch.cuda.is_initialized()
-        self.assertFalse(
-            initialized,
-            "CUDA has been initialized somewhere after importing. Please see the imports in `test_accelerator.py` to try and locate the problem.",
-        )
+    def test_cuda_initialization_on_import(self):
+        mod_file = inspect.getfile(accelerate.test_utils)
+        script = os.path.sep.join(mod_file.split(os.path.sep)[:-1] + ["scripts", "test_notebook.py"])
+        cmd = ["accelerate", "launch", script]
+        with patch_environment(omp_num_threads=1):
+            execute_subprocess_async(cmd, env=os.environ.copy())