make AveragedModel OSS compatible

Summary: Certain imports from pytorch AveragedModel are not in torch version 2.0.0, so this diff guards against potential import errors at runtime

Reviewed By: galrotem

Differential Revision: D56534614

fbshipit-source-id: fe6a5b56eabf51c101fa0ebc0f1eb2870df10c97

tests/framework/test_auto_unit.py

@@ -41,6 +41,7 @@
 from torchtnt.utils.lr_scheduler import TLRScheduler
 from torchtnt.utils.prepare_module import DDPStrategy
 from torchtnt.utils.progress import Progress
+from torchtnt.utils.swa import _AVERAGED_MODEL_AVAIL
 from torchtnt.utils.test_utils import skip_if_not_distributed
 from torchtnt.utils.timer import Timer
 
@@ -149,6 +150,9 @@ def test_predict_step(self) -> None:
             predict(auto_unit, pred_dataloader, max_steps_per_epoch=1)
             mock_predict_step_end.assert_called_once()
 
+    @unittest.skipUnless(
+        _AVERAGED_MODEL_AVAIL, "AveragedModel needed in version of Pytorch"
+    )
     def test_stochastic_weight_averaging_basic(self) -> None:
         """
         Basic stochastic weight averaging tests
@@ -182,6 +186,9 @@ def test_stochastic_weight_averaging_basic(self) -> None:
         self.assertIn("swa_scheduler", auto_unit2.app_state())
         self.assertIn("swa_scheduler", auto_unit2.tracked_lr_schedulers())
 
+    @unittest.skipUnless(
+        _AVERAGED_MODEL_AVAIL, "AveragedModel needed in version of Pytorch"
+    )
     def test_stochastic_weight_averaging_update_freq(self) -> None:
         """
         e2e stochastic weight averaging test to ensure that the SWA model is updated at the correct frequency
@@ -295,11 +302,12 @@ def test_auto_unit_ddp(self) -> None:
         Launch tests of AutoUnit with DDP strategy
         """
 
-        spawn_multi_process(
-            2,
-            "gloo",
-            self._test_stochastic_weight_averaging_with_ddp,
-        )
+        if _AVERAGED_MODEL_AVAIL:
+            spawn_multi_process(
+                2,
+                "gloo",
+                self._test_stochastic_weight_averaging_with_ddp,
+            )
         spawn_multi_process(
             2,
             "gloo",

tests/utils/test_swa.py

@@ -13,7 +13,10 @@
 
 import torch
 
-from torchtnt.utils.swa import AveragedModel
+from torchtnt.utils.swa import _AVERAGED_MODEL_AVAIL, AveragedModel
+
+if not _AVERAGED_MODEL_AVAIL:
+    raise unittest.SkipTest("Latest Pytorch is required to run SWA tests")
 
 
 class TestSWA(unittest.TestCase):

torchtnt/utils/swa.py

@@ -10,11 +10,17 @@
 
 import torch
 
-from torch.optim.swa_utils import (
-    AveragedModel as PyTorchAveragedModel,
-    get_ema_multi_avg_fn,
-    get_swa_multi_avg_fn,
-)
+_AVERAGED_MODEL_AVAIL: bool = True
+
+try:
+    from torch.optim.swa_utils import (
+        AveragedModel as PyTorchAveragedModel,
+        get_ema_multi_avg_fn,
+        get_swa_multi_avg_fn,
+    )
+except ImportError:
+    _AVERAGED_MODEL_AVAIL = False
+
 
 TSWA_avg_fn = Callable[[torch.Tensor, torch.Tensor, int], torch.Tensor]
 TSWA_multi_avg_fn = Callable[[List[torch.Tensor], List[torch.Tensor], int], None]
@@ -49,6 +55,12 @@ def __init__(
                 number of updates. The EMA decay will start small and will approach the
                 specified ema_decay as more updates occur.
         """
+        if not _AVERAGED_MODEL_AVAIL:
+            raise ImportError(
+                "AveragedModel is not available in this version of PyTorch. \
+                Please install the latest version of PyTorch."
+            )
+
         # setup averaging method
         if averaging_method == "ema":
             if ema_decay < 0.0 or ema_decay > 1.0: