[fix] Fix test_nucleus_sampling after pytorch multinomial fix (#684)

Summary:
Pull Request resolved: #684

Fix test_nucleus_sampling after pytorch multinomial fix

Reviewed By: vedanuj

Differential Revision: D24997596

fbshipit-source-id: 27b98d1289a36d151abf60e3a2ad0fb150da4139

Author: xiaomengy

tests/utils/test_text.py

@@ -10,6 +10,7 @@
 from mmf.utils.configuration import Configuration
 from mmf.utils.env import setup_imports
 from mmf.utils.general import get_mmf_root
+from packaging.version import LegacyVersion
 from tests.test_utils import dummy_args
 from tests.utils.test_model import TestDecoderModel
 
@@ -152,19 +153,26 @@ def test_nucleus_sampling(self):
         tokens = model(sample_list)["captions"]
 
         # these are expected tokens for sum_threshold = 0.5
-        expected_tokens = [
-            1.0,
-            29.0,
-            11.0,
-            11.0,
-            39.0,
-            10.0,
-            31.0,
-            4.0,
-            19.0,
-            39.0,
-            2.0,
-        ]
+
+        # Because of a bug fix in https://github.com/pytorch/pytorch/pull/47386
+        # the torch.Tensor.multinomail will generate different random sequence.
+        # TODO: Remove this hack after OSS uses later version of PyTorch.
+        if LegacyVersion(torch.__version__) > LegacyVersion("1.7.1"):
+            expected_tokens = [1.0, 23.0, 38.0, 30.0, 5.0, 11.0, 2.0]
+        else:
+            expected_tokens = [
+                1.0,
+                29.0,
+                11.0,
+                11.0,
+                39.0,
+                10.0,
+                31.0,
+                4.0,
+                19.0,
+                39.0,
+                2.0,
+            ]
 
         self.assertEqual(tokens[0].tolist(), expected_tokens)