Add option to skip certain unit tests (#2878)

Summary:
Pull Request resolved: #2878

Add option to skip certain tests - useful when a few tests are broken and we need to skip them to gain visibility in other unit tests. This usually happens on our CPU unit tests, so only modifying this script.

To use, simply add the name of the tests to skip in this txt file. such as:

```
test_sharding_fused_ebc_as_top_level
```

You can also use the class name for the test: e.g.
```
ModelParallelSparseOnlyTestGloo
```

Differential Revision: D72815908

Author: aporialiao

.github/scripts/tests_to_skip.txt

@@ -0,0 +1 @@
+_disabled_in_oss_compatibility

.github/workflows/unittest_ci_cpu.yml

@@ -72,6 +72,14 @@ jobs:
           python -c "import numpy"
         echo "numpy succeeded"
         conda install -n build_binary -y pytest
+        # Read the list of tests to skip from a file, ignoring empty lines and comments
+        skip_expression=$(awk '!/^($|#)/ {printf " and not %s", $0}' ./.github/scripts/tests_to_skip.txt)
+        # Check if skip_expression is effectively empty
+        if [ -z "$skip_expression" ]; then
+          skip_expression=""
+        else
+          skip_expression=${skip_expression:5}  # Remove the leading " and "
+        fi
         conda run -n build_binary \
           python -m pytest torchrec -v -s -W ignore::pytest.PytestCollectionWarning --continue-on-collection-errors \
-          --ignore-glob=**/test_utils/
+          --ignore-glob=**/test_utils/ -k "$skip_expression"

torchrec/ir/tests/test_serializer.py

@@ -253,13 +253,7 @@ def test_serialize_deserialize_ebc(self) -> None:
             self.assertEqual(deserialized.shape, orginal.shape)
             self.assertTrue(torch.allclose(deserialized, orginal))
 
-    # pyre-ignore[56]: Pyre was not able to infer the type of argument
-    @unittest.skipIf(
-        torch.cuda.device_count() == 0,
-        "skip this test in OSS (no GPU available) because torch.export uses training ir in OSS",
-    )
-    def test_dynamic_shape_ebc(self) -> None:
-        # TODO: https://fb.workplace.com/groups/1028545332188949/permalink/1138699244506890/
+    def test_dynamic_shape_ebc_disabled_in_oss_compatibility(self) -> None:
         model = self.generate_model()
         feature1 = KeyedJaggedTensor.from_offsets_sync(
             keys=["f1", "f2", "f3"],

torchrec/models/experimental/test_transformerdlrm.py

@@ -61,12 +61,7 @@ def test_larger(self) -> None:
         concat_dense = inter_arch(dense_features, sparse_features)
         self.assertEqual(concat_dense.size(), (B, D * (F + 1)))
 
-    # pyre-ignore[56]: Pyre was not able to infer the type of argument
-    @unittest.skipIf(
-        torch.cuda.device_count() == 0,
-        "skip this test in OSS (no GPU available) because seed might be different in OSS",
-    )
-    def test_correctness(self) -> None:
+    def test_correctness_disabled_in_oss_compatibility(self) -> None:
         D = 4
         B = 3
         # multi-head attentions
@@ -170,12 +165,7 @@ def test_correctness(self) -> None:
             )
         )
 
-    # pyre-ignore[56]: Pyre was not able to infer the type of argument
-    @unittest.skipIf(
-        torch.cuda.device_count() == 0,
-        "skip this test in OSS (no GPU available) because seed might be different in OSS",
-    )
-    def test_numerical_stability(self) -> None:
+    def test_numerical_stability_disabled_in_oss_compatibility(self) -> None:
         D = 4
         B = 3
         # multi-head attentions
@@ -204,12 +194,7 @@ def test_numerical_stability(self) -> None:
 
 
 class DLRMTransformerTest(unittest.TestCase):
-    # pyre-ignore[56]: Pyre was not able to infer the type of argument
-    @unittest.skipIf(
-        torch.cuda.device_count() == 0,
-        "skip this test in OSS (no GPU available) because seed might be different in OSS",
-    )
-    def test_basic(self) -> None:
+    def test_basic_disabled_in_oss_compatibility(self) -> None:
         torch.manual_seed(0)
         B = 2
         D = 8