Include stride_per_key_per_rank in KJT's PyTree flatten/unflatten logic (#2903)

Summary:
Pull Request resolved: #2903

# Context
* Currently torchrec IR serializer can't handle variable batch use case.
* `torch.export` only captures the keys, values, lengths, weights, offsets of a KJT, however, some variable-batch related parameters like `stride_per_rank` or `inverse_indices` would be ignored.
* This test case (expected failure right now) covers the vb-KJT scenario for verifying that the serialize_deserialize_ebc use case works fine with KJTs with variable batch size.

# Ref

Differential Revision: D73051959

Author: jd7-tr

torchrec/ir/tests/test_serializer.py

@@ -206,8 +206,14 @@ def forward(
             num_embeddings=10,
             feature_names=["f2"],
         )
+        config3 = EmbeddingBagConfig(
+            name="t3",
+            embedding_dim=5,
+            num_embeddings=10,
+            feature_names=["f3"],
+        )
         ebc = EmbeddingBagCollection(
-            tables=[config1, config2],
+            tables=[config1, config2, config3],
             is_weighted=False,
         )
 
@@ -292,15 +298,17 @@ def test_serialize_deserialize_ebc(self) -> None:
             self.assertEqual(deserialized.shape, orginal.shape)
             self.assertTrue(torch.allclose(deserialized, orginal))
 
-    @unittest.skip("Adding test for demonstrating VBE KJT flattening issue for now.")
     def test_serialize_deserialize_ebc_with_vbe_kjt(self) -> None:
         model = self.generate_model_for_vbe_kjt()
         id_list_features = KeyedJaggedTensor(
-            keys=["f1", "f2"],
-            values=torch.tensor([5, 6, 7, 1, 2, 3, 0, 1]),
-            lengths=torch.tensor([3, 3, 2]),
-            stride_per_key_per_rank=[[2], [1]],
-            inverse_indices=(["f1", "f2"], torch.tensor([[0, 1, 0], [0, 0, 0]])),
+            keys=["f1", "f2", "f3"],
+            values=torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 0]),
+            lengths=torch.tensor([1, 2, 3, 2, 1, 1]),
+            stride_per_key_per_rank=[[3], [2], [1]],
+            inverse_indices=(
+                ["f1", "f2", "f3"],
+                torch.tensor([[0, 1, 2], [0, 1, 0], [0, 0, 0]]),
+            ),
         )
 
         eager_out = model(id_list_features)
@@ -319,15 +327,16 @@ def test_serialize_deserialize_ebc_with_vbe_kjt(self) -> None:
         # Run forward on ExportedProgram
         ep_output = ep.module()(id_list_features)
 
+        self.assertEqual(len(ep_output), len(id_list_features.keys()))
         for i, tensor in enumerate(ep_output):
-            self.assertEqual(eager_out[i].shape, tensor.shape)
+            self.assertEqual(eager_out[i].shape[1], tensor.shape[1])
 
         # Deserialize EBC
         unflatten_ep = torch.export.unflatten(ep)
         deserialized_model = decapsulate_ir_modules(unflatten_ep, JsonSerializer)
 
         # check EBC config
-        for i in range(5):
+        for i in range(1):
             ebc_name = f"ebc{i + 1}"
             self.assertIsInstance(
                 getattr(deserialized_model, ebc_name), EmbeddingBagCollection
@@ -342,29 +351,9 @@ def test_serialize_deserialize_ebc_with_vbe_kjt(self) -> None:
                 self.assertEqual(deserialized.num_embeddings, orginal.num_embeddings)
                 self.assertEqual(deserialized.feature_names, orginal.feature_names)
 
-        # check FPEBC config
-        for i in range(2):
-            fpebc_name = f"fpebc{i + 1}"
-            assert isinstance(
-                getattr(deserialized_model, fpebc_name),
-                FeatureProcessedEmbeddingBagCollection,
-            )
-
-            for deserialized, orginal in zip(
-                getattr(
-                    deserialized_model, fpebc_name
-                )._embedding_bag_collection.embedding_bag_configs(),
-                getattr(
-                    model, fpebc_name
-                )._embedding_bag_collection.embedding_bag_configs(),
-            ):
-                self.assertEqual(deserialized.name, orginal.name)
-                self.assertEqual(deserialized.embedding_dim, orginal.embedding_dim)
-                self.assertEqual(deserialized.num_embeddings, orginal.num_embeddings)
-                self.assertEqual(deserialized.feature_names, orginal.feature_names)
-
         # Run forward on deserialized model and compare the output
         deserialized_model.load_state_dict(model.state_dict())
+
         deserialized_out = deserialized_model(id_list_features)
 
         self.assertEqual(len(deserialized_out), len(eager_out))
@@ -385,6 +374,7 @@ def test_dynamic_shape_ebc_disabled_in_oss_compatibility(self) -> None:
             values=torch.tensor([0, 1, 2, 3, 2, 3, 4]),
             offsets=torch.tensor([0, 2, 2, 3, 4, 5, 7]),
         )
+
         eager_out = model(feature2)
 
         # Serialize EBC

torchrec/modules/embedding_modules.py

@@ -27,7 +27,7 @@ def reorder_inverse_indices(
     inverse_indices: Optional[Tuple[List[str], torch.Tensor]],
     feature_names: List[str],
 ) -> torch.Tensor:
-    if inverse_indices is None:
+    if inverse_indices is None or inverse_indices[1].numel() == 0:
         return torch.empty(0)
     index_per_name = {name: i for i, name in enumerate(inverse_indices[0])}
     index = torch.tensor(

torchrec/sparse/jagged_tensor.py

@@ -1756,6 +1756,7 @@ class KeyedJaggedTensor(Pipelineable, metaclass=JaggedTensorMeta):
         "_weights",
         "_lengths",
         "_offsets",
+        "_inverse_indices",
     ]
 
     def __init__(
@@ -1800,7 +1801,6 @@ def __init__(
         self._inverse_indices: Optional[Tuple[List[str], torch.Tensor]] = (
             inverse_indices
         )
-
         # legacy attribute, for backward compatabilibity
         self._variable_stride_per_key: Optional[bool] = None
 
@@ -3032,13 +3032,36 @@ def dist_init(
 
 def _kjt_flatten(
     t: KeyedJaggedTensor,
-) -> Tuple[List[Optional[torch.Tensor]], List[str]]:
-    return [getattr(t, a) for a in KeyedJaggedTensor._fields], t._keys
+) -> Tuple[
+    List[Optional[torch.Tensor]],
+    Tuple[List[str], Optional[List[str]], Optional[List[List[int]]]],
+]:
+    field_values = [getattr(t, a) for a in KeyedJaggedTensor._fields[:-1]]
+    inverse_index_keys: Optional[List[str]] = None
+    # Init to an empty tensor so it will be exported as FakeTensor by torch.export.
+    # Otherwise it will be exported as a ConstantArgument when the KJT's _inverse_indices None, which causes Unsupported data type error.
+    inverse_indices: torch.Tensor = torch.empty(0)
+
+    if t._inverse_indices is not None:
+        inverse_indices = t._inverse_indices[1]
+        # pyre-fixme: [16]: `Optional` has no attribute `__getitem__`.
+        inverse_index_keys = t._inverse_indices[0]
+
+    field_values.append(inverse_indices)
+
+    return field_values, (
+        t._keys,
+        inverse_index_keys,
+        t._stride_per_key_per_rank,
+    )
 
 
 def _kjt_flatten_with_keys(
     t: KeyedJaggedTensor,
-) -> Tuple[List[Tuple[KeyEntry, Optional[torch.Tensor]]], List[str]]:
+) -> Tuple[
+    List[Tuple[KeyEntry, Optional[torch.Tensor]]],
+    Tuple[List[str], Optional[List[str]], Optional[List[List[int]]]],
+]:
     values, context = _kjt_flatten(t)
     # pyre can't tell that GetAttrKey implements the KeyEntry protocol
     return [  # pyre-ignore[7]
@@ -3047,9 +3070,17 @@ def _kjt_flatten_with_keys(
 
 
 def _kjt_unflatten(
-    values: List[Optional[torch.Tensor]], context: List[str]  # context is the _keys
+    values: List[Optional[torch.Tensor]],
+    context: Tuple[
+        List[str], Optional[List[str]], Optional[List[List[int]]]
+    ],  # context is the (_keys, inverse_index_keys, _stride_per_key_per_rank) tuple
 ) -> KeyedJaggedTensor:
-    return KeyedJaggedTensor(context, *values)
+    return KeyedJaggedTensor(
+        context[0],
+        *values[:-1],
+        stride_per_key_per_rank=context[2],
+        inverse_indices=(context[1], values[-1]) if context[1] is not None else None,
+    )
 
 
 def _kjt_flatten_spec(
@@ -3070,7 +3101,10 @@ def _kjt_flatten_spec(
 
 def flatten_kjt_list(
     kjt_arr: List[KeyedJaggedTensor],
-) -> Tuple[List[Optional[torch.Tensor]], List[List[str]]]:
+) -> Tuple[
+    List[Optional[torch.Tensor]],
+    List[Tuple[List[str], Optional[List[str]], Optional[List[List[int]]]]],
+]:
     _flattened_data = []
     _flattened_context = []
     for t in kjt_arr:
@@ -3081,7 +3115,8 @@ def flatten_kjt_list(
 
 
 def unflatten_kjt_list(
-    values: List[Optional[torch.Tensor]], contexts: List[List[str]]
+    values: List[Optional[torch.Tensor]],
+    contexts: List[Tuple[List[str], Optional[List[str]], Optional[List[List[int]]]]],
 ) -> List[KeyedJaggedTensor]:
     num_kjt_fields = len(KeyedJaggedTensor._fields)
     length = len(values)

torchrec/sparse/tests/test_keyed_jagged_tensor.py

@@ -1017,6 +1017,23 @@ def test_meta_device_compatibility(self) -> None:
             lengths=torch.tensor([], device=torch.device("meta")),
         )
 
+    def test_flatten_unflatten_with_vbe(self) -> None:
+        kjt = KeyedJaggedTensor(
+            keys=["f1", "f2"],
+            values=torch.tensor([5, 6, 7, 1, 2, 3, 0, 1]),
+            lengths=torch.tensor([3, 3, 2]),
+            stride_per_key_per_rank=[[2], [1]],
+            inverse_indices=(["f1", "f2"], torch.tensor([[0, 1, 0], [0, 0, 0]])),
+        )
+
+        flat_kjt, spec = pytree.tree_flatten(kjt)
+        unflattened_kjt = pytree.tree_unflatten(flat_kjt, spec)
+
+        self.assertEqual(
+            kjt.stride_per_key_per_rank(), unflattened_kjt.stride_per_key_per_rank()
+        )
+        self.assertEqual(kjt.inverse_indices(), unflattened_kjt.inverse_indices())
+
 
 class TestKeyedJaggedTensorScripting(unittest.TestCase):
     def test_scriptable_forward(self) -> None: