remove features_to_dict (#1851)

Summary:
Pull Request resolved: #1851

Use ComputeKJTToJTDict instead of features_to_dict in the tagging rule.

Reviewed By: gnahzg

Differential Revision: D55842584

fbshipit-source-id: 599fae110c76250fca1cb85d4a2b6ccfbcdf948b

torchrec/quant/embedding_modules.py

@@ -46,7 +46,12 @@
 )
 
 from torchrec.modules.utils import construct_jagged_tensors_inference
-from torchrec.sparse.jagged_tensor import JaggedTensor, KeyedJaggedTensor, KeyedTensor
+from torchrec.sparse.jagged_tensor import (
+    ComputeKJTToJTDict,
+    JaggedTensor,
+    KeyedJaggedTensor,
+    KeyedTensor,
+)
 from torchrec.tensor_types import UInt2Tensor, UInt4Tensor
 from torchrec.types import ModuleNoCopyMixin
 
@@ -230,13 +235,6 @@ def _update_embedding_configs(
         )
 
 
-@torch.fx.wrap
-def features_to_dict(
-    features: KeyedJaggedTensor,
-) -> Dict[str, JaggedTensor]:
-    return features.to_dict()
-
-
 class EmbeddingBagCollection(EmbeddingBagCollectionInterface, ModuleNoCopyMixin):
     """
     EmbeddingBagCollection represents a collection of pooled embeddings (EmbeddingBags).
@@ -332,6 +330,7 @@ def __init__(
             Dict[str, Tuple[Tensor, Tensor]]
         ] = None
         self.row_alignment = row_alignment
+        self._kjt_to_jt_dict = ComputeKJTToJTDict()
 
         table_names = set()
         for table in self._embedding_bag_configs:
@@ -463,7 +462,7 @@ def forward(
             KeyedTensor
         """
 
-        feature_dict = features_to_dict(features)
+        feature_dict = self._kjt_to_jt_dict(features)
         embeddings = []
 
         # TODO ideally we can accept KJTs with any feature order. However, this will require an order check + permute, which will break torch.script.

torchrec/quant/tests/test_embedding_modules.py

@@ -29,10 +29,14 @@
 from torchrec.quant.embedding_modules import (
     EmbeddingBagCollection as QuantEmbeddingBagCollection,
     EmbeddingCollection as QuantEmbeddingCollection,
-    features_to_dict,
     quant_prep_enable_quant_state_dict_split_scale_bias,
 )
-from torchrec.sparse.jagged_tensor import JaggedTensor, KeyedJaggedTensor, KeyedTensor
+from torchrec.sparse.jagged_tensor import (
+    ComputeKJTToJTDict,
+    JaggedTensor,
+    KeyedJaggedTensor,
+    KeyedTensor,
+)
 
 
 class EmbeddingBagCollectionTest(unittest.TestCase):
@@ -445,7 +449,7 @@ def test_trace_and_script(self) -> None:
 
         from torchrec.fx import symbolic_trace
 
-        gm = symbolic_trace(qebc)
+        gm = symbolic_trace(qebc, leaf_modules=[ComputeKJTToJTDict.__name__])
 
         non_placeholder_nodes = [
             node for node in gm.graph.nodes if node.op != "placeholder"
@@ -455,13 +459,13 @@ def test_trace_and_script(self) -> None:
         )
         self.assertEqual(
             non_placeholder_nodes[0].op,
-            "call_function",
-            f"First non-placeholder node must be call_function, got {non_placeholder_nodes[0].op} instead",
+            "call_module",
+            f"First non-placeholder node must be call_method, got {non_placeholder_nodes[0].op} instead",
         )
         self.assertEqual(
             non_placeholder_nodes[0].name,
-            features_to_dict.__name__,
-            f"First non-placeholder node must be features_to_dict, got {non_placeholder_nodes[0].name} instead",
+            "_kjt_to_jt_dict",
+            f"First non-placeholder node must be _kjt_to_jt_dict, got {non_placeholder_nodes[0].name} instead",
         )
 
         features = KeyedJaggedTensor(