Support skipping offsets computation in KJT to_dict (#3503)

Summary:

This diff optimizes KeyedJaggedTensor.to_dict by optionally removing several unnecessary offset computation. It can evidently improve latency if offsets are not needed in the outputs.

Differential Revision: D85995629

Author: coufon

torchrec/sparse/jagged_tensor.py

@@ -1402,6 +1402,7 @@ def _maybe_compute_kjt_to_jt_dict(
     variable_stride_per_key: bool,
     weights: Optional[torch.Tensor],
     jt_dict: Optional[Dict[str, JaggedTensor]],
+    compute_offsets: bool = True,
 ) -> Dict[str, JaggedTensor]:
     if not length_per_key:
         return {}
@@ -1418,50 +1419,50 @@ def _maybe_compute_kjt_to_jt_dict(
             torch._check(cat_size <= total_size)
         torch._check(cat_size == total_size)
         torch._check_is_size(stride)
+
     values_list = torch.split(values, length_per_key)
+    split_offsets: list[torch.Tensor] | None = None
     if variable_stride_per_key:
         split_lengths = torch.split(lengths, stride_per_key)
-        split_offsets = [
-            torch.ops.fbgemm.asynchronous_complete_cumsum(lengths)
-            for lengths in split_lengths
-        ]
+        if compute_offsets:
+            split_offsets = [
+                torch.ops.fbgemm.asynchronous_complete_cumsum(lengths)
+                for lengths in split_lengths
+            ]
     elif pt2_guard_size_oblivious(lengths.numel() > 0):
         strided_lengths = lengths.view(len(keys), stride)
         if not torch.jit.is_scripting() and is_torchdynamo_compiling():
             torch._check(strided_lengths.size(0) > 0)
             torch._check(strided_lengths.size(1) > 0)
-        split_lengths = torch.unbind(
-            strided_lengths,
-            dim=0,
-        )
-        split_offsets = torch.unbind(
-            _batched_lengths_to_offsets(strided_lengths),
-            dim=0,
-        )
+
+        split_lengths = torch.unbind(strided_lengths, dim=0)
+        if compute_offsets:
+            split_offsets = torch.unbind(  # pyre-ignore
+                _batched_lengths_to_offsets(strided_lengths), dim=0
+            )
     else:
         split_lengths = torch.unbind(lengths, dim=0)
-        split_offsets = torch.unbind(lengths, dim=0)
+        if compute_offsets:
+            split_offsets = split_lengths  # pyre-ignore
 
     if weights is not None:
         weights_list = torch.split(weights, length_per_key)
         for idx, key in enumerate(keys):
-            length = split_lengths[idx]
-            offset = split_offsets[idx]
             _jt_dict[key] = JaggedTensor(
-                lengths=length,
-                offsets=offset,
+                lengths=split_lengths[idx],
+                # pyre-ignore[16]: Undefined attribute: None has no attribute __getitem__
+                offsets=split_offsets[idx] if split_offsets is not None else None,
                 values=values_list[idx],
                 weights=weights_list[idx],
             )
     else:
         for idx, key in enumerate(keys):
-            length = split_lengths[idx]
-            offset = split_offsets[idx]
             _jt_dict[key] = JaggedTensor(
-                lengths=length,
-                offsets=offset,
+                lengths=split_lengths[idx],
+                offsets=split_offsets[idx] if split_offsets is not None else None,
                 values=values_list[idx],
             )
+
     return _jt_dict
 
 
@@ -2698,11 +2699,14 @@ def __getitem__(self, key: str) -> JaggedTensor:
                 offsets=None,
             )
 
-    def to_dict(self) -> Dict[str, JaggedTensor]:
+    def to_dict(self, compute_offsets: bool = True) -> Dict[str, JaggedTensor]:
         """
         Returns a dictionary of JaggedTensor for each key.
         Will cache result in self._jt_dict.
 
+        Args:
+            compute_offsets (str): compute offsets when true.
+
         Returns:
             Dict[str, JaggedTensor]: dictionary of JaggedTensor for each key.
         """
@@ -2720,6 +2724,7 @@ def to_dict(self) -> Dict[str, JaggedTensor]:
             variable_stride_per_key=self.variable_stride_per_key(),
             weights=self.weights_or_none(),
             jt_dict=self._jt_dict,
+            compute_offsets=compute_offsets,
         )
         self._jt_dict = _jt_dict
         return _jt_dict

torchrec/sparse/tests/test_jagged_tensor.py

@@ -811,6 +811,54 @@ def test_from_jt_dict_vb(self) -> None:
             torch.equal(j1.values(), torch.Tensor([3.0, 4.0, 5.0, 6.0, 7.0, 8.0]))
         )
 
+    def test_to_dict_compute_offsets_false(self) -> None:
+        # Setup: KJT with two keys and standard stride
+        values = torch.Tensor([1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0])
+        keys = ["f1", "f2"]
+        lengths = torch.IntTensor([2, 0, 1, 1, 1, 3])
+
+        kjt = KeyedJaggedTensor(values=values, keys=keys, lengths=lengths)
+
+        # Execute: call to_dict with compute_offsets=False
+        jt_dict = kjt.to_dict(compute_offsets=False)
+
+        # Assert: offsets_or_none() should be None for each JaggedTensor
+        self.assertIsNone(jt_dict["f1"].offsets_or_none())
+        self.assertIsNone(jt_dict["f2"].offsets_or_none())
+        # Lengths should still be available
+        self.assertTrue(
+            torch.equal(jt_dict["f1"].lengths(), torch.IntTensor([2, 0, 1]))
+        )
+        self.assertTrue(
+            torch.equal(jt_dict["f2"].lengths(), torch.IntTensor([1, 1, 3]))
+        )
+
+    def test_to_dict_compute_offsets_false_variable_stride(self) -> None:
+        # Setup: KJT with variable stride per key (reusing test_from_jt_dict_vb data)
+        values = torch.Tensor([1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0])
+        keys = ["f1", "f2"]
+        lengths = torch.IntTensor([2, 0, 1, 1, 1, 3])
+        stride_per_key_per_rank = [[2], [4]]
+
+        kjt = KeyedJaggedTensor(
+            values=values,
+            keys=keys,
+            lengths=lengths,
+            stride_per_key_per_rank=stride_per_key_per_rank,
+        )
+
+        # Execute: call to_dict with compute_offsets=False
+        jt_dict = kjt.to_dict(compute_offsets=False)
+
+        # Assert: offsets_or_none() should be None for each JaggedTensor
+        self.assertIsNone(jt_dict["f1"].offsets_or_none())
+        self.assertIsNone(jt_dict["f2"].offsets_or_none())
+        # Lengths should still be available
+        self.assertTrue(torch.equal(jt_dict["f1"].lengths(), torch.IntTensor([2, 0])))
+        self.assertTrue(
+            torch.equal(jt_dict["f2"].lengths(), torch.IntTensor([1, 1, 1, 3]))
+        )
+
 
 class TestJaggedTensorTracing(unittest.TestCase):
     def test_jagged_tensor(self) -> None: