Add VBE KJT support to EmbeddingCollection (#2047)

Summary:

- pad VBE kjt lengths to final batch size so that it's compatible with EC kernel.
- works with index dedup
- expands embeddings with vbe inverse indices
- long term solution is to fix seq TBE to not need lengths/batch size info, just length per key

Differential Revision: D51600051

torchrec/distributed/embedding.py

@@ -12,7 +12,7 @@
 import logging
 import warnings
 from collections import defaultdict, deque, OrderedDict
-from dataclasses import dataclass, field
+from dataclasses import dataclass
 from itertools import accumulate
 from typing import Any, cast, Dict, List, MutableMapping, Optional, Tuple, Type, Union
 
@@ -72,10 +72,15 @@
     EmbeddingCollection,
     EmbeddingCollectionInterface,
 )
-from torchrec.modules.utils import construct_jagged_tensors
+from torchrec.modules.utils import construct_jagged_tensors, SequenceVBEContext
 from torchrec.optim.fused import EmptyFusedOptimizer, FusedOptimizerModule
 from torchrec.optim.keyed import CombinedOptimizer, KeyedOptimizer
-from torchrec.sparse.jagged_tensor import JaggedTensor, KeyedJaggedTensor
+from torchrec.sparse.jagged_tensor import (
+    _pin_and_move,
+    _to_offsets,
+    JaggedTensor,
+    KeyedJaggedTensor,
+)
 
 try:
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
@@ -323,6 +328,33 @@ def create_sharding_infos_by_sharding_device_group(
     return sharding_type_device_group_to_sharding_infos
 
 
+def pad_vbe_kjt_lengths(features: KeyedJaggedTensor) -> KeyedJaggedTensor:
+    final_stride = features.inverse_indices()[1].numel() // len(
+        features.inverse_indices()[0]
+    )
+    new_lengths = torch.zeros(
+        final_stride * len(features.keys()),
+        device=features.device(),
+        dtype=features.lengths().dtype,
+    )
+    cum_stride = 0
+    for i, stride in enumerate(features.stride_per_key()):
+        new_lengths[i * final_stride : i * final_stride + stride] = features.lengths()[
+            cum_stride : cum_stride + stride
+        ]
+        cum_stride += stride
+
+    return KeyedJaggedTensor(
+        keys=features.keys(),
+        values=features.values(),
+        lengths=new_lengths,
+        stride=final_stride,
+        length_per_key=features.length_per_key(),
+        offset_per_key=features.offset_per_key(),
+    )
+
+
+@dataclass
 class EmbeddingCollectionContext(Multistreamable):
     # Torch Dynamo does not support default_factory=list:
     # https://github.com/pytorch/pytorch/issues/120108
@@ -333,11 +365,13 @@ def __init__(
         sharding_contexts: Optional[List[SequenceShardingContext]] = None,
         input_features: Optional[List[KeyedJaggedTensor]] = None,
         reverse_indices: Optional[List[torch.Tensor]] = None,
+        seq_vbe_ctx: Optional[List[SequenceVBEContext]] = None,
     ) -> None:
         super().__init__()
         self.sharding_contexts: List[SequenceShardingContext] = sharding_contexts or []
         self.input_features: List[KeyedJaggedTensor] = input_features or []
         self.reverse_indices: List[torch.Tensor] = reverse_indices or []
+        self.seq_vbe_ctx: List[SequenceVBEContext] = seq_vbe_ctx or []
 
     def record_stream(self, stream: torch.cuda.streams.Stream) -> None:
         for ctx in self.sharding_contexts:
@@ -346,6 +380,8 @@ def record_stream(self, stream: torch.cuda.streams.Stream) -> None:
             f.record_stream(stream)
         for r in self.reverse_indices:
             r.record_stream(stream)
+        for s in self.seq_vbe_ctx:
+            s.record_stream(stream)
 
 
 class EmbeddingCollectionAwaitable(LazyAwaitable[Dict[str, JaggedTensor]]):
@@ -385,6 +421,9 @@ def _wait_impl(self) -> Dict[str, JaggedTensor]:
                 if i >= len(self._ctx.reverse_indices)
                 else self._ctx.reverse_indices[i]
             )
+            seq_vbe_ctx = (
+                None if i >= len(self._ctx.seq_vbe_ctx) else self._ctx.seq_vbe_ctx[i]
+            )
             jt_dict.update(
                 construct_jagged_tensors(
                     embeddings=w.wait(),
@@ -394,6 +433,7 @@ def _wait_impl(self) -> Dict[str, JaggedTensor]:
                     features_to_permute_indices=self._features_to_permute_indices,
                     original_features=original_features,
                     reverse_indices=reverse_indices,
+                    seq_vbe_ctx=seq_vbe_ctx,
                 )
             )
         return jt_dict
@@ -506,6 +546,7 @@ def __init__(
                 module.embedding_configs(), table_name_to_parameter_sharding
             )
         self._need_indices: bool = module.need_indices()
+        self._inverse_indices_permute_per_sharding: Optional[List[torch.Tensor]] = None
 
         for index, (sharding, lookup) in enumerate(
             zip(
@@ -847,8 +888,8 @@ def _create_output_dist(
 
     def _dedup_indices(
         self,
-        input_feature_splits: List[KeyedJaggedTensor],
         ctx: EmbeddingCollectionContext,
+        input_feature_splits: List[KeyedJaggedTensor],
     ) -> List[KeyedJaggedTensor]:
         if not self._use_index_dedup:
             return input_feature_splits
@@ -874,37 +915,113 @@ def _dedup_indices(
                     offsets=offsets,
                     values=unique_indices,
                 )
-
                 ctx.input_features.append(input_feature)
                 ctx.reverse_indices.append(reverse_indices)
                 features_by_shards.append(dedup_features)
 
         return features_by_shards
 
+    def _create_inverse_indices_permute_per_sharding(
+        self, inverse_indices: Tuple[List[str], torch.Tensor]
+    ) -> None:
+        if (
+            len(self._embedding_names_per_sharding) == 1
+            and self._embedding_names_per_sharding[0] == inverse_indices[0]
+        ):
+            return
+        index_per_name = {name: i for i, name in enumerate(inverse_indices[0])}
+        permute_per_sharding = []
+        for emb_names in self._embedding_names_per_sharding:
+            permute = _pin_and_move(
+                torch.tensor(
+                    [index_per_name[name.split("@")[0]] for name in emb_names]
+                ),
+                inverse_indices[1].device,
+            )
+            permute_per_sharding.append(permute)
+        self._inverse_indices_permute_per_sharding = permute_per_sharding
+
+    def _compute_sequence_vbe_context(
+        self,
+        ctx: EmbeddingCollectionContext,
+        unpadded_features: KeyedJaggedTensor,
+    ) -> None:
+        assert (
+            unpadded_features.inverse_indices_or_none() is not None
+        ), "inverse indices must be provided from KJT if using variable batch size per feature."
+
+        inverse_indices = unpadded_features.inverse_indices()
+        stride = inverse_indices[1].numel() // len(inverse_indices[0])
+        if self._inverse_indices_permute_per_sharding is None:
+            self._create_inverse_indices_permute_per_sharding(inverse_indices)
+
+        if self._features_order:
+            features = unpadded_features.permute(
+                self._features_order,
+                self._features_order_tensor,
+            )
+
+        features_by_sharding = features.split(self._feature_splits)
+        for i, feature in enumerate(features_by_sharding):
+            if self._inverse_indices_permute_per_sharding is not None:
+                permute = self._inverse_indices_permute_per_sharding[i]
+                permuted_indices = torch.index_select(inverse_indices[1], 0, permute)
+            else:
+                permuted_indices = inverse_indices[1]
+            stride_per_key = _pin_and_move(
+                torch.tensor(feature.stride_per_key()), feature.device()
+            )
+            offsets = _to_offsets(stride_per_key)[:-1].unsqueeze(-1)
+            recat = (permuted_indices + offsets).flatten().int()
+
+            if self._need_indices:
+                reindexed_lengths, reindexed_values, _ = (
+                    torch.ops.fbgemm.permute_1D_sparse_data(
+                        recat,
+                        feature.lengths(),
+                        feature.values(),
+                    )
+                )
+            else:
+                reindexed_lengths = torch.index_select(feature.lengths(), 0, recat)
+                reindexed_values = None
+
+            reindexed_lengths = reindexed_lengths.view(-1, stride)
+            reindexed_length_per_key = torch.sum(reindexed_lengths, dim=1).tolist()
+
+            ctx.seq_vbe_ctx.append(
+                SequenceVBEContext(
+                    recat=recat,
+                    unpadded_lengths=feature.lengths(),
+                    reindexed_lengths=reindexed_lengths,
+                    reindexed_length_per_key=reindexed_length_per_key,
+                    reindexed_values=reindexed_values,
+                )
+            )
+
     # pyre-ignore [14]
     def input_dist(
         self,
         ctx: EmbeddingCollectionContext,
         features: KeyedJaggedTensor,
     ) -> Awaitable[Awaitable[KJTList]]:
-        if features.variable_stride_per_key():
-            raise ValueError(
-                "Variable batch per feature is not supported with EmbeddingCollection"
-            )
         if self._has_uninitialized_input_dist:
             self._create_input_dist(input_feature_names=features.keys())
             self._has_uninitialized_input_dist = False
         with torch.no_grad():
+            unpadded_features = None
+            if features.variable_stride_per_key():
+                unpadded_features = features
+                features = pad_vbe_kjt_lengths(unpadded_features)
+
             if self._features_order:
                 features = features.permute(
                     self._features_order,
                     self._features_order_tensor,
                 )
-
-            input_feature_splits = features.split(
-                self._feature_splits,
-            )
-            features_by_shards = self._dedup_indices(input_feature_splits, ctx)
+            features_by_shards = features.split(self._feature_splits)
+            if self._use_index_dedup:
+                features_by_shards = self._dedup_indices(ctx, features_by_shards)
 
             awaitables = []
             for input_dist, features in zip(self._input_dists, features_by_shards):
@@ -919,6 +1036,8 @@ def input_dist(
                         ),
                     )
                 )
+            if unpadded_features is not None:
+                self._compute_sequence_vbe_context(ctx, unpadded_features)
         return KJTListSplitsAwaitable(awaitables, ctx)
 
     def compute(

torchrec/distributed/tests/test_sequence_model.py

@@ -305,6 +305,7 @@ def __init__(
         kernel_type: str,
         qcomms_config: Optional[QCommsConfig] = None,
         fused_params: Optional[Dict[str, Any]] = None,
+        use_index_dedup: bool = False,
     ) -> None:
         self._sharding_type = sharding_type
         self._kernel_type = kernel_type
@@ -321,6 +322,7 @@ def __init__(
         super().__init__(
             fused_params=fused_params,
             qcomm_codecs_registry=qcomm_codecs_registry,
+            use_index_dedup=use_index_dedup,
         )
 
     """

torchrec/distributed/tests/test_sequence_model_parallel.py

@@ -62,8 +62,9 @@ class SequenceModelParallelTest(MultiProcessTestBase):
                 },
             ]
         ),
+        variable_batch_size=st.booleans(),
     )
-    @settings(verbosity=Verbosity.verbose, max_examples=2, deadline=None)
+    @settings(verbosity=Verbosity.verbose, max_examples=3, deadline=None)
     def test_sharding_nccl_rw(
         self,
         sharding_type: str,
@@ -72,6 +73,7 @@ def test_sharding_nccl_rw(
         apply_optimizer_in_backward_config: Optional[
             Dict[str, Tuple[Type[torch.optim.Optimizer], Dict[str, Any]]]
         ],
+        variable_batch_size: bool,
     ) -> None:
         assume(
             apply_optimizer_in_backward_config is None
@@ -88,6 +90,7 @@ def test_sharding_nccl_rw(
             backend="nccl",
             qcomms_config=qcomms_config,
             apply_optimizer_in_backward_config=apply_optimizer_in_backward_config,
+            variable_batch_size=variable_batch_size,
         )
 
     @unittest.skipIf(
@@ -152,8 +155,9 @@ def test_sharding_nccl_dp(
                 },
             ]
         ),
+        variable_batch_size=st.booleans(),
     )
-    @settings(verbosity=Verbosity.verbose, max_examples=2, deadline=None)
+    @settings(verbosity=Verbosity.verbose, max_examples=3, deadline=None)
     def test_sharding_nccl_tw(
         self,
         sharding_type: str,
@@ -162,6 +166,7 @@ def test_sharding_nccl_tw(
         apply_optimizer_in_backward_config: Optional[
             Dict[str, Tuple[Type[torch.optim.Optimizer], Dict[str, Any]]]
         ],
+        variable_batch_size: bool,
     ) -> None:
         assume(
             apply_optimizer_in_backward_config is None
@@ -178,7 +183,7 @@ def test_sharding_nccl_tw(
             backend="nccl",
             qcomms_config=qcomms_config,
             apply_optimizer_in_backward_config=apply_optimizer_in_backward_config,
-            variable_batch_size=False,
+            variable_batch_size=variable_batch_size,
         )
 
     @unittest.skipIf(
@@ -203,15 +208,17 @@ def test_sharding_nccl_tw(
                 },
             ]
         ),
+        variable_batch_size=st.booleans(),
     )
-    @settings(verbosity=Verbosity.verbose, max_examples=2, deadline=None)
+    @settings(verbosity=Verbosity.verbose, max_examples=3, deadline=None)
     def test_sharding_nccl_cw(
         self,
         sharding_type: str,
         kernel_type: str,
         apply_optimizer_in_backward_config: Optional[
             Dict[str, Tuple[Type[torch.optim.Optimizer], Dict[str, Any]]]
         ],
+        variable_batch_size: bool,
     ) -> None:
         assume(
             apply_optimizer_in_backward_config is None
@@ -230,7 +237,7 @@ def test_sharding_nccl_cw(
                 for table in self.tables
             },
             apply_optimizer_in_backward_config=apply_optimizer_in_backward_config,
-            variable_batch_size=False,
+            variable_batch_size=variable_batch_size,
         )
 
     @unittest.skipIf(
@@ -246,25 +253,28 @@ def test_sharding_nccl_cw(
                 ShardingType.ROW_WISE.value,
             ]
         ),
+        index_dedup=st.booleans(),
     )
     @settings(verbosity=Verbosity.verbose, max_examples=3, deadline=None)
     def test_sharding_variable_batch(
         self,
         sharding_type: str,
+        index_dedup: bool,
     ) -> None:
         self._test_sharding(
             sharders=[
                 TestEmbeddingCollectionSharder(
                     sharding_type=sharding_type,
                     kernel_type=EmbeddingComputeKernel.FUSED.value,
+                    use_index_dedup=index_dedup,
                 )
             ],
             backend="nccl",
             constraints={
                 table.name: ParameterConstraints(min_partition=4)
                 for table in self.tables
             },
-            variable_batch_size=True,
+            variable_batch_per_feature=True,
         )
 
     # pyre-fixme[56]
@@ -347,6 +357,7 @@ def _test_sharding(
             Dict[str, Tuple[Type[torch.optim.Optimizer], Dict[str, Any]]]
         ] = None,
         variable_batch_size: bool = False,
+        variable_batch_per_feature: bool = False,
     ) -> None:
         self._run_multi_process_test(
             callable=sharding_single_rank_test,
@@ -362,4 +373,6 @@ def _test_sharding(
             qcomms_config=qcomms_config,
             apply_optimizer_in_backward_config=apply_optimizer_in_backward_config,
             variable_batch_size=variable_batch_size,
+            variable_batch_per_feature=variable_batch_per_feature,
+            global_constant_batch=True,
         )