Reland of D65489998 Optimize sharding performance of embeddings" (#2664)

Summary:
Pull Request resolved: #2664

X-link: pytorch/FBGEMM#3549

X-link: facebookresearch/FBGEMM#634

This diff is a reland of D65489998 after backout in D66800554.

Reviewed By: iamzainhuda

Differential Revision: D66828907

fbshipit-source-id: ab6e6a9faa8255c4847a69a8efb46182bedc9737

Author: Boris Sarana

torchrec/distributed/embedding_types.py

@@ -9,6 +9,7 @@
 
 import abc
 import copy
+import os
 from dataclasses import dataclass
 from enum import Enum, unique
 from typing import Any, Dict, Generic, Iterator, List, Optional, Tuple, TypeVar, Union
@@ -343,6 +344,12 @@ def __init__(
         self._lookups: List[nn.Module] = []
         self._output_dists: List[nn.Module] = []
 
+        # option to construct ShardedTensor from metadata avoiding expensive all-gather
+        self._construct_sharded_tensor_from_metadata: bool = (
+            os.environ.get("TORCHREC_CONSTRUCT_SHARDED_TENSOR_FROM_METADATA", "0")
+            == "1"
+        )
+
     def prefetch(
         self,
         dist_input: KJTList,

torchrec/distributed/embeddingbag.py

@@ -29,6 +29,7 @@
 from fbgemm_gpu.permute_pooled_embedding_modules import PermutePooledEmbeddings
 from torch import distributed as dist, nn, Tensor
 from torch.autograd.profiler import record_function
+from torch.distributed._shard.sharded_tensor import TensorProperties
 from torch.distributed._tensor import DTensor
 from torch.nn.modules.module import _IncompatibleKeys
 from torch.nn.parallel import DistributedDataParallel
@@ -81,6 +82,7 @@
     optimizer_type_to_emb_opt_type,
 )
 from torchrec.modules.embedding_configs import (
+    data_type_to_dtype,
     EmbeddingBagConfig,
     EmbeddingTableConfig,
     PoolingType,
@@ -945,17 +947,48 @@ def _initialize_torch_state(self) -> None:  # noqa
                 # created ShardedTensors once in init, use in post_state_dict_hook
                 # note: at this point kvstore backed tensors don't own valid snapshots, so no read
                 # access is allowed on them.
-                self._model_parallel_name_to_sharded_tensor[table_name] = (
-                    ShardedTensor._init_from_local_shards(
-                        local_shards,
-                        self._name_to_table_size[table_name],
-                        process_group=(
-                            self._env.sharding_pg
-                            if isinstance(self._env, ShardingEnv2D)
-                            else self._env.process_group
+
+                # create ShardedTensor from local shards and metadata avoding all_gather collective
+                if self._construct_sharded_tensor_from_metadata:
+                    sharding_spec = none_throws(
+                        self.module_sharding_plan[table_name].sharding_spec
+                    )
+
+                    tensor_properties = TensorProperties(
+                        dtype=(
+                            data_type_to_dtype(
+                                self._table_name_to_config[table_name].data_type
+                            )
                         ),
                     )
-                )
+
+                    self._model_parallel_name_to_sharded_tensor[table_name] = (
+                        ShardedTensor._init_from_local_shards_and_global_metadata(
+                            local_shards=local_shards,
+                            sharded_tensor_metadata=sharding_spec.build_metadata(
+                                tensor_sizes=self._name_to_table_size[table_name],
+                                tensor_properties=tensor_properties,
+                            ),
+                            process_group=(
+                                self._env.sharding_pg
+                                if isinstance(self._env, ShardingEnv2D)
+                                else self._env.process_group
+                            ),
+                        )
+                    )
+                else:
+                    # create ShardedTensor from local shards using all_gather collective
+                    self._model_parallel_name_to_sharded_tensor[table_name] = (
+                        ShardedTensor._init_from_local_shards(
+                            local_shards,
+                            self._name_to_table_size[table_name],
+                            process_group=(
+                                self._env.sharding_pg
+                                if isinstance(self._env, ShardingEnv2D)
+                                else self._env.process_group
+                            ),
+                        )
+                    )
 
         def extract_sharded_kvtensors(
             module: ShardedEmbeddingBagCollection,