2025-09-24 nightly release (c26367f)

Author: pytorchbot

torchrec/distributed/embedding_sharding.py

@@ -565,6 +565,7 @@ def _group_tables_per_rank(
                 ),
                 _prefetch_and_cached(table),
                 table.use_virtual_table if is_inference else None,
+                table.enable_embedding_update,
             )
             # micromanage the order of we traverse the groups to ensure backwards compatibility
             if grouping_key not in groups:
@@ -581,6 +582,7 @@ def _group_tables_per_rank(
                 _,
                 _,
                 use_virtual_table,
+                enable_embedding_update,
             ) = grouping_key
             grouped_tables = groups[grouping_key]
             # remove non-native fused params
@@ -602,6 +604,7 @@ def _group_tables_per_rank(
                     compute_kernel=compute_kernel_type,
                     embedding_tables=grouped_tables,
                     fused_params=per_tbe_fused_params,
+                    enable_embedding_update=enable_embedding_update,
                 )
             )
         return grouped_embedding_configs

torchrec/distributed/embedding_types.py

@@ -251,6 +251,8 @@ class GroupedEmbeddingConfig:
     compute_kernel: EmbeddingComputeKernel
     embedding_tables: List[ShardedEmbeddingTable]
     fused_params: Optional[Dict[str, Any]] = None
+    # Write-enabled Embedding Tables cannot be grouped with read-only Embedding Tables TBE needs to be separate.
+    enable_embedding_update: bool = False
 
     def feature_hash_sizes(self) -> List[int]:
         feature_hash_sizes = []

torchrec/distributed/sharding/rw_sharding.py

@@ -223,6 +223,7 @@ def _shard(
                         total_num_buckets=info.embedding_config.total_num_buckets,
                         use_virtual_table=info.embedding_config.use_virtual_table,
                         virtual_table_eviction_policy=info.embedding_config.virtual_table_eviction_policy,
+                        enable_embedding_update=info.embedding_config.enable_embedding_update,
                     )
                 )
         return tables_per_rank
@@ -278,6 +279,20 @@ def _get_feature_hash_sizes(self) -> List[int]:
             feature_hash_sizes.extend(group_config.feature_hash_sizes())
         return feature_hash_sizes
 
+    def _get_num_writable_features(self) -> int:
+        return sum(
+            group_config.num_features()
+            for group_config in self._grouped_embedding_configs
+            if group_config.enable_embedding_update
+        )
+
+    def _get_writable_feature_hash_sizes(self) -> List[int]:
+        feature_hash_sizes: List[int] = []
+        for group_config in self._grouped_embedding_configs:
+            if group_config.enable_embedding_update:
+                feature_hash_sizes.extend(group_config.feature_hash_sizes())
+        return feature_hash_sizes
+
 
 class RwSparseFeaturesDist(BaseSparseFeaturesDist[KeyedJaggedTensor]):
     """

torchrec/modules/embedding_configs.py

@@ -370,6 +370,7 @@ class BaseEmbeddingConfig:
     total_num_buckets: Optional[int] = None
     use_virtual_table: bool = False
     virtual_table_eviction_policy: Optional[VirtualTableEvictionPolicy] = None
+    enable_embedding_update: bool = False
 
     def get_weight_init_max(self) -> float:
         if self.weight_init_max is None:

torchrec/schema/api_tests/test_embedding_config_schema.py

@@ -43,6 +43,7 @@ class StableEmbeddingBagConfig:
     total_num_buckets: Optional[int] = None
     use_virtual_table: bool = False
     virtual_table_eviction_policy: Optional[VirtualTableEvictionPolicy] = None
+    enable_embedding_update: bool = False
     pooling: PoolingType = PoolingType.SUM