(1/n) Allow DI sharding for models with FPE_EBC (#2968)

Summary:
Pull Request resolved: #2968

Support models that have FeatureProcessedEmbeddingBagCollection to be DI sharded. However, conservatively enforce that FPE itself can only be sharded on HBM and not across CPU as well.

These changes make sure we add QuantFeatureProcessedEmbeddingBagCollectionSharder as a recognized sharder, handle multiple envs needed for specifying DI sharding, and propagate TBE properly when processing the sharding plan.

This doesn't support true hybrid sharding for FPE.

Reviewed By: faran928

Differential Revision: D74671655

Author: Hanna Xu

torchrec/distributed/quant_embeddingbag.py

@@ -430,7 +430,7 @@ def __init__(
         self,
         module: EmbeddingBagCollectionInterface,
         table_name_to_parameter_sharding: Dict[str, ParameterSharding],
-        env: ShardingEnv,
+        env: Union[ShardingEnv, Dict[str, ShardingEnv]],  # support for hybrid sharding
         fused_params: Optional[Dict[str, Any]] = None,
         device: Optional[torch.device] = None,
         feature_processor: Optional[FeatureProcessorsCollection] = None,
@@ -462,11 +462,30 @@ def __init__(
                     f"Feature processor has inconsistent devices. Expected {feature_processor_device}, got {param.device}"
                 )
 
+        if isinstance(env, Dict):
+            expected_device_type = "cuda"
+            for (
+                embedding_configs
+            ) in self._sharding_type_device_group_to_sharding_infos.values():
+                # throws if not all shards only have the expected device type
+                shard_device_type = get_device_from_sharding_infos(embedding_configs)
+                if isinstance(shard_device_type, tuple):
+                    raise RuntimeError(
+                        f"Sharding across multiple device types for FeatureProcessedEmbeddingBagCollection is not supported yet, got {shard_device_type}",
+                    )
+                assert (
+                    shard_device_type == expected_device_type
+                ), f"Expected {expected_device_type} but got {shard_device_type} for FeatureProcessedEmbeddingBagCollection sharding device type"
+
+            # TODO(hcxu): support hybrid sharding with feature_processors_per_rank: ModuleList(ModuleList()), if compatible
+            world_size = env[expected_device_type].world_size
+        else:
+            world_size = env.world_size
+
         if feature_processor_device is None:
-            for _ in range(env.world_size):
-                self.feature_processors_per_rank.append(feature_processor)
+            self.feature_processors_per_rank += [feature_processor] * world_size
         else:
-            for i in range(env.world_size):
+            for i in range(world_size):
                 # Generic copy, for example initailized on cpu but -> sharding as meta
                 self.feature_processors_per_rank.append(
                     copy.deepcopy(feature_processor)

torchrec/distributed/quant_state.py

@@ -419,6 +419,16 @@ def get_bucket_offsets_per_virtual_table(
     }
 
 
+def get_param_id_from_type(is_sqebc: bool, is_sqmcec: bool, is_sfpebc: bool) -> str:
+    if is_sqebc:
+        return "embedding_bags"
+    elif is_sqmcec:
+        return "_embedding_module.embeddings"
+    elif is_sfpebc:
+        return "_embedding_bag_collection.embedding_bags"
+    return "embeddings"
+
+
 def sharded_tbes_weights_spec(
     sharded_model: torch.nn.Module,
     virtual_table_name_to_bucket_lengths: Optional[Dict[str, list[int]]] = None,
@@ -454,11 +464,14 @@ def sharded_tbes_weights_spec(
         is_sqebc: bool = "ShardedQuantEmbeddingBagCollection" in type_name
         is_sqec: bool = "ShardedQuantEmbeddingCollection" in type_name
         is_sqmcec: bool = "ShardedQuantManagedCollisionEmbeddingCollection" in type_name
+        is_sfpebc: bool = (
+            "ShardedQuantFeatureProcessedEmbeddingBagCollection" in type_name
+        )
 
-        if is_sqebc or is_sqec or is_sqmcec:
+        if is_sqebc or is_sqec or is_sqmcec or is_sfpebc:
             assert (
-                is_sqec + is_sqebc + is_sqmcec == 1
-            ), "Cannot have any two of ShardedQuantEmbeddingBagCollection, ShardedQuantEmbeddingCollection and ShardedQuantManagedCollisionEmbeddingCollection are true"
+                is_sqec + is_sqebc + is_sqmcec + is_sfpebc == 1
+            ), "Cannot have any two of ShardedQuantEmbeddingBagCollection, ShardedQuantEmbeddingCollection, ShardedQuantManagedCollisionEmbeddingCollection and ShardedQuantFeatureProcessedEmbeddingBagCollection are true"
             tbes_configs: Dict[
                 IntNBitTableBatchedEmbeddingBagsCodegen, GroupedEmbeddingConfig
             ] = module.tbes_configs()
@@ -550,8 +563,7 @@ def sharded_tbes_weights_spec(
                         row_offsets,
                         table_metadata.shard_offsets[1],
                     ]
-                    s: str = "embedding_bags" if is_sqebc else "embeddings"
-                    s = ("_embedding_module." if is_sqmcec else "") + s
+                    s: str = get_param_id_from_type(is_sqebc, is_sqmcec, is_sfpebc)
                     unsharded_fqn_weight_prefix: str = f"{module_fqn}.{s}.{table_name}"
                     unsharded_fqn_weight: str = unsharded_fqn_weight_prefix + ".weight"