Make changes to proposer design (pytorch#1505)

Summary:

Allowing proposer to call the enumerator. Together with enumerate.populate_estimates, this allows the proposer to change the sharding options and re-estimate their perfs and storages.

Reviewed By: ge0405

Differential Revision: D50514266

torchrec/distributed/planner/planners.py

@@ -229,7 +229,7 @@ def plan(
         ] = {}
 
         for proposer in self._proposers:
-            proposer.load(search_space=search_space)
+            proposer.load(search_space=search_space, enumerator=self._enumerator)
 
         for proposer in self._proposers:
             proposal = proposer.propose()
@@ -242,6 +242,7 @@ def plan(
                         partitionable=partitionable,
                         plan=plan,
                         perf_rating=perf_rating,
+                        storage_constraint=storage_constraint,
                     )
                     proposal = proposer.propose()
                     continue
@@ -260,7 +261,10 @@ def plan(
                         best_plan = copy.deepcopy(plan)
                     proposal_cache[proposal_key] = (True, plan, perf_rating)
                     proposer.feedback(
-                        partitionable=True, plan=plan, perf_rating=perf_rating
+                        partitionable=True,
+                        plan=plan,
+                        perf_rating=perf_rating,
+                        storage_constraint=storage_constraint,
                     )
                 except PlannerError as planner_error:
                     last_planner_error = planner_error
@@ -280,7 +284,9 @@ def plan(
                     if current_storage < lowest_storage:
                         lowest_storage = current_storage
                     proposal_cache[proposal_key] = (False, None, None)
-                    proposer.feedback(partitionable=False)
+                    proposer.feedback(
+                        partitionable=False, storage_constraint=storage_constraint
+                    )
 
                 # clear shard.rank for each sharding_option
                 reset_shard_rank(proposal)

torchrec/distributed/planner/proposers.py

@@ -10,7 +10,13 @@
 from decimal import Decimal
 from typing import cast, Dict, List, Optional, Set, Tuple
 
-from torchrec.distributed.planner.types import Perf, Proposer, ShardingOption
+from torchrec.distributed.planner.types import (
+    Enumerator,
+    Perf,
+    Proposer,
+    ShardingOption,
+    Topology,
+)
 from torchrec.distributed.planner.utils import prod
 
 logger: logging.Logger = logging.getLogger(__name__)
@@ -43,7 +49,11 @@ def __init__(self, use_depth: bool = True, threshold: Optional[int] = None) -> N
         self._best_perf_rating: float = float("inf")
         self._num_inferior_perf: int = 0
 
-    def load(self, search_space: List[ShardingOption]) -> None:
+    def load(
+        self,
+        search_space: List[ShardingOption],
+        enumerator: Optional[Enumerator] = None,
+    ) -> None:
         self._reset()
         for sharding_option in search_space:
             fqn = sharding_option.fqn
@@ -78,6 +88,7 @@ def feedback(
         partitionable: bool,
         plan: Optional[List[ShardingOption]] = None,
         perf_rating: Optional[float] = None,
+        storage_constraint: Optional[Topology] = None,
     ) -> None:
         # When threshold is passed, observe the perf_rating trend. If the perf_rating
         # of the newly proposed plans have worse perf_rating, stop proposing.
@@ -126,7 +137,11 @@ def __init__(self, use_depth: bool = True) -> None:
         self._grouped_sharding_options: List[List[ShardingOption]] = []
         self._proposal_index: int = 0
 
-    def load(self, search_space: List[ShardingOption]) -> None:
+    def load(
+        self,
+        search_space: List[ShardingOption],
+        enumerator: Optional[Enumerator] = None,
+    ) -> None:
         self._reset()
         all_fqns = set()
         sharding_options_by_type_and_fqn: Dict[
@@ -175,6 +190,7 @@ def feedback(
         partitionable: bool,
         plan: Optional[List[ShardingOption]] = None,
         perf_rating: Optional[float] = None,
+        storage_constraint: Optional[Topology] = None,
     ) -> None:
         # static strategy, ignore feedback and just provide next proposal
         self._proposal_index += 1
@@ -187,7 +203,11 @@ def __init__(self, max_proposals: int = MAX_PROPOSALS) -> None:
         self._proposal_index: int = 0
         self._proposals: List[List[int]] = []
 
-    def load(self, search_space: List[ShardingOption]) -> None:
+    def load(
+        self,
+        search_space: List[ShardingOption],
+        enumerator: Optional[Enumerator] = None,
+    ) -> None:
         self._reset()
         for sharding_option in search_space:
             fqn = sharding_option.fqn
@@ -246,6 +266,7 @@ def feedback(
         partitionable: bool,
         plan: Optional[List[ShardingOption]] = None,
         perf_rating: Optional[float] = None,
+        storage_constraint: Optional[Topology] = None,
     ) -> None:
         # static strategy, ignore feedback and just provide next proposal
         self._proposal_index += 1

torchrec/distributed/planner/tests/test_proposers.py

@@ -19,7 +19,12 @@
     proposers_to_proposals_list,
     UniformProposer,
 )
-from torchrec.distributed.planner.types import Proposer, ShardingOption, Topology
+from torchrec.distributed.planner.types import (
+    Enumerator,
+    Proposer,
+    ShardingOption,
+    Topology,
+)
 from torchrec.distributed.test_utils.test_model import TestSparseNN
 from torchrec.distributed.types import ModuleSharder, ShardingType
 from torchrec.modules.embedding_configs import EmbeddingBagConfig
@@ -29,6 +34,7 @@ class MockProposer(Proposer):
     def load(
         self,
         search_space: List[ShardingOption],
+        enumerator: Optional[Enumerator] = None,
     ) -> None:
         pass
 
@@ -37,6 +43,7 @@ def feedback(
         partitionable: bool,
         plan: Optional[List[ShardingOption]] = None,
         perf_rating: Optional[float] = None,
+        storage_constraint: Optional[Topology] = None,
     ) -> None:
         pass
 

torchrec/distributed/planner/types.py

@@ -496,6 +496,13 @@ def enumerate(
         """
         ...
 
+    @abc.abstractmethod
+    def populate_estimates(self, sharding_options: List[ShardingOption]) -> None:
+        """
+        See class description.
+        """
+        ...
+
 
 class Proposer(abc.ABC):
     """
@@ -507,6 +514,7 @@ class Proposer(abc.ABC):
     def load(
         self,
         search_space: List[ShardingOption],
+        enumerator: Optional[Enumerator] = None,
     ) -> None:
         ...
 
@@ -516,6 +524,7 @@ def feedback(
         partitionable: bool,
         plan: Optional[List[ShardingOption]] = None,
         perf_rating: Optional[float] = None,
+        storage_constraint: Optional[Topology] = None,
     ) -> None:
         ...