Recurse partitioner through branches

pytorch · Jan 6, 2025 · ed5b952 · ed5b952
1 parent 73591f1
commit ed5b952
Show file tree

Hide file tree

Showing 2 changed files with 21 additions and 12 deletions.
diff --git a/backends/xnnpack/partition/xnnpack_partitioner.py b/backends/xnnpack/partition/xnnpack_partitioner.py
@@ -21,6 +21,7 @@
     ConfigerationBasedPartitioner,
 )
 from executorch.exir.backend.partitioner import DelegationSpec
+from torch.fx import GraphModule
 from torch.fx.passes.infra.partitioner import Partition
 
 logging.basicConfig(level=logging.WARNING)
@@ -65,25 +66,25 @@ def __init__(
         self.per_op_mode = per_op_mode
         super().__init__(delegation_spec, initialized_configs)
 
-    def generate_partitions(self, ep: ExportedProgram) -> List[Partition]:
+    def generate_partitions(self, ep: ExportedProgram, gm: Optional[GraphModule] = None) -> List[Partition]:
         """
         generate_partitions is different if partitioner is set to per_op_mode
         for per_op_mode we only need to generate unmerged partitions instead
         of using the default generate_partitions method.
         """
         if self.per_op_mode:
-            return self.generate_per_op_partitions(ep)
+            return self.generate_per_op_partitions(ep, gm)
         else:
-            return super().generate_partitions(ep)
+            return super().generate_partitions(ep, gm)
 
-    def generate_per_op_partitions(self, ep: ExportedProgram) -> List[Partition]:
+    def generate_per_op_partitions(self, ep: ExportedProgram, gm: Optional[GraphModule] = None) -> List[Partition]:
         """
         Uses configs to generate per_op_partitions. That is no partitions are
         merged together. All partitions (node + deps) returned by PartitionerConfigs
         are put into their own partition.
         """
         partitions = []
-        matched_nodes = self.get_matched_nodes_from_configs(ep)
+        matched_nodes = self.get_matched_nodes_from_configs(ep, gm)
         partition_id = itertools.count()
         nodes_seen = {}
         for match in matched_nodes:

diff --git a/exir/backend/canonical_partitioners/config_partitioner.py b/exir/backend/canonical_partitioners/config_partitioner.py
@@ -17,7 +17,9 @@
     Partitioner,
     PartitionResult,
 )
+from torch.fx import GraphModule
 from torch.fx.passes.infra.partitioner import Partition
+from executorch.exir.graph_module import get_control_flow_submodules
 
 
 def format_target_name(target_name: str) -> str:
@@ -160,11 +162,11 @@ def filter_fn(node: torch.fx.Node) -> bool:
         return (do_not_decomp, filter_fn)
 
     def get_matched_nodes_from_configs(
-        self, ep: ExportedProgram
+        self, ep: ExportedProgram, gm: Optional[GraphModule] = None
     ) -> List[List[torch.fx.Node]]:
         # gather supported nodes
         matched_nodes = []
-        gm = ep.graph_module
+        gm = gm or ep.graph_module
         for node in gm.graph.nodes:
             if node.op == "call_function":
                 target = format_target_name(node.target.__name__)
@@ -175,17 +177,19 @@ def get_matched_nodes_from_configs(
 
         return matched_nodes
 
-    def generate_partitions(self, ep: ExportedProgram) -> List[Partition]:
-        matched_nodes = self.get_matched_nodes_from_configs(ep)
+    def generate_partitions(self, ep: ExportedProgram, gm: Optional[GraphModule] = None) -> List[Partition]:
+        gm = gm or ep.graph_module
+        matched_nodes = self.get_matched_nodes_from_configs(ep, gm)
         # create partitions
         partitions = generate_partitions_from_list_of_nodes(
-            ep.graph_module,
+            gm,
             matched_nodes,
         )
         return partitions
 
-    def partition(self, exported_program: ExportedProgram) -> PartitionResult:
-        partitions = self.generate_partitions(exported_program)
+    def partition(self, exported_program: ExportedProgram, graph_module: Optional[GraphModule] = None) -> PartitionResult:
+        graph_module = graph_module or exported_program.graph_module
+        partitions = self.generate_partitions(exported_program, graph_module)
 
         # tag nodes
         partition_tags: Dict[str, DelegationSpec] = {}
@@ -199,6 +203,10 @@ def partition(self, exported_program: ExportedProgram) -> PartitionResult:
                 node.meta["delegation_tag"] = delegation_tag
                 partition_tags[delegation_tag] = self.delegation_spec
 
+        for _, submodule, _ in get_control_flow_submodules(graph_module):
+            # pyre-ignore
+            self.partition(exported_program, submodule)
+
         return PartitionResult(
             tagged_exported_program=exported_program, partition_tags=partition_tags
         )