Change memory planning API to accept full algorithm as argument as opposed to string name (#6130)

pytorchbot · Eashan Garg · web-flow · commit a2bc6bdf54bb · 2024-10-10T20:45:31.000-07:00
Change memory planning API to accept full algorithm as argument as opposed to string name (#4727) Summary: Pull Request resolved: #4727 Executorch memory planning currently accepts a string identifier to represent the desired algorithm. However, this makes it difficult to pass custom arguments to write more customized memory planning algorithms. This change allows users to pass the full memory planning function as an argument as opposed to just the string identifier. Core changes in: - fbcode/executorch/exir/passes/memory_planning_pass.py - fbcode/executorch/exir/tests/test_memory_planning.py Remaining changes are just to enforce compliance with new API at all call sites in codebase NOTE: A less intrusive change could be to allow argument to be either string or entire custom functions. I opted for just passing only functions to simplify and avoid confusion Reviewed By: zonglinpeng, hsharma35, mcremon-meta Differential Revision: D60433641 fbshipit-source-id: 0fe3677b7c3f4c3763cb1b4fe6d28ef814f2ecf9 (cherry picked from commit 618466e) Co-authored-by: Eashan Garg <skrtskrt@meta.com>
diff --git a/backends/qualcomm/tests/utils.py b/backends/qualcomm/tests/utils.py
@@ -350,7 +350,6 @@ def lower_module_and_test_output(
                 # Therefore, won't want to pre-allocate
                 # by memory manager in runtime.
                 memory_planning_pass=MemoryPlanningPass(
-                    memory_planning_algo="greedy",
                     alloc_graph_input=not self.shared_buffer,
                     alloc_graph_output=not self.shared_buffer,
                 ),
diff --git a/backends/vulkan/vulkan_preprocess.py b/backends/vulkan/vulkan_preprocess.py
@@ -57,7 +57,7 @@ def preprocess(  # noqa: C901
             MeanToSumDiv(),
             SpecPropPass(),
             ConstraintBasedSymShapeEvalPass(),
-            MemoryPlanningPass("greedy"),
+            MemoryPlanningPass(),
         ]
 
         new_gm = program.graph_module
diff --git a/docs/source/compiler-memory-planning.md b/docs/source/compiler-memory-planning.md
@@ -32,7 +32,6 @@ The `MemoryPlanningPass` exposes the option to not memory plan program inputs an
 program = edge_program.to_executorch(
             exir.ExecutorchBackendConfig(
                 memory_planning_pass=MemoryPlanningPass(
-                    memory_planning_algo="greedy",
                     alloc_graph_input=False, # Inputs will not be memory planned, the data_ptr for input tensors after model load will be nullptr
                     alloc_graph_output=True, # Outputs will be memory planned, the data_ptr for output tensors after model load will be in the `planned_memory`.
                 )
@@ -77,7 +76,7 @@ Then later when lowering to ExecuTorch you can use your custom plan in the follo
 program = edge_program.to_executorch(
             exir.ExecutorchBackendConfig(
                 memory_planning_pass=CustomPoolMemoryPlanningPass(
-                    memory_planning_algo="greedy",
+                    memory_planning_algo=greedy,
                 )
             )
         )
diff --git a/docs/source/tutorials_source/export-to-executorch-tutorial.py b/docs/source/tutorials_source/export-to-executorch-tutorial.py
@@ -523,9 +523,7 @@ def forward(self, a, x, b):
 executorch_program: ExecutorchProgramManager = edge_program.to_executorch(
     ExecutorchBackendConfig(
         passes=[],  # User-defined passes
-        memory_planning_pass=MemoryPlanningPass(
-            "greedy"
-        ),  # Default memory planning pass
+        memory_planning_pass=MemoryPlanningPass(),  # Default memory planning pass
     )
 )
 
diff --git a/examples/mediatek/model_export_scripts/llama.py b/examples/mediatek/model_export_scripts/llama.py
@@ -365,7 +365,6 @@ def export_to_et_ir(
         executorch_program = delegated_program.to_executorch(
             config=exir.ExecutorchBackendConfig(
                 memory_planning_pass=exir.passes.MemoryPlanningPass(
-                    memory_planning_algo="greedy",
                     alloc_graph_input=False,
                     alloc_graph_output=False,
                 ),
diff --git a/examples/models/llava/export_llava.py b/examples/models/llava/export_llava.py
@@ -233,7 +233,7 @@ def export_all(llava_model: LlavaModel):
             passes=[
                 QuantFusionPass(),
             ],
-            memory_planning_pass=MemoryPlanningPass("greedy", alloc_graph_input=False),
+            memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
             sym_shape_eval_pass={
                 "image_encoder": ConstraintBasedSymShapeEvalPass(),
                 "text_model": ConstraintBasedSymShapeEvalPass(),
diff --git a/examples/qualcomm/oss_scripts/llama2/llama.py b/examples/qualcomm/oss_scripts/llama2/llama.py
@@ -311,7 +311,6 @@ def lowering_modules(
             # Therefore, won't want to pre-allocate
             # by memory manager in runtime.
             memory_planning_pass=MemoryPlanningPass(
-                memory_planning_algo="greedy",
                 alloc_graph_input=False,
                 alloc_graph_output=False,
             ),
diff --git a/examples/qualcomm/qaihub_scripts/utils/export.py b/examples/qualcomm/qaihub_scripts/utils/export.py
@@ -220,7 +220,6 @@ def compile(args):
     )
     # setup memory planning
     memory_planning_pass = MemoryPlanningPass(
-        memory_planning_algo="greedy",
         alloc_graph_input=args.allocate_graph_io,
         alloc_graph_output=args.allocate_graph_io,
     )
diff --git a/examples/qualcomm/utils.py b/examples/qualcomm/utils.py
@@ -285,7 +285,6 @@ def build_executorch_binary(
         # Therefore, won't want to pre-allocate
         # by memory manager in runtime.
         memory_planning_pass=MemoryPlanningPass(
-            memory_planning_algo="greedy",
             alloc_graph_input=not shared_buffer,
             alloc_graph_output=not shared_buffer,
         ),
diff --git a/exir/capture/_config.py b/exir/capture/_config.py
@@ -56,9 +56,7 @@ class ExecutorchBackendConfig:
 
     # A single memory planning pass can be defined for all the programs in the
     # EdgeProgramManager or can be defined per program.
-    memory_planning_pass: Union[PassType, Dict[str, PassType]] = MemoryPlanningPass(
-        "greedy"
-    )
+    memory_planning_pass: Union[PassType, Dict[str, PassType]] = MemoryPlanningPass()
     to_out_var_pass: PassType = ToOutVarPass(ignore_to_out_var_failure=False)
     dynamic_memory_planning_mode: DynamicMemoryPlanningMode = (
         DynamicMemoryPlanningMode.UPPER_BOUND
diff --git a/exir/emit/test/test_emit.py b/exir/emit/test/test_emit.py
@@ -1145,7 +1145,6 @@ def forward(self, k: torch.Tensor) -> torch.Tensor:
         config = exir.ExecutorchBackendConfig(
             sym_shape_eval_pass=ConstraintBasedSymShapeEvalPass(),
             memory_planning_pass=MemoryPlanningPass(
-                memory_planning_algo="greedy",
                 # allow_lifetime_and_storage_overlap: bool = False,
                 alloc_graph_input=True,
                 alloc_graph_output=False,
@@ -1606,9 +1605,7 @@ def forward(self, x):
         )
         model = model.to_executorch(
             config=ExecutorchBackendConfig(
-                memory_planning_pass=MemoryPlanningPass(
-                    "greedy", alloc_graph_input=False
-                ),
+                memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
                 sym_shape_eval_pass=ConstraintBasedSymShapeEvalPass(),
             )
         )
diff --git a/exir/lowered_backend_module.py b/exir/lowered_backend_module.py
@@ -326,7 +326,7 @@ def program(
             verifiers=[lowered_exported_program.verifier],
         )
         if memory_planning is None:
-            memory_planning = MemoryPlanningPass("greedy")
+            memory_planning = MemoryPlanningPass()
         exported_program = _transform(exported_program, SpecPropPass(), memory_planning)
         emitted_program = emit_program(
             exported_program, emit_stacktrace=emit_stacktrace
diff --git a/exir/memory_planning.py b/exir/memory_planning.py
@@ -18,12 +18,7 @@
 from executorch.exir import memory
 from executorch.exir.control_flow import while_loop as exir_while
 from executorch.exir.delegate import executorch_call_delegate
-from executorch.exir.error import (
-    ExportError,
-    ExportErrorType,
-    internal_assert,
-    InternalError,
-)
+from executorch.exir.error import internal_assert, InternalError
 from executorch.exir.operator.convert import is_inplace_variant, is_out_variant
 from executorch.exir.schema import TensorShapeDynamism
 from executorch.exir.tensor import TensorSpec
@@ -255,17 +250,6 @@ def verify_graph_input_output(self) -> None:
             ), f"Misallocate graph output {graph_output_allocated} v.s. {self.alloc_graph_output}"
 
 
-def register_algo(fn: Callable[..., List[int]]) -> Callable[..., List[int]]:
-    algo_name = fn.__name__
-    if algo_name in REGISTERED_ALGOS:
-        raise ExportError(
-            ExportErrorType.VIOLATION_OF_SPEC,
-            f"Re-registering memory planning algorithm {algo_name}",
-        )
-    REGISTERED_ALGOS[algo_name] = fn
-    return fn
-
-
 def _is_out_var_node(node: torch.fx.Node) -> bool:
     return (
         node.op == "call_function"
@@ -561,7 +545,6 @@ def get_node_tensor_specs(
         ]
 
 
-@register_algo
 def greedy(
     graph_module: torch.fx.GraphModule,
     alignment: int,
@@ -615,7 +598,6 @@ def greedy(
     return total_sizes
 
 
-@register_algo
 def naive(
     graph_module: torch.fx.GraphModule,
     alignment: int,
@@ -656,15 +638,6 @@ def _allocate_buf(bufsizes: List[int], mem_id: int, allocated: int) -> int:
     return bufsizes
 
 
-def get_algo(algo_name: str) -> Callable[..., List[int]]:
-    if algo_name not in REGISTERED_ALGOS:
-        raise ExportError(
-            ExportErrorType.NOT_SUPPORTED,
-            f"Memory planning algorithm '{algo_name}' not found",
-        )
-    return REGISTERED_ALGOS[algo_name]
-
-
 def get_cond_nodes(graph_module: torch.fx.GraphModule) -> Iterable[Node]:
     for nd in graph_module.graph.nodes:
         if nd.target is torch.ops.higher_order.cond:
diff --git a/exir/passes/memory_planning_pass.py b/exir/passes/memory_planning_pass.py
@@ -6,16 +6,16 @@
 
 import logging
 import warnings
-from typing import Optional
+from typing import Callable, List, Optional
 
 import torch
 from executorch.exir.error import internal_assert
 from executorch.exir.memory import alloc
 from executorch.exir.memory_planning import (
     _is_out_var_node,
     apply_algo,
-    get_algo,
     get_node_tensor_specs,
+    greedy,
     Verifier,
 )
 from executorch.exir.operator.convert import get_out_args_from_opoverload
@@ -27,7 +27,7 @@
 class MemoryPlanningPass(PassBase):
     def __init__(
         self,
-        memory_planning_algo: str = "greedy",
+        memory_planning_algo: Callable[..., List[int]] = greedy,
         allow_lifetime_and_storage_overlap: bool = False,
         alloc_graph_input: bool = True,
         alloc_graph_output: bool = True,
@@ -96,14 +96,13 @@ def run(
         memory_planning_algo
         """
         self._set_alloc_node_spec(graph_module)
-        algo = get_algo(self.memory_planning_algo)
         # TODO(shunting) if people have concern of adding a field to GraphModule
         # directly, we should define a GraphModule subclass that we can add our
         # customized fields. Using the graph_module object to convey information across
         # passes/stages is quite natural and avoid yet another 'context' data structure
         # to do the job.
         _ = apply_algo(
-            algo,
+            self.memory_planning_algo,
             graph_module,
             self.alignment,
             graph_signature,
@@ -125,7 +124,7 @@ def run(
                 self.allow_lifetime_and_storage_overlap
             )
             logging.debug(
-                f"The {self.memory_planning_algo} algorithm reuses storage for {num_reuse_pairs} pair of tensors"
+                f"The {getattr(self.memory_planning_algo, '__name__', repr(self.memory_planning_algo))} algorithm reuses storage for {num_reuse_pairs} pair of tensors"
             )
         verifier.verify_graph_input_output()
         return PassResult(graph_module, True)
diff --git a/exir/program/TARGETS b/exir/program/TARGETS
@@ -22,6 +22,7 @@ python_library(
         "//caffe2:torch",
         "//executorch/exir:error",
         "//executorch/exir:graph_module",
+        "//executorch/exir:pass_base",
         "//executorch/exir:pass_manager",
         "//executorch/exir:print_program",
         "//executorch/exir:schema",
diff --git a/exir/program/test/test_program.py b/exir/program/test/test_program.py
@@ -250,12 +250,10 @@ def test_executorch_manager_multi_config(self):
         def get_executorch_memory_planning_passes() -> Dict[str, MemoryPlanningPass]:
             return {
                 "forward": MemoryPlanningPass(
-                    memory_planning_algo="greedy",
                     alloc_graph_input=True,
                     alloc_graph_output=False,
                 ),
                 "foo": MemoryPlanningPass(
-                    memory_planning_algo="greedy",
                     alloc_graph_input=False,
                     alloc_graph_output=True,
                 ),
diff --git a/exir/tests/test_memory_planning.py b/exir/tests/test_memory_planning.py
@@ -17,6 +17,8 @@
 from executorch.exir.memory_planning import (
     filter_nodes,
     get_node_tensor_specs,
+    greedy,
+    naive,
     Verifier,
 )
 from executorch.exir.pass_base import PassResult
@@ -208,7 +210,7 @@ def forward(self, a: torch.Tensor) -> torch.Tensor:
 
 def maketest(
     module_cls: Type[torch.nn.Module],
-    criteria: Optional[List[Tuple[str, bool]]] = None,
+    criteria: Optional[List[Tuple[Callable[..., List[int]], bool]]] = None,
     extra_check: Optional[Callable[..., None]] = None,
     use_functionalization: bool = True,
     alloc_graph_input: bool = True,
@@ -222,13 +224,15 @@ def wrapper(self: "TestMemoryPlanning") -> None:
         if not criteria:
             criteria = [
                 # naive algorithm does not reuse tensor storages
-                ("naive", False),
+                (naive, False),
                 # greedy algorithm should reuse tensor storages in the testing model
-                ("greedy", True),
+                (greedy, True),
             ]
 
         for algo, expect_reuse in criteria:
-            print(f"algo {algo}, expect_reuse {expect_reuse}")
+            print(
+                f"algo {getattr(algo, '__name__', repr(algo))}, expect_reuse {expect_reuse}"
+            )
             eager_module = module_cls().eval()
             inputs = eager_module.get_random_inputs()
             graph_module = (
@@ -353,8 +357,8 @@ def verify_overlap_placeholders(
     test_return_two: Callable[..., None] = maketest(
         ModuleReturnTwo,
         criteria=[
-            ("naive", False),
-            ("greedy", True),
+            (naive, False),
+            (greedy, True),
         ],
     )
 
@@ -363,8 +367,8 @@ def verify_overlap_placeholders(
     test_list_arg: Callable[..., None] = maketest(
         ModuleListArg,
         criteria=[
-            ("naive", False),
-            ("greedy", True),
+            (naive, False),
+            (greedy, True),
         ],
         extra_check=ModuleListArg.extra_check,
     )
@@ -466,20 +470,20 @@ def quantize(self, eager_model: nn.Module) -> nn.Module:
     @parameterized.expand(
         [
             (
-                "naive",
+                naive,
                 [(1, 0), (3, 0), (1, 4), (3, 4), (1, 8)],
                 [0, 12, 0, 8],
             ),
             (
-                "greedy",
+                greedy,
                 [(1, 0), (3, 0), (1, 4), (3, 4), (1, 0)],
                 [0, 8, 0, 8],
             ),
         ]
     )
     def test_multiple_pools(
         self,
-        algo: str,
+        algo: Callable[..., List[int]],
         expected_allocs: List[Tuple[int, int]],
         expected_bufsizes: List[int],
     ) -> None:
@@ -550,9 +554,7 @@ def count_planned_inputs(
 
         ep_no_input_planning = to_edge(export(model, inputs)).to_executorch(
             config=ExecutorchBackendConfig(
-                memory_planning_pass=MemoryPlanningPass(
-                    "greedy", alloc_graph_input=False
-                ),
+                memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
                 sym_shape_eval_pass=ConstraintBasedSymShapeEvalPass(),
             )
         )
@@ -572,9 +574,7 @@ def count_planned_inputs(
 
         ep_input_planning = to_edge(export(model, inputs)).to_executorch(
             config=ExecutorchBackendConfig(
-                memory_planning_pass=MemoryPlanningPass(
-                    "greedy", alloc_graph_input=True
-                ),
+                memory_planning_pass=MemoryPlanningPass(alloc_graph_input=True),
                 sym_shape_eval_pass=ConstraintBasedSymShapeEvalPass(),
             )
         )
diff --git a/exir/tests/test_passes.py b/exir/tests/test_passes.py
@@ -713,7 +713,7 @@ def test_alloc_node_spec(self) -> None:
         self.assertIsNotNone(new_gm_res)
         new_gm = new_gm_res.graph_module
 
-        new_gm_res = MemoryPlanningPass("greedy")(new_gm)
+        new_gm_res = MemoryPlanningPass()(new_gm)
         self.assertIsNotNone(new_gm_res)
         new_gm = new_gm_res.graph_module
 
diff --git a/exir/tests/test_remove_view_copy.py b/exir/tests/test_remove_view_copy.py
diff --git a/extension/llm/export/builder.py b/extension/llm/export/builder.py
diff --git a/test/end2end/exported_module.py b/test/end2end/exported_module.py
diff --git a/test/models/export_program.py b/test/models/export_program.py

Original file line number	Diff line number	Diff line change
`@@ -57,7 +57,7 @@ def preprocess( # noqa: C901`
`57`	`57`	`MeanToSumDiv(),`
`58`	`58`	`SpecPropPass(),`
`59`	`59`	`ConstraintBasedSymShapeEvalPass(),`
`60`		`- MemoryPlanningPass("greedy"),`
	`60`	`+ MemoryPlanningPass(),`
`61`	`61`	`]`
`62`	`62`
`63`	`63`	`new_gm = program.graph_module`
Original file line number	Diff line number	Diff line change
@@ -32,7 +32,6 @@ The `MemoryPlanningPass` exposes the option to not memory plan program inputs an
`32`	`32`	`program = edge_program.to_executorch(`
`33`	`33`	`exir.ExecutorchBackendConfig(`
`34`	`34`	`memory_planning_pass=MemoryPlanningPass(`
`35`		`- memory_planning_algo="greedy",`
`36`	`35`	`alloc_graph_input=False, # Inputs will not be memory planned, the data_ptr for input tensors after model load will be nullptr`
`37`	`36`	alloc_graph_output=True, # Outputs will be memory planned, the data_ptr for output tensors after model load will be in the `planned_memory`.
`38`	`37`	`)`
`@@ -77,7 +76,7 @@ Then later when lowering to ExecuTorch you can use your custom plan in the follo`
`77`	`76`	`program = edge_program.to_executorch(`
`78`	`77`	`exir.ExecutorchBackendConfig(`
`79`	`78`	`memory_planning_pass=CustomPoolMemoryPlanningPass(`
`80`		`- memory_planning_algo="greedy",`
	`79`	`+ memory_planning_algo=greedy,`
`81`	`80`	`)`
`82`	`81`	`)`
`83`	`82`	`)`
Original file line number	Diff line number	Diff line change
`@@ -523,9 +523,7 @@ def forward(self, a, x, b):`
`523`	`523`	`executorch_program: ExecutorchProgramManager = edge_program.to_executorch(`
`524`	`524`	`ExecutorchBackendConfig(`
`525`	`525`	`passes=[], # User-defined passes`
`526`		`- memory_planning_pass=MemoryPlanningPass(`
`527`		`- "greedy"`
`528`		`- ), # Default memory planning pass`
	`526`	`+ memory_planning_pass=MemoryPlanningPass(), # Default memory planning pass`
`529`	`527`	`)`
`530`	`528`	`)`
`531`	`529`
Original file line number	Diff line number	Diff line change
`@@ -220,7 +220,6 @@ def compile(args):`
`220`	`220`	`)`
`221`	`221`	`# setup memory planning`
`222`	`222`	`memory_planning_pass = MemoryPlanningPass(`
`223`		`- memory_planning_algo="greedy",`
`224`	`223`	`alloc_graph_input=args.allocate_graph_io,`
`225`	`224`	`alloc_graph_output=args.allocate_graph_io,`
`226`	`225`	`)`