[Executorch][target recipes] Add target based recipes for lowering models to a target device (#13983)

Co-authored-by: Abhinay Kukkadapu <[email protected]>

Author: pytorchbot

backends/apple/coreml/TARGETS

@@ -61,16 +61,21 @@ runtime.python_library(
 )
 
 runtime.python_library(
-    name = "recipes",
-    srcs = glob([
-        "recipes/*.py",
-    ]),
+    name = "coreml_recipes",
+    srcs = [
+        "recipes/__init__.py",
+        "recipes/coreml_recipe_provider.py"
+    ],
     visibility = [
         "@EXECUTORCH_CLIENTS",
+        "//executorch/export/...",
     ],
     deps = [
         "fbsource//third-party/pypi/coremltools:coremltools",
+        ":coreml_recipe_types",
         ":backend",
+        ":partitioner",
+        ":quantizer",
         "//caffe2:torch",
         "//executorch/exir:lib",
         "//executorch/exir/backend:compile_spec_schema",
@@ -80,6 +85,20 @@ runtime.python_library(
     ],
 )
 
+runtime.python_library(
+    name = "coreml_recipe_types",
+    srcs = [
+        "recipes/coreml_recipe_types.py",
+    ],
+    visibility = [
+        "@EXECUTORCH_CLIENTS",
+        "//executorch/export/...",
+    ],
+    deps = [
+        "//executorch/export:recipe",
+    ],
+)
+
 runtime.cxx_python_extension(
     name = "executorchcoreml",
     srcs = [
@@ -124,7 +143,7 @@ runtime.python_test(
         "fbsource//third-party/pypi/pytest:pytest",
         ":partitioner",
         ":quantizer",
-        ":recipes",
+        ":coreml_recipes",
         "//caffe2:torch",
         "//pytorch/vision:torchvision",
         "fbsource//third-party/pypi/scikit-learn:scikit-learn",

backends/apple/coreml/recipes/coreml_recipe_provider.py

@@ -3,6 +3,7 @@
 # Please refer to the license found in the LICENSE file in the root directory of the source tree.
 
 
+import logging
 from typing import Any, Optional, Sequence
 
 import coremltools as ct
@@ -111,8 +112,9 @@ def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> Non
 
         unexpected = set(kwargs.keys()) - expected_keys
         if unexpected:
-            raise ValueError(
-                f"Recipe '{recipe_type.value}' received unexpected parameters: {list(unexpected)}"
+            logging.warning(
+                f"CoreML recipe '{recipe_type.value}' ignoring unexpected parameters: {list(unexpected)}. "
+                f"Expected parameters: {list(expected_keys)}"
             )
 
         self._validate_base_parameters(kwargs)
@@ -121,7 +123,13 @@ def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> Non
 
     def _get_expected_keys(self, recipe_type: RecipeType) -> set:
         """Get expected parameter keys for a recipe type"""
-        common_keys = {"minimum_deployment_target", "compute_unit"}
+        common_keys = {
+            "minimum_deployment_target",
+            "compute_unit",
+            "skip_ops_for_coreml_delegation",
+            "lower_full_graph",
+            "take_over_constant_data",
+        }
 
         if recipe_type in [
             CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_GROUP,
@@ -377,9 +385,19 @@ def _get_coreml_lowering_recipe(
         if minimum_deployment_target and minimum_deployment_target < ct.target.iOS18:
             take_over_mutable_buffer = False
 
+        # Extract additional partitioner parameters
+        skip_ops_for_coreml_delegation = kwargs.get(
+            "skip_ops_for_coreml_delegation", None
+        )
+        lower_full_graph = kwargs.get("lower_full_graph", False)
+        take_over_constant_data = kwargs.get("take_over_constant_data", True)
+
         partitioner = CoreMLPartitioner(
             compile_specs=compile_specs,
             take_over_mutable_buffer=take_over_mutable_buffer,
+            skip_ops_for_coreml_delegation=skip_ops_for_coreml_delegation,
+            lower_full_graph=lower_full_graph,
+            take_over_constant_data=take_over_constant_data,
         )
 
         edge_compile_config = EdgeCompileConfig(

backends/apple/coreml/test/test_coreml_recipes.py

@@ -185,14 +185,6 @@ def test_int4_weight_only_per_group_validation(self):
             )
         self.assertIn("must be positive", str(cm.exception))
 
-        # Test unexpected parameter
-        with self.assertRaises(ValueError) as cm:
-            self.provider.create_recipe(
-                CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_CHANNEL,
-                group_size=32,  # group_size not valid for per-channel
-            )
-        self.assertIn("unexpected parameters", str(cm.exception))
-
     def test_int8_weight_only_per_channel(self):
         """Test INT8 weight-only per-channel quantization"""
         model = TestHelperModules.TwoLinearModule().eval()
@@ -385,23 +377,6 @@ def forward(self, x):
         self._compare_eager_quantized_model_outputs(session, example_inputs, atol=1e-2)
         self._compare_eager_unquantized_model_outputs(session, model, example_inputs)
 
-    def test_pt2e_recipes_parameter_rejection(self):
-        """Test that PT2E recipes reject TorchAO-specific parameters"""
-        # PT2E recipes should reject TorchAO-specific parameters
-        pt2e_recipes = [
-            CoreMLRecipeType.PT2E_INT8_STATIC,
-            CoreMLRecipeType.PT2E_INT8_WEIGHT_ONLY,
-        ]
-        torchao_params = ["filter_fn", "group_size", "bits", "block_size"]
-
-        for recipe_type in pt2e_recipes:
-            for param in torchao_params:
-                with self.subTest(recipe=recipe_type.value, param=param):
-                    kwargs = {param: "dummy_value"}
-                    with self.assertRaises(ValueError) as cm:
-                        self.provider.create_recipe(recipe_type, **kwargs)
-                    self.assertIn("unexpected parameters", str(cm.exception).lower())
-
     def test_filter_fn_comprehensive(self):
         """Comprehensive test for filter_fn parameter functionality"""
 

backends/xnnpack/TARGETS

@@ -36,10 +36,7 @@ runtime.python_library(
     ],
     deps = [
         ":xnnpack_preprocess",
-        "//executorch/export:lib",
         "//executorch/backends/xnnpack/partition:xnnpack_partitioner",
         "//executorch/backends/xnnpack/utils:xnnpack_utils",
-        "//executorch/backends/xnnpack/recipes:xnnpack_recipe_provider",
-        "//executorch/backends/xnnpack/recipes:xnnpack_recipe_types",
     ],
 )

backends/xnnpack/__init__.py

@@ -4,18 +4,11 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from executorch.export import recipe_registry
-
 # Exposed Partitioners in XNNPACK Package
 from .partition.xnnpack_partitioner import (
     XnnpackDynamicallyQuantizedPartitioner,
     XnnpackPartitioner,
 )
-from .recipes.xnnpack_recipe_provider import XNNPACKRecipeProvider
-from .recipes.xnnpack_recipe_types import XNNPackRecipeType
-
-# Auto-register XNNPACK recipe provider
-recipe_registry.register_backend_recipe_provider(XNNPACKRecipeProvider())
 
 # Exposed Configs in XNNPACK Package
 from .utils.configs import (
@@ -34,7 +27,6 @@
     "XnnpackDynamicallyQuantizedPartitioner",
     "XnnpackPartitioner",
     "XnnpackBackend",
-    "XNNPackRecipeType",
     "capture_graph_for_xnnpack",
     "get_xnnpack_capture_config",
     "get_xnnpack_edge_compile_config",

backends/xnnpack/recipes/TARGETS

@@ -2,6 +2,22 @@ load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
 
 oncall("executorch")
 
+runtime.python_library(
+    name = "xnnpack_recipes",
+    srcs = [
+        "__init__.py",
+    ],
+    visibility = [
+        "//executorch/...",
+        "@EXECUTORCH_CLIENTS",
+    ],
+    deps = [
+        "//executorch/export:recipe_registry",
+        ":xnnpack_recipe_provider",
+        ":xnnpack_recipe_types",
+    ],
+)
+
 runtime.python_library(
     name = "xnnpack_recipe_provider",
     srcs = [
@@ -30,6 +46,6 @@ runtime.python_library(
         "@EXECUTORCH_CLIENTS",
     ],
     deps = [
-        "//executorch/export:lib",
+        "//executorch/export:recipe",
     ],
 )

backends/xnnpack/recipes/__init__.py

@@ -0,0 +1,19 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from executorch.export import recipe_registry
+
+from .xnnpack_recipe_provider import XNNPACKRecipeProvider
+from .xnnpack_recipe_types import XNNPackRecipeType
+
+# Auto-register XNNPACK recipe provider
+recipe_registry.register_backend_recipe_provider(XNNPACKRecipeProvider())
+
+
+__all__ = [
+    "XNNPACKRecipeProvider",
+    "XNNPackRecipeType",
+]

backends/xnnpack/recipes/xnnpack_recipe_provider.py

@@ -6,6 +6,7 @@
 
 # pyre-strict
 
+import logging
 from typing import Any, Optional, Sequence
 
 import torch
@@ -180,9 +181,9 @@ def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> Non
             expected_keys = {"group_size"}
             unexpected = set(kwargs.keys()) - expected_keys
             if unexpected:
-                raise ValueError(
-                    f"Recipe '{recipe_type.value}' only accepts 'group_size' parameter. "
-                    f"Unexpected parameters: {list(unexpected)}"
+                logging.warning(
+                    f"XNNPACK recipe '{recipe_type.value}' ignoring unexpected parameters: {list(unexpected)}. "
+                    f"Only 'group_size' is supported for this recipe."
                 )
             if "group_size" in kwargs:
                 group_size = kwargs["group_size"]
@@ -193,7 +194,7 @@ def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> Non
         elif kwargs:
             # All other recipes don't expect any kwargs
             unexpected = list(kwargs.keys())
-            raise ValueError(
-                f"Recipe '{recipe_type.value}' does not accept any parameters. "
-                f"Unexpected parameters: {unexpected}"
+            logging.warning(
+                f"XNNPACK recipe '{recipe_type.value}' ignoring unexpected parameters: {unexpected}. "
+                f"This recipe does not accept any parameters."
             )

backends/xnnpack/recipes/xnnpack_recipe_types.py

@@ -12,23 +12,25 @@
 class XNNPackRecipeType(RecipeType):
     """XNNPACK-specific recipe types"""
 
-    FP32 = "fp32"
+    FP32 = "xnnpack_fp32"
 
     ## PT2E-based quantization recipes
     # INT8 Dynamic Quantization
-    PT2E_INT8_DYNAMIC_PER_CHANNEL = "pt2e_int8_dynamic_per_channel"
+    PT2E_INT8_DYNAMIC_PER_CHANNEL = "xnnpack_pt2e_int8_dynamic_per_channel"
     # INT8 Static Quantization, needs calibration dataset
-    PT2E_INT8_STATIC_PER_CHANNEL = "pt2e_int8_static_per_channel"
-    PT2E_INT8_STATIC_PER_TENSOR = "pt2e_int8_static_per_tensor"
+    PT2E_INT8_STATIC_PER_CHANNEL = "xnnpack_pt2e_int8_static_per_channel"
+    PT2E_INT8_STATIC_PER_TENSOR = "xnnpack_pt2e_int8_static_per_tensor"
 
     ## TorchAO-based quantization recipes
     # INT8 Dynamic Activations INT4 Weight Quantization, Axis = 0
     TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL = (
-        "torchao_int8da_int4w_per_channel"
+        "xnnpack_torchao_int8da_int4w_per_channel"
     )
     # INT8 Dynamic Activations INT4 Weight Quantization, default group_size = 32
     # can be overriden by group_size kwarg
-    TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR = "torchao_int8da_int4w_per_tensor"
+    TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR = (
+        "xnnpack_torchao_int8da_int4w_per_tensor"
+    )
 
     @classmethod
     def get_backend_name(cls) -> str:

backends/xnnpack/test/TARGETS

@@ -105,7 +105,7 @@ runtime.python_test(
         "HTTPS_PROXY": "http://fwdproxy:8080",
     },
     deps = [
-        "//executorch/backends/xnnpack:xnnpack_delegate",
+       "//executorch/backends/xnnpack/recipes:xnnpack_recipes",
         "//executorch/export:lib",
         "//pytorch/vision:torchvision",  # @manual
         "//executorch/backends/xnnpack/test/tester:tester",