Add TorchAO wrapper config to allow filter_fn for quantize_

ghstack-source-id: 53ff5b6
ghstack-comment-id: 3172341537
Pull-Request: #13264

Author: abhinaykukkadapu

backends/xnnpack/recipes/xnnpack_recipe_provider.py

@@ -25,6 +25,7 @@
     get_xnnpack_executorch_backend_config,
 )
 from executorch.export import (
+    AOQuantizationConfig,
     BackendRecipeProvider,
     ExportRecipe,
     LoweringRecipe,
@@ -57,31 +58,37 @@ def create_recipe(
         if recipe_type == XNNPackRecipeType.FP32:
             return self._build_fp32_recipe(recipe_type)
 
-        elif recipe_type == XNNPackRecipeType.INT8_DYNAMIC_PER_CHANNEL:
+        elif recipe_type == XNNPackRecipeType.PT2E_INT8_DYNAMIC_PER_CHANNEL:
             return self._build_quantized_recipe(
                 recipe_type, is_per_channel=True, is_dynamic=True
             )
 
-        elif recipe_type == XNNPackRecipeType.INT8_STATIC_PER_CHANNEL:
+        elif recipe_type == XNNPackRecipeType.PT2E_INT8_STATIC_PER_CHANNEL:
             return self._build_quantized_recipe(
                 recipe_type, is_per_channel=True, is_dynamic=False
             )
 
-        elif recipe_type == XNNPackRecipeType.INT8_STATIC_PER_TENSOR:
+        elif recipe_type == XNNPackRecipeType.PT2E_INT8_STATIC_PER_TENSOR:
             return self._build_quantized_recipe(
                 recipe_type, is_per_channel=False, is_dynamic=False
             )
 
-        elif recipe_type == XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL:
-            return self._build_int8da_intx_weight_recipe(
+        elif (
+            recipe_type
+            == XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL
+        ):
+            return self._build_torchao_quantized_recipe(
                 recipe_type=recipe_type,
                 is_per_channel=True,
                 weight_dtype=torch.int4,
             )
 
-        elif recipe_type == XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR:
+        elif (
+            recipe_type
+            == XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR
+        ):
             group_size = kwargs.get("group_size", 32)
-            return self._build_int8da_intx_weight_recipe(
+            return self._build_torchao_quantized_recipe(
                 recipe_type=recipe_type,
                 is_per_channel=False,
                 weight_dtype=torch.int4,
@@ -132,7 +139,7 @@ def _build_quantized_recipe(
             executorch_backend_config=get_xnnpack_executorch_backend_config(),
         )
 
-    def _build_int8da_intx_weight_recipe(
+    def _build_torchao_quantized_recipe(
         self,
         recipe_type: RecipeType,
         is_per_channel: bool = True,
@@ -141,17 +148,21 @@ def _build_int8da_intx_weight_recipe(
     ) -> ExportRecipe:
         if is_per_channel:
             weight_granularity = PerAxis(axis=0)
+            assert weight_dtype == torch.int4 or weight_dtype == torch.int8
         else:
             weight_granularity = PerGroup(group_size=group_size)
+            assert weight_dtype == torch.int4
 
-        config = Int8DynamicActivationIntxWeightConfig(
-            weight_dtype=weight_dtype,
-            weight_granularity=weight_granularity,
+        config = AOQuantizationConfig(
+            Int8DynamicActivationIntxWeightConfig(
+                weight_dtype=weight_dtype,
+                weight_granularity=weight_granularity,
+            )
         )
 
         quant_recipe = QuantizationRecipe(
             quantizers=None,
-            ao_base_config=[config],
+            ao_quantization_configs=[config],
         )
 
         return ExportRecipe(
@@ -162,7 +173,10 @@ def _build_int8da_intx_weight_recipe(
         )
 
     def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> None:
-        if recipe_type == XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR:
+        if (
+            recipe_type
+            == XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR
+        ):
             expected_keys = {"group_size"}
             unexpected = set(kwargs.keys()) - expected_keys
             if unexpected:

backends/xnnpack/recipes/xnnpack_recipe_types.py

@@ -13,19 +13,22 @@ class XNNPackRecipeType(RecipeType):
     """XNNPACK-specific recipe types"""
 
     FP32 = "fp32"
+
+    ## PT2E-based quantization recipes
     # INT8 Dynamic Quantization
-    INT8_DYNAMIC_PER_CHANNEL = "int8_dynamic_per_channel"
+    PT2E_INT8_DYNAMIC_PER_CHANNEL = "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"
+
+    ## TorchAO-based quantization recipes
     # INT8 Dynamic Activations INT4 Weight Quantization, Axis = 0
-    INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL = "int8da_int4w_per_channel"
+    TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL = (
+        "torchao_int8da_int4w_per_channel"
+    )
     # INT8 Dynamic Activations INT4 Weight Quantization, default group_size = 32
     # can be overriden by group_size kwarg
-    INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR = "int8da_int4w_per_tensor"
-    # INT8 Static Activations INT4 Weight Quantization
-    INT8_STATIC_ACT_INT4_WEIGHT_PER_CHANNEL = "int8a_int4w_per_channel"
-    INT8_STATIC_ACT_INT4_WEIGHT_PER_TENSOR = "int8a_int44w_per_tensor"
-    # INT8 Static Quantization, needs calibration dataset
-    INT8_STATIC_PER_CHANNEL = "int8_static_per_channel"
-    INT8_STATIC_PER_TENSOR = "int8_static_per_tensor"
+    TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR = "torchao_int8da_int4w_per_tensor"
 
     @classmethod
     def get_backend_name(cls) -> str:

backends/xnnpack/test/recipes/test_xnnpack_recipes.py

@@ -6,6 +6,7 @@
 
 # pyre-strict
 
+import logging
 import unittest
 
 import torch
@@ -19,8 +20,10 @@
 from executorch.examples.xnnpack import MODEL_NAME_TO_OPTIONS, QuantType
 from executorch.exir.schema import DelegateCall, Program
 from executorch.export import export, ExportRecipe, recipe_registry
+from export.types import StageType
 from torch import nn
 from torch.testing._internal.common_quantization import TestHelperModules
+from torchao.quantization.utils import compute_error
 
 
 class TestXnnpackRecipes(unittest.TestCase):
@@ -38,6 +41,29 @@ def check_fully_delegated(self, program: Program) -> None:
         self.assertEqual(len(instructions), 1)
         self.assertIsInstance(instructions[0].instr_args, DelegateCall)
 
+    # pyre-ignore
+    def _compare_eager_quantized_model_outputs(
+        self, session, example_inputs, atol: float
+    ) -> None:
+        """Utility to compare eager quantized model output with session output after xnnpack lowering"""
+        torch_export_stage_output = session.get_stage_artifacts()[
+            StageType.TORCH_EXPORT
+        ]
+        eager_quantized_model = torch_export_stage_output.data["forward"].module()
+        output = session.run_method("forward", example_inputs[0])[0]
+        expected = eager_quantized_model(*example_inputs[0])
+        Tester._assert_outputs_equal(output, expected, atol=atol)
+
+    def _compare_eager_unquantized_model_outputs(
+        self, session, eager_unquantized_model, example_inputs, sqnr_threshold=20
+    ):
+        """Utility to compare eager unquantized model output with session output using SQNR"""
+        quantized_output = session.run_method("forward", example_inputs[0])[0]
+        original_output = eager_unquantized_model(*example_inputs[0])
+        error = compute_error(original_output, quantized_output)
+        print(f"{self._testMethodName} - SQNR: {error} dB")
+        self.assertTrue(error > sqnr_threshold)
+
     def test_basic_recipe(self) -> None:
         m_eager = TestHelperModules.TwoLinearModule().eval()
         example_inputs = [(torch.randn(9, 8),)]
@@ -46,18 +72,13 @@ def test_basic_recipe(self) -> None:
             example_inputs=example_inputs,
             export_recipe=ExportRecipe.get_recipe(XNNPackRecipeType.FP32),
         )
-        self.assertTrue(
-            torch.allclose(
-                session.run_method("forward", example_inputs[0])[0],
-                m_eager(*example_inputs[0]),
-                atol=1e-3,
-            )
-        )
+        self._compare_eager_quantized_model_outputs(session, example_inputs, 1e-3)
         self.check_fully_delegated(session.get_executorch_program())
+        self._compare_eager_unquantized_model_outputs(session, m_eager, example_inputs)
 
     def test_int8_dynamic_quant_recipe(self) -> None:
         test_cases = [
-            ExportRecipe.get_recipe(XNNPackRecipeType.INT8_DYNAMIC_PER_CHANNEL),
+            ExportRecipe.get_recipe(XNNPackRecipeType.PT2E_INT8_DYNAMIC_PER_CHANNEL),
         ]
 
         for export_recipe in test_cases:
@@ -70,19 +91,18 @@ def test_int8_dynamic_quant_recipe(self) -> None:
                         example_inputs=example_inputs,
                         export_recipe=export_recipe,
                     )
-                    self.assertTrue(
-                        torch.allclose(
-                            session.run_method("forward", example_inputs[0])[0],
-                            m_eager(*example_inputs[0]),
-                            atol=1e-1,
-                        )
+                    self._compare_eager_quantized_model_outputs(
+                        session, example_inputs, 1e-1
                     )
                     self.check_fully_delegated(session.get_executorch_program())
+                    self._compare_eager_unquantized_model_outputs(
+                        session, m_eager, example_inputs
+                    )
 
     def test_int8_static_quant_recipe(self) -> None:
         test_cases = [
-            ExportRecipe.get_recipe(XNNPackRecipeType.INT8_STATIC_PER_CHANNEL),
-            ExportRecipe.get_recipe(XNNPackRecipeType.INT8_STATIC_PER_TENSOR),
+            ExportRecipe.get_recipe(XNNPackRecipeType.PT2E_INT8_STATIC_PER_CHANNEL),
+            ExportRecipe.get_recipe(XNNPackRecipeType.PT2E_INT8_STATIC_PER_TENSOR),
         ]
 
         for export_recipe in test_cases:
@@ -95,14 +115,13 @@ def test_int8_static_quant_recipe(self) -> None:
                         example_inputs=example_inputs,
                         export_recipe=export_recipe,
                     )
-                    self.assertTrue(
-                        torch.allclose(
-                            session.run_method("forward", example_inputs[0])[0],
-                            m_eager(*example_inputs[0]),
-                            atol=1e-1,
-                        )
+                    self._compare_eager_quantized_model_outputs(
+                        session, example_inputs, 1e-2
                     )
                     self.check_fully_delegated(session.get_executorch_program())
+                    self._compare_eager_unquantized_model_outputs(
+                        session, m_eager, example_inputs
+                    )
 
     def test_8a4w_recipe(self) -> None:
         class SimpleLinearModel(nn.Module):
@@ -116,10 +135,10 @@ def forward(self, x) -> torch.Tensor:
 
         test_cases = [
             ExportRecipe.get_recipe(
-                XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL,
+                XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL,
             ),
             ExportRecipe.get_recipe(
-                XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR,
+                XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR,
                 group_size=32,
             ),
         ]
@@ -133,23 +152,22 @@ def forward(self, x) -> torch.Tensor:
                     example_inputs=example_inputs,
                     export_recipe=export_recipe,
                 )
-                self.assertTrue(
-                    torch.allclose(
-                        session.run_method("forward", example_inputs[0])[0],
-                        model(*example_inputs[0]),
-                        atol=1e-2,
-                    )
-                )
                 self.check_fully_delegated(session.get_executorch_program())
+                self._compare_eager_quantized_model_outputs(
+                    session, example_inputs, 1e-3
+                )
+                self._compare_eager_unquantized_model_outputs(
+                    session, model, example_inputs, sqnr_threshold=15
+                )
 
     def _get_recipe_for_quant_type(self, quant_type: QuantType) -> XNNPackRecipeType:
         # Map QuantType to corresponding recipe name.
         if quant_type == QuantType.STATIC_PER_CHANNEL:
-            return XNNPackRecipeType.INT8_STATIC_PER_CHANNEL
+            return XNNPackRecipeType.PT2E_INT8_STATIC_PER_CHANNEL
         elif quant_type == QuantType.DYNAMIC_PER_CHANNEL:
-            return XNNPackRecipeType.INT8_DYNAMIC_PER_CHANNEL
+            return XNNPackRecipeType.PT2E_INT8_DYNAMIC_PER_CHANNEL
         elif quant_type == QuantType.STATIC_PER_TENSOR:
-            return XNNPackRecipeType.INT8_STATIC_PER_TENSOR
+            return XNNPackRecipeType.PT2E_INT8_STATIC_PER_TENSOR
         elif quant_type == QuantType.NONE:
             return XNNPackRecipeType.FP32
         else:
@@ -224,12 +242,13 @@ def test_validate_recipe_kwargs_int4_tensor_with_valid_group_size(
 
         # Should not raise any exception
         recipe_w_default_group = provider.create_recipe(
-            XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR
+            XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR
         )
         self.assertIsNotNone(recipe_w_default_group)
 
         recipe = provider.create_recipe(
-            XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR, group_size=64
+            XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR,
+            group_size=64,
         )
         self.assertIsNotNone(recipe)
 
@@ -240,7 +259,7 @@ def test_validate_recipe_kwargs_int4_tensor_with_invalid_group_size(
 
         with self.assertRaises(ValueError) as cm:
             provider.create_recipe(
-                XNNPackRecipeType.INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR,
+                XNNPackRecipeType.TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR,
                 group_size="32",  # String instead of int
             )
 

export/__init__.py

@@ -15,12 +15,19 @@
 """
 
 from .export import export, ExportSession
-from .recipe import ExportRecipe, LoweringRecipe, QuantizationRecipe, RecipeType
+from .recipe import (
+    AOQuantizationConfig,
+    ExportRecipe,
+    LoweringRecipe,
+    QuantizationRecipe,
+    RecipeType,
+)
 from .recipe_provider import BackendRecipeProvider
 from .recipe_registry import recipe_registry
 from .types import StageType
 
 __all__ = [
+    "AOQuantizationConfig",
     "StageType",
     "ExportRecipe",
     "LoweringRecipe",

export/recipe.py

@@ -6,7 +6,9 @@
 from abc import ABCMeta, abstractmethod
 from dataclasses import dataclass
 from enum import Enum, EnumMeta
-from typing import List, Optional, Sequence
+from typing import Callable, List, Optional, Sequence
+
+import torch
 
 from executorch.exir._warnings import experimental
 
@@ -64,6 +66,20 @@ class Mode(str, Enum):
     RELEASE = "release"
 
 
+@dataclass
+class AOQuantizationConfig:
+    """
+    Configuration for torchao quantization with optional filter function.
+
+    Attributes:
+        ao_base_config: The AOBaseConfig for quantization
+        filter_fn: Optional filter function to selectively apply quantization
+    """
+
+    ao_base_config: AOBaseConfig
+    filter_fn: Optional[Callable[[torch.nn.Module, str], bool]] = None
+
+
 @dataclass
 class QuantizationRecipe:
     """
@@ -73,11 +89,12 @@ class QuantizationRecipe:
 
     Attributes:
         quantizers: Optional list of quantizers for model quantization
-        ao_base_config: Optional list of AO base configurations
+        ao_quantization_configs: Optional list of AOQuantizationConfig objects that pair
+                                 AOBaseConfig with optional filter functions
     """
 
     quantizers: Optional[List[Quantizer]] = None
-    ao_base_config: Optional[List[AOBaseConfig]] = None
+    ao_quantization_configs: Optional[List[AOQuantizationConfig]] = None
 
     def get_quantizers(self) -> Optional[List[Quantizer]]:
         """