[executorch] Add coreml quant recipes (#13441)

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

Author: pytorchbot

backends/apple/coreml/TARGETS

@@ -120,11 +120,13 @@ runtime.python_test(
         "test/*.py",
     ]),
     deps = [
+        "fbsource//third-party/pypi/coremltools:coremltools",
         "fbsource//third-party/pypi/pytest:pytest",
         ":partitioner",
         ":quantizer",
         ":recipes",
         "//caffe2:torch",
         "//pytorch/vision:torchvision",
+        "fbsource//third-party/pypi/scikit-learn:scikit-learn",
     ],
 )

backends/apple/coreml/recipes/coreml_recipe_provider.py

@@ -6,6 +6,7 @@
 from typing import Any, Optional, Sequence
 
 import coremltools as ct
+import torch
 
 from executorch.backends.apple.coreml.compiler import CoreMLBackend
 from executorch.backends.apple.coreml.partition.coreml_partitioner import (
@@ -18,11 +19,15 @@
 
 from executorch.exir import EdgeCompileConfig
 from executorch.export import (
+    AOQuantizationConfig,
     BackendRecipeProvider,
     ExportRecipe,
     LoweringRecipe,
+    QuantizationRecipe,
     RecipeType,
 )
+from torchao.quantization.granularity import PerAxis, PerGroup
+from torchao.quantization.quant_api import IntxWeightOnlyConfig
 
 
 class CoreMLRecipeProvider(BackendRecipeProvider):
@@ -50,66 +55,321 @@ def create_recipe(
         # Validate kwargs
         self._validate_recipe_kwargs(recipe_type, **kwargs)
 
-        # Parse recipe type to get precision and compute unit
-        precision = None
         if recipe_type == CoreMLRecipeType.FP32:
-            precision = ct.precision.FLOAT32
+            return self._build_fp_recipe(recipe_type, ct.precision.FLOAT32, **kwargs)
         elif recipe_type == CoreMLRecipeType.FP16:
-            precision = ct.precision.FLOAT16
-
-        if precision is None:
-            raise ValueError(f"Unknown precision for recipe: {recipe_type.value}")
+            return self._build_fp_recipe(recipe_type, ct.precision.FLOAT16, **kwargs)
+        elif recipe_type == CoreMLRecipeType.PT2E_INT8_STATIC:
+            return self._build_pt2e_quantized_recipe(
+                recipe_type, activation_dtype=torch.quint8, **kwargs
+            )
+        elif recipe_type == CoreMLRecipeType.PT2E_INT8_WEIGHT_ONLY:
+            return self._build_pt2e_quantized_recipe(
+                recipe_type, activation_dtype=torch.float32, **kwargs
+            )
+        elif recipe_type == CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_CHANNEL:
+            return self._build_torchao_quantized_recipe(
+                recipe_type,
+                weight_dtype=torch.int4,
+                is_per_channel=True,
+                **kwargs,
+            )
+        elif recipe_type == CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_GROUP:
+            group_size = kwargs.pop("group_size", 32)
+            return self._build_torchao_quantized_recipe(
+                recipe_type,
+                weight_dtype=torch.int4,
+                is_per_channel=False,
+                group_size=group_size,
+                **kwargs,
+            )
+        elif recipe_type == CoreMLRecipeType.TORCHAO_INT8_WEIGHT_ONLY_PER_CHANNEL:
+            return self._build_torchao_quantized_recipe(
+                recipe_type, weight_dtype=torch.int8, is_per_channel=True, **kwargs
+            )
+        elif recipe_type == CoreMLRecipeType.TORCHAO_INT8_WEIGHT_ONLY_PER_GROUP:
+            group_size = kwargs.pop("group_size", 32)
+            return self._build_torchao_quantized_recipe(
+                recipe_type,
+                weight_dtype=torch.int8,
+                is_per_channel=False,
+                group_size=group_size,
+                **kwargs,
+            )
+        elif recipe_type == CoreMLRecipeType.CODEBOOK_WEIGHT_ONLY:
+            bits = kwargs.pop("bits")
+            block_size = kwargs.pop("block_size")
+            return self._build_codebook_quantized_recipe(
+                recipe_type, bits=bits, block_size=block_size, **kwargs
+            )
 
-        return self._build_recipe(recipe_type, precision, **kwargs)
+        return None
 
     def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> None:
-        if not kwargs:
-            return
-        expected_keys = {"minimum_deployment_target", "compute_unit"}
+        """Validate kwargs for each recipe type"""
+        expected_keys = self._get_expected_keys(recipe_type)
+
         unexpected = set(kwargs.keys()) - expected_keys
         if unexpected:
             raise ValueError(
-                f"CoreML Recipes only accept 'minimum_deployment_target' or 'compute_unit' as parameter. "
-                f"Unexpected parameters: {list(unexpected)}"
+                f"Recipe '{recipe_type.value}' received unexpected parameters: {list(unexpected)}"
             )
+
+        self._validate_base_parameters(kwargs)
+        self._validate_group_size_parameter(recipe_type, kwargs)
+        self._validate_codebook_parameters(recipe_type, kwargs)
+
+    def _get_expected_keys(self, recipe_type: RecipeType) -> set:
+        """Get expected parameter keys for a recipe type"""
+        common_keys = {"minimum_deployment_target", "compute_unit"}
+
+        if recipe_type in [
+            CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_GROUP,
+            CoreMLRecipeType.TORCHAO_INT8_WEIGHT_ONLY_PER_GROUP,
+        ]:
+            return common_keys | {"group_size", "filter_fn"}
+        elif recipe_type in [
+            CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_CHANNEL,
+            CoreMLRecipeType.TORCHAO_INT8_WEIGHT_ONLY_PER_CHANNEL,
+        ]:
+            return common_keys | {"filter_fn"}
+        elif recipe_type == CoreMLRecipeType.CODEBOOK_WEIGHT_ONLY:
+            return common_keys | {"bits", "block_size", "filter_fn"}
+        else:
+            return common_keys
+
+    def _validate_base_parameters(self, kwargs: Any) -> None:
+        """Validate minimum_deployment_target and compute_unit parameters"""
         if "minimum_deployment_target" in kwargs:
             minimum_deployment_target = kwargs["minimum_deployment_target"]
             if not isinstance(minimum_deployment_target, ct.target):
                 raise ValueError(
                     f"Parameter 'minimum_deployment_target' must be an enum of type ct.target, got {type(minimum_deployment_target)}"
                 )
+
         if "compute_unit" in kwargs:
             compute_unit = kwargs["compute_unit"]
             if not isinstance(compute_unit, ct.ComputeUnit):
                 raise ValueError(
                     f"Parameter 'compute_unit' must be an enum of type ct.ComputeUnit, got {type(compute_unit)}"
                 )
 
-    def _build_recipe(
+    def _validate_group_size_parameter(
+        self, recipe_type: RecipeType, kwargs: Any
+    ) -> None:
+        """Validate group_size parameter for applicable recipe types"""
+        if (
+            recipe_type
+            in [
+                CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_GROUP,
+                CoreMLRecipeType.TORCHAO_INT8_WEIGHT_ONLY_PER_GROUP,
+            ]
+            and "group_size" in kwargs
+        ):
+            group_size = kwargs["group_size"]
+            if not isinstance(group_size, int):
+                raise ValueError(
+                    f"Parameter 'group_size' must be an integer, got {type(group_size).__name__}: {group_size}"
+                )
+            if group_size <= 0:
+                raise ValueError(
+                    f"Parameter 'group_size' must be positive, got: {group_size}"
+                )
+
+    def _validate_codebook_parameters(
+        self, recipe_type: RecipeType, kwargs: Any
+    ) -> None:
+        """Validate bits and block_size parameters for codebook recipe type"""
+        if recipe_type != CoreMLRecipeType.CODEBOOK_WEIGHT_ONLY:
+            return
+
+        # Both bits and block_size must be present
+        if not ("bits" in kwargs and "block_size" in kwargs):
+            raise ValueError(
+                "Parameters 'bits' and 'block_size' must be present for codebook recipes"
+            )
+
+        if "bits" in kwargs:
+            bits = kwargs["bits"]
+            if not isinstance(bits, int):
+                raise ValueError(
+                    f"Parameter 'bits' must be an integer, got {type(bits).__name__}: {bits}"
+                )
+            if not (1 <= bits <= 8):
+                raise ValueError(
+                    f"Parameter 'bits' must be between 1 and 8, got: {bits}"
+                )
+
+        if "block_size" in kwargs:
+            block_size = kwargs["block_size"]
+            if not isinstance(block_size, list):
+                raise ValueError(
+                    f"Parameter 'block_size' must be a list, got {type(block_size).__name__}: {block_size}"
+                )
+
+    def _validate_and_set_deployment_target(
+        self, kwargs: Any, min_target: ct.target, quantization_type: str
+    ) -> None:
+        """Validate or set minimum deployment target for quantization recipes"""
+        minimum_deployment_target = kwargs.get("minimum_deployment_target", None)
+        if minimum_deployment_target and minimum_deployment_target < min_target:
+            raise ValueError(
+                f"minimum_deployment_target must be {str(min_target)} or higher for {quantization_type} quantization"
+            )
+        else:
+            # Default to the minimum target for this quantization type
+            kwargs["minimum_deployment_target"] = min_target
+
+    def _build_fp_recipe(
         self,
         recipe_type: RecipeType,
         precision: ct.precision,
         **kwargs: Any,
     ) -> ExportRecipe:
+        """Build FP32/FP16 recipe"""
         lowering_recipe = self._get_coreml_lowering_recipe(
             compute_precision=precision,
             **kwargs,
         )
 
         return ExportRecipe(
             name=recipe_type.value,
-            quantization_recipe=None,  # TODO - add quantization recipe
+            lowering_recipe=lowering_recipe,
+        )
+
+    def _build_pt2e_quantized_recipe(
+        self,
+        recipe_type: RecipeType,
+        activation_dtype: torch.dtype,
+        **kwargs: Any,
+    ) -> ExportRecipe:
+        """Build PT2E-based quantization recipe"""
+        from executorch.backends.apple.coreml.quantizer import CoreMLQuantizer
+
+        self._validate_and_set_deployment_target(kwargs, ct.target.iOS17, "pt2e")
+
+        # Validate activation_dtype
+        assert activation_dtype in [
+            torch.quint8,
+            torch.float32,
+        ], f"activation_dtype must be torch.quint8 or torch.float32, got {activation_dtype}"
+
+        # Create quantization config
+        config = ct.optimize.torch.quantization.LinearQuantizerConfig(
+            global_config=ct.optimize.torch.quantization.ModuleLinearQuantizerConfig(
+                quantization_scheme="symmetric",
+                activation_dtype=activation_dtype,
+                weight_dtype=torch.qint8,
+                weight_per_channel=True,
+            )
+        )
+
+        quantizer = CoreMLQuantizer(config)
+        quantization_recipe = QuantizationRecipe(quantizers=[quantizer])
+
+        lowering_recipe = self._get_coreml_lowering_recipe(**kwargs)
+
+        return ExportRecipe(
+            name=recipe_type.value,
+            quantization_recipe=quantization_recipe,
+            lowering_recipe=lowering_recipe,
+        )
+
+    def _build_torchao_quantized_recipe(
+        self,
+        recipe_type: RecipeType,
+        weight_dtype: torch.dtype,
+        is_per_channel: bool,
+        group_size: int = 32,
+        **kwargs: Any,
+    ) -> ExportRecipe:
+        """Build TorchAO-based quantization recipe"""
+        if is_per_channel:
+            weight_granularity = PerAxis(axis=0)
+        else:
+            weight_granularity = PerGroup(group_size=group_size)
+
+        # Use user-provided filter_fn if provided
+        filter_fn = kwargs.get("filter_fn", None)
+        config = AOQuantizationConfig(
+            ao_base_config=IntxWeightOnlyConfig(
+                weight_dtype=weight_dtype,
+                granularity=weight_granularity,
+            ),
+            filter_fn=filter_fn,
+        )
+
+        quantization_recipe = QuantizationRecipe(
+            quantizers=None,
+            ao_quantization_configs=[config],
+        )
+
+        # override minimum_deployment_target to ios18 for torchao (GH issue #13122)
+        self._validate_and_set_deployment_target(kwargs, ct.target.iOS18, "torchao")
+        lowering_recipe = self._get_coreml_lowering_recipe(**kwargs)
+
+        return ExportRecipe(
+            name=recipe_type.value,
+            quantization_recipe=quantization_recipe,
+            lowering_recipe=lowering_recipe,
+        )
+
+    def _build_codebook_quantized_recipe(
+        self,
+        recipe_type: RecipeType,
+        bits: int,
+        block_size: list,
+        **kwargs: Any,
+    ) -> ExportRecipe:
+        """Build codebook/palettization quantization recipe"""
+        from torchao.prototype.quantization.codebook_coreml import (
+            CodebookWeightOnlyConfig,
+        )
+
+        self._validate_and_set_deployment_target(kwargs, ct.target.iOS18, "codebook")
+
+        # Get the appropriate dtype (torch.uint1 through torch.uint8)
+        dtype = getattr(torch, f"uint{bits}")
+
+        # Use user-provided filter_fn or default to Linear/Embedding layers
+        filter_fn = kwargs.get(
+            "filter_fn",
+            lambda m, fqn: (
+                isinstance(m, torch.nn.Embedding) or isinstance(m, torch.nn.Linear)
+            ),
+        )
+
+        config = AOQuantizationConfig(
+            ao_base_config=CodebookWeightOnlyConfig(
+                dtype=dtype,
+                block_size=block_size,
+            ),
+            filter_fn=filter_fn,
+        )
+
+        quantization_recipe = QuantizationRecipe(
+            quantizers=None,
+            ao_quantization_configs=[config],
+        )
+
+        lowering_recipe = self._get_coreml_lowering_recipe(**kwargs)
+
+        return ExportRecipe(
+            name=recipe_type.value,
+            quantization_recipe=quantization_recipe,
             lowering_recipe=lowering_recipe,
         )
 
     def _get_coreml_lowering_recipe(
         self,
-        compute_precision: ct.precision,
+        compute_precision: ct.precision = ct.precision.FLOAT16,
         **kwargs: Any,
     ) -> LoweringRecipe:
+        """Get CoreML lowering recipe with optional precision"""
         compile_specs = CoreMLBackend.generate_compile_specs(
             compute_precision=compute_precision,
-            **kwargs,
+            compute_unit=kwargs.get("compute_unit", ct.ComputeUnit.ALL),
+            minimum_deployment_target=kwargs.get("minimum_deployment_target", None),
         )
 
         minimum_deployment_target = kwargs.get("minimum_deployment_target", None)