Table of Contents
Shortcuts

Core ML Backend

Core ML delegate is the ExecuTorch solution to take advantage of Apple’s Core ML framework for on-device ML. With Core ML, a model can run on CPU, GPU, and the Apple Neural Engine (ANE).

Features

  • Dynamic dispatch to the CPU, GPU, and ANE.

  • Supports fp32 and fp16 computation.

Target Requirements

Below are the minimum OS requirements on various hardware for running a Core ML-delegated ExecuTorch model:

Development Requirements

To develop you need:

Before starting, make sure you install the Xcode Command Line Tools:

xcode-select --install

Using the Core ML Backend

To target the Core ML backend during the export and lowering process, pass an instance of the CoreMLPartitioner to to_edge_transform_and_lower. The example below demonstrates this process using the MobileNet V2 model from torchvision.

import torch
import torchvision.models as models
from torchvision.models.mobilenetv2 import MobileNet_V2_Weights
from executorch.backends.apple.coreml.partition import CoreMLPartitioner
from executorch.exir import to_edge_transform_and_lower

mobilenet_v2 = models.mobilenetv2.mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT).eval()
sample_inputs = (torch.randn(1, 3, 224, 224), )

et_program = to_edge_transform_and_lower(
    torch.export.export(mobilenet_v2, sample_inputs),
    partitioner=[CoreMLPartitioner()],
).to_executorch()

with open("mv2_coreml.pte", "wb") as file:
    et_program.write_to_file(file)

Partitioner API

The Core ML partitioner API allows for configuration of the model delegation to Core ML. Passing a CoreMLPartitioner instance with no additional parameters will run as much of the model as possible on the Core ML backend with default settings. This is the most common use case. For advanced use cases, the partitioner exposes the following options via the constructor:

  • skip_ops_for_coreml_delegation: Allows you to skip ops for delegation by Core ML. By default, all ops that Core ML supports will be delegated. See here for an example of skipping an op for delegation.

  • compile_specs: A list of CompileSpecs for the Core ML backend. These control low-level details of Core ML delegation, such as the compute unit (CPU, GPU, ANE), the iOS deployment target, and the compute precision (FP16, FP32). These are discussed more below.

  • take_over_mutable_buffer: A boolean that indicates whether PyTorch mutable buffers in stateful models should be converted to Core ML MLState. If set to False, mutable buffers in the PyTorch graph are converted to graph inputs and outputs to the Core ML lowered module under the hood. Generally, setting take_over_mutable_buffer to true will result in better performance, but using MLState requires iOS >= 18.0, macOS >= 15.0, and Xcode >= 16.0.

  • take_over_constant_data: A boolean that indicates whether PyTorch constant data like model weights should be consumed by the Core ML delegate. If set to False, constant data is passed to the Core ML delegate as inputs. By deafault, take_over_constant_data=True.

  • lower_full_graph: A boolean that indicates whether the entire graph must be lowered to Core ML. If set to True and Core ML does not support an op, an error is raised during lowering. If set to False and Core ML does not support an op, the op is executed on the CPU by ExecuTorch. Although setting lower_full_graph=False can allow a model to lower where it would otherwise fail, it can introduce performance overhead in the model when there are unsupported ops. You will see warnings about unsupported ops during lowering if there are any. By default, lower_full_graph=False.

Core ML CompileSpec

A list of CompileSpecs is constructed with CoreMLBackend.generate_compile_specs. Below are the available options:

  • compute_unit: this controls the compute units (CPU, GPU, ANE) that are used by Core ML. The default value is coremltools.ComputeUnit.ALL. The available options from coremltools are:

    • coremltools.ComputeUnit.ALL (uses the CPU, GPU, and ANE)

    • coremltools.ComputeUnit.CPU_ONLY (uses the CPU only)

    • coremltools.ComputeUnit.CPU_AND_GPU (uses both the CPU and GPU, but not the ANE)

    • coremltools.ComputeUnit.CPU_AND_NE (uses both the CPU and ANE, but not the GPU)

  • minimum_deployment_target: The minimum iOS deployment target (e.g., coremltools.target.iOS18). By default, the smallest deployment target needed to deploy the model is selected. During export, you will see a warning about the “Core ML specification version” that was used for the model, which maps onto a deployment target as discussed here. If you need to control the deployment target, please specify it explicitly.

  • compute_precision: The compute precision used by Core ML (coremltools.precision.FLOAT16 or coremltools.precision.FLOAT32). The default value is coremltools.precision.FLOAT16. Note that the compute precision is applied no matter what dtype is specified in the exported PyTorch model. For example, an FP32 PyTorch model will be converted to FP16 when delegating to the Core ML backend by default. Also note that the ANE only supports FP16 precision.

  • model_type: Whether the model should be compiled to the Core ML mlmodelc format during .pte creation (CoreMLBackend.MODEL_TYPE.COMPILED_MODEL), or whether it should be compiled to mlmodelc on device (CoreMLBackend.MODEL_TYPE.MODEL). Using CoreMLBackend.MODEL_TYPE.COMPILED_MODEL and doing compilation ahead of time should improve the first time on-device model load time.

Dynamic and Enumerated Shapes in Core ML Export

When exporting an ExportedProgram to Core ML, dynamic shapes are mapped to RangeDim. This enables Core ML .pte files to accept inputs with varying dimensions at runtime.

⚠️ Note: The Apple Neural Engine (ANE) does not support true dynamic shapes. If a model relies on RangeDim, Core ML will fall back to scheduling the model on the CPU or GPU instead of the ANE.

Enumerated Shapes

To enable limited flexibility on the ANE—and often achieve better performance overall—you can export models using enumerated shapes.

  • Enumerated shapes are not fully dynamic.

  • Instead, they define a finite set of valid input shapes that Core ML can select from at runtime.

  • This approach allows some adaptability while still preserving ANE compatibility.

Specifying Enumerated Shapes

Unlike RangeDim, enumerated shapes are not part of the ExportedProgram itself. They must be provided through a compile spec.

For reference on how to do this, see:

class Model(torch.nn.Module):
        def __init__(self):
                super().__init__()
                self.linear1 = torch.nn.Linear(10, 5)
                self.linear2 = torch.nn.Linear(11, 5)

        def forward(self, x, y):
            return self.linear1(x).sum() + self.linear2(y)

model = Model()
example_inputs = (
    torch.randn((4, 6, 10)),
    torch.randn((5, 11)),
)

# Specify the enumerated shapes.  Below we specify that:
#
# * x can take shape [1, 5, 10] and y can take shape [3, 11], or
# * x can take shape [4, 6, 10] and y can take shape [5, 11]
#
# Any other input shapes will result in a runtime error.
#
# Note that we must export x and y with dynamic shapes in the ExportedProgram
# because some of their dimensions are dynamic
enumerated_shapes = {"x": [[1, 5, 10], [4, 6, 10]], "y": [[3, 11], [5, 11]]}
dynamic_shapes = [
    {
        0: torch.export.Dim.AUTO(min=1, max=4),
        1: torch.export.Dim.AUTO(min=5, max=6),
    },
    {0: torch.export.Dim.AUTO(min=3, max=5)},
]
ep = torch.export.export(
    model.eval(), example_inputs, dynamic_shapes=dynamic_shapes
)

# If enumerated shapes are specified for multiple inputs, we must export
# for iOS18+
compile_specs = CoreMLBackend.generate_compile_specs(
    minimum_deployment_target=ct.target.iOS18
)
compile_specs.append(
    CoreMLBackend.generate_enumerated_shapes_compile_spec(
        ep,
        enumerated_shapes,
    )
)

# When using an enumerated shape compile spec, you must specify lower_full_graph=True
# in the CoreMLPartitioner.  We do not support using enumerated shapes
# for partially exported models
partitioner = CoreMLPartitioner(
    compile_specs=compile_specs, lower_full_graph=True
)
delegated_program = executorch.exir.to_edge_transform_and_lower(
    ep,
    partitioner=[partitioner],
)
et_prog = delegated_program.to_executorch()

Backward compatibility

Core ML supports backward compatibility via the minimum_deployment_target option. A model exported with a specific deployment target is guaranteed to work on all deployment targets >= the specified deployment target. For example, a model exported with coremltools.target.iOS17 will work on iOS 17 or higher.

Testing the Model

After generating the Core ML-delegated .pte, the model can be tested from Python using the ExecuTorch runtime Python bindings. This can be used to quickly check the model and evaluate numerical accuracy. See Testing the Model for more information.

Quantization

To quantize a PyTorch model for the Core ML backend, use the CoreMLQuantizer.

8-bit Quantization using the PT2E Flow

Quantization with the Core ML backend requires exporting the model for iOS 17 or later. To perform 8-bit quantization with the PT2E flow, follow these steps:

  1. Create a coremltools.optimize.torch.quantization.LinearQuantizerConfig and use to to create an instance of a CoreMLQuantizer.

  2. Use torch.export.export to export a graph module that will be prepared for quantization.

  3. Call prepare_pt2e to prepare the model for quantization.

  4. Run the prepared model with representative samples to calibrate the quantizated tensor activation ranges.

  5. Call convert_pt2e to quantize the model.

  6. Export and lower the model using the standard flow.

The output of convert_pt2e is a PyTorch model which can be exported and lowered using the normal flow. As it is a regular PyTorch model, it can also be used to evaluate the accuracy of the quantized model using standard PyTorch techniques.

import torch
import coremltools as ct
import torchvision.models as models
from torchvision.models.mobilenetv2 import MobileNet_V2_Weights
from executorch.backends.apple.coreml.quantizer import CoreMLQuantizer
from executorch.backends.apple.coreml.partition import CoreMLPartitioner
from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e
from executorch.exir import to_edge_transform_and_lower
from executorch.backends.apple.coreml.compiler import CoreMLBackend

mobilenet_v2 = models.mobilenetv2.mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT).eval()
sample_inputs = (torch.randn(1, 3, 224, 224), )

# Step 1: Define a LinearQuantizerConfig and create an instance of a CoreMLQuantizer
# Note that "linear" here does not mean only linear layers are quantized, but that linear (aka affine) quantization
# is being performed
static_8bit_config = ct.optimize.torch.quantization.LinearQuantizerConfig(
    global_config=ct.optimize.torch.quantization.ModuleLinearQuantizerConfig(
        quantization_scheme="symmetric",
        activation_dtype=torch.quint8,
        weight_dtype=torch.qint8,
        weight_per_channel=True,
    )
)
quantizer = CoreMLQuantizer(static_8bit_config)

# Step 2: Export the model for training
training_gm = torch.export.export(mobilenet_v2, sample_inputs).module()

# Step 3: Prepare the model for quantization
prepared_model = prepare_pt2e(training_gm, quantizer)

# Step 4: Calibrate the model on representative data
# Replace with your own calibration data
for calibration_sample in [torch.randn(1, 3, 224, 224)]:
	prepared_model(calibration_sample)

# Step 5: Convert the calibrated model to a quantized model
quantized_model = convert_pt2e(prepared_model)

# Step 6: Export the quantized model to Core ML
et_program = to_edge_transform_and_lower(
    torch.export.export(quantized_model, sample_inputs),
    partitioner=[
        CoreMLPartitioner(
             # iOS17 is required for the quantized ops in this example
            compile_specs=CoreMLBackend.generate_compile_specs(
                minimum_deployment_target=ct.target.iOS17
            )
        )
    ],
).to_executorch()

The above does static quantization (activations and weights are quantized).

You can see a full description of available quantization configs in the coremltools documentation. For example, the config below will perform weight-only quantization:

weight_only_8bit_config = ct.optimize.torch.quantization.LinearQuantizerConfig(
    global_config=ct.optimize.torch.quantization.ModuleLinearQuantizerConfig(
        quantization_scheme="symmetric",
        activation_dtype=torch.float32,
        weight_dtype=torch.qint8,
        weight_per_channel=True,
    )
)
quantizer = CoreMLQuantizer(weight_only_8bit_config)

Quantizing activations requires calibrating the model on representative data. Also note that PT2E currently requires passing at least 1 calibration sample before calling convert_pt2e, even for data-free weight-only quantization.

See PyTorch 2 Export Post Training Quantization for more information.

LLM quantization with quantize_

The Core ML backend also supports quantizing models with the torchao quantize_ API. This is most commonly used for LLMs, requiring more advanced quantization. Since quantize_ is not backend aware, it is important to use a config that is compatible with Core ML:

  • Quantize embedding/linear layers with IntxWeightOnlyConfig (with weight_dtype torch.int4 or torch.int8, using PerGroup or PerAxis granularity). Using 4-bit or PerGroup quantization requires exporting with minimum_deployment_target >= ct.target.iOS18. Using 8-bit quantization with per-axis granularity is supported on ct.target.IOS16+. See Core ML CompileSpec for more information on setting the deployment target.

  • Quantize embedding/linear layers with CodebookWeightOnlyConfig (with dtype torch.uint1 through torch.uint8, using various block sizes). Quantizing with CodebookWeightOnlyConfig requires exporting with minimum_deployment_target >= ct.target.iOS18, see Core ML CompileSpec for more information on setting the deployment target.

Below is an example that quantizes embeddings to 8-bits per-axis and linear layers to 4-bits using group_size=32 with affine quantization:

from torchao.quantization.granularity import PerGroup, PerAxis
from torchao.quantization.quant_api import (
    IntxWeightOnlyConfig,
    quantize_,
)

# Quantize embeddings with 8-bits, per channel
embedding_config = IntxWeightOnlyConfig(
    weight_dtype=torch.int8,
    granularity=PerAxis(0),
)
quantize_(
    eager_model,
    embedding_config,
    lambda m, fqn: isinstance(m, torch.nn.Embedding),
)

# Quantize linear layers with 4-bits, per-group
linear_config = IntxWeightOnlyConfig(
    weight_dtype=torch.int4,
    granularity=PerGroup(32),
)
quantize_(
    eager_model,
    linear_config,
)

Below is another example that uses codebook quantization to quantize both embeddings and linear layers to 3-bits. In the coremltools documentation, this is called palettization:

from torchao.quantization.quant_api import (
    quantize_,
)
from torchao.prototype.quantization.codebook_coreml import CodebookWeightOnlyConfig

quant_config = CodebookWeightOnlyConfig(
    dtype=torch.uint3,
    # There is one LUT per 16 rows
    block_size=[16, -1],
)

quantize_(
    eager_model,
    quant_config,
    lambda m, fqn: isinstance(m, torch.nn.Embedding) or isinstance(m, torch.nn.Linear),
)

Both of the above examples will export and lower to Core ML with the to_edge_transform_and_lower API.

Runtime integration

To run the model on device, use the standard ExecuTorch runtime APIs. See Running on Device for more information, including building the iOS frameworks.

When building from source, pass -DEXECUTORCH_BUILD_COREML=ON when configuring the CMake build to compile the Core ML backend.

Due to the use of static initializers for registration, it may be necessary to use whole-archive to link against the coremldelegate target. This can typically be done by passing "$<LINK_LIBRARY:WHOLE_ARCHIVE,coremldelegate>" to target_link_libraries.

# CMakeLists.txt
add_subdirectory("executorch")
...
target_link_libraries(
    my_target
    PRIVATE executorch
    extension_module_static
    extension_tensor
    optimized_native_cpu_ops_lib
    $<LINK_LIBRARY:WHOLE_ARHIVE,coremldelegate>)

No additional steps are necessary to use the backend beyond linking the target. A Core ML-delegated .pte file will automatically run on the registered backend.

Advanced

Extracting the mlpackage

Core ML *.mlpackage files can be extracted from a Core ML-delegated *.pte file. This can help with debugging and profiling for users who are more familiar with *.mlpackage files:

python examples/apple/coreml/scripts/extract_coreml_models.py -m /path/to/model.pte

Note that if the ExecuTorch model has graph breaks, there may be multiple extracted *.mlpackage files.

Common issues and what to do

During lowering

  1. “ValueError: In op, of type [X], named [Y], the named input [Z] must have the same data type as the named input x. However, [Z] has dtype fp32 whereas x has dtype fp16.”

This happens because the model is in FP16, but Core ML interprets some of the arguments as FP32, which leads to a type mismatch. The solution is to keep the PyTorch model in FP32. Note that the model will be still be converted to FP16 during lowering to Core ML unless specified otherwise in the compute_precision Core ML CompileSpec. Also see the related issue in coremltools.

  1. coremltools/converters/mil/backend/mil/load.py”, line 499, in export raise RuntimeError(“BlobWriter not loaded”)

If you’re using Python 3.13, try reducing your python version to Python 3.12. coremltools does not support Python 3.13 per coremltools issue #2487.

At runtime

  1. [ETCoreMLModelCompiler.mm:55] [Core ML] Failed to compile model, error = Error Domain=com.apple.mlassetio Code=1 “Failed to parse the model specification. Error: Unable to parse ML Program: at unknown location: Unknown opset ‘CoreML7’.” UserInfo={NSLocalizedDescription=Failed to par$

This means the model requires the the Core ML opset ‘CoreML7’, which requires running the model on iOS >= 17 or macOS >= 14.

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources

To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: Cookies Policy.