9.0 Release

BUILDING.md

@@ -19,7 +19,7 @@ Follow these steps:
 1. Fork and clone the GitHub [coremltools repository](https://github.com/apple/coremltools).
 
 2. Run the [build.sh](scripts/build.sh) script to build `coremltools`.
-	* By default this script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`, `3.11`, `3.12`) as a argument to change the Python version.
+	* By default this script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`, `3.11`, `3.12`, `3.13`) as a argument to change the Python version.
 	* The script creates a new `build` folder with the coremltools distribution, and a `dist` folder with Python wheel files.
 
 3. Run the [test.sh](scripts/test.sh) script to test the build.
@@ -45,7 +45,7 @@ The following build targets help you configure the development environment. If y
 * `test_slow` | Run all non-fast tests.
 * `wheel` | Build wheels in release mode.
 
-The script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`, `3.11`, `3.12`) as a argument to change the Python version.
+The script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`, `3.11`, `3.12`, `3.13`) as a argument to change the Python version.
 
 ## Resources
 

CMakeLists.txt

@@ -70,7 +70,7 @@ add_library(modelpackage
   modelpackage/src/utils/JsonMap.cpp
   modelpackage/src/ModelPackagePython.cpp
   )
-  
+
 target_compile_definitions(modelpackage
   PRIVATE
   CPU_ONLY=1
@@ -197,8 +197,13 @@ set(KMEANS_DIR "${PROJECT_SOURCE_DIR}/deps/kmeans1d")
 execute_process(
   COMMAND python3 setup.py build_ext --inplace
   WORKING_DIRECTORY ${KMEANS_DIR}
+  RESULT_VARIABLE KMEANS1D_BUILD_STATUS
 )
 
+if(NOT KMEANS1D_BUILD_STATUS EQUAL 0)
+  message(FATAL_ERROR "Could not build kmeans1d dependency")
+endif()
+
 # Somehow Python's setuptools is building this shared object file so that it tries to load the C++
 # standard library using an rpath that only exist on the build machine. Change that so it gets
 # loaded from the standard location.

coremltools/converters/mil/frontend/_utils.py

@@ -126,7 +126,12 @@ def pymil_broadcast_to(tensor: Var, shape: Union[Var, VARIABLE_SHAPE_TYPE], name
 
     if any_symbolic(tensor.shape) or shape_var.val is None:
         tensor_shape = mb.shape(x=tensor)
-        reps = mb.real_div(x=shape_var, y=tensor_shape)
+        reps = mb.select(
+            cond = mb.equal(x=shape_var, y=-1),
+            a = tensor_shape,
+            b = shape_var,
+        )
+        reps = mb.real_div(x=reps, y=tensor_shape)
         reps = mb.cast(x=reps, dtype="int32")
         res = mb.tile(x=tensor, reps=reps, name=name)
     else:

coremltools/converters/mil/frontend/torch/ssa_passes/torch_upsample_to_core_upsample.py

@@ -42,12 +42,67 @@ def _torch_upsample_to_core_upsample_block(block):
 
         if op.op_type in target_ops:
             if _try_replace_with_core_upsample(op):
-                logger.info("Successfully map {} to core upsample".format(op.op_type))
+                msg = f"Successfully map {op.op_type} to core upsample"
+                logger.info(msg)
             else:
                 raise ValueError("Unable to map {} to core upsample".format(op.op_type))
 
-
-def _try_get_upsample_factor(output_size):
+def _try_get_upsample_factor_pattern_2(output_size, expected_gather_indices, target_op):
+    """
+    This is the pattern corresponds to the python source code:
+
+    class UpsampleBilinear(nn.Module):
+        def forward(self, x):
+            b, c, h, w = x.shape
+            return F.interpolate(x, size=(h*2, w*2), mode='bilinear', align_corners=False)
+
+    The resulting pymil program is:
+
+    function[CoreML5](%x: (1, 3, is0, is1, fp32)(Tensor)) {
+      block0() {
+        %3_shape: (4,int32)^(Tensor) = shape(x=%x, name="3_shape")
+        %gather_0: (int32)^(Scalar) = gather(x=%3_shape, indices=2, axis=0, name="gather_0")
+        %6_shape: (4,int32)^(Tensor) = shape(x=%x, name="6_shape")
+        %gather_1: (int32)^(Scalar) = gather(x=%6_shape, indices=3, axis=0, name="gather_1")
+        %9: (int32)(Scalar) = mul(x=%gather_0, y=2, name="9")
+        %10: (int32)(Scalar) = cast(x=%9, dtype="int32", name="10")
+        %12: (int32)(Scalar) = mul(x=%gather_1, y=2, name="12")
+        %13: (int32)(Scalar) = cast(x=%12, dtype="int32", name="13")
+        %17: (1, 3, is2, is3, fp32)(Tensor) = torch_upsample_bilinear(x=%x, output_height=%10, output_width=%13, align_corners=False, name="17")
+
+    We do a pattern matching to extract the scale value.
+    """
+    # cast op
+    op = output_size
+    if op.op_type != "cast" or op.dtype.val != "int32":
+        return None
+
+    # mul op
+    mul_op = op.x.op
+    if mul_op.op_type != "mul":
+        return None
+    mul_op_y = mul_op.y
+
+    # gather op
+    gather_op = mul_op.x.op
+    if gather_op.op_type != "gather":
+        return None
+    if gather_op.indices.val != expected_gather_indices:
+        return None
+    if gather_op.axis.val != 0:
+        return None
+
+    # shape op
+    shape_op = gather_op.x.op
+    if shape_op.op_type != "shape":
+        return None
+    if shape_op.x != target_op:
+        return None
+
+    return mul_op_y.val
+
+
+def _try_get_upsample_factor_pattern_1(output_size):
     op = output_size
     # If the output has value, then the upsample op itself is derived from the upsample_1d op,
     # so we can just return scale factor 1 for that case
@@ -103,9 +158,15 @@ def _try_replace_with_core_upsample(op):
     assert op.op_type in target_ops
 
     # 2d upsampling
-    if op.op_type in ["torch_upsample_nearest_neighbor", "torch_upsample_bilinear"]:
-        scales_h = _try_get_upsample_factor(op.output_height.op)
-        scales_w = _try_get_upsample_factor(op.output_width.op)
+    if op.op_type in target_ops:
+
+        # try to resolve the scaling factor - pattern 1
+        scales_h = _try_get_upsample_factor_pattern_1(op.output_height.op)
+        scales_w = _try_get_upsample_factor_pattern_1(op.output_width.op)
+
+        if scales_h is None or scales_w is None:
+            scales_h = _try_get_upsample_factor_pattern_2(op.output_height.op, 2, op.x)
+            scales_w = _try_get_upsample_factor_pattern_2(op.output_width.op, 3, op.x)
 
         if scales_h is None or scales_w is None:
             return False

coremltools/converters/mil/frontend/torch/test/test_torch_export_conversion_api.py

@@ -334,6 +334,7 @@ def forward(self, x):
         assert mlmodel.user_defined_metadata[_METADATA_SOURCE_DIALECT] == dialect_name
 
 
+@pytest.mark.skipif((version_info.major, version_info.minor) == (3, 13), reason="rdar://158079341")
 class TestExecuTorchExamples(TorchBaseTest):
     @pytest.mark.parametrize(
         "compute_unit, backend, frontend, dynamic",

coremltools/converters/mil/frontend/torch/test/test_torch_ops.py

@@ -2668,6 +2668,41 @@ def forward(self, x):
             compute_unit=compute_unit,
         )
 
+    @pytest.mark.parametrize(
+        "compute_unit, backend, frontend",
+        itertools.product(compute_units, backends, frontends),
+    )
+    def test_upsample_with_shape_gather_pattern(self, compute_unit, backend, frontend):
+        if frontend == TorchFrontend.TORCHEXPORT:
+            pytest.xfail("CoreML model not runnable for the torch export frontend.")
+
+        input_shape = (1, 3, 32, 32)
+
+        class UpsampleBilinear(nn.Module):
+            def forward(self, x):
+                b, c, h, w = x.shape
+                return nn.functional.interpolate(
+                    x, size=(h * 2, w * 2), mode="bilinear", align_corners=False
+                )
+
+        model = UpsampleBilinear().eval()
+
+        h_dim = torch.export.Dim(name="height", min=16, max=128)
+        w_dim = torch.export.Dim(name="width", min=16, max=128)
+        torch_export_dynamic_shapes = {"x": {2: h_dim, 3: w_dim}}
+
+        self.run_compare_torch(
+            input_shape,
+            model,
+            frontend=frontend,
+            backend=backend,
+            compute_unit=compute_unit,
+            converter_input_type=[
+                TensorType(shape=(1, 3, ct.RangeDim(16, 128), ct.RangeDim(16, 128))),
+            ],
+            torch_export_dynamic_shapes=torch_export_dynamic_shapes,
+        )
+
     @pytest.mark.parametrize(
         "compute_unit, backend, frontend, output_size",
         itertools.product(compute_units, backends, frontends, [10, 170]),
@@ -5234,7 +5269,42 @@ def forward(self, x, y):
             input_shapes, model, compute_unit=compute_unit, backend=backend, frontend=frontend
         )
 
+    @pytest.mark.parametrize(
+        "compute_unit, backend, frontend, input_shape",
+        itertools.product(
+            compute_units,
+            backends,
+            frontends,
+            [
+                (ct.RangeDim(3, 21), ),
+                (15, )
+            ]
+        ),
+    )
+    def test_expand_dynamic_shape4(self, compute_unit, backend, frontend, input_shape):
+        if frontend in TORCH_EXPORT_BASED_FRONTENDS:
+            pytest.xfail(
+                "torch.export refuses to make size-1 dim dynamic, "
+                "and cannot expand one dynamic dimension into another dynamic dimension"
+            )
+
+        class TestModel(nn.Module):
+            def forward(self, x):
+                return x.reshape(-1, 1, 3).expand(-1, 7, -1)
+
+        converter_input_type = [ct.TensorType(name = "x", shape=input_shape, dtype=types.fp32)]
+        model = TestModel()
 
+        self.run_compare_torch(
+            torch.rand(15),
+            model,
+            input_as_shape=False,
+            converter_input_type=converter_input_type,
+            frontend=frontend,
+            backend=backend,
+            compute_unit=compute_unit,
+        )
+
 class TestExpandDims(TorchBaseTest):
     @pytest.mark.parametrize(
         "compute_unit, backend, frontend, rank_and_axis",

coremltools/test/ml_program/test_utils.py

@@ -9,6 +9,7 @@
 import platform
 import shutil
 import tempfile
+from sys import version_info
 from typing import Dict, Tuple
 
 import numpy as np
@@ -1496,8 +1497,11 @@ def validate_inference(multifunction_mlpackage_path: str) -> None:
         shutil.rmtree(multifunction_mlpackage_path)
 
 
+@pytest.mark.skipif(
+    (version_info.major, version_info.minor) == (3, 13),
+    reason="rdar://157488825 (Python 3.13 Unit Test Segmentation Fault)",
+)
 class TestBisectModel:
-
     @staticmethod
     def check_spec_op_type(model_path, expected_ops):
         spec = load_spec(model_path)

coremltools/version.py

@@ -4,4 +4,4 @@
 # found in the LICENSE.txt file or at https://opensource.org/licenses/BSD-3-Clause
 
 
-__version__ = "9.0b1"  # VERSION_STRING
+__version__ = "9.0"  # VERSION_STRING

docs-guides/source/installing-coremltools.md

@@ -4,7 +4,7 @@ This page describes how to install the [`coremltools`](https://github.com/apple/
 
 ```{admonition} Supported Python and MacOS Versions
 
-The current version of coremltools ([version 8.0](https://github.com/apple/coremltools)) includes wheels for Python 3.7, 3.8, 3.9, 3.10, 3.11, and 3.12. The last stable release of coremltools to support Python 2 is version 4.0.
+The current version of coremltools ([version 9.0b1](https://github.com/apple/coremltools)) includes wheels for Python 3.7, 3.8, 3.9, 3.10, 3.11, 3.12, 3.13. The last stable release of coremltools to support Python 2 is version 4.0.
 
 The supported MacOS versions are as follows:
 
@@ -19,7 +19,7 @@ The supported MacOS versions are as follows:
 If you are using macOS, you should already be familiar with the [Mac Terminal app command line](https://developer.apple.com/library/archive/documentation/OpenSource/Conceptual/ShellScripting/CommandLInePrimer/CommandLine.html#//apple_ref/doc/uid/TP40004268-CH271-BBCBEAJD "Command Line Primer") to perform tasks such as installations and updates. If you are using Linux, you should already be familiar with [basic Shell commands in Linux](https://www.geeksforgeeks.org/basic-shell-commands-in-linux/).
 ```
 
-Before installing coremltools, you need [Python](https://www.python.org/downloads/ "Python Downloads") and the [`pip`](https://pip.pypa.io/en/stable/) installer. 
+Before installing coremltools, you need [Python](https://www.python.org/downloads/ "Python Downloads") and the [`pip`](https://pip.pypa.io/en/stable/) installer.
 
 The `coremltools` package supports [Python 3](https://www.python.org/download/releases/3.0/). We recommend that you install Python 3.6 or newer. Use a Python package manager such as [Conda](https://docs.conda.io/en/latest/index.html) or [venv](https://docs.python.org/3/library/venv.html) to install the newest version of Python and other dependencies. [Conda](https://docs.conda.io/en/latest/index.html) is recommended because it is the most reliable way to install all required dependencies.
 
@@ -82,7 +82,7 @@ python -m venv coremltools-venv
 source coremltools-venv/bin/activate
 ```
 
-4. Follow the instructions in [Install Core ML Tools](#install-core-ml-tools). 
+4. Follow the instructions in [Install Core ML Tools](#install-core-ml-tools).
 
 ## Install Core ML Tools
 
@@ -103,17 +103,17 @@ The continuous integration (CI) system linked to the `coremltools` repo builds a
 To access the wheel for a particular `coremltools` release, follow these steps:
 
 1. Go to the [`coremltools` repository](https://github.com/apple/coremltools) on GitHub, scroll down to the **README.md** heading, and click the **build passing** button. The **Branches** tab appears:
-    
+
     ![Branches tab](images/repo-readme-build-passing-button-annot.png)
-    
+
     ![Branches passed](images/repo-branches-passed-button.png)
 
 2. Click the **passed** button to show the **Pipeline** tab:
-    
+
     ![Pipeline tab](images/repo-build-wheel-selected.png)
 
 3. Click a wheel in the **Build** column. For example, in the previous figure, the **build_wheel_macos_py38** wheel is highlighted for clicking. After clicking a wheel, the raw job log appears, with the **Download** and **Browse** buttons in the right column:
-    
+
     ![Download and Browse](images/repo-job-artifacts.png)
 
 4. Click the **Download** button to download the `dist` folder with the wheel files.

reqs/build.pip

@@ -1,8 +1,11 @@
 numpy==1.21.0; platform_machine == "arm64" and python_version < "3.9"
 numpy<1.20; platform_machine != "arm64" and python_version < "3.9"
-numpy==2.0.0; python_version >= "3.9"
+numpy==2.0.0; python_version >= "3.9" and python_version < "3.13"
+numpy==2.1.0; python_version >= "3.13"
 
+protobuf
 pytest
+setuptools; python_version >= "3.13"
 six
 sympy
 tqdm

reqs/common_test_packages.pip

@@ -1,13 +1,13 @@
 boto3==1.14.8; python_version < '3.12'
-boto3==1.24.22; python_version >= '3.12'
+boto3==1.24.22; python_version == '3.12'
+boto3==1.39.3; python_version >= '3.13'
 
 configparser
 future
 
 olefile==0.44
 pandas
 parameterized==0.8.1
-protobuf
 pillow
 pytest==7.1.2
 pytest-cov

reqs/pytorch.pip

@@ -5,7 +5,7 @@ torchaudio==2.2.0; platform_machine != "arm64"
 torchvision==0.17.0; platform_machine != "arm64"
 
 # Torch dependencies for ARM
-torch==2.7.0; platform_machine == "arm64"
+torch>=2.2.0,<=2.7.0; platform_machine == "arm64"
 torchaudio>=2.2.0; platform_machine == "arm64"
 torchvision>=0.17.0; platform_machine == "arm64"
 torchsr==1.0.4; platform_machine == "arm64"

reqs/test_executorch.pip

@@ -3,4 +3,4 @@
 
 # Warning: Starting from ExecuTorch 0.6.0, coremltools is added as a dependency
 # so we need to re-install built-from-source coremltools after pip install ExecuTorch
-executorch>=0.6.0; platform_machine == "arm64" and python_version >= '3.10'
+executorch>=0.6.0; platform_machine == "arm64" and python_version >= '3.10' and python_version < '3.13'

setup.py

@@ -106,6 +106,7 @@
         "Programming Language :: Python :: 3.10",
         "Programming Language :: Python :: 3.11",
         "Programming Language :: Python :: 3.12",
+        "Programming Language :: Python :: 3.13",
         "Topic :: Scientific/Engineering",
         "Topic :: Software Development",
     ],