Pybind merge fix

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-295f2ed4d103017f7e19a7b8263ece606cd629db
+7ae0ce6360b6e4f944906502d20da24c04debee5

.ci/scripts/gather_test_models.py

@@ -14,7 +14,7 @@
 from typing import Any
 
 from examples.models import MODEL_NAME_TO_MODEL
-from examples.xnnpack import MODEL_NAME_TO_OPTIONS
+from examples.xnnpack import MODEL_NAME_TO_OPTIONS, QuantType
 
 DEFAULT_RUNNERS = {
     "linux": "linux.2xlarge",
@@ -154,7 +154,7 @@ def export_models_for_ci() -> dict[str, dict]:
         if backend == "xnnpack":
             if name not in MODEL_NAME_TO_OPTIONS:
                 continue
-            if MODEL_NAME_TO_OPTIONS[name].quantization:
+            if MODEL_NAME_TO_OPTIONS[name].quantization != QuantType.NONE:
                 backend += "-quantization"
 
             if MODEL_NAME_TO_OPTIONS[name].delegation:

.ci/scripts/unittest-linux.sh

@@ -21,8 +21,7 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
     source .ci/scripts/setup-vulkan-linux-deps.sh
 
     PYTHON_EXECUTABLE=python \
-    EXECUTORCH_BUILD_PYBIND=ON \
-    CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
     .ci/scripts/setup-linux.sh "$@"
 
     # Install llama3_2_vision dependencies.

.ci/scripts/unittest-macos.sh

@@ -21,8 +21,7 @@ trap 'rm -rfv ${TMP_DIR}' EXIT
 
 # Setup MacOS dependencies as there is no Docker support on MacOS atm
 PYTHON_EXECUTABLE=python \
-EXECUTORCH_BUILD_PYBIND=ON \
-CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
 ${CONDA_RUN} --no-capture-output \
 .ci/scripts/setup-macos.sh "$@"
 

.ci/scripts/wheel/envvar_base.sh

@@ -8,13 +8,10 @@
 # should typically only contain shell variable assignments. Be sure to export
 # any variables so that subprocesses will see them.
 
-# Enable pybindings so that users can execute ExecuTorch programs from python.
-export EXECUTORCH_BUILD_PYBIND=1
-
 # Ensure that CMAKE_ARGS is defined before referencing it. Defaults to empty
 # if not defined.
 export CMAKE_ARGS="${CMAKE_ARGS:-}"
 
 # Link the XNNPACK backend into the pybindings runtime so that users can execute
 # ExecuTorch programs that delegate to it.
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_XNNPACK=ON"
+CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON"

.github/workflows/pull.yml

@@ -365,8 +365,7 @@ jobs:
         # build module for executorch.extension.pybindings.portable_lib
         BUILD_TOOL="cmake"
         PYTHON_EXECUTABLE=python \
-        EXECUTORCH_BUILD_XNNPACK=ON \
-        EXECUTORCH_BUILD_PYBIND=ON \
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON" \
         bash .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
 
         # see if we can import the module successfully
@@ -504,7 +503,7 @@ jobs:
 
         # Setup MacOS dependencies as there is no Docker support on MacOS atm
         PYTHON_EXECUTABLE=python \
-        EXECUTORCH_BUILD_PYBIND=ON \
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON" \
         EXECUTORCH_BUILD_ARM_BAREMETAL=ON \
         .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
 

.github/workflows/trunk.yml

@@ -65,22 +65,29 @@ jobs:
       matrix:
         model: [linear, add, add_mul, ic3, ic4, mv2, mv3, resnet18, resnet50, vit, w2l, mobilebert, emformer_join, emformer_transcribe]
         backend: [portable, xnnpack-quantization-delegation]
+        runner: [linux.arm64.2xlarge]
         include:
           - model: lstm
             backend: portable
+            runner: linux.arm64.2xlarge
           - model: mul
             backend: portable
+            runner: linux.arm64.2xlarge
           - model: softmax
             backend: portable
+            runner: linux.arm64.2xlarge
           - model: phi_4_mini
             backend: portable
+            runner: linux.arm64.m7g.4xlarge
           - model: qwen2_5
             backend: portable
+            runner: linux.arm64.2xlarge
           - model: llama3_2_vision_encoder
             backend: portable
+            runner: linux.arm64.2xlarge
       fail-fast: false
     with:
-      runner: linux.arm64.2xlarge
+      runner: ${{ matrix.runner }}
       docker-image: executorch-ubuntu-22.04-gcc11-aarch64
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -261,7 +268,7 @@ jobs:
 
         # build module for executorch.extension.pybindings.portable_lib
         BUILD_TOOL=${{ matrix.build-tool }}
-        EXECUTORCH_BUILD_PYBIND=ON PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh --build-tool "${BUILD_TOOL}"
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON" PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh --build-tool "${BUILD_TOOL}"
 
         # see if we can import the module successfully
         ${CONDA_RUN} python -c "from executorch.extension.pybindings import portable_lib; print('success!')"
@@ -536,9 +543,8 @@ jobs:
         git clone https://github.com/huggingface/optimum-executorch
         cd optimum-executorch
         # There is no release yet, for CI stability, always test from the same commit on main
-        git checkout 6a7e83f3eee2976fa809335bfb78a45b1ea1cb25
-        pip install .
-        pip install accelerate sentencepiece
+        git checkout 577a2b19670e4c643a5c6ecb09bf47b9a699e7c6
+        pip install .[tests]
         pip list
         echo "::endgroup::"
 

Package.swift

@@ -82,5 +82,24 @@ let package = Package(
           (value["libraries"] as? [String] ?? []).map { .linkedLibrary($0) }
       ),
     ]
-  }
+  } + [
+    .testTarget(
+      name: "tests",
+      dependencies: [
+        .target(name: "executorch_debug"),
+        .target(name: "kernels_portable"),
+      ],
+      path: "extension/apple/ExecuTorch/__tests__",
+      resources: [
+        .copy("resources/add.pte")
+      ],
+      linkerSettings: [
+        .linkedLibrary("c++"),
+        .unsafeFlags([
+          "-Xlinker", "-force_load",
+          "-Xlinker", "cmake-out/kernels_portable.xcframework/macos-arm64/libkernels_portable_macos.a",
+        ])
+      ]
+    )
+  ]
 )

backends/apple/coreml/compiler/coreml_preprocess.py

@@ -16,8 +16,8 @@
 
 import coremltools as ct
 import coremltools.optimize as cto
-import executorchcoreml
 
+from executorch.backends.apple.coreml import executorchcoreml
 from executorch.exir.backend.backend_details import (
     BackendDetails,
     ExportedProgram,

backends/apple/coreml/runtime/delegate/ETCoreMLModelManager.mm

@@ -208,7 +208,7 @@ void set_outputs(std::vector<executorchcoreml::MultiArray>& outputs,
                                     const inmemoryfs::InMemoryFileSystem *inmemory_fs,
                                     NSError * __autoreleasing *error) {
     NSError *local_error = nil;
-    if (![fm createDirectoryAtURL:dst_url withIntermediateDirectories:NO attributes:@{} error:error]) {
+    if (![fm createDirectoryAtURL:dst_url withIntermediateDirectories:YES attributes:@{} error:error]) {
         ETCoreMLLogUnderlyingErrorAndSetNSError(error,
                                                 ETCoreMLErrorModelSaveFailed,
                                                 local_error,

backends/apple/coreml/scripts/install_requirements.sh

@@ -10,6 +10,10 @@ SCRIPT_DIR_PATH="$(
     pwd -P
 )"
 
+# TODO(jathu): remove the need to fetch coremltools to build deps for coreml_executor_runner.
+# Keep this version in sync with: pyproject.toml
+COREMLTOOLS_VERSION="8.1"
+
 red=`tput setaf 1`
 green=`tput setaf 2`
 
@@ -24,7 +28,7 @@ rm -rf "$COREML_DIR_PATH/third-party"
 mkdir "$COREML_DIR_PATH/third-party"
 
 echo "${green}ExecuTorch: Cloning coremltools."
-git clone --depth 1 --branch 8.1 "https://github.com/apple/coremltools.git" $COREMLTOOLS_DIR_PATH
+git clone --depth 1 --branch "${COREMLTOOLS_VERSION}" "https://github.com/apple/coremltools.git" $COREMLTOOLS_DIR_PATH
 cd $COREMLTOOLS_DIR_PATH
 
 STATUS=$?
@@ -43,16 +47,7 @@ fi
 
 mkdir "$COREMLTOOLS_DIR_PATH/build"
 cmake -S "$COREMLTOOLS_DIR_PATH" -B "$COREMLTOOLS_DIR_PATH/build"
-cmake --build "$COREMLTOOLS_DIR_PATH/build" --parallel
-
-echo "${green}ExecuTorch: Installing coremltools."
-pip install "$COREMLTOOLS_DIR_PATH"
-
-STATUS=$?
-if [ $STATUS -ne 0 ]; then
-    echo "${red}ExecuTorch: Failed to install coremltools."
-    exit 1
-fi
+cmake --build "$COREMLTOOLS_DIR_PATH/build" --parallel --target mlmodel
 
 echo "${green}ExecuTorch: Cloning nlohmann."
 git clone https://github.com/nlohmann/json.git "$COREML_DIR_PATH/third-party/nlohmann_json"
@@ -62,8 +57,6 @@ if [ $STATUS -ne 0 ]; then
     exit 1
 fi
 
-sh "$COREML_DIR_PATH/scripts/install_inmemoryfs.sh"
-
 echo "${green}ExecuTorch: Copying protobuf files."
 mkdir -p "$COREML_DIR_PATH/runtime/sdk/format/"
 cp -rf "$PROTOBUF_FILES_DIR_PATH" "$COREML_DIR_PATH/runtime/sdk/format/"

backends/apple/coreml/setup.md

@@ -6,16 +6,8 @@ This is a tutorial for setting up the Core ML backend.
 
 1. Follow the instructions described in [Setting Up ExecuTorch](/docs/source/getting-started-setup.md) to set up ExecuTorch environment.
 
-2. Run `install_requirements.sh` to install dependencies required by the **Core ML** backend.
 
-```
-cd executorch
-
-./backends/apple/coreml/scripts/install_requirements.sh
-
-```
-
-3. Run the example script to validate that the **Core ML** backend is set up correctly.
+2. Run the example script to validate that the **Core ML** backend is set up correctly.
 
 ```
 cd executorch
@@ -26,7 +18,7 @@ python3 -m examples.apple.coreml.scripts.export --model_name add
 
 ```
 
-4. You can now integrate the **Core ML** backend in code.
+3. You can now integrate the **Core ML** backend in code.
 
 ```python
 # Delegate to Core ML backend

backends/arm/operator_support/convolution_support.py

@@ -55,7 +55,7 @@ def _is_node_supported_u55(self, node: fx.Node):
 
         C_in = shape_in[1]
         C_out = shape_out[1]
-        if (C_in == group) and (C_out % C_in) == 0:
+        if (C_in == group) and (C_out % C_in) == 0 and len(shape_in) <= 4:
             # Depthwise convolution
             for dim in shape_in[1:]:
                 if not 1 <= dim <= 65536:
@@ -74,13 +74,19 @@ def _is_node_supported_u55(self, node: fx.Node):
 
         kernel_w = kernel[2]
         kernel_h = kernel[3] if len(kernel) > 3 else 1
+        kernel_z = kernel[4] if len(kernel) > 4 else 1
         # Kernel condition misses constraint on sum of absolute weights
         if not 1 <= kernel_h <= 64 or not 1 <= kernel_w * kernel_h <= 4096:
             self.reporter.report_reject(
                 node,
                 f"Convolution needs to have kernel_y<=64, kernel_x*kernel_y<=4096, got kernel ({kernel_w}, {kernel_h})",
             )
             return False
+        if kernel_z != 1:
+            self.reporter.report_reject(
+                node, f"Convolution3d needs to have kernel_z==1, got {kernel_z}."
+            )
+            return False
 
         if not self._stride_condition(node):
             self.reporter.report_reject(
@@ -107,6 +113,14 @@ def _stride_condition(self, node: fx.Node) -> bool:
         if len(strides) == 1:
             strides = [strides[0]] * 2
 
+        if len(strides) > 2:
+            stride_z = strides[2]
+            if stride_z > 1:
+                self.reporter.report_reject(
+                    node, f"Convolution3d only supports stride_z<=1, got {stride_z}."
+                )
+                return False
+
         for stride, dilation in zip(strides, dilations):
             stride_condition = 1 <= stride <= 3
             dilation_condition = (not has_padding) and (dilation == 1)