Update

[ghstack-poisoned]

Author: swolchok

.ci/docker/ci_commit_pins/pytorch.txt

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

.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

@@ -261,7 +261,7 @@ 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!')"

CMakeLists.txt

@@ -646,7 +646,7 @@ target_link_options_shared_lib(executorch)
 if(EXECUTORCH_BUILD_KERNELS_OPTIMIZED)
   # find pytorch lib here to make it available to all
   # sub-directories. Find it before including portable so that
-  # optimized_portabale_kernels can use it.
+  # optimized_portable_kernels can use it.
   find_package_torch_headers()
 endif()
 

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/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/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)