Update

[ghstack-poisoned]

Author: swolchok

.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/setup-openvino.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -ex
+
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+git clone https://github.com/openvinotoolkit/openvino.git
+cd openvino && git checkout releases/2025/1
+git submodule update --init --recursive
+sudo ./install_build_dependencies.sh
+mkdir build && cd build
+cmake .. -DCMAKE_BUILD_TYPE=Release -DENABLE_PYTHON=ON
+make -j$(nproc)
+
+cd ..
+cmake --install build --prefix dist
+
+source dist/setupvars.sh
+cd ../backends/openvino
+pip install -r requirements.txt
+cd scripts
+./openvino_build.sh --enable_python

.ci/scripts/test_llama.sh

@@ -269,7 +269,7 @@ $PYTHON_EXECUTABLE -m examples.models.llama.export_llama ${EXPORT_ARGS}
 
 # Create tokenizer.bin.
 echo "Creating tokenizer.bin"
-$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 
 
 RUNTIME_ARGS="--model_path=${EXPORTED_MODEL_NAME} --tokenizer_path=tokenizer.bin --prompt=Once --temperature=0 --seq_len=10 --warmup=1"

.ci/scripts/test_llama_torchao_lowbit.sh

@@ -55,7 +55,7 @@ cmake --build cmake-out/examples/models/llama -j16 --config Release
 download_stories_model_artifacts
 
 echo "Creating tokenizer.bin"
-$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 
 # Export model
 LLAMA_CHECKPOINT=stories110M.pt

.ci/scripts/test_model.sh

@@ -96,15 +96,15 @@ test_model() {
       bash examples/models/llama/install_requirements.sh
       # Test export_llama script: python3 -m examples.models.llama.export_llama.
       # Use Llama random checkpoint with Qwen 2.5 1.5b model configuration.
-      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/qwen2_5/1_5b_config.json
+      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -p examples/models/qwen2_5/1_5b_config.json
       rm "./${MODEL_NAME}.pte"
       return  # Skip running with portable executor runnner since portable doesn't support Qwen's biased linears.
   fi
   if [[ "${MODEL_NAME}" == "phi_4_mini" ]]; then
       # Install requirements for export_llama
       bash examples/models/llama/install_requirements.sh
       # Test export_llama script: python3 -m examples.models.llama.export_llama.
-      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/phi_4_mini/config.json
+      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -p examples/models/phi_4_mini/config.json
       run_portable_executor_runner
       rm "./${MODEL_NAME}.pte"
       return
@@ -224,19 +224,22 @@ test_model_with_coreml() {
 
   "${PYTHON_EXECUTABLE}" -m examples.apple.coreml.scripts.export --model_name="${MODEL_NAME}" --compute_precision "${DTYPE}"
   EXPORTED_MODEL=$(find "." -type f -name "${MODEL_NAME}*.pte" -print -quit)
-  # TODO:
+
   if [ -n "$EXPORTED_MODEL" ]; then
     EXPORTED_MODEL_WITH_DTYPE="${EXPORTED_MODEL%.pte}_${DTYPE}.pte"
     mv "$EXPORTED_MODEL" "$EXPORTED_MODEL_WITH_DTYPE"
     EXPORTED_MODEL="$EXPORTED_MODEL_WITH_DTYPE"
-    echo "Renamed file path: $EXPORTED_MODEL"
+    echo "OK exported model: $EXPORTED_MODEL"
   else
-    echo "No .pte file found"
+    echo "[error] failed to export model: no .pte file found"
     exit 1
   fi
 
   # Run the model
   if [ "${should_test}" = true ]; then
+    echo "Installing requirements needed to build coreml_executor_runner..."
+    backends/apple/coreml/scripts/install_requirements.sh
+
     echo "Testing exported model with coreml_executor_runner..."
     local out_dir=$(mktemp -d)
     COREML_EXECUTOR_RUNNER_OUT_DIR="${out_dir}" examples/apple/coreml/scripts/build_executor_runner.sh

.ci/scripts/test_openvino.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -ex
+
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+source openvino/dist/setupvars.sh
+cd backends/openvino/tests
+python test_runner.py --test_type ops
+python test_runner.py --test_type models

.ci/scripts/test_phi_3_mini.sh

@@ -56,7 +56,7 @@ cmake_build_phi_3_mini() {
 prepare_tokenizer() {
   echo "Downloading and converting tokenizer.model"
   wget -O tokenizer.model "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/tokenizer.model?download=true"
-  $PYTHON_EXECUTABLE -m executorch.extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+  $PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 }
 
 # Export phi-3-mini model to pte

.ci/scripts/test_qnn_static_llama.sh

@@ -5,7 +5,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-set -exu
+set -euxo pipefail
 
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
@@ -30,7 +30,7 @@ pip install graphviz
 # Download stories llama110m artifacts
 download_stories_model_artifacts
 echo "Creating tokenizer.bin"
-$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 
 set +e
 # Compile only as weight sharing is not applicable on x86
@@ -56,4 +56,3 @@ if [ $exit_code1 -ne 0 ] || [ $exit_code2 -ne 0 ]; then
 else
     exit 0
 fi
-set -e

.ci/scripts/utils.sh

@@ -32,7 +32,7 @@ install_executorch() {
   which pip
   # Install executorch, this assumes that Executorch is checked out in the
   # current directory.
-  ./install_executorch.sh --pybind xnnpack "$@"
+  ./install_executorch.sh "$@"
   # Just print out the list of packages for debugging
   pip list
 }

.ci/scripts/wheel/test_macos.py

@@ -15,11 +15,9 @@
                 model=Model.Mv3,
                 backend=Backend.XnnpackQuantizationDelegation,
             ),
-            # Enable this once CoreML is suppported out-of-the-box
-            # https://github.com/pytorch/executorch/issues/9019
-            # test_base.ModelTest(
-            #     model=Model.Mv3,
-            #     backend=Backend.CoreMlTest,
-            # )
+            test_base.ModelTest(
+                model=Model.Mv3,
+                backend=Backend.CoreMlTest,
+            ),
         ]
     )

.github/workflows/apple.yml

@@ -37,7 +37,7 @@ jobs:
         id: set_version
         shell: bash
         run: |
-          VERSION="0.5.0.$(TZ='PST8PDT' date +%Y%m%d)"
+          VERSION="0.7.0.$(TZ='PST8PDT' date +%Y%m%d)"
           echo "version=$VERSION" >> "$GITHUB_OUTPUT"
 
   build-demo-ios:

.github/workflows/build-wheels-linux.yml

@@ -6,6 +6,9 @@ on:
     paths:
       - .ci/**/*
       - .github/workflows/build-wheels-linux.yml
+      - examples/**/*
+      - pyproject.toml
+      - setup.py
   push:
     branches:
       - nightly

.github/workflows/build-wheels-macos.yml

@@ -6,6 +6,9 @@ on:
     paths:
       - .ci/**/*
       - .github/workflows/build-wheels-macos.yml
+      - examples/**/*
+      - pyproject.toml
+      - setup.py
   push:
     branches:
       - nightly
@@ -57,6 +60,8 @@ jobs:
       pre-script: ${{ matrix.pre-script }}
       post-script: ${{ matrix.post-script }}
       package-name: ${{ matrix.package-name }}
-      runner-type: macos-m1-stable
+      # Meta's macOS runners do not have Xcode, so use GitHub's runners.
+      runner-type: macos-latest-xlarge
+      setup-miniconda: true
       smoke-test-script: ${{ matrix.smoke-test-script }}
       trigger-event: ${{ github.event_name }}

.github/workflows/pull.yml

@@ -573,7 +573,6 @@ jobs:
 
         BUILD_TOOL="cmake"
 
-        ./install_requirements.sh --use-pt-pinned-commit
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
         PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
 
@@ -736,3 +735,25 @@ jobs:
         conda activate "${CONDA_ENV}"
 
         # placeholder for mediatek to add more tests
+
+  test-openvino-linux:
+    name: test-openvino-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-gcc9
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-openvino.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_openvino.sh

.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 }}
@@ -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::"
 

.gitmodules

@@ -1,9 +1,9 @@
 [submodule "backends/arm/third-party/ethos-u-core-driver"]
 	path = backends/arm/third-party/ethos-u-core-driver
-	url = https://github.com/pytorch-labs/ethos-u-core-driver-mirror
+	url = https://git.gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-core-driver.git
 [submodule "backends/arm/third-party/serialization_lib"]
 	path = backends/arm/third-party/serialization_lib
-	url = https://github.com/pytorch-labs/tosa_serialization_lib-mirror
+	url = https://git.gitlab.arm.com/tosa/tosa-serialization.git
 [submodule "backends/vulkan/third-party/Vulkan-Headers"]
 	path = backends/vulkan/third-party/Vulkan-Headers
 	url = https://github.com/KhronosGroup/Vulkan-Headers

.lintrunner.toml

@@ -266,6 +266,8 @@ exclude_patterns = [
     'extension/**',
     # Uses properly-gated (ET_USE_PYTORCH_HEADERS) ATen include.
     'kernels/portable/cpu/util/elementwise_util.h',
+    'kernels/portable/cpu/util/math_util.h',
+    'kernels/portable/cpu/util/vectorized_math.h',
     'kernels/optimized/**',
     'runtime/core/exec_aten/**',
     # Want to be able to keep c10 in sync with PyTorch core.
@@ -301,12 +303,14 @@ include_patterns = [
     # TODO(https://github.com/pytorch/executorch/issues/7441): Gradually start enabling all folders.
     # 'backends/**/*.py',
     'backends/arm/**/*.py',
+    'backends/openvino/**/*.py',
     'build/**/*.py',
     'codegen/**/*.py',
     # 'devtools/**/*.py',
     'devtools/visualization/**/*.py',
     'docs/**/*.py',
     # 'examples/**/*.py',
+    'examples/openvino/**/*.py',
     # 'exir/**/*.py',
     # 'extension/**/*.py',
     'kernels/**/*.py',