Update on "[Executorch] optimized sigmoid"

basically use exp approximation using sleef instead of std::exp

Differential Revision: [D64156864](https://our.internmc.facebook.com/intern/diff/D64156864/)

[ghstack-poisoned]

Author: kimishpatel

.ci/docker/build.sh

@@ -37,6 +37,10 @@ case "${IMAGE_NAME}" in
     ARM_SDK=yes
     CLANG_VERSION=12
     ;;
+  executorch-ubuntu-22.04-qnn-sdk)
+    QNN_SDK=yes
+    CLANG_VERSION=12
+    ;;
   executorch-ubuntu-22.04-clang12-android)
     LINTRUNNER=""
     CLANG_VERSION=12
@@ -72,6 +76,7 @@ docker build \
   --build-arg "LINTRUNNER=${LINTRUNNER:-}" \
   --build-arg "BUILD_DOCS=${BUILD_DOCS}" \
   --build-arg "ARM_SDK=${ARM_SDK:-}" \
+  --build-arg "QNN_SDK=${QNN_SDK:-}" \
   --build-arg "ANDROID_NDK_VERSION=${ANDROID_NDK_VERSION:-}" \
   -f "${OS}"/Dockerfile \
   "$@" \

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-c8a648d4dffb9f0133ff4a2ea0e660b42105d3ad
+19eff28ff3f19b50da46f5a9ff5f4d4d213806fe

.ci/docker/ci_commit_pins/torchao.txt

@@ -82,5 +82,7 @@ COPY --chown=ci-user:ci-user ./arm /opt/arm
 # Set up ARM SDK if needed
 RUN if [ -n "${ARM_SDK}" ]; then git config --global user.email "[email protected]"; git config --global user.name "OSS CI"; bash /opt/arm/setup.sh --i-agree-to-the-contained-eula /opt/arm-sdk; chown -R ci-user:ci-user /opt/arm-sdk; fi
 
+ARG QNN_SDK
+
 USER ci-user
 CMD ["bash"]

.ci/docker/ubuntu/Dockerfile

@@ -24,6 +24,8 @@
         "ic4": "linux.12xlarge",
         "resnet50": "linux.12xlarge",
         "llava": "linux.12xlarge",
+        "llama3_2_vision_encoder": "linux.12xlarge",
+        # "llama3_2_text_decoder": "linux.12xlarge",  # TODO: re-enable test when Huy's change is in / model gets smaller.
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
         "dl3": "linux.12xlarge",
         "emformer_join": "linux.12xlarge",
@@ -88,8 +90,8 @@ def model_should_run_on_event(model: str, event: str) -> bool:
     if event == "pull_request":
         return model in ["mv3", "vit"]
     elif event == "push":
-        # 'emformer_predict' is running super slow. Only run it periodically
-        return model not in ["emformer_predict"]
+        # These are super slow. Only run it periodically
+        return model not in ["dl3", "edsr", "emformer_predict"]
     else:
         return True
 

.ci/scripts/gather_test_models.py

@@ -23,3 +23,4 @@ fi
 # of nightly. This allows CI to test against latest commits from PyTorch
 install_executorch "use-pt-pinned-commit"
 build_executorch_runner "${BUILD_TOOL}"
+do_not_use_nightly_on_ci

.ci/scripts/setup-linux.sh

@@ -77,7 +77,7 @@ test_model() {
     # Install requirements for export_llama
     bash examples/models/llama/install_requirements.sh
     # Test export_llama script: python3 -m examples.models.llama.export_llama
-    "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama -c examples/models/llama/params/demo_rand_params.pth -p examples/models/llama/params/demo_config.json
+    "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/llama/params/demo_config.json
     run_portable_executor_runner
     rm "./${MODEL_NAME}.pte"
   fi
@@ -87,6 +87,10 @@ test_model() {
     bash examples/models/llava/install_requirements.sh
     STRICT="--no-strict"
   fi
+  if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" || "$MODEL_NAME" == "llama3_2_text_decoder" ]]; then
+    # Install requirements for llama vision.
+    bash examples/models/llama3_2_vision/install_requirements.sh
+  fi
   # python3 -m examples.portable.scripts.export --model_name="llama2" should works too
   "${PYTHON_EXECUTABLE}" -m examples.portable.scripts.export --model_name="${MODEL_NAME}" "${STRICT}"
   run_portable_executor_runner

.ci/scripts/test_model.sh

@@ -113,10 +113,26 @@ cmake_install_executorch_lib() {
 }
 
 download_stories_model_artifacts() {
-    # Download stories110M.pt and tokenizer from Github
+  # Download stories110M.pt and tokenizer from Github
   curl -Ls "https://huggingface.co/karpathy/tinyllamas/resolve/main/stories110M.pt" --output stories110M.pt
   curl -Ls "https://raw.githubusercontent.com/karpathy/llama2.c/master/tokenizer.model" --output tokenizer.model
   # Create params.json file
   touch params.json
   echo '{"dim": 768, "multiple_of": 32, "n_heads": 12, "n_layers": 12, "norm_eps": 1e-05, "vocab_size": 32000}' > params.json
 }
+
+do_not_use_nightly_on_ci() {
+  # An assert to make sure that we are not using PyTorch nightly on CI to prevent
+  # regression as documented in https://github.com/pytorch/executorch/pull/6564
+  TORCH_VERSION=$(pip list | grep -w 'torch ' | awk -F ' ' {'print $2'} | tr -d '\n')
+
+  # The version of PyTorch building from source looks like 2.6.0a0+gitc8a648d that
+  # includes the commit while nightly (2.6.0.dev20241019+cpu) or release (2.6.0)
+  # won't have that. Note that we couldn't check for the exact commit from the pin
+  # ci_commit_pins/pytorch.txt here because the value will be different when running
+  # this on PyTorch CI
+  if [[ "${TORCH_VERSION}" != *"+git"* ]]; then
+    echo "Unexpected torch version. Expected binary built from source, got ${TORCH_VERSION}"
+    exit 1
+  fi
+}

.ci/scripts/utils.sh

@@ -73,6 +73,8 @@ jobs:
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug-androidTest.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/fp32-xnnpack-custom/model.zip
+          curl -o android-test-debug.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug.apk
+          curl -o android-test-debug-androidTest.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug-androidTest.apk
           unzip model.zip
           mv *.pte model.pte
 

.github/workflows/_android.yml

@@ -7,7 +7,6 @@ on:
       - release/*
   pull_request:
     paths:
-      - .ci/docker/**
       - .ci/scripts/setup-ios.sh
       - .github/workflows/apple.yml
       - install_requirements.sh
@@ -21,12 +20,26 @@ on:
       - extension/benchmark/apple/**
       - extension/module/**
   workflow_dispatch:
+  schedule:
+    - cron: '0 10 * * *'  # Runs daily at 2 AM PST
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
   cancel-in-progress: true
 
 jobs:
+  set-version:
+    runs-on: ubuntu-22.04
+    outputs:
+      version: ${{ steps.set_version.outputs.version }}
+    steps:
+      - name: Set VERSION variable
+        id: set_version
+        shell: bash
+        run: |
+          VERSION="0.4.0.$(TZ='PST8PDT' date +%Y%m%d)"
+          echo "version=$VERSION" >> "$GITHUB_OUTPUT"
+
   build-demo-ios:
     name: build-demo-ios
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
@@ -40,6 +53,8 @@ jobs:
       secrets-env: BUILD_CERTIFICATE_BASE64 EXECUTORCH_DEMO_BUILD_PROVISION_PROFILE_BASE64 KEYCHAIN_PASSWORD
       upload-artifact: ios-apps
       script: |
+        set -eux
+
         BUILD_TOOL=cmake
 
         .ci/scripts/setup-conda.sh
@@ -58,7 +73,7 @@ jobs:
 
         # Build and test iOS Demo App
         PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-        build/test_ios_ci.sh ${ARTIFACTS_DIR_NAME}
+        build/test_ios_ci.sh "${ARTIFACTS_DIR_NAME}"
 
   # Upload the test demo app to S3
   upload-demo-ios:
@@ -76,6 +91,7 @@ jobs:
         shell: bash
         working-directory: ${{ runner.temp }}/artifacts/
         run: |
+          set -eux
           ls -lah ./
 
       - name: Upload the artifacts to S3
@@ -113,6 +129,7 @@ jobs:
 
   build-frameworks-ios:
     name: build-frameworks-ios
+    needs: set-version
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with:
       runner: macos-latest-xlarge
@@ -122,8 +139,10 @@ jobs:
       upload-artifact: executorch-frameworks-ios
       timeout: 90
       script: |
+        set -eux
+
         BUILD_TOOL=cmake
-        VERSION="latest"
+        VERSION="${{ needs.set-version.outputs.version }}"
         FRAMEWORKS=(
           "executorch"
           "backend_coreml"
@@ -172,13 +191,17 @@ jobs:
 
   upload-frameworks-ios:
     runs-on: ubuntu-22.04
-    needs: build-frameworks-ios
+    needs: [build-frameworks-ios, set-version]
     timeout-minutes: 30
+    environment: ${{ github.ref == 'refs/heads/main' && 'cherry-pick-bot' || '' }}
     permissions:
       id-token: write
-      contents: read
+      contents: write
     steps:
       - uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+          token: ${{ secrets.GH_PYTORCHBOT_CHERRY_PICK_TOKEN || secrets.GITHUB_TOKEN }}
       - uses: actions/setup-python@v4
         with:
           python-version: '3.11'
@@ -195,15 +218,15 @@ jobs:
           name: executorch-frameworks-ios
           path: ${{ runner.temp }}/frameworks-ios/
       - name: Only push to S3 when running the workflow manually from main branch
-        if: ${{ github.event_name == 'workflow_dispatch' && github.ref == 'refs/heads/main' }}
+        if: ${{ (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && github.ref == 'refs/heads/main' }}
         shell: bash
         run: |
-          set -eux
           echo "UPLOAD_ON_MAIN=1" >> "${GITHUB_ENV}"
       - name: Upload the artifact to ossci-ios S3 bucket
         shell: bash
         run: |
           set -eux
+          VERSION="${{ needs.set-version.outputs.version }}"
 
           pip install awscli==1.32.18
 
@@ -214,9 +237,44 @@ jobs:
 
           for FILENAME in "${RUNNER_TEMP}"/frameworks-ios/*.zip; do
             [ -e "${FILENAME}" ] || continue
-            shasum -a 256 "${FILENAME}"
+            FRAMEWORK_NAME=$(basename "${FILENAME}" | sed "s/-${VERSION}.zip//")
+            CHECKSUM=$(shasum -a 256 "${FILENAME}" | cut -d ' ' -f1)
+            echo "${FRAMEWORK_NAME} ${CHECKSUM}" >> "${RUNNER_TEMP}/checksums.txt"
             ${AWS_CMD} "${FILENAME}" s3://ossci-ios/executorch/ --acl public-read
           done
+      - name: Update SwiftPM
+        shell: bash
+        run: |
+          set -eux
+          VERSION="${{ needs.set-version.outputs.version }}"
+          BRANCH="swiftpm-${VERSION}"
+
+          git checkout swiftpm
+
+          if git show-ref --verify --quiet refs/heads/${BRANCH}; then
+            git checkout "${BRANCH}"
+          else
+            git checkout -b "${BRANCH}"
+          fi
+
+          [[ -f Package.swift ]] || mv Package.swift.template Package.swift
+
+          sed -i "s/__VERSION__/${VERSION}/g" Package.swift
+
+          while read -r FRAMEWORK CHECKSUM; do
+            sed -i "s/__SHA256_${FRAMEWORK}__/${CHECKSUM}/g" Package.swift
+          done < "${RUNNER_TEMP}/checksums.txt"
+
+          if [[ "${UPLOAD_ON_MAIN:-0}" == "1" ]]; then
+            git config --global user.name "PyTorch Bot"
+            git config --global user.email "[email protected]"
+            git add Package.swift
+            git commit -am "${VERSION}"
+            git push -f origin "${BRANCH}"
+          else
+            echo "Draft Package.swift:"
+            cat Package.swift
+          fi
 
   build-benchmark-app:
     name: build-benchmark-app
@@ -282,5 +340,5 @@ jobs:
         echo "::group::Build ExecuTorch benchmark app"
         mkdir -p extension/benchmark/apple/Benchmark/Models
         ${CONDA_RUN} --no-capture-output \
-          build/build_apple_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+          build/build_apple_llm_demo.sh "${ARTIFACTS_DIR_NAME}"
         echo "::endgroup::"

.github/workflows/apple.yml

@@ -38,6 +38,7 @@ jobs:
           - docker-image-name: executorch-ubuntu-22.04-clang12
           - docker-image-name: executorch-ubuntu-22.04-linter
           - docker-image-name: executorch-ubuntu-22.04-arm-sdk
+          - docker-image-name: executorch-ubuntu-22.04-qnn-sdk
           - docker-image-name: executorch-ubuntu-22.04-clang12-android
     env:
       DOCKER_IMAGE: 308535385114.dkr.ecr.us-east-1.amazonaws.com/executorch/${{ matrix.docker-image-name }}

.github/workflows/docker-builds.yml

@@ -372,7 +372,7 @@ jobs:
       fail-fast: false
     with:
       runner: linux.2xlarge
-      docker-image: executorch-ubuntu-22.04-clang12-android
+      docker-image: executorch-ubuntu-22.04-qnn-sdk
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 900

.github/workflows/pull.yml

@@ -58,7 +58,7 @@ jobs:
         bash .ci/scripts/setup-conda.sh
         # Setup MacOS dependencies as there is no Docker support on MacOS atm
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh "${BUILD_TOOL}"
-        # Build and test xecutorch
+        # Build and test executorch
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_model.sh "${MODEL_NAME}" "${BUILD_TOOL}" "${BACKEND}" "${DEMO_BACKEND_DELEGATION}"
 
   test-custom-ops-macos:
@@ -411,8 +411,6 @@ jobs:
         pip install -U "huggingface_hub[cli]"
         huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
         pip install accelerate sentencepiece
-        # TODO(guangyang): Switch to use released transformers library after all required patches are included
-        pip install "git+https://github.com/huggingface/transformers.git@6cc4dfe3f1e8d421c6d6351388e06e9b123cbfe1"
         pip list
         echo "::endgroup::"