Update on "Arm backend: Add INT16 support to rescale operation"

Add INT16 support for RequantizeNode rescale operations in ExecutorTorch ARM backend.

This follows the pattern established for linear, mul, sigmoid, tanh, slice, view/transpose, cat, and FCNode operations, extending int16 support to RequantizeNode rescale operations.

Changes:
- Add INT16 dtype validation support in op_rescale.py
- Enable rescale operations for 16A8W quantization configuration

The 16A8W configuration uses 16-bit activations with 8-bit weights, enabling higher precision for activations while maintaining weight efficiency. RequantizeNode rescale operations are essential for proper quantization scaling in the 16A8W pipeline.

Differential Revision: [D80513725](https://our.internmc.facebook.com/intern/diff/D80513725/)

cc digantdesai freddan80 per zingo oscarandersson8218

[ghstack-poisoned]

Author: Ninja91

.ci/scripts/build-qnn-sdk.sh

@@ -11,8 +11,10 @@ set -o xtrace
 
 build_qnn_backend() {
   echo "Start building qnn backend."
-  export ANDROID_NDK_ROOT=${ANDROID_NDK_ROOT:-/opt/ndk}
-  export QNN_SDK_ROOT=${QNN_SDK_ROOT:-/tmp/qnn/2.28.0.241029}
+  # Source QNN configuration
+  source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/install_qnn_sdk.sh"
+  setup_android_ndk
+  install_qnn
   export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." && pwd)"
 
   parallelism=$(( $(nproc) - 1 ))

.ci/scripts/setup-qnn-deps.sh

@@ -10,4 +10,5 @@ set -ex
 source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/install_qnn_sdk.sh"
 
 setup_libcpp 12
+setup_android_ndk
 install_qnn

.ci/scripts/setup-vulkan-linux-deps.sh

@@ -43,7 +43,7 @@ install_vulkan_sdk() {
   export PATH="${PATH}:${_vulkan_sdk_dir}/${VULKAN_SDK_VERSION}/x86_64/bin/"
 }
 
-VULKAN_SDK_VERSION="1.3.296.0"
+VULKAN_SDK_VERSION="1.4.321.1"
 
 install_swiftshader
 install_vulkan_sdk "${VULKAN_SDK_VERSION}"

.ci/scripts/setup-windows.ps1

@@ -0,0 +1,24 @@
+param (
+    [string]$editable = $false
+)
+
+conda create --yes --quiet -n et python=3.12
+conda activate et
+
+# Activate the VS environment - this is required for Dynamo to work, as it uses MSVC.
+# There are a bunch of environment variables that it requires.
+# See https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line.
+& "C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Launch-VsDevShell.ps1" -Arch amd64
+
+# Install test dependencies
+pip install -r .ci/docker/requirements-ci.txt
+
+if ($editable -eq 'true') {
+    install_executorch.bat --editable
+} else {
+    install_executorch.bat
+}
+if ($LASTEXITCODE -ne 0) {
+    Write-Host "Installation was unsuccessful. Exit code: $LASTEXITCODE."
+    exit $LASTEXITCODE
+}

.ci/scripts/test_llama.sh

@@ -119,8 +119,12 @@ echo "COREML option ${COREML}"
 
 if [[ "${MODE}" =~ .*qnn.* ]]; then
   QNN=ON
+
+  # Download QNN_SDK. If already downloaded, export environment path
+  source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/install_qnn_sdk.sh"
+  install_qnn
+
   export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
-  export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
   export LD_LIBRARY_PATH="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
   export PYTHONPATH=".."
   cp schema/program.fbs exir/_serialize/program.fbs
@@ -150,6 +154,7 @@ cmake_install_executorch_libraries() {
     echo "Installing libexecutorch.a, libextension_module.so, libportable_ops_lib.a"
     rm -rf cmake-out
     retry cmake --preset llm \
+        -DEXECUTORCH_BUILD_TESTS=ON \
         -DBUILD_TESTING=OFF \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
         -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
@@ -166,6 +171,7 @@ cmake_build_llama_runner() {
     popd
     dir="examples/models/llama"
     retry cmake \
+        -DEXECUTORCH_BUILD_TESTS=ON \
         -DBUILD_TESTING=OFF \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
         -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \

.ci/scripts/test_model.ps1

@@ -0,0 +1,89 @@
+param (
+    [string]$modelName,
+    [string]$backend,
+    [string]$buildDir = "cmake-out",
+    [bool]$strict = $false
+)
+
+Set-PSDebug -Trace 1
+$ErrorActionPreference = 'Stop'
+$PSNativeCommandUseErrorActionPreference = $true
+
+function ExportModel-Portable {
+    param (
+        [string]$model_name,
+        [bool]$strict
+    )
+
+    $exportParams = "--model_name", "$modelName"
+    if ($strict) {
+        $exportParams += "--strict"
+    }
+    python -m examples.portable.scripts.export @exportParams | Write-Host
+    if ($LASTEXITCODE -ne 0) {
+        Write-Host "Model export failed. Exit code: $LASTEXITCODE."
+        exit $LASTEXITCODE
+    }
+
+    "$modelName.pte"
+}
+
+function ExportModel-Xnnpack {
+    param (
+        [string]$model_name,
+        [bool]$quantize
+    )
+
+    if $(quantize) {
+        python -m examples.xnnpack.aot_compiler --model_name="${MODEL_NAME}" --delegate --quantize | Write-Host
+        $modelFile = "$($modelName)_xnnpack_q8.pte"
+    } else {
+        python -m examples.xnnpack.aot_compiler --model_name="${MODEL_NAME}" --delegate | Write-Host
+        $modelFile = "$($modelName)_xnnpack_fp32.pte"
+    }
+    if ($LASTEXITCODE -ne 0) {
+        Write-Host "Model export failed. Exit code: $LASTEXITCODE."
+        exit $LASTEXITCODE
+    }
+
+    $modelFile
+}
+
+# Build the runner
+if (Test-Path -Path $buildDir) {
+    Remove-Item -Path $buildDir -Recurse -Force
+}
+New-Item -Path $buildDir -ItemType Directory
+Push-Location $buildDir
+cmake .. --preset windows
+cmake --build . -t executor_runner -j16 --config Release
+if ($LASTEXITCODE -ne 0) {
+    Write-Host "Runner build failed. Exit code: $LASTEXITCODE."
+    exit $LASTEXITCODE
+}
+$executorBinaryPath = Join-Path -Path $buildDir -ChildPath "Release\executor_runner.exe"
+Pop-Location
+
+# Export the model
+switch ($backend) {
+    "portable" {
+        $model_path = ExportModel-Portable -model_name $modelName -strict $strict
+    }
+    "xnnpack-f32" {
+        $model_path = ExportModel-Xnnpack -model_name $modelName -quantize $false
+    }
+    "xnnpack-q8" {
+        $model_path = ExportModel-Xnnpack -model_name $modelName -quantize $true
+    }
+    default {
+        Write-Host "Unknown backend $backend."
+        exit 1
+    }
+}
+
+# Run the runner
+& "$executorBinaryPath" --model_path="$model_path"
+if ($LASTEXITCODE -ne 0) {
+    Write-Host "Model execution failed. Exit code: $LASTEXITCODE."
+    exit $LASTEXITCODE
+}

.ci/scripts/test_qnn_static_llama.sh

@@ -9,8 +9,13 @@ set -euxo pipefail
 
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
+# Source QNN configuration
+source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/qnn_config.sh"
+# Download QNN_SDK. If already downloaded, export environment path
+source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/install_qnn_sdk.sh"
+install_qnn
+
 export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
-export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
 export LD_LIBRARY_PATH="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
 export PYTHONPATH=".."
 cp schema/program.fbs exir/_serialize/program.fbs

.ci/scripts/unittest-windows.ps1

@@ -0,0 +1,15 @@
+param (
+    [string]$editable = $false
+)
+
+Set-PSDebug -Trace 1
+$ErrorActionPreference = 'Stop'
+$PSNativeCommandUseErrorActionPreference = $true
+
+# Run pytest with coverage
+# pytest -n auto --cov=./ --cov-report=xml
+pytest -v --full-trace -c pytest-windows.ini
+if ($LASTEXITCODE -ne 0) {
+    Write-Host "Pytest invocation was unsuccessful. Exit code: $LASTEXITCODE."
+    exit $LASTEXITCODE
+}

.github/workflows/_unittest.yml

@@ -19,6 +19,7 @@ on:
         required: false
         type: string
         description: Install ExecuTorch in editable mode or not.
+        default: 'false'
       python-version:
         required: false
         type: string
@@ -52,3 +53,23 @@ jobs:
         # This is needed to get the prebuilt PyTorch wheel from S3
         ${CONDA_RUN} --no-capture-output pip install awscli==1.37.21
         .ci/scripts/unittest-macos.sh --build-tool "${{ inputs.build-tool }}" --build-mode "${{ inputs.build-mode }}" --editable "${{ inputs.editable }}"
+
+  windows:
+    if: ${{ inputs.build-tool == 'cmake' }}
+    uses: pytorch/test-infra/.github/workflows/windows_job.yml@main
+    with:
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      script: |
+        conda init powershell
+
+        powershell -Command "& {
+          Set-PSDebug -Trace 1
+          \$ErrorActionPreference = 'Stop'
+          \$PSNativeCommandUseErrorActionPreference = \$true
+
+          .ci/scripts/setup-windows.ps1       
+
+          powershell .ci/scripts/unittest-windows.ps1 -editable "${{ inputs.editable }}"
+        }"

.github/workflows/android-perf.yml

@@ -72,7 +72,7 @@ jobs:
           # Separate default values from the workflow dispatch. To ensure defaults are accessible
           # during scheduled runs and to provide flexibility for different defaults between
           # on-demand and periodic benchmarking.
-          CRON_DEFAULT_MODELS: ${{ github.event_name == 'schedule' && 'mv3,mv2,ic4,ic3,resnet50,edsr,mobilebert,w2l,meta-llama/Llama-3.2-1B,meta-llama/Llama-3.2-1B-Instruct-SpinQuant_INT4_EO8,meta-llama/Llama-3.2-1B-Instruct-QLORA_INT4_EO8,Qwen/Qwen3-0.6B,HuggingFaceTB/SmolLM2-135M,allenai/OLMo-1B-hf,google/gemma-3-1b-it' || 'Qwen/Qwen3-0.6B' }}
+          CRON_DEFAULT_MODELS: ${{ github.event_name == 'schedule' && 'mv3,mv2,ic4,ic3,resnet50,mobilebert,w2l,meta-llama/Llama-3.2-1B,meta-llama/Llama-3.2-1B-Instruct-SpinQuant_INT4_EO8,meta-llama/Llama-3.2-1B-Instruct-QLORA_INT4_EO8,Qwen/Qwen3-0.6B,HuggingFaceTB/SmolLM2-135M,allenai/OLMo-1B-hf,google/gemma-3-1b-it' || 'Qwen/Qwen3-0.6B' }}
           CRON_DEFAULT_DEVICES: samsung_galaxy_s22+public
         run: |
           set -eux
@@ -292,7 +292,7 @@ jobs:
                       export.output_name="${OUT_ET_MODEL_NAME}.pte"
                     ls -lh "${OUT_ET_MODEL_NAME}.pte"
                 elif [[ ${{ matrix.config }} == "llama3_qnn_htp" ]]; then
-                    export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
+                    export QNN_SDK_ROOT=/tmp/qnn/2.37.0.25072
                     export LD_LIBRARY_PATH=$QNN_SDK_ROOT/lib/x86_64-linux-clang/
                     export PYTHONPATH=$(pwd)/..
 
@@ -432,7 +432,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
 
         mkdir -p aar-out
-        PYTHON_EXECUTABLE=python ANDROID_ABIS="arm64-v8a" BUILD_AAR_DIR=aar-out EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029 EXECUTORCH_ANDROID_PROFILING=ON bash scripts/build_android_library.sh
+        PYTHON_EXECUTABLE=python ANDROID_ABIS="arm64-v8a" BUILD_AAR_DIR=aar-out EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.37.0.25072 EXECUTORCH_ANDROID_PROFILING=ON bash scripts/build_android_library.sh
         mkdir -p extension/benchmark/android/benchmark/app/libs
         cp aar-out/executorch.aar extension/benchmark/android/benchmark/app/libs
         pushd extension/benchmark/android/benchmark