[QNN EP] set up py packaging pipeline for Linux x64

cmake/CMakeLists.txt

@@ -838,7 +838,10 @@ if (onnxruntime_USE_QNN)
     endif()
 
     if (MSVC OR ${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
-      file(GLOB QNN_LIB_FILES LIST_DIRECTORIES false "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/libQnn*.so" "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/Qnn*.dll")
+      file(GLOB QNN_LIB_FILES LIST_DIRECTORIES false "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/libQnn*.so"
+	                                     "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/Qnn*.dll"
+                                             "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/libHtpPrepare.so"
+                                             "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/HtpPrepare.dll")
       if (${QNN_ARCH_ABI} STREQUAL "aarch64-windows-msvc" OR ${QNN_ARCH_ABI} STREQUAL "arm64x-windows-msvc")
         file(GLOB EXTRA_HTP_LIB LIST_DIRECTORIES false "${onnxruntime_QNN_HOME}/lib/hexagon-v68/unsigned/libQnnHtpV68Skel.so"
                                                "${onnxruntime_QNN_HOME}/lib/hexagon-v73/unsigned/libQnnHtpV73Skel.so"

cmake/onnxruntime_python.cmake

@@ -1027,6 +1027,12 @@ if (onnxruntime_USE_QNN)
         ${QNN_LIB_FILES}
         $<TARGET_FILE_DIR:${build_output_target}>/onnxruntime/capi/
   )
+  add_custom_command(
+    TARGET onnxruntime_pybind11_state POST_BUILD
+    COMMAND ${CMAKE_COMMAND} -E copy
+        $<TARGET_FILE:onnxruntime_qnn_ctx_gen>
+        $<TARGET_FILE_DIR:${build_output_target}>/onnxruntime/capi/
+  )
   if (EXISTS "${onnxruntime_QNN_HOME}/Qualcomm AI Hub Proprietary License.pdf")
     add_custom_command(
       TARGET onnxruntime_pybind11_state POST_BUILD

setup.py

@@ -59,6 +59,7 @@ def parse_arg_remove_string(argv, arg_name_equal):
 is_migraphx = False
 is_rocm = False
 is_openvino = False
+is_qnn = False
 # The following arguments are mutually exclusive
 if wheel_name_suffix == "gpu":
     # TODO: how to support multiple CUDA versions?
@@ -87,7 +88,8 @@ def parse_arg_remove_string(argv, arg_name_equal):
     # keep the same name since AzureEP will release with CpuEP by default.
     pass
 elif parse_arg_remove_boolean(sys.argv, "--use_qnn"):
-    package_name = "onnxruntime-qnn"
+    is_qnn = True
+    package_name = "onnxruntime-qnn" if not nightly_build else "ort-qnn-nightly"
 
 if is_rocm or is_migraphx:
     package_name = "onnxruntime-rocm" if not nightly_build else "ort-rocm-nightly"
@@ -278,7 +280,7 @@ def run(self):
                 pass
 
             _bdist_wheel.run(self)
-            if is_manylinux and not disable_auditwheel_repair and not is_openvino:
+            if is_manylinux and not disable_auditwheel_repair and not is_openvino and not is_qnn:
                 assert self.dist_dir is not None
                 file = glob(path.join(self.dist_dir, "*linux*.whl"))[0]
                 logger.info("repairing %s for manylinux1", file)
@@ -348,6 +350,16 @@ def finalize_options(self):
     libs.append(providers_cuda_or_rocm)
     libs.append(providers_tensorrt_or_migraphx)
     libs.append(providers_cann)
+    # QNN
+    qnn_deps = [
+        "libQnnCpu.so",
+        "libQnnHtp.so",
+        "libQnnSaver.so",
+        "libQnnSystem.so",
+        "libHtpPrepare.so",
+        "onnxruntime_qnn_ctx_gen",
+    ]
+    dl_libs.extend(qnn_deps)
     if nightly_build:
         libs.extend(["libonnxruntime_pywrapper.so"])
 elif platform.system() == "Darwin":

tools/ci_build/build.py

@@ -2277,10 +2277,10 @@ def build_python_wheel(
             args.append("--wheel_name_suffix=directml")
         elif use_cann:
             args.append("--use_cann")
-        elif use_azure:
-            args.append("--use_azure")
         elif use_qnn:
             args.append("--use_qnn")
+        elif use_azure:
+            args.append("--use_azure")
 
         run_subprocess(args, cwd=cwd)
 

tools/ci_build/github/azure-pipelines/orttraining-py-packaging-pipeline-cpu.yml

@@ -25,3 +25,4 @@ stages:
     enable_windows_arm64_qnn: false
     enable_windows_arm64ec_qnn: false
     enable_windows_x64_qnn: false
+    enable_linux_x64_qnn: false

tools/ci_build/github/azure-pipelines/py-packaging-pipeline.yml

@@ -44,6 +44,11 @@ parameters:
   type: boolean
   default: true
 
+- name: enable_linux_x64_qnn
+  displayName: 'Whether Linux x86_64 package with QNN EP is built.'
+  type: boolean
+  default: true
+
 - name: build_py_parameters
   displayName: 'Specify extra build parameters'
   type: string
@@ -80,6 +85,7 @@ stages:
     enable_windows_arm64_qnn: ${{ parameters.enable_windows_arm64_qnn }}
     enable_windows_arm64ec_qnn: ${{ parameters.enable_windows_arm64ec_qnn }}
     enable_windows_x64_qnn: ${{ parameters.enable_windows_x64_qnn }}
+    enable_linux_x64_qnn: ${{ parameters.enable_linux_x64_qnn }}
     build_py_parameters: ${{ parameters.build_py_parameters }}
     cmake_build_type: ${{ parameters.cmake_build_type }}
     qnn_sdk_version: ${{ parameters.qnn_sdk_version }}

tools/ci_build/github/azure-pipelines/templates/py-linux-qnn.yml

@@ -0,0 +1,97 @@
+parameters:
+- name: machine_pool
+  type: string
+
+- name: cmake_build_type
+  type: string
+  default: 'Release'
+  values:
+   - Debug
+   - Release
+   - RelWithDebInfo
+   - MinSizeRel
+
+- name: device
+  type: string
+  default: 'NPU'
+
+- name: with_cache
+  type: boolean
+  default: false
+
+- name: extra_build_arg
+  type: string
+  default: ''
+
+- name: QnnSdk
+  displayName: QNN SDK version
+  type: string
+  default: 2.26.0.240828
+
+jobs:
+- job: Linux_py_qnn_Wheels_x64
+  timeoutInMinutes: 240
+  workspace:
+    clean: all
+  pool: ${{ parameters.machine_pool }}
+  variables:
+    # The build machine pool doesn't have dotnet, so it can't run CG.
+    - name: skipComponentGovernanceDetection
+      value: true
+    - name: ORT_CACHE_DIR
+      value: $(Agent.TempDirectory)/ort_ccache
+    - name: TODAY
+      value: $[format('{0:dd}{0:MM}{0:yyyy}', pipeline.startTime)]
+    - name: extra_build_args
+      ${{ if ne(parameters.extra_build_arg, '') }}:
+        value: -x ${{ parameters.extra_build_arg }}
+      ${{ if eq(parameters.extra_build_arg, '') }}:
+        value: ''
+  steps:
+    - task: mspremier.PostBuildCleanup.PostBuildCleanup-task.PostBuildCleanup@3
+      displayName: 'Clean Agent Directories'
+      condition: always()
+
+    - checkout: self
+      clean: true
+      submodules: none
+
+    - template: jobs/download_linux_qnn_sdk.yml
+      parameters:
+        QnnSDKVersion: ${{ parameters.QnnSdk }}
+
+    - template: set-nightly-build-option-variable-step.yml
+
+    - template: get-docker-image-steps.yml
+      parameters:
+        Dockerfile: tools/ci_build/github/linux/docker/inference/x86_64/python/cpu/Dockerfile
+        Context: tools/ci_build/github/linux/docker/inference/x86_64/python/cpu
+        DockerBuildArgs: "--build-arg BUILD_UID=$( id -u )"
+        Repository: onnxruntimecpubuildpythonx86_64_qnn
+
+    - template: linux-build-step-with-cache.yml
+      parameters:
+        WithCache: ${{parameters.with_cache}}
+        Today: $(TODAY)
+        AdditionalKey: Linux_py_qnn_Wheels_x64
+        CacheDir: $(ORT_CACHE_DIR)
+        ChangeEveryCommit: true
+        BuildStep:
+          - task: Bash@3
+            displayName: 'Build Python Wheel'
+            inputs:
+              targetType: filePath
+              filePath: tools/ci_build/github/linux/run_python_dockerbuild.sh
+              arguments: -i onnxruntimecpubuildpythonx86_64_qnn -d "${{ parameters.device }}" -c ${{ parameters.cmake_build_type }} $(extra_build_args)
+            env:
+              ADDITIONAL_DOCKER_PARAMETER: "--volume $(QnnSDKRootDir):/qnn_sdk"
+
+    - task: PublishBuildArtifacts@1
+      displayName: 'Publish Artifact: Linux ONNXRuntime QNN python wheel'
+      inputs:
+        PathtoPublish: '$(Build.BinariesDirectory)/dist'
+        ArtifactName: onnxruntime-linux-qnn-x64
+
+    - template: component-governance-component-detection-steps.yml
+      parameters :
+        condition : 'succeeded'

tools/ci_build/github/azure-pipelines/templates/py-packaging-stage.yml

@@ -50,6 +50,11 @@ parameters:
   type: boolean
   default: true
 
+- name: enable_linux_x64_qnn
+  displayName: 'Whether Linux x86_64 package with QNN EP is built.'
+  type: boolean
+  default: true
+
 - name: cmake_build_type
   type: string
   displayName: 'Linux packages cmake build type. Linux Only.'
@@ -533,3 +538,13 @@ stages:
             MACHINE_POOL: 'Onnxruntime-QNNEP-Windows-2022-CPU'
             QNN_SDK: ${{ parameters.qnn_sdk_version }}
             BUILD_PY_PARAMETERS: ${{ parameters.build_py_parameters }}
+
+  - ${{ if eq(parameters.enable_linux_x64_qnn, true) }}:
+    - stage: Python_Packaging_Linux_x64_QNN
+      dependsOn: []
+      jobs:
+      - template: py-linux-qnn.yml
+        parameters:
+          machine_pool: 'onnxruntime-Ubuntu2204-AMD-CPU'
+          extra_build_arg: ${{ parameters.build_py_parameters }}
+          cmake_build_type: ${{ parameters.cmake_build_type }}

tools/ci_build/github/linux/build_linux_python_package.sh

@@ -13,12 +13,12 @@ PYTHON_EXES=("/opt/python/cp38-cp38/bin/python3.8" "/opt/python/cp39-cp39/bin/py
 while getopts "d:p:x:c:" parameter_Option
 do case "${parameter_Option}"
 in
-#GPU or CPU.
+#GPU|CPU|NPU.
 d) BUILD_DEVICE=${OPTARG};;
 p) PYTHON_EXES=${OPTARG};;
 x) EXTRA_ARG=${OPTARG};;
 c) BUILD_CONFIG=${OPTARG};;
-*) echo "Usage: $0 -d <GPU|CPU> [-p <python_exe_path>] [-x <extra_build_arg>] [-c <build_config>]"
+*) echo "Usage: $0 -d <GPU|CPU|NPU> [-p <python_exe_path>] [-x <extra_build_arg>] [-c <build_config>]"
    exit 1;;
 esac
 done
@@ -55,6 +55,11 @@ if [ "$BUILD_DEVICE" == "GPU" ]; then
     BUILD_ARGS+=("--nvcc_threads=1" "--use_cuda" "--use_tensorrt" "--cuda_version=$SHORT_CUDA_VERSION" "--tensorrt_home=/usr" "--cuda_home=/usr/local/cuda-$SHORT_CUDA_VERSION" "--cudnn_home=/usr/local/cuda-$SHORT_CUDA_VERSION" "--cmake_extra_defines" "CMAKE_CUDA_ARCHITECTURES=52;60;61;70;75;80")
 fi
 
+if [ "$BUILD_DEVICE" == "NPU" ]; then
+    #Enable QNN EP
+    BUILD_ARGS+=("--use_qnn" "--qnn_home=/qnn_sdk")
+fi
+
 for PYTHON_EXE in "${PYTHON_EXES[@]}"
 do
   rm -rf /build/"$BUILD_CONFIG"