Windows test changes (#7503)

Co-authored-by: Francesco Petrini <francescogpetrini@gmail.com>

Author: mc-nv

Dockerfile.win10.min

@@ -37,7 +37,7 @@ RUN choco install unzip -y
 #
 # Installing TensorRT
 #
-ARG TENSORRT_VERSION=10.2.0.19
+ARG TENSORRT_VERSION=10.3.0.26
 ARG TENSORRT_ZIP="TensorRT-${TENSORRT_VERSION}.Windows10.x86_64.cuda-12.5.zip"
 ARG TENSORRT_SOURCE=https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.2.0/zip/TensorRT-10.2.0.19.Windows10.x86_64.cuda-12.5.zip
 # COPY ${TENSORRT_ZIP} /tmp/${TENSORRT_ZIP}
@@ -51,7 +51,7 @@ LABEL TENSORRT_VERSION="${TENSORRT_VERSION}"
 #
 # Installing cuDNN
 #
-ARG CUDNN_VERSION=9.2.1.18
+ARG CUDNN_VERSION=9.3.0.75
 ARG CUDNN_ZIP=cudnn-windows-x86_64-${CUDNN_VERSION}_cuda12-archive.zip
 ARG CUDNN_SOURCE=https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-9.2.1.18_cuda12-archive.zip
 ADD ${CUDNN_SOURCE} /tmp/${CUDNN_ZIP}
@@ -107,14 +107,14 @@ ENV VCPKG_TARGET_TRIPLET x64-windows
 #
 # Installing Visual Studio BuildTools: VS17 2022
 #
-ARG BUILDTOOLS_VERSION=17.9.34622.214
+ARG BUILDTOOLS_VERSION=17.10.35201.131
 # Download collect.exe in case of an install failure.
 ADD https://aka.ms/vscollect.exe "C:\tmp\collect.exe"
 
 # Use the latest release channel. For more control, specify the location of an internal layout.
 # Download the Build Tools bootstrapper.
 # ARG BUILD_TOOLS_SOURCE=https://aka.ms/vs/17/release/vs_buildtools.exe
-ARG BUILD_TOOLS_SOURCE=https://download.visualstudio.microsoft.com/download/pr/5e7b923b-7d89-4e14-95b8-a84ab168e243/96b21d216c7954aaf606c6d7ba59a3de991884a8a86c578c767ba349c23188a9/vs_BuildTools.exe
+ARG BUILD_TOOLS_SOURCE=https://download.visualstudio.microsoft.com/download/pr/28626b4b-f88f-4b55-a0cf-f3eaa2c643fb/e6c43d4dfb36338d954cdb3ad9010ab2a479e712088f4f6b016eadcc721bab28/vs_BuildTools.exe
 ADD ${BUILD_TOOLS_SOURCE} vs_buildtools.exe
 # Install Build Tools with the Microsoft.VisualStudio.Workload.VCTools workload, including recommended.
 ARG VS_INSTALL_PATH_WP="C:\BuildTools"
@@ -181,15 +181,15 @@ RUN copy "%CUDA_INSTALL_ROOT_WP%\extras\visual_studio_integration\MSBuildExtensi
 
 RUN setx PATH "%CUDA_INSTALL_ROOT_WP%\bin;%PATH%"
 
-ARG CUDNN_VERSION=9.2.1.18
+ARG CUDNN_VERSION=9.3.0.75
 ENV CUDNN_VERSION ${CUDNN_VERSION}
 COPY --from=dependency_base /cudnn /cudnn
 RUN copy cudnn\bin\cudnn*.dll "%CUDA_INSTALL_ROOT_WP%\bin\."
 RUN copy cudnn\lib\x64\cudnn*.lib "%CUDA_INSTALL_ROOT_WP%\lib\x64\."
 RUN copy cudnn\include\cudnn*.h "%CUDA_INSTALL_ROOT_WP%\include\."
 LABEL CUDNN_VERSION="${CUDNN_VERSION}"
 
-ARG TENSORRT_VERSION=10.2.0.19
+ARG TENSORRT_VERSION=10.3.0.26
 ENV TRT_VERSION ${TENSORRT_VERSION}
 COPY --from=dependency_base /TensorRT /TensorRT
 RUN setx PATH "c:\TensorRT\lib;%PATH%"

build.py

@@ -203,6 +203,8 @@ def header(self, desc=None):
 
         self.comment("Exit script immediately if any command fails")
         if target_platform() == "windows":
+            self._file.write("$UseStructuredOutput = $false\n")
+            self.blankln()
             self._file.write("function ExitWithCode($exitcode) {\n")
             self._file.write("    $host.SetShouldExit($exitcode)\n")
             self._file.write("    exit $exitcode\n")
@@ -644,7 +646,9 @@ def onnxruntime_cmake_args(images, library_paths):
             "onnxruntime",
             "TRITON_BUILD_ONNXRUNTIME_VERSION",
             None,
-            TRITON_VERSION_MAP[FLAGS.version][2],
+            os.getenv("TRITON_BUILD_ONNXRUNTIME_VERSION")
+            if os.getenv("TRITON_BUILD_ONNXRUNTIME_VERSION")
+            else TRITON_VERSION_MAP[FLAGS.version][2],
         )
     ]
 

qa/L0_backend_python/test.sh

@@ -39,18 +39,18 @@ fi
 # /mnt/c when needed but the paths on the tritonserver command-line
 # must be C:/ style.
 export TEST_WINDOWS=0
-if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
     export DATADIR=${DATADIR:="/c/data/inferenceserver/${REPO_VERSION}"}
     export TRITON_DIR=${TRITON_DIR:=c:/tritonserver}
     # This will run in WSL, but Triton will run in windows, so environment
     # variables meant for loaded models must be exported using WSLENV.
     # The /w flag indicates the value should only be included when invoking
     # Win32 from WSL.
-    export WSLENV=TRITON_DIR/w
+    export WSLENV=TRITON_DIR
     export SERVER=${SERVER:=c:/tritonserver/bin/tritonserver.exe}
     export BACKEND_DIR=${BACKEND_DIR:=c:/tritonserver/backends}
     export MODELDIR=${MODELDIR:=c:/}
-    TEST_WINDOWS=1
+    export TEST_WINDOWS=1
 else
     export DATADIR=${DATADIR:="/data/inferenceserver/${REPO_VERSION}"}
     export TRITON_DIR=${TRITON_DIR:="/opt/tritonserver"}

qa/L0_batcher/test.sh

@@ -79,7 +79,7 @@ TF_VERSION=${TF_VERSION:=2}
 # On windows the paths invoked by the script (running in WSL) must use
 # /mnt/c when needed but the paths on the tritonserver command-line
 # must be C:/ style.
-if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
     MODELDIR=${MODELDIR:=C:/models}
     DATADIR=${DATADIR:="/mnt/c/data/inferenceserver/${REPO_VERSION}"}
     BACKEND_DIR=${BACKEND_DIR:=C:/tritonserver/backends}
@@ -601,7 +601,7 @@ done
 TEST_CASE=test_multi_batch_preserve_ordering
 
 # Skip test for Windows. Trace file concats at 8192 chars on Windows.
-if [[ "$(< /proc/sys/kernel/osrelease)" != *microsoft* ]]; then
+if  [[ ! -v WSL_DISTRO_NAME ]] || [[ ! -v MSYSTEM ]]; then
     rm -fr ./custom_models && mkdir ./custom_models && \
         cp -r ../custom_models/custom_zero_1_float32 ./custom_models/. && \
         mkdir -p ./custom_models/custom_zero_1_float32/1

qa/L0_grpc/test.sh

@@ -48,7 +48,7 @@ NGINX_CONF="./nginx.conf"
 # On windows the paths invoked by the script (running in WSL) must use
 # /mnt/c when needed but the paths on the tritonserver command-line
 # must be C:/ style.
-if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
     SDKDIR=${SDKDIR:=C:/sdk}
     MODELDIR=${MODELDIR:=C:/models}
     CLIENT_PLUGIN_MODELDIR=${MODELDIR:=C:/client_plugin_models}

qa/L0_http/test.sh

@@ -49,7 +49,7 @@ NGINX_CONF="./nginx.conf"
 # On windows the paths invoked by the script (running in WSL) must use
 # /mnt/c when needed but the paths on the tritonserver command-line
 # must be C:/ style.
-if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
     SDKDIR=${SDKDIR:=C:/sdk}
     MODELDIR=${MODELDIR:=C:/models}
     DATADIR=${DATADIR:="/mnt/c/data/inferenceserver/${REPO_VERSION}"}

qa/L0_infer/test.sh

@@ -87,7 +87,7 @@ DEFAULT_SHM_SIZE_BYTES=$((1024*1024*$DEFAULT_SHM_SIZE_MB))
 # On windows the paths invoked by the script (running in WSL) must use
 # /mnt/c when needed but the paths on the tritonserver command-line
 # must be C:/ style.
-if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
     MODELDIR=${MODELDIR:=C:/models}
     DATADIR=${DATADIR:="/mnt/c/data/inferenceserver/${REPO_VERSION}"}
     BACKEND_DIR=${BACKEND_DIR:=C:/tritonserver/backends}

qa/L0_sequence_batcher/test.sh

@@ -93,7 +93,7 @@ TF_VERSION=${TF_VERSION:=2}
 # /mnt/c when needed but the paths on the tritonserver command-line
 # must be C:/ style.
 WINDOWS=0
-if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
     MODELDIR=${MODELDIR:=C:/models}
     DATADIR=${DATADIR:="/mnt/c/data/inferenceserver/${REPO_VERSION}"}
     BACKEND_DIR=${BACKEND_DIR:=C:/tritonserver/backends}

qa/L0_trt_plugin/test.sh

@@ -47,7 +47,7 @@ PLUGIN_TEST=trt_plugin_test.py
 # On windows the paths invoked by the script (running in WSL) must use
 # /mnt/c when needed but the paths on the tritonserver command-line
 # must be C:/ style.
-if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
     DATADIR=${DATADIR:="/mnt/c/data/inferenceserver/${REPO_VERSION}"}
     MODELDIR=${MODELDIR:=C:/models}
     CUSTOMPLUGIN=${CUSTOMPLUGIN:=$MODELDIR/HardmaxPlugin.dll}
@@ -135,7 +135,7 @@ SERVER_LD_PRELOAD=$CUSTOMPLUGIN
 SERVER_ARGS=$SERVER_ARGS_BASE
 SERVER_LOG="./inference_server_$LOG_IDX.log"
 
-if [[ "$(< /proc/sys/kernel/osrelease)" != *microsoft* ]]; then
+if  [[ ! -v WSL_DISTRO_NAME ]] || [[ ! -v MSYSTEM ]]; then
     run_server
     if [ "$SERVER_PID" == "0" ]; then
         echo -e "\n***\n*** Failed to start $SERVER\n***"

qa/common/util.sh

@@ -257,7 +257,7 @@ function run_server_nowait () {
         return
     fi
 
-    if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+    if [[ -v WSL_DISTRO_NAME ]] || [[ -v MSYSTEM ]]; then
         # LD_PRELOAD not yet supported on windows
         if [ -z "$SERVER_LD_PRELOAD" ]; then
             echo "=== Running $SERVER $SERVER_ARGS"
@@ -329,7 +329,7 @@ function kill_server () {
     # causes the entire WSL shell to just exit. So instead we must use
     # taskkill.exe which can only forcefully kill tritonserver which
     # means that it does not gracefully exit.
-    if [[ "$(< /proc/sys/kernel/osrelease)" == *microsoft* ]]; then
+    if [[ -v WSL_DISTRO_NAME ]]; then
         # Disable -x as it makes output below hard to read
         oldstate="$(set +o)"; [[ -o errexit ]] && oldstate="$oldstate; set -e"
         set +x
@@ -353,6 +353,8 @@ function kill_server () {
         fi
 
         set +vx; eval "$oldstate"
+    elif [[ -v MSYSTEM ]] ; then
+        taskkill //F //IM tritonserver.exe
     else
         # Non-windows...
         kill $SERVER_PID
@@ -512,17 +514,23 @@ remove_array_outliers() {
 
 function setup_virtualenv() {
     # Create and activate virtual environment
-    virtualenv --system-site-packages venv
-    source venv/bin/activate
-    pip install pytest
+    if [[ -v MSYSTEM ]]; then
+      pip3 install pytest
+    else
+      virtualenv --system-site-packages venv
+      source venv/bin/activate
+      pip install pytest
+    fi
 
     if [[ ${TEST_WINDOWS} == 1 ]]; then
-        pip3 install "numpy<2" tritonclient[all]
+      pip3 install "numpy<2" tritonclient[all]
     fi
 }
 
 function deactivate_virtualenv() {
     # Deactivate virtual environment and clean up
+  if [[ ! -v MSYSTEM ]]; then
     deactivate
     rm -fr venv
+  fi
 }