Update CUDA archs in ORT

CMakeLists.txt

@@ -418,7 +418,7 @@ install(
 # that tar file. We copy over the libraries and other requirements
 # prior to running this build and therefore these set of install
 # commands are not needed.
-if(TRITON_ONNXRUNTIME_DOCKER_BUILD)
+if(TRITON_ONNXRUNTIME_DOCKER_BUILD OR DEFINED TRITON_ONNXRUNTIME_PACKAGE_URL)
   install(
     DIRECTORY
       ${CMAKE_CURRENT_BINARY_DIR}/onnxruntime/

cmake/download_onnxruntime.cmake

@@ -1,7 +1,7 @@
 if(DEFINED TRITON_ONNXRUNTIME_PACKAGE_URL)
 
   set(DOWNLOAD_PATH "${CMAKE_BINARY_DIR}/_deps/downloads/onnxruntime.zip")
-  set(EXTRACT_DIR "${CMAKE_BINARY_DIR}/_deps/downloads/onnxruntime")
+  set(EXTRACT_DIR "${CMAKE_BINARY_DIR}/onnxruntime")
 
   message(NOTICE "Downloading onnxruntime: ${TRITON_ONNXRUNTIME_PACKAGE_URL}")
 

tools/gen_ort_dockerfile.py

@@ -56,6 +56,10 @@
         "2024.4",  # OpenVINO short version
         "2024.4.0.16579.c3152d32c9c",  # OpenVINO version with build number
     ),
+    "2024.5.0": (
+        "2024.5",  # OpenVINO short version
+        "2024.5.0.17288.7975fa5da0c",  # OpenVINO version with build number
+    ),
 }
 
 
@@ -262,7 +266,7 @@ def dockerfile_for_linux(output_file):
         if FLAGS.cudnn_home is not None:
             ep_flags += ' --cudnn_home "{}"'.format(FLAGS.cudnn_home)
         elif target_platform() == "igpu":
-            ep_flags += ' --cudnn_home "/usr/lib/aarch64-linux-gnu"'
+            ep_flags += ' --cudnn_home "/usr/include"'
         if FLAGS.ort_tensorrt:
             ep_flags += " --use_tensorrt"
             if FLAGS.ort_version >= "1.12.1":
@@ -281,9 +285,9 @@ def dockerfile_for_linux(output_file):
         ep_flags += (
             " --skip_tests --cmake_extra_defines 'onnxruntime_BUILD_UNIT_TESTS=OFF'"
         )
-        cuda_archs = "53;62;72;87"
+        cuda_archs = "87;101"
     else:
-        cuda_archs = "75;80;86;90"
+        cuda_archs = "75;80;86;90;100;120"
 
     df += """
 WORKDIR /workspace/onnxruntime
@@ -395,7 +399,7 @@ def dockerfile_for_linux(output_file):
 """
     df += """
 RUN cd /opt/onnxruntime/lib && \
-    for i in `find . -mindepth 1 -maxdepth 1 -type f -name '*\.so*'`; do \
+    for i in `find . -mindepth 1 -maxdepth 1 -type f -name '*\\.so*'`; do \
         patchelf --set-rpath '$ORIGIN' $i; \
     done
 
@@ -476,9 +480,9 @@ def dockerfile_for_windows(output_file):
 
     df += """
 WORKDIR /workspace/onnxruntime
-ARG VS_DEVCMD_BAT="\BuildTools\VC\Auxiliary\Build\vcvars64.bat"
+ARG VS_DEVCMD_BAT="\\BuildTools\\VC\\Auxiliary\\Build\\vcvars64.bat"
 RUN powershell Set-Content 'build.bat' -value 'call %VS_DEVCMD_BAT%',(Get-Content 'build.bat')
-RUN build.bat --cmake_generator "Visual Studio 17 2022" --config Release --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=75;80;86;90" --skip_submodule_sync --parallel --build_shared_lib --compile_no_warning_as_error --skip_tests --update --build --build_dir /workspace/build {}
+RUN build.bat --cmake_generator "Visual Studio 17 2022" --config Release --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=75;80;86;90;100;120" --skip_submodule_sync --parallel --build_shared_lib --compile_no_warning_as_error --skip_tests --update --build --build_dir /workspace/build {}
 """.format(
         ep_flags
     )
@@ -559,12 +563,8 @@ def preprocess_gpu_flags():
             FLAGS.tensorrt_home = "/tensorrt"
     else:
         if "CUDNN_VERSION" in os.environ:
-            version = None
-            m = re.match(r"([0-9]\.[0-9])\.[0-9]\.[0-9]", os.environ["CUDNN_VERSION"])
-            if m:
-                version = m.group(1)
             if FLAGS.cudnn_home is None:
-                FLAGS.cudnn_home = "/usr/local/cudnn-{}/cuda".format(version)
+                FLAGS.cudnn_home = "/usr"
 
         if FLAGS.cuda_home is None:
             FLAGS.cuda_home = "/usr/local/cuda"