prototype e2e works on latest ET

Author: Gasoonjia

CMakeLists.txt

@@ -49,6 +49,8 @@
 # https://github.com/google/XNNPACK/commit/c690daa67f883e1b627aadf7684c06797e9a0684
 cmake_minimum_required(VERSION 3.29)
 project(executorch)
+# project(executorch LANGUAGES CXX CUDA)
+
 
 include(${PROJECT_SOURCE_DIR}/tools/cmake/common/preset.cmake)
 include(${PROJECT_SOURCE_DIR}/tools/cmake/Utils.cmake)

backends/aoti/CMakeLists.txt

@@ -21,27 +21,36 @@ if(NOT EXECUTORCH_ROOT)
   set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../..)
 endif()
 
-include(${EXECUTORCH_ROOT}/build/Utils.cmake)
+# include(${EXECUTORCH_ROOT}/build/Utils.cmake)
 
-find_package(CUDA)
-
-set(_common_include_directories ${EXECUTORCH_ROOT}/..)
+find_package(CUDAToolkit REQUIRED)
 
 set(_aoti_sources runtime/AotiBackend.cpp)
-
 add_library(aoti_backend STATIC ${_aoti_sources})
 target_include_directories(
-  aoti_backend PUBLIC ${_common_include_directories} ${CUDA_INCLUDE_DIRS}
+  aoti_backend
+  PUBLIC
+    ${CUDAToolkit_INCLUDE_DIRS}
+    $<BUILD_INTERFACE:${EXECUTORCH_ROOT}>
+    $<INSTALL_INTERFACE:include>
 )
-
 target_compile_options(aoti_backend PUBLIC -fexceptions -frtti -fPIC)
-target_link_libraries(aoti_backend PUBLIC extension_tensor ${CUDA_LIBRARIES})
+# Ensure symbols are exported properly
+target_link_options(aoti_backend PUBLIC -Wl,--export-dynamic)
+# Link against CUDA::cudart (the CUDA runtime library)
+target_link_libraries(
+  aoti_backend
+  PUBLIC
+    extension_tensor
+    CUDA::cudart
+    ${CMAKE_DL_LIBS}
+)
+# If you need other CUDA libraries, link them similarly:
+# target_link_libraries(aoti_backend PUBLIC CUDA::cublas CUDA::cufft ...)
+# If you have a custom function, keep it
 executorch_target_link_options_shared_lib(aoti_backend)
-
 install(
   TARGETS aoti_backend
   EXPORT ExecuTorchTargets
   DESTINATION lib
-  INCLUDES
-  DESTINATION ${_common_include_directories}
 )

backends/aoti/aoti_backend.py

@@ -34,7 +34,7 @@ def preprocess(
         so_path = torch._inductor.aot_compile(graph_module, args, kwargs, options={})  # type: ignore[arg-type]
         print(so_path)
         check_call(
-            f"patchelf --remove-needed libtorch.so --remove-needed libtorch_cuda.so --remove-needed libc10_cuda.so --remove-needed libtorch_cpu.so --add-needed libcudart.so {so_path}",
+            f"patchelf --remove-needed libtorch.so --remove-needed libc10.so --remove-needed libtorch_cuda.so --remove-needed libc10_cuda.so --remove-needed libtorch_cpu.so --add-needed libcudart.so {so_path}",
             shell=True,
         )
 

backends/aoti/runtime/AotiBackend.cpp

@@ -6,7 +6,6 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include <executorch/backends/arm/runtime/VelaBinStream.h>
 #include <executorch/extension/tensor/tensor.h>
 #include <executorch/runtime/backend/interface.h>
 #include <executorch/runtime/core/error.h>

export_and_run_aoti.sh

@@ -0,0 +1,11 @@
+./install_executorch.sh
+python export_add.py
+./install_executorch.sh --clean
+mkdir -p cmake-out
+cd cmake-out
+cmake -DEXECUTORCH_BUILD_AOTI=ON \
+      -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
+      ..
+cd ..
+cmake --build cmake-out -j9
+./cmake-out/executor_runner --model_path add.pte

requirements-dev.txt

@@ -10,3 +10,4 @@ zstd  # Imported by resolve_buck.py.
 certifi  # Imported by resolve_buck.py.
 lintrunner==0.12.7
 lintrunner-adapters==0.12.6
+patchelf