[slimtensor] integration into backend

backends/aoti/CMakeLists.txt

@@ -25,34 +25,25 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 find_package_torch()
 
-# Common AOTI functionality - combines all AOTI common components
-set(_aoti_common_sources common_shims.cpp)
-add_library(aoti_common STATIC ${_aoti_common_sources})
+# Common AOTI functionality - header-only library for common shims
+add_library(aoti_common INTERFACE)
 target_include_directories(
   aoti_common
-  PUBLIC $<BUILD_INTERFACE:${EXECUTORCH_ROOT}> $<INSTALL_INTERFACE:include>
-         $<BUILD_INTERFACE:${EXECUTORCH_ROOT}/..>
+  INTERFACE $<BUILD_INTERFACE:${EXECUTORCH_ROOT}>
+            $<INSTALL_INTERFACE:include>
+            $<BUILD_INTERFACE:${EXECUTORCH_ROOT}/..>
 )
 target_compile_options(
   aoti_common
-  PUBLIC $<$<CXX_COMPILER_ID:MSVC>:/EHsc /GR>
-         $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions -frtti -fPIC>
+  INTERFACE $<$<CXX_COMPILER_ID:MSVC>:/EHsc /GR>
+            $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions -frtti -fPIC>
 )
 target_compile_definitions(
-  aoti_common PRIVATE $<$<PLATFORM_ID:Windows>:EXPORT_AOTI_FUNCTIONS>
+  aoti_common INTERFACE $<$<PLATFORM_ID:Windows>:EXPORT_AOTI_FUNCTIONS>
 )
-# Ensure symbols are exported properly
-if(APPLE)
-  target_link_options(aoti_common PUBLIC -Wl,-export_dynamic)
-else()
-  target_link_options(
-    aoti_common PUBLIC $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-Wl,--export-dynamic>
-  )
-endif()
 
 # Link against ExecuTorch libraries and standard libraries
-target_link_libraries(aoti_common PUBLIC extension_tensor ${CMAKE_DL_LIBS})
-executorch_target_link_options_shared_lib(aoti_common)
+target_link_libraries(aoti_common INTERFACE extension_tensor ${CMAKE_DL_LIBS})
 
 install(
   TARGETS aoti_common

backends/aoti/aoti_delegate_handle.h

@@ -11,6 +11,11 @@
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/evalue.h>
 #include <string>
+#include <vector>
+
+#ifdef CUDA_AVAILABLE
+#include <executorch/backends/aoti/slim/core/slim_tensor.h>
+#endif
 
 namespace executorch {
 namespace backends {
@@ -95,6 +100,13 @@ struct AOTIDelegateHandle {
   AOTInductorModelContainerGetNumOutputsFunc get_num_outputs;
   AOTInductorModelContainerRunFunc run;
   AOTInductorModelUpdateConstantsFromBlobFunc update_constants_from_blob;
+
+#ifdef CUDA_AVAILABLE
+  // Cached output tensors for skip-copy optimization.
+  // When copy-skip is enabled, output SlimTensors are cached here to keep
+  // GPU memory alive while the caller processes the results.
+  std::vector<slim::SlimTensor> cached_outputs;
+#endif
 };
 
 } // namespace aoti