Update on "update cuda delegate resource free pipeline for safety and segfault-free"

This diff survives `clear_all_tensors()` function and enable it during backend destroy stage. Furthermore, we defer the container handle deletion to OS to avoid potential segfault if there's more than one .so files.

Differential Revision: [D84135792](https://our.internmc.facebook.com/intern/diff/D84135792/)

[ghstack-poisoned]

Author: Gasoonjia

backends/aoti/CMakeLists.txt

@@ -41,12 +41,7 @@ target_compile_options(aoti_common PUBLIC -fexceptions -frtti -fPIC)
 target_link_options(aoti_common PUBLIC -Wl,--export-dynamic)
 
 # Link against PyTorch libraries and standard libraries
-target_link_libraries(
-  aoti_common
-  PUBLIC extension_tensor ${CMAKE_DL_LIBS}
-         # Link PyTorch libraries for AOTI functions
-         ${TORCH_LIBRARIES}
-)
+target_link_libraries(aoti_common PUBLIC extension_tensor ${CMAKE_DL_LIBS})
 executorch_target_link_options_shared_lib(aoti_common)
 
 install(

backends/aoti/targets.bzl

@@ -51,7 +51,7 @@ def define_common_targets():
         link_whole = True,
         supports_python_dlopen = True,
         visibility = ["@EXECUTORCH_CLIENTS"],
-        deps = [
+        exported_deps = [
             ":common_shims",
             ":model_container",
         ],

backends/cuda/runtime/TARGETS

@@ -34,3 +34,25 @@ runtime.cxx_library(
         ("cuda", None, "cuda-lazy"),
     ],
 )
+
+runtime.cxx_library(
+    name = "cuda_backend",
+    srcs = [
+        "cuda_backend.cpp",
+    ],
+    # @lint-ignore BUCKLINT: Avoid `link_whole=True` (https://fburl.com/avoid-link-whole)
+    link_whole = True,
+    supports_python_dlopen = True,
+    # Constructor needed for backend registration.
+    compiler_flags = ["-Wno-global-constructors"],
+    visibility = ["@EXECUTORCH_CLIENTS"],
+    deps = [
+        ":runtime_shims",
+        "//executorch/backends/aoti:aoti_common",
+        "//executorch/runtime/backend:interface",
+        "//executorch/runtime/core/exec_aten/util:tensor_util",
+    ],
+    external_deps = [
+        ("cuda", None, "cuda-lazy"),
+    ],
+)

backends/cuda/runtime/cuda_backend.cpp

@@ -97,10 +97,10 @@ class ET_EXPERIMENTAL CudaBackend final
     auto aoti_cuda_buffer = named_data_map->get_data(so_blob_key.c_str());
     ET_CHECK_OR_RETURN_ERROR(
         aoti_cuda_buffer.ok(),
-        aoti_cuda_buffer.error(),
+        Internal,
         "Failed to get data for key %s: 0x%x",
         so_blob_key.c_str(),
-        aoti_cuda_buffer.error());
+        static_cast<uint32_t>(aoti_cuda_buffer.error()));
 
     // Generate dynamic temporary file path
     filesystem::path temp_dir = filesystem::temp_directory_path();
@@ -299,20 +299,19 @@ class ET_EXPERIMENTAL CudaBackend final
     if (handle->cuda_stream != nullptr) {
       cudaStream_t cuda_stream = static_cast<cudaStream_t>(handle->cuda_stream);
       cudaError_t stream_err = cudaStreamDestroy(cuda_stream);
-      ET_CHECK_OR_LOG(
+      ET_CHECK_OR_LOG_ERROR(
           stream_err == cudaSuccess,
           "Failed to destroy CUDA stream: %s",
           cudaGetErrorString(stream_err));
       handle->cuda_stream = nullptr;
     }
 
-    // We noticed that AOTInductorModelContainerDelete doesn't work well with
-    // mutitple .so files when we tried to use it to delete container handle,
-    // since freeing one of them will free some sharing resources, leading to
-    // segfault when trying to free the other .so files. Now we do not explicted
-    // delete the container and defer to OS to handle them.
-    // TODO(gasoonjia): find a better and safer solution to delete the
-    // container.
+    // NOTE: AOTInductorModelContainerDelete does not work correctly with
+    // multiple .so files. Deleting one container frees shared resources,
+    // which causes segmentation faults when attempting to delete other
+    // containers. As a workaround, we skip explicit container deletion
+    // and defer cleanup to the OS.
+    // TODO(gasoonjia): Find a proper solution for safe container deletion.
     // AOTInductorModelContainerDelete(handle->container_handle);
 
     // Now close the shared library

runtime/platform/log.h

@@ -188,11 +188,11 @@ using ::executorch::runtime::LogLevel;
  * @param[in] _condition The condition to check.
  * @param[in] _format Log message format string.
  */
-#define ET_CHECK_OR_LOG(_condition, _format, ...) \
-  do {                                            \
-    if (!(_condition)) {                          \
-      ET_LOG(Error, _format, ##__VA_ARGS__);      \
-    }                                             \
+#define ET_CHECK_OR_LOG_ERROR(_condition, _format, ...) \
+  do {                                                  \
+    if (!(_condition)) {                                \
+      ET_LOG(Error, _format, ##__VA_ARGS__);            \
+    }                                                   \
   } while (0)
 
 #else // ET_LOG_ENABLED
@@ -211,6 +211,6 @@ using ::executorch::runtime::LogLevel;
  * @param[in] _condition The condition to check.
  * @param[in] _format Log message format string.
  */
-#define ET_CHECK_OR_LOG(_condition, _format, ...) ((void)0)
+#define ET_CHECK_OR_LOG_ERROR(_condition, _format, ...) ((void)0)
 
 #endif // ET_LOG_ENABLED