[fix] Assign CUDAEvent external member properly (pytorch#167711)

# Motivation
This PR aims to fix the bug that the moved-to object's `external_` member is not assigned correctly.

# Additional Context
It's not fine to swap the valid value and the invalid value. We'd just need to prevent double-free.

Pull Request resolved: pytorch#167711
Approved by: https://github.com/albanD

Author: guangyey

aten/src/ATen/cuda/CUDAEvent.h

@@ -238,11 +238,18 @@ struct TORCH_CUDA_CPP_API CUDAEvent {
   }
 
   void moveHelper(CUDAEvent&& other) {
-    std::swap(flags_, other.flags_);
-    std::swap(is_created_, other.is_created_);
-    std::swap(was_recorded_, other.was_recorded_);
-    std::swap(device_index_, other.device_index_);
-    std::swap(event_, other.event_);
+    // Transfer ownership of all state from other to this
+    flags_ = other.flags_;
+    is_created_ = other.is_created_;
+    was_recorded_ = other.was_recorded_;
+    external_ = other.external_;
+    device_index_ = other.device_index_;
+    event_ = other.event_;
+
+    // Reset other to a valid empty state to prevent double-free
+    // The moved-from object must not attempt to destroy the event
+    other.is_created_ = false;
+    other.event_ = cudaEvent_t{};
   }
 };
 

aten/src/ATen/test/CMakeLists.txt

@@ -65,6 +65,7 @@ list(APPEND ATen_CUDA_TEST_SRCS
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_device_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_distributions_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_dlconvertor_test.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/cuda_event_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_exchange_device_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_generator_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_half_test.cu

aten/src/ATen/test/cuda_event_test.cpp

@@ -0,0 +1,36 @@
+#include <gtest/gtest.h>
+
+#include <ATen/cuda/CUDAEvent.h>
+#include <ATen/cuda/CUDAGraph.h>
+#include <ATen/cuda/Sleep.h>
+
+TEST(CUDAEventTest, testCUDAExternalEvent) {
+  if (!at::cuda::is_available()) {
+    return;
+  }
+
+  // Create two external CUDA events
+  unsigned int flags = cudaEventDefault | cudaEventExternal;
+  auto event1 = at::cuda::CUDAEvent(flags);
+  auto event2 = at::cuda::CUDAEvent(flags);
+  // Ensure external CUDAEvent remain valid and functional after being moved.
+  auto start_event = std::move(event1);
+  auto end_event = std::move(event2);
+
+  auto stream = at::cuda::getStreamFromPool();
+  at::cuda::setCurrentCUDAStream(stream);
+
+  auto graph = at::cuda::CUDAGraph();
+  graph.capture_begin();
+  start_event.record();
+  at::cuda::sleep(100000);
+  end_event.record();
+  graph.capture_end();
+
+  // External events should correctly record timestamps even when used inside
+  // CUDA graphs, and elapsed_time() between them should be positive.
+  stream.synchronize();
+  graph.replay();
+  at::cuda::device_synchronize();
+  EXPECT_TRUE(start_event.elapsed_time(end_event) > 0);
+}