[slimtensor] Add item() methods with CUDA and CPU support for scalar value extraction (#16840)

This PR was created by the merge bot to help merge the original PR into
the main branch.
ghstack PR number: #16445 by
@Gasoonjia
^ Please use this as the source of truth for the PR details, comments,
and reviews
ghstack PR base:
https://github.com/pytorch/executorch/tree/gh/gasoonjia/87/base
ghstack PR head:
https://github.com/pytorch/executorch/tree/gh/gasoonjia/87/head
Merge bot PR base:
https://github.com/pytorch/executorch/tree/gh/gasoonjia/86/orig
Merge bot PR head:
https://github.com/pytorch/executorch/tree/gh/gasoonjia/87/orig
Differential Revision:
[D90034340](https://our.internmc.facebook.com/intern/diff/D90034340/)
@diff-train-skip-merge

---------

Co-authored-by: gasoonjia <gasoonjia@icloud.com>

Author: pytorchbot

backends/aoti/slim/core/SlimTensor.h

@@ -504,6 +504,34 @@ class SlimTensor {
     return *this;
   }
 
+  /**
+   * Extract the scalar value from a tensor with exactly 1 element.
+   * Automatically handles CUDA tensors by copying data to CPU.
+   *
+   * @tparam T The type to extract (must match tensor dtype).
+   * @return The scalar value.
+   */
+  template <typename T>
+  T item() const {
+    ET_CHECK_MSG(
+        this->numel() == 1,
+        "item() requires tensor to have exactly 1 element, got %zu",
+        this->numel());
+
+    T result;
+    if (this->is_cpu()) {
+      result = *static_cast<const T*>(this->data_ptr());
+    } else {
+#if defined(CUDA_AVAILABLE)
+      DeviceTraits<c10::DeviceType::CUDA>::memcpy(
+          &result, this->data_ptr(), sizeof(T), CPU_DEVICE, this->device());
+#else
+      ET_CHECK_MSG(false, "item(): CUDA tensor but CUDA support not available");
+#endif
+    }
+    return result;
+  }
+
  private:
   SlimTensor _clone_impl(
       c10::IntArrayRef sizes,

backends/aoti/slim/core/test/test_slimtensor_basic.cpp

@@ -365,6 +365,111 @@ TEST(SlimTensorBasicTest, CopyConstructor) {
   EXPECT_EQ(copy.dtype(), c10::ScalarType::Float);
 }
 
+// =============================================================================
+// Item Tests (Device-Parameterized)
+// =============================================================================
+
+// Helper to set value in storage (handles both CPU and CUDA)
+template <typename T>
+void set_storage_value(
+    Storage& storage,
+    const T& value,
+    const c10::Device& dev) {
+  if (dev.is_cpu()) {
+    *static_cast<T*>(storage->data()) = value;
+  } else {
+#if defined(CUDA_AVAILABLE)
+    DeviceTraits<c10::DeviceType::CUDA>::memcpy(
+        storage->data(), &value, sizeof(T), dev, CPU_DEVICE);
+#endif
+  }
+}
+
+// Template function for testing item<T>() with explicit type
+template <typename T>
+void test_item_typed(
+    const c10::Device& dev,
+    c10::ScalarType dtype,
+    T input_value,
+    T expected_value) {
+  std::vector<int64_t> sizes = {1};
+  std::vector<int64_t> strides = {1};
+  Storage storage(new MaybeOwningStorage(dev, sizeof(T)));
+  set_storage_value(storage, input_value, dev);
+
+  SlimTensor tensor(
+      std::move(storage), makeArrayRef(sizes), makeArrayRef(strides), dtype);
+
+  T result = tensor.item<T>();
+  if constexpr (std::is_floating_point_v<T>) {
+    EXPECT_FLOAT_EQ(result, expected_value);
+  } else {
+    EXPECT_EQ(result, expected_value);
+  }
+}
+
+// Tests for item<T>() with explicit type
+TEST_P(SlimTensorBasicDeviceTest, ItemTypedFloat) {
+  test_item_typed<float>(device(), c10::ScalarType::Float, 42.5f, 42.5f);
+}
+
+TEST_P(SlimTensorBasicDeviceTest, ItemTypedInt) {
+  test_item_typed<int32_t>(device(), c10::ScalarType::Int, 123, 123);
+}
+
+TEST_P(SlimTensorBasicDeviceTest, ItemTypedLong) {
+  test_item_typed<int64_t>(
+      device(), c10::ScalarType::Long, 9876543210LL, 9876543210LL);
+}
+
+TEST_P(SlimTensorBasicDeviceTest, ItemTypedShort) {
+  test_item_typed<int16_t>(device(), c10::ScalarType::Short, 1234, 1234);
+}
+
+TEST_P(SlimTensorBasicDeviceTest, ItemTypedChar) {
+  test_item_typed<int8_t>(device(), c10::ScalarType::Char, -42, -42);
+}
+
+TEST_P(SlimTensorBasicDeviceTest, ItemTypedBool) {
+  test_item_typed<bool>(device(), c10::ScalarType::Bool, true, true);
+}
+
+// Can't reuse test_item_typed() because we need to cast to float explictly for
+// comparison.
+TEST_P(SlimTensorBasicDeviceTest, ItemTypedBFloat16) {
+  c10::BFloat16 input{3.14f};
+  c10::BFloat16 expected{3.14f};
+  std::vector<int64_t> sizes = {1};
+  std::vector<int64_t> strides = {1};
+  Storage storage(new MaybeOwningStorage(device(), sizeof(c10::BFloat16)));
+  set_storage_value(storage, input, device());
+
+  SlimTensor tensor(
+      std::move(storage),
+      makeArrayRef(sizes),
+      makeArrayRef(strides),
+      c10::ScalarType::BFloat16);
+
+  c10::BFloat16 result = tensor.item<c10::BFloat16>();
+  EXPECT_FLOAT_EQ(static_cast<float>(result), static_cast<float>(expected));
+}
+
+// Test item() fails on non-scalar tensor (numel > 1)
+TEST_P(SlimTensorBasicDeviceTest, ItemFailsOnNonScalarTensor) {
+  std::vector<int64_t> sizes = {2, 3};
+  std::vector<int64_t> strides = {3, 1};
+  Storage storage = make_storage(6 * sizeof(float));
+
+  SlimTensor tensor(
+      std::move(storage),
+      makeArrayRef(sizes),
+      makeArrayRef(strides),
+      c10::ScalarType::Float);
+
+  EXPECT_EQ(tensor.numel(), 6u);
+  EXPECT_DEATH(tensor.item<float>(), "");
+}
+
 // CPU-only test for DataPtrWithOffset (requires reading data back)
 TEST(SlimTensorBasicTest, DataPtrWithOffset) {
   std::vector<int64_t> sizes = {2, 3};