[slimtensor] Add aoti_torch_copy_ for SlimTensor

backends/cuda/runtime/shims/memory_slim.cpp

@@ -158,6 +158,67 @@ AOTITorchError aoti_torch_new_tensor_handle(
   return Error::Ok;
 }
 
+AOTITorchError aoti_torch__reinterpret_tensor(
+    Tensor* self,
+    int64_t ndim,
+    const int64_t* sizes_ptr,
+    const int64_t* strides_ptr,
+    int64_t storage_offset,
+    Tensor** ret_new_tensor) {
+  ET_CHECK_OR_RETURN_ERROR(
+      self != nullptr,
+      InvalidArgument,
+      "aoti_torch__reinterpret_tensor: self is null");
+
+  ET_CHECK_OR_RETURN_ERROR(
+      ret_new_tensor != nullptr,
+      InvalidArgument,
+      "aoti_torch__reinterpret_tensor: ret_new_tensor is null");
+
+  ET_CHECK_OR_RETURN_ERROR(
+      ndim >= 0,
+      InvalidArgument,
+      "aoti_torch__reinterpret_tensor: ndim must be non-negative, got %lld",
+      static_cast<long long>(ndim));
+
+  ET_CHECK_OR_RETURN_ERROR(
+      !(sizes_ptr == nullptr && ndim > 0),
+      InvalidArgument,
+      "aoti_torch__reinterpret_tensor: sizes_ptr is null but ndim > 0");
+
+  IntArrayRef sizes(sizes_ptr, static_cast<size_t>(ndim));
+  IntArrayRef strides(strides_ptr, static_cast<size_t>(ndim));
+
+  // Create a new tensor view using as_strided. This creates a tensor that
+  // shares the same underlying storage but with different sizes, strides,
+  // and storage offset. SlimTensor::as_strided() handles this via copy
+  // constructor which shares the SharedPtr<Storage>.
+  *ret_new_tensor =
+      new Tensor(self->as_strided(sizes, strides, storage_offset));
+
+  return Error::Ok;
+}
+
+AOTITorchError
+aoti_torch_copy_(Tensor* self, Tensor* src, int32_t non_blocking) {
+  (void)non_blocking; // SlimTensor::copy_() is always synchronous for now
+
+  ET_CHECK_OR_RETURN_ERROR(
+      self != nullptr, InvalidArgument, "aoti_torch_copy_: self is null");
+
+  ET_CHECK_OR_RETURN_ERROR(
+      src != nullptr, InvalidArgument, "aoti_torch_copy_: src is null");
+
+  // SlimTensor::copy_() handles:
+  // - Same numel validation
+  // - Same dtype validation
+  // - CPU-CPU, CPU-CUDA, CUDA-CPU, CUDA-CUDA copies
+  // - Contiguous fast path and non-contiguous element-wise copy
+  self->copy_(*src);
+
+  return Error::Ok;
+}
+
 } // extern "C"
 
 } // namespace executorch::backends::cuda

backends/cuda/runtime/shims/memory_slim.h

@@ -106,6 +106,43 @@ AOTI_SHIM_EXPORT AOTITorchError aoti_torch_delete_tensor_object(Tensor* tensor);
 AOTI_SHIM_EXPORT AOTITorchError
 aoti_torch_new_tensor_handle(Tensor* orig_handle, Tensor** new_handle);
 
+/**
+ * Creates a reinterpreted view of a tensor with new sizes, strides, and offset.
+ *
+ * This is equivalent to torch.as_strided() - it creates a new tensor that
+ * shares the same underlying storage but with different view parameters.
+ *
+ * @param self Original tensor to reinterpret (must not be null)
+ * @param ndim Number of dimensions for the new view
+ * @param sizes_ptr Pointer to array of dimension sizes
+ * @param strides_ptr Pointer to array of strides for each dimension
+ * @param storage_offset Storage offset in number of elements
+ * @param ret_new_tensor Output parameter for the reinterpreted tensor view
+ * @return AOTITorchError error code (Error::Ok on success)
+ */
+AOTI_SHIM_EXPORT AOTITorchError aoti_torch__reinterpret_tensor(
+    Tensor* self,
+    int64_t ndim,
+    const int64_t* sizes_ptr,
+    const int64_t* strides_ptr,
+    int64_t storage_offset,
+    Tensor** ret_new_tensor);
+
+/**
+ * Copies data from source tensor to destination tensor.
+ *
+ * Handles all device combinations (CPU-CPU, CPU-CUDA, CUDA-CPU, CUDA-CUDA)
+ * and supports tensors with different strides. The destination tensor must
+ * already be allocated with sufficient storage.
+ *
+ * @param self Destination tensor (must not be null)
+ * @param src Source tensor to copy from (must not be null)
+ * @param non_blocking If true, the copy may be asynchronous (currently ignored)
+ * @return AOTITorchError error code (Error::Ok on success)
+ */
+AOTI_SHIM_EXPORT AOTITorchError
+aoti_torch_copy_(Tensor* self, Tensor* src, int32_t non_blocking);
+
 } // extern "C"
 
 } // namespace executorch::backends::cuda

backends/cuda/runtime/shims/tests/targets.bzl

@@ -75,3 +75,5 @@ def define_common_targets():
     cuda_shim_slim_cpp_unittest("aoti_torch_create_tensor_from_blob_v2")
     cuda_shim_slim_cpp_unittest("aoti_torch_delete_tensor_object")
     cuda_shim_slim_cpp_unittest("aoti_torch_new_tensor_handle")
+    cuda_shim_slim_cpp_unittest("aoti_torch__reinterpret_tensor")
+    cuda_shim_slim_cpp_unittest("aoti_torch_copy_")