ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
    TF_DeleteAbstractTensor(add_output1);
    TF_DeleteAbstractTensor(add_output2);
    TF_DeleteOutputList(func_outputs);
  }

  /**
   * We traced so far this function:
   *
   *   def two_adds(a, b):
   *     my_add1 = a + b
   *     my_add2 = b + b
   *     return my_add1, my_add2
   *
   * Now we will execute this function with an eager context:

    ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
    ASSERT_TRUE(absl::StrContains(device_name, "GPU:0")) << device_name;

    // .backing_device of shape is CPU since the tensor is backed by CPU
    backing_device_name =
        TFE_TensorHandleBackingDeviceName(retvals[0], status.get());
    ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

}

void NoOpDeallocator(void* data, size_t, void*) {}

TEST(CAPI, MalformedTensor) {
  // See https://github.com/tensorflow/tensorflow/issues/7394
  // num_dims = 0 implies a scalar, so should be backed by at least 4 bytes of
  // data.
  TF_Tensor* t =
      TF_NewTensor(TF_FLOAT, nullptr, 0, nullptr, 0, &NoOpDeallocator, nullptr);
  ASSERT_TRUE(t == nullptr);
}

TEST(CAPI, AllocateTensor) {

TF_CAPI_EXPORT TFE_TensorHandle* TFE_NewTensorHandleFromTensor(
    TFE_Context* ctx, TF_Tensor* t, TF_Status* status);

// Create a packed TensorHandle with the given list of TensorHandles.
// If `handles` are on the same device, assign the same device to the packed
// handle; if `handles` are on different deivces, assign a CompositeDevice to
// it.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_CreatePackedTensorHandle(

TensorFlow process effectively executes arbitrary code.

The risk of loading untrusted checkpoints depends on the code or graph that you
are working with. When loading untrusted checkpoints, the values of the traced
variables from your model are also going to be untrusted. That means that if
your code interacts with the filesystem, network, etc. and uses checkpointed
variables as part of those interactions (ex: using a string variable to build a

  // Same as `AddFunctionDef`, but additionally saves the `stack_traces` under
  // the key of the function definition name (to be retrieved during function
  // instantiation).
  virtual Status AddFunctionDefWithStackTraces(
      const FunctionDef& fdef, const StackTracesMap& stack_traces) = 0;

  // Find and return a added function by its name.

    and a [python tracer](https://www.tensorflow.org/guide/profiler#events)
    which allows you to trace python function calls in your model. Usability
    improvements include better diagnostic messages and
    [profile options](https://tensorflow.org/guide/profiler#collect_performance_data)
    to customize the host and device trace verbosity level.

#include "tensorflow/core/platform/intrusive_ptr.h"
#include "tensorflow/core/platform/refcount.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

class FunctionRecord;

// A traced function: this hides the complexity of converting the serialized
// representation between various supported formats e.g. FunctionDef and Mlir
// function.
class AbstractFunction : public core::RefCounted {
 protected:

                                      unneeded_gradients, out_gradients,
                                      absl::MakeSpan(in_gradients));
      if (!persistent_) {
        trace.backward_function_deleter(trace.backward_function);
      }
      if (!s.ok()) {
        return s;
      }
    } else {
      if (!persistent_) {
        trace.backward_function_deleter(trace.backward_function);

//
// This serves as a factory for creating `AbstractOperation`s and for
// registering traced functions.
// Operations creation within a context can only be executed in that context
// (for now at least).
// Implementations of the context may contain some state e.g. an execution
// environment, a traced representation etc.
class AbstractContext {
 protected: