for (auto* h : handles_task2) {
    TFE_DeleteTensorHandle(h);
  }

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  if (func) {
    TFE_ContextRemoveFunction(ctx, "MatMulFunction", status);
  }
  TFE_DeleteContext(ctx);

                           bool strict_scope_placement, bool* arrived_flag,
                           bool* executed_flag, TF_Status* status);
void AllocateLoggingDevice(const char* name, bool* arrived_flag,
                           bool* executed_flag, TFE_CustomDevice** device,
                           void** device_info);
TFE_TensorHandle* UnpackTensorHandle(TFE_TensorHandle* logged_tensor_handle,

  TFE_DeleteTensorHandle(h1_task1);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_DeleteContext(ctx);

  TF_DeleteStatus(status);

// and `underlying_devices` is
//   {"/job:localhost/replica:0/task:0/device:GPU:0",
//    "/job:localhost/replica:0/task:0/device:GPU:1"}
// Then executing an operation on CUSTOM:0 will execute it on GPU:0 and GPU:1.
//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//

  }
  TFE_OpSetDevice(op.get(), device, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  TFE_TensorHandle* result_handle;
  int num_retvals = 1;
  TFE_Execute(op.get(), &result_handle, &num_retvals, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  return TensorHandlePtr(result_handle);
}

template <typename value_type>

  GradientRegistry registry;
  s = RegisterGradients(&registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto tape = std::make_unique<Tape>(/*persistent=*/false);
  s = Execute(check_numerics_op.get(), ctx.get(), absl::MakeSpan(outputs),
              &num_retvals, &forward_op, tape.get(), registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

  string read_message;

  explicit TFE_Executor(tensorflow::EagerExecutor* executor)
      : owned_executor(nullptr), unowned_executor(executor) {}

  tensorflow::EagerExecutor* executor() {
    return owned_executor == nullptr ? unowned_executor : owned_executor.get();
  }

  std::unique_ptr<tensorflow::EagerExecutor> owned_executor;
  tensorflow::EagerExecutor* unowned_executor;
};

// `struct_size` is used for struct member compatibility check between core TF
// and plug-ins with the same C API minor version. More info here:
// https://github.com/tensorflow/community/blob/master/rfcs/20200612-stream-executor-c-api/C_API_versioning_strategy.md
#define TF_VALIDATE_STRUCT_SIZE(STRUCT_NAME, STRUCT_OBJ, SIZE_VALUE_NAME) \
  do {                                                                    \

    return ptr->getKind() == kGraph || ptr->getKind() == kMlir;
  }
};

// An abstract operation describes an operation by its type, name, and
// attributes. It can be "executed" by the context with some input tensors.
// It is allowed to reusing the same abstract operation for multiple execution
// on a given context, with the same or different input tensors.

// a map (`tensorflow::eager::TensorTape`) from the wrapped tensor to the id of
// the op that produced it (or -1 if this tensor was watched using
// `GradientTape::Watch`.) The op_id is simply a unique index assigned to each
// op executed under the tape. A separate map (`tensorflow::eager::OpTape`)
// maintains the map from `op_id` to a `OpTapeEntry` which stores the `op_type`,
// inputs and outputs and the gradient function These data structures combined