changes to be made to PR to pass quality checks and send it back or on a
    rare occasion we may reject it.

**3. Review**

-   For a valid PR, reviewer (person familiar with the code/functionality)
    checks if the PR looks good or needs additional changes.
-   If all looks good, the reviewer will approve the PR.
-   If a change is needed, the contributor is requested to make the suggested
    change.

    int64_t target_id = target_tensor_ids[i];
    if (sources_set.find(target_id) != sources_set.end()) {
      auto tensor = targets[i];
      sources_that_are_targets.insert(
          std::make_pair(target_id, TapeTensor(tensor)));
    }
  }

  TF_RETURN_IF_ERROR(GradientTape::ComputeGradient(
      vspace, target_tensor_ids, source_tensor_ids, sources_that_are_targets,

// Deletes the eager Executor without waiting for enqueued nodes. Please call
// TFE_ExecutorWaitForAllPendingNodes before calling this API if you want to
// make sure all nodes are finished.
TF_CAPI_EXPORT extern void TFE_DeleteExecutor(TFE_Executor*);

// Returns true if the executor is in async mode.
TF_CAPI_EXPORT extern bool TFE_ExecutorIsAsync(TFE_Executor*);

      /*run_eager_op_as_function=*/opts->run_eager_op_as_function,
      /*jit_compile_rewrite=*/opts->jit_compile_rewrite);
#if !defined(IS_MOBILE_PLATFORM)
  eager_context->SetDistributedManager(
      std::make_unique<tensorflow::EagerContextDistributedManager>(
          eager_context));
#endif  // !IS_MOBILE_PLATFORM
  return tensorflow::wrap(eager_context);
}

void TFE_DeleteContext(TFE_Context* ctx) {

        "GradientDef is not supported in reading Dataset related functions: ",
        text_proto);
    return {};
  }
  std::vector<UniqueFuncPtr> ret;
  for (const FunctionDef& fdef : fdef_lib.function()) {
    // Make a copy so that we can mutate it.
    FunctionDef fdef_to_load = fdef;
    if (mutate_proto_func) {
      (*mutate_proto_func)(&fdef_to_load);
    }
    VLOG(1) << "Adding func to graph: " << fdef_to_load.DebugString();

  }
  if (!ShouldRecord(input_tensor_id, input_dtypes)) {
    return absl::OkStatus();
  }

  // We may need to allocate zero inputs for trainable dtypes we don't have JVPs
  // for. Make sure they get cleaned up.
  std::vector<Gradient*> new_zeros;
  auto delete_new_zeros = gtl::MakeCleanup([&new_zeros, this] {
    for (Gradient* tensor : new_zeros) {
      this->vspace_.DeleteGradient(tensor);
    }

  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

ParallelDevice::ParallelDevice(const std::vector<std::string>& devices,
                               bool is_async, int in_flight_nodes_limit)
    : underlying_devices_(devices),
      default_cancellation_manager_(absl::make_unique<CancellationManager>()) {
  device_threads_.reserve(devices.size());
  for (int device_index = 0; device_index < devices.size(); ++device_index) {
    device_threads_.emplace_back(new DeviceThread(

}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(
    const TF_Graph* fn_body, const char* fn_name, int ninputs,
    const TF_Output* inputs, std::vector<OutputTensor>* input_tensors,