device_name, "\")");
}

Status ImmediateExecutionTensorHandle::SummarizeValue(
    std::string& summary) const {
  Status status;
  AbstractTensorPtr resolved(
      // TODO(allenl): Resolve should be const, and the caches that get updated
      // marked mutable.
      const_cast<ImmediateExecutionTensorHandle*>(this)->Resolve(&status));
  if (!status.ok()) {
    return status;
  }

                                                TF_Buffer* buf);

// APIs for generically dealing with op attributes (e.g. when forwarding them
// through custom device implementations).
//
// TODO(allenl): Currently these are black boxes, but we should have some way to
// inspect values. This would let people e.g. copy over most attributes and then
// modify some based on their values.

This release contains contributions from many people at Google, as well as:

Aadhitya A, Abhilash Mahendrakar, Abhishek Varma, Abin Shahab, Adam Hillier,
Aditya Kane, AdityaKane2001, ag.ramesh, Amogh Joshi, Armen Poghosov,
armkevincheng, Avrosh K, Ayan Moitra, azazhu, Banikumar Maiti, Bas Aarts, bhack,
Bhanu Prakash Bandaru Venkata, Billy Cao, Bohumir Zamecnik, Bradley Reece,

      // Note: Needs to be zeros rather than ones since there's currently no
      // ones_like for variants.
      aid = output_tensor.ZerosLike();
    } else {
      // TODO(allenl): Figure out why using zeros_like everywhere causes issues
      // for some gradient functions and if there's another way to work around
      // it (e.g. conds instead of ifs). The value shouldn't really matter.

    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;
  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {

  // acquires each thread in order (and so only one Execute will schedule
  // blocking collective operations at a time), and avoids some dynamic
  // allocation/scheduling.
  //
  // TODO(allenl): Keep a map from outer thread to list of inner threads rather
  // than a single list of threads so aliased nested parallel devices don't
  // re-use a thread.
  std::vector<std::unique_ptr<DeviceThread>> device_threads_;

#include "tensorflow/c/c_api.h"
#include "tensorflow/c/c_api_experimental.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(
            TFE_NewExecutor(is_async, /*enable_streaming_enqueue=*/true,

#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"
#include "tensorflow/c/tf_status_internal.h"
#include "tensorflow/core/lib/core/status_test_util.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.