//    "/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.
//
// All component tensors must have the same dtype. Currently they must also have

      : device_(device),
        tensors_(std::move(tensors)),
        shape_(std::vector<int64_t>(shape.begin(), shape.end())),
        dtype_(dtype) {}
  ParallelTensor(const ParallelDevice& device,
                 std::vector<TensorHandlePtr> tensors, const TF_DataType dtype)
      : device_(device),
        tensors_(std::move(tensors)),
        shape_(absl::nullopt),
        dtype_(dtype) {}

  Cleanup(Cleanup&& src)  // NOLINT
      : released_(src.is_released()), f_(src.release()) {}

  // Implicitly move-constructible from any compatible Cleanup<G>.
  // The source will be released as if src.release() were called.
  // A moved-from Cleanup can be safely destroyed or reassigned.
  template <typename G>
  Cleanup(Cleanup<G>&& src)  // NOLINT

      const char* description, LabelDesc&&... label) {
    sampler = absl::WrapUnique(tensorflow::monitoring::Sampler<NumLabels>::New(
        {name, description, label...}, std::move(buckets)));
  }

  std::unique_ptr<tensorflow::monitoring::Sampler<NumLabels>> sampler;
};

struct TFE_MonitoringSampler0 : TFE_MonitoringSampler<0> {
  using TFE_MonitoringSampler::TFE_MonitoringSampler;

  // will be blocked till the copy completes. This is the default placement
  // policy.
  TFE_DEVICE_PLACEMENT_SILENT = 2,
  // Placement policy which silently copies int32 tensors but not other dtypes.
  TFE_DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
} TFE_ContextDevicePlacementPolicy;
// LINT.ThenChange(//tensorflow/c/eager/immediate_execution_context.h)

// in tests only.)
TF_CAPI_EXPORT int TF_PickUnusedPortOrDie(void);

// Fast path method that makes constructing a single scalar tensor require less
// overhead and copies.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandleFromScalar(
    TF_DataType data_type, void* data, size_t len, TF_Status* status);

// Specify the server_def that enables collective ops.

                                         TF_AbstractFunction*, TF_Status*);

// -----------------------------------------------------------------------------
// APIs specific to Eager modes
// -----------------------------------------------------------------------------

// Temporary APIs till we figure out how to create scalar valued Eager
// tensors and how to get value out of eager abstract tensors.

TF_CAPI_EXPORT extern bool TFE_ExecutorIsAsync(TFE_Executor*);

// Causes the calling thread to block till all ops dispatched in this executor
// have been executed. Note that "execution" here refers to kernel execution /
// scheduling of copies, etc. Similar to sync execution, it doesn't guarantee
// that lower level device queues (like GPU streams) have been flushed.
//
// This call may not block for execution of ops enqueued concurrently with this
// call.

#define TENSORFLOW_C_EXPERIMENTAL_FILESYSTEM_PLUGINS_POSIX_POSIX_FILESYSTEM_HELPER_H_

#include <dirent.h>
#include <sys/stat.h>

namespace tf_posix_filesystem {

// Copies up to `size` of `src` to `dst`, creating destination if needed.
//
// Callers should pass size of `src` in `size` and the permissions of `src` in
// `mode`. The later is only used if `dst` needs to be created.

  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.
  DEVICE_PLACEMENT_SILENT = 2,
  // Placement policy which silently copies int32 tensors but not other dtypes.
  DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
};
// LINT.ThenChange(//tensorflow/c/eager/c_api.h)

// Abstract interface to a context.
//