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

  // `StartExecute`s will deadlock).
  //
  // If `is_async=false` (constructor argument), `cancellation_manager` must
  // live until `Join` finishes. If `is_async=true` it must live until `Join` is
  // followed by `TFE_ContextAsyncWait` to clear pending operations. It will be
  // used to cancel all other operations if any fails.
  //
  // Set step_id to configure the step id used for rendezvous creation. step id

#define TENSORFLOW_C_EAGER_C_API_EXPERIMENTAL_READER_H_

#include "tensorflow/c/eager/c_api.h"

#ifdef __cplusplus
extern "C" {
#endif

// Test only exports of the monitoring Cell Reader API which allows tests to
// read current values from streamz counters defined in other modules.
//
// The code under test will have created streamz counters like this:
// auto* streamz = tensorflow::monitoring::Counter<1>::New("name",

// TF_SessionRun will no longer extend the session on every call.
//
// We expose this here to allow fine-grained synchronization in multi-threaded
// workloads, which is required since the Python implementation depends on the
// above mutation methods. This allows us to prevent modifications to nodes in
// the graph after the session has been made aware of them.

#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Abstract interface to an operation.
// This interface allows building and executing an operation in either
// tracing or immediate execution mode.
class AbstractOperation {
 protected:
  enum AbstractOperationKind {
    kGraph,
    kMlir,
    kEager,
    kTfrt,

// `enable_xla_compilation` is non-zero, and OFF otherwise.
// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.
// c) ConfigProto.device_count is set to `num_cpu_devices`.
TF_CAPI_EXPORT extern TF_Buffer* TF_CreateConfig(
    unsigned char enable_xla_compilation, unsigned char gpu_memory_allow_growth,
    unsigned int num_cpu_devices);

  }
};

// 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.
class TracingOperation : public AbstractOperation {
 protected:

  bool ShouldRecord(absl::Span<const int64_t> tensor_ids,
                    absl::Span<const tensorflow::DataType> dtypes);

  // Temporarily push or pop transient state for this accumulator.
  //
  // Allows an accumulator which is currently processing an operation to
  // temporarily reset its state. Without pushing and popping, accumulators
  // ignore operations executed as a direct result of their own jvp
  // computations.

// Stores a function representation that can be used for execution or for
// setting functional attributes of other composite ops e.g. control flow.
typedef struct TF_AbstractFunction TF_AbstractFunction;

// This allows the client to swap the implementation of the tracing engine.
// Any future call to TF_CreateFunction will use the implementation defined
// here.
void TF_SetTracingImplementation(const char* name, TF_Status*);

#include "tensorflow/c/c_api_macros.h"
#include "tensorflow/c/tf_status.h"

// --------------------------------------------------------------------------
// C API for Graph. The API is under active development and eventually
// should allow registering a plugin graph optimizer with TensorFlow.
//
// Conventions:
//   * Struct prefix indicates whether struct fields should be filled by the
//     plugin or core implementation: