//    "/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

//   like shadowing a numpy array.
// * We do not provide TF_OK since it is not strictly necessary and we
//   are not optimizing for convenience.
// * We make assumption that one session has one graph.  This should be
//   fine since we have the ability to run sub-graphs.
// * We could allow NULL for some arguments (e.g., NULL options arg).
//   However since convenience is not a primary goal, we don't do this.

int TF_OutputListNumOutputs(TF_OutputList* o);
// Return the `i`th output in the list.
TF_AbstractTensor* TF_OutputListGet(TF_OutputList* o, int i);
// Append a tensor at the end of the output list, growing its size by one.
void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,
                           TF_Status*);

// TF_ExecuteOperation will, if in eager mode, execute, if in graph mode, maybe

    TF_Status* status);

// -----------------------------------------------------------------------------
// Eager Executor APIs.
typedef struct TFE_Executor TFE_Executor;

// Creates a new eager Executor. Nodes in one executor are guaranteed to be
// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the

    // Note: we hold onto mu_ for the rest of this function. In practice, this
    // is okay, as stat requests are typically fast, and concurrent requests are
    // often for the same file. Future work can split this up into one lock per
    // key if this proves to be a significant performance bottleneck.
    absl::MutexLock lock(&mu_);
    if (LookupLocked(key, value)) {
      return TF_SetStatus(status, TF_OK, "");
    }

// TF_TriState is the C API typedef for tri-state.
typedef enum TF_TriState {
  TF_TriState_Default = 0,
  TF_TriState_Off,
  TF_TriState_On,
} TF_TriState;

// TF_GrapplerItem represents a combination of a graph, one of more fetch nodes,
// and potentially a set of nodes to feed.
typedef struct TF_GrapplerItem TF_GrapplerItem;

// Flags indicating whether existing optimizers should be turned off.

  const std::vector<std::string> underlying_devices_;
  // A sequence of thread wrappers, one per device, for executing operations in
  // parallel.
  //
  // Conceptually this is a thread pool with one thread per device. It requires
  // less synchronization than a thread pool would for this task, since Execute
  // acquires each thread in order (and so only one Execute will schedule
  // blocking collective operations at a time), and avoids some dynamic

  void DeleteTrace(int64_t tensor_id);

  // Consumes the internal state of the tape (so cannot be called more than
  // once) and produces the gradient of the target tensors with respect to the
  // source tensors. The output gradients are used if not empty and not
  // null. The result is populated with one tensor per target element.
  // When running backward functions, builds zeros-like tensors for

                                                  TF_Function* function,
                                                  TF_Status* status);

// Removes a function from the context. Once removed, you can no longer
// TFE_Execute it or TFE_Execute any TFE_Op which has it as an attribute or any
// other function which calls it as an attribute.
TF_CAPI_EXPORT extern void TFE_ContextRemoveFunction(TFE_Context* ctx,

  // Consumes the internal state of the tape (so cannot be called more than
  // once unless the tape is persistent) and produces the gradient of the target
  // tensors with respect to the source tensors. The output gradients are used
  // if not empty and not null. The result is populated with one tensor per
  // target element.
  Status ComputeGradient(