int keep_alive_secs) = 0;

  // Set up a multi-client distributed execution environment. Must be called
  // on all tasks in the cluster. This call internally coordinates with other
  // tasks to initialize the eager context and TF server for multi-client
  // execution.
  virtual Status EnableCollectiveOps(const ServerDef& server_def) = 0;

                                   int num_tasks);

// Create a ServerDef with job name "localhost" and add `num_tasks` tasks in it.
tensorflow::ServerDef GetServerDef(int num_tasks);

// Create a multi-client ServerDef with the given `job_name`, add `num_tasks`
// tasks and `num_virtual_gpus` virtual GPUs in it.
tensorflow::ServerDef GetMultiClientServerDef(const std::string& job_name,

TF_CAPI_EXPORT extern void TFE_AbortCollectiveOps(TFE_Context* ctx,
                                                  TF_Status* status);

// Checks the health of collective ops peers. Explicit health check is needed in
// multi worker collective ops to detect failures in the cluster.  If a peer is
// down, collective ops may hang.
TF_CAPI_EXPORT extern void TFE_CollectiveOpsCheckPeerHealth(
    TFE_Context* ctx, const char* task, int64_t timeout_in_ms,

#if !defined(IS_MOBILE_PLATFORM)
  // Set up a multi-client distributed execution environment. Must be called on
  // all tasks in the cluster.
  // This call internally coordinates with other tasks to initialize the eager
  // context and TF server for multi-client execution.
  virtual Status EnableCollectiveOps(const ServerDef& server_def) = 0;

// Usually this happens automatically in TF_SessionRun. After this is called,
// 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.