syntax = "proto3";

package tensorflow;

option cc_enable_arenas = true;
option java_outer_classname = "ConfigProtos";
option java_multiple_files = true;
option java_package = "org.tensorflow.framework";

// add go_package externally with copybara
import "tensorflow/core/framework/cost_graph.proto";
import "tensorflow/core/framework/graph.proto";
import "tensorflow/core/framework/step_stats.proto";
import "tensorflow/core/protobuf/cluster.proto";
import "tensorflow/core/protobuf/debug.proto";
import "tensorflow/core/protobuf/rewriter_config.proto";

message GPUOptions {
  // Fraction of the available GPU memory to allocate for each process.
  // 1 means to allocate all of the GPU memory, 0.5 means the process
  // allocates up to ~50% of the available GPU memory.
  //
  // GPU memory is pre-allocated unless the allow_growth option is enabled.
  //
  // If greater than 1.0, uses CUDA unified memory to potentially oversubscribe
  // the amount of memory available on the GPU device by using host memory as a
  // swap space. Accessing memory not available on the device will be
  // significantly slower as that would require memory transfer between the host
  // and the device. Options to reduce the memory requirement should be
  // considered before enabling this option as this may come with a negative
  // performance impact. Oversubscription using the unified memory requires
  // Pascal class or newer GPUs and it is currently only supported on the Linux
  // operating system. See
  // https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-requirements
  // for the detailed requirements.
  double per_process_gpu_memory_fraction = 1;

  // If true, the allocator does not pre-allocate the entire specified
  // GPU memory region, instead starting small and growing as needed.
  bool allow_growth = 4;

  // The type of GPU allocation strategy to use.
  //
  // Allowed values:
  // "": The empty string (default) uses a system-chosen default
  //     which may change over time.
  //
  // "BFC": A "Best-fit with coalescing" algorithm, simplified from a
  //        version of dlmalloc.
  string allocator_type = 2;

  // Delay deletion of up to this many bytes to reduce the number of
  // interactions with gpu driver code.  If 0, the system chooses
  // a reasonable default (several MBs).
  int64 deferred_deletion_bytes = 3;

  // A comma-separated list of GPU ids that determines the 'visible'
  // to 'virtual' mapping of GPU devices.  For example, if TensorFlow
  // can see 8 GPU devices in the process, and one wanted to map
  // visible GPU devices 5 and 3 as "/device:GPU:0", and "/device:GPU:1",
  // then one would specify this field as "5,3".  This field is similar in
  // spirit to the CUDA_VISIBLE_DEVICES environment variable, except
  // it applies to the visible GPU devices in the process.
  //
  // NOTE:
  // 1. The GPU driver provides the process with the visible GPUs
  //    in an order which is not guaranteed to have any correlation to
  //    the *physical* GPU id in the machine.  This field is used for
  //    remapping "visible" to "virtual", which means this operates only
  //    after the process starts.  Users are required to use vendor
  //    specific mechanisms (e.g., CUDA_VISIBLE_DEVICES) to control the
  //    physical to visible device mapping prior to invoking TensorFlow.
  // 2. In the code, the ids in this list are also called "platform GPU id"s,
  //    and the 'virtual' ids of GPU devices (i.e. the ids in the device
  //    name "/device:GPU:<id>") are also called "TF GPU id"s. Please
  //    refer to third_party/tensorflow/core/common_runtime/gpu/gpu_id.h
  //    for more information.
  string visible_device_list = 5;

  // In the event polling loop sleep this many microseconds between
  // PollEvents calls, when the queue is not empty.  If value is not
  // set or set to 0, gets set to a non-zero default.
  int32 polling_active_delay_usecs = 6;

  // This field is deprecated and ignored.
  int32 polling_inactive_delay_msecs = 7;

  // Force all tensors to be gpu_compatible. On a GPU-enabled TensorFlow,
  // enabling this option forces all CPU tensors to be allocated with Cuda
  // pinned memory. Normally, TensorFlow will infer which tensors should be
  // allocated as the pinned memory. But in case where the inference is
  // incomplete, this option can significantly speed up the cross-device memory
  // copy performance as long as it fits the memory.
  // Note that this option is not something that should be
  // enabled by default for unknown or very large models, since all Cuda pinned
  // memory is unpageable, having too much pinned memory might negatively impact
  // the overall host system performance.
  bool force_gpu_compatible = 8;

  message Experimental {
    // Configuration for breaking down a visible GPU into multiple "virtual"
    // devices.
    message VirtualDevices {
      // Per "virtual" device memory limit, in MB. The number of elements in
      // the list is the number of virtual devices to create on the
      // corresponding visible GPU (see "virtual_devices" below).
      // If empty, it will create single virtual device taking all available
      // memory from the device.
      //
      // For the concept of "visible" and "virtual" GPU, see the comments for
      // "visible_device_list" above for more information.
      repeated float memory_limit_mb = 1;
    }

    // The multi virtual device settings. If empty (not set), it will create
    // single virtual device on each visible GPU, according to the settings
    // in "visible_device_list" above. Otherwise, the number of elements in the
    // list must be the same as the number of visible GPUs (after
    // "visible_device_list" filtering if it is set), and the string represented
    // device names (e.g. /device:GPU:<id>) will refer to the virtual
    // devices and have the <id> field assigned sequentially starting from 0,
    // according to the order they appear in this list and the "memory_limit"
    // list inside each element. For example,
    //   visible_device_list = "1,0"
    //   virtual_devices { memory_limit: 1GB memory_limit: 2GB }
    //   virtual_devices {}
    // will create three virtual devices as:
    //   /device:GPU:0 -> visible GPU 1 with 1GB memory
    //   /device:GPU:1 -> visible GPU 1 with 2GB memory
    //   /device:GPU:2 -> visible GPU 0 with all available memory
    //
    // NOTE:
    // 1. It's invalid to set both this and "per_process_gpu_memory_fraction"
    //    at the same time.
    // 2. Currently this setting is per-process, not per-session. Using
    //    different settings in different sessions within same process will
    //    result in undefined behavior.
    repeated VirtualDevices virtual_devices = 1;

    // If true, uses CUDA unified memory for memory allocations. If
    // per_process_gpu_memory_fraction option is greater than 1.0, then unified
    // memory is used regardless of the value for this field. See comments for
    // per_process_gpu_memory_fraction field for more details and requirements
    // of the unified memory. This option is useful to oversubscribe memory if
    // multiple processes are sharing a single GPU while individually using less
    // than 1.0 per process memory fraction.
    bool use_unified_memory = 2;

    // If > 1, the number of device-to-device copy streams to create
    // for each GPUDevice.  Default value is 0, which is automatically
    // converted to 1.
    int32 num_dev_to_dev_copy_streams = 3;

    // If non-empty, defines a good GPU ring order on a single worker based on
    // device interconnect.  This assumes that all workers have the same GPU
    // topology.  Specify as a comma-separated string, e.g. "3,2,1,0,7,6,5,4".
    // This ring order is used by the RingReducer implementation of
    // CollectiveReduce, and serves as an override to automatic ring order
    // generation in OrderTaskDeviceMap() during CollectiveParam resolution.
    string collective_ring_order = 4;

    // If true then extra work is done by GPUDevice and GPUBFCAllocator to
    // keep track of when GPU memory is freed and when kernels actually
    // complete so that we can know when a nominally free memory chunk
    // is really not subject to pending use.
    bool timestamped_allocator = 5;

    // reserved id: 6

    // Parameters for GPUKernelTracker.  By default no kernel tracking is done.
    // Note that timestamped_allocator is only effective if some tracking is
    // specified.
    //
    // If kernel_tracker_max_interval = n > 0, then a tracking event
    // is inserted after every n kernels without an event.
    int32 kernel_tracker_max_interval = 7;
    // If kernel_tracker_max_bytes = n > 0, then a tracking event is
    // inserted after every series of kernels allocating a sum of
    // memory >= n.  If one kernel allocates b * n bytes, then one
    // event will be inserted after it, but it will count as b against
    // the pending limit.
    int32 kernel_tracker_max_bytes = 8;
    // If kernel_tracker_max_pending > 0 then no more than this many
    // tracking events can be outstanding at a time.  An attempt to
    // launch an additional kernel will stall until an event
    // completes.
    int32 kernel_tracker_max_pending = 9;
  }

  // Everything inside experimental is subject to change and is not subject
  // to API stability guarantees in
  // https://www.tensorflow.org/guide/version_compat.
  Experimental experimental = 9;
}

// Options passed to the graph optimizer
message OptimizerOptions {
  // If true, optimize the graph using common subexpression elimination.
  bool do_common_subexpression_elimination = 1;

  // If true, perform constant folding optimization on the graph.
  bool do_constant_folding = 2;

  // Constant folding optimization replaces tensors whose values can be
  // predetermined, with constant nodes. To avoid inserting too large constants,
  // the size of each constant created can be limited. If this value is zero, a
  // default limit of 10 MiB will be applied. If constant folding optimization
  // is disabled, this value is ignored.
  int64 max_folded_constant_in_bytes = 6;

  // If true, perform function inlining on the graph.
  bool do_function_inlining = 4;

  // Optimization level
  enum Level {
    // L1 is the default level.
    // Optimization performed at L1 :
    // 1. Common subexpression elimination
    // 2. Constant folding
    L1 = 0;

    // No optimizations
    L0 = -1;
  }

  // Overall optimization level. The actual optimizations applied will be the
  // logical OR of the flags that this level implies and any flags already set.
  Level opt_level = 3;

  // Control the use of the compiler/jit.  Experimental.
  enum GlobalJitLevel {
    DEFAULT = 0;  // Default setting ("off" now, but later expected to be "on")
    OFF = -1;
    // The following settings turn on compilation, with higher values being
    // more aggressive.  Higher values may reduce opportunities for parallelism
    // and may use more memory.  (At present, there is no distinction, but this
    // is expected to change.)
    ON_1 = 1;
    ON_2 = 2;
  }
  GlobalJitLevel global_jit_level = 5;
}

message GraphOptions {
  // Removed, use optimizer_options below.
  reserved "skip_common_subexpression_elimination";
  reserved 1;

  // If true, use control flow to schedule the activation of Recv nodes.
  // (Currently ignored.)
  bool enable_recv_scheduling = 2;

  // Options controlling how graph is optimized.
  OptimizerOptions optimizer_options = 3;

  // The number of steps to run before returning a cost model detailing
  // the memory usage and performance of each node of the graph. 0 means
  // no cost model.
  int64 build_cost_model = 4;

  // The number of steps to skip before collecting statistics for the
  // cost model.
  int64 build_cost_model_after = 9;

  // Annotate each Node with Op output shape data, to the extent it can
  // be statically inferred.
  bool infer_shapes = 5;

  // Only place the subgraphs that are run, rather than the entire graph.
  //
  // This is useful for interactive graph building, where one might
  // produce graphs that cannot be placed during the debugging
  // process.  In particular, it allows the client to continue work in
  // a session after adding a node to a graph whose placement
  // constraints are unsatisfiable.
  bool place_pruned_graph = 6;

  // If true, transfer float values between processes as bfloat16.
  bool enable_bfloat16_sendrecv = 7;

  // If > 0, record a timeline every this many steps.
  // EXPERIMENTAL: This currently has no effect in MasterSession.
  int32 timeline_step = 8;

  // Options that control the type and amount of graph rewriting.
  // Not currently configurable via the public Python API (i.e. there is no API
  // stability guarantee if you import RewriterConfig explicitly).
  RewriterConfig rewrite_options = 10;
}

message ThreadPoolOptionProto {
  // The number of threads in the pool.
  //
  // 0 means the system picks a value based on where this option proto is used
  // (see the declaration of the specific field for more info).
  int32 num_threads = 1;

  // The global name of the threadpool.
  //
  // If empty, then the threadpool is made and used according to the scope it's
  // in - e.g., for a session threadpool, it is used by that session only.
  //
  // If non-empty, then:
  // - a global threadpool associated with this name is looked
  //   up or created. This allows, for example, sharing one threadpool across
  //   many sessions (e.g., like the default behavior, if
  //   inter_op_parallelism_threads is not configured), but still partitioning
  //   into a large and small pool.
  // - if the threadpool for this global_name already exists, then it is an
  //   error if the existing pool was created using a different num_threads
  //   value as is specified on this call.
  // - threadpools created this way are never garbage collected.
  string global_name = 2;
}

message RPCOptions {
  // If true, always use RPC to contact the session target.
  //
  // If false (the default option), TensorFlow may use an optimized
  // transport for client-master communication that avoids the RPC
  // stack. This option is primarily for used testing the RPC stack.
  bool use_rpc_for_inprocess_master = 1;

  // The compression algorithm to be used. One of "deflate", "gzip".
  string compression_algorithm = 2;

  // If compression_algorithm is set, the compression level to be used.
  // From 0 (no compression), up to 3.
  int32 compression_level = 3;
}

// Session configuration parameters.
// The system picks appropriate values for fields that are not set.
message ConfigProto {
  // Map from device type name (e.g., "CPU" or "GPU" ) to maximum
  // number of devices of that type to use.  If a particular device
  // type is not found in the map, the system picks an appropriate
  // number.
  map<string, int32> device_count = 1;

  // The execution of an individual op (for some op types) can be
  // parallelized on a pool of intra_op_parallelism_threads.
  // 0 means the system picks an appropriate number.
  int32 intra_op_parallelism_threads = 2;

  // Nodes that perform blocking operations are enqueued on a pool of
  // inter_op_parallelism_threads available in each process.
  //
  // 0 means the system picks an appropriate number.
  // Negative means all operations are performed in caller's thread.
  //
  // Note that the first Session created in the process sets the
  // number of threads for all future sessions unless use_per_session_threads is
  // true or session_inter_op_thread_pool is configured.
  int32 inter_op_parallelism_threads = 5;

  // If true, use a new set of threads for this session rather than the global
  // pool of threads. Only supported by direct sessions.
  //
  // If false, use the global threads created by the first session, or the
  // per-session thread pools configured by session_inter_op_thread_pool.
  //
  // This option is deprecated. The same effect can be achieved by setting
  // session_inter_op_thread_pool to have one element, whose num_threads equals
  // inter_op_parallelism_threads.
  bool use_per_session_threads = 9;

  // This option is experimental - it may be replaced with a different mechanism
  // in the future.
  //
  // Configures session thread pools. If this is configured, then RunOptions for
  // a Run call can select the thread pool to use.
  //
  // The intended use is for when some session invocations need to run in a
  // background pool limited to a small number of threads:
  // - For example, a session may be configured to have one large pool (for
  // regular compute) and one small pool (for periodic, low priority work);
  // using the small pool is currently the mechanism for limiting the inter-op
  // parallelism of the low priority work.  Note that it does not limit the
  // parallelism of work spawned by a single op kernel implementation.
  // - Using this setting is normally not needed in training, but may help some
  // serving use cases.
  // - It is also generally recommended to set the global_name field of this
  // proto, to avoid creating multiple large pools. It is typically better to
  // run the non-low-priority work, even across sessions, in a single large
  // pool.
  repeated ThreadPoolOptionProto session_inter_op_thread_pool = 12;

  // Assignment of Nodes to Devices is recomputed every placement_period
  // steps until the system warms up (at which point the recomputation
  // typically slows down automatically).
  int32 placement_period = 3;

  // When any filters are present sessions will ignore all devices which do not
  // match the filters. Each filter can be partially specified, e.g. "/job:ps"
  // "/job:worker/replica:3", etc.
  repeated string device_filters = 4;

  // Options that apply to all GPUs.
  GPUOptions gpu_options = 6;

  // Whether soft placement is allowed. If allow_soft_placement is true,
  // an op will be placed on CPU if
  //   1. there's no GPU implementation for the OP
  // or
  //   2. no GPU devices are known or registered
  // or
  //   3. need to co-locate with reftype input(s) which are from CPU.
  bool allow_soft_placement = 7;

  // Whether device placements should be logged.
  bool log_device_placement = 8;

  // Options that apply to all graphs.
  GraphOptions graph_options = 10;

  // Global timeout for all blocking operations in this session.  If non-zero,
  // and not overridden on a per-operation basis, this value will be used as the
  // deadline for all blocking operations.
  int64 operation_timeout_in_ms = 11;

  // Options that apply when this session uses the distributed runtime.
  RPCOptions rpc_options = 13;

  // Optional list of all workers to use in this session.
  ClusterDef cluster_def = 14;

  // If true, any resources such as Variables used in the session will not be
  // shared with other sessions. However, when clusterspec propagation is
  // enabled, this field is ignored and sessions are always isolated.
  bool isolate_session_state = 15;

  // Everything inside Experimental is subject to change and is not subject
  // to API stability guarantees in
  // https://www.tensorflow.org/guide/version_compat.
  message Experimental {
    // Task name for group resolution.
    string collective_group_leader = 1;

    // We removed the flag client_handles_error_formatting. Marking the tag
    // number as reserved.
    // TODO(shikharagarwal): Should we just remove this tag so that it can be
    // used in future for other purpose?
    reserved 2;

    // Which executor to use, the default executor will be used
    // if it is an empty string or "DEFAULT"
    string executor_type = 3;

    // Guidance to formatting of large RecvBuf fields for transfer.
    // Any positive value sets the max chunk size.  0 defaults to 4096.
    // Any negative value indicates no max, i.e. one chunk only.
    int32 recv_buf_max_chunk = 4;

    // If true, and supported by the platform, the runtime will attempt to
    // use NUMA affinity where applicable.  One consequence will be the
    // existence of as many CPU devices as there are available NUMA nodes.
    bool use_numa_affinity = 5;

    // If true, make collective op execution order sequential and deterministic
    // for potentially concurrent collective instances.
    bool collective_deterministic_sequential_execution = 6;

    // If true, use NCCL for CollectiveOps.  This feature is highly
    // experimental.
    bool collective_nccl = 7;

    // In the following, session state means the value of a variable, elements
    // in a hash table, or any other resource, accessible by worker sessions
    // held by a TF server.
    //
    // When ClusterSpec propagation is enabled, the value of
    // isolate_session_state is ignored when deciding whether to share session
    // states in a TF server (for backwards compatibility reasons).
    // - If share_session_state_in_clusterspec_propagation is true, the session
    // states are shared.
    // - If share_session_state_in_clusterspec_propagation is false, session
    // states are isolated.
    //
    // When clusterspec propagation is not used, the value of
    // share_session_state_in_clusterspec_propagation is ignored when deciding
    // whether to share session states in a TF server.
    // - If isolate_session_state is true, session states are isolated.
    // - If isolate_session_state is false, session states are shared.
    //
    // TODO(b/129330037): Add a single API that consistently treats
    // isolate_session_state and ClusterSpec propagation.
    bool share_session_state_in_clusterspec_propagation = 8;

    // If using a direct session, disable spinning while waiting for work in
    // the thread pool. This may result in higher latency for completing ops,
    // but in the case where there is a lot of spinning may result in lower
    // CPU usage.
    bool disable_thread_spinning = 9;

    // When true, WorkerSessions are created with device attributes from the
    // full cluster.
    // This is helpful when a worker wants to partition a graph
    // (for example during a PartitionedCallOp).
    bool share_cluster_devices_in_session = 10;
  };

  Experimental experimental = 16;

  // Next: 17
}

// Options for a single Run() call.
message RunOptions {
  // TODO(pbar) Turn this into a TraceOptions proto which allows
  // tracing to be controlled in a more orthogonal manner?
  enum TraceLevel {
    NO_TRACE = 0;
    SOFTWARE_TRACE = 1;
    HARDWARE_TRACE = 2;
    FULL_TRACE = 3;
  }
  TraceLevel trace_level = 1;

  // Time to wait for operation to complete in milliseconds.
  int64 timeout_in_ms = 2;

  // The thread pool to use, if session_inter_op_thread_pool is configured.
  // To use the caller thread set this to -1 - this uses the caller thread
  // to execute Session::Run() and thus avoids a context switch. Using the
  // caller thread to execute Session::Run() should be done ONLY for simple
  // graphs, where the overhead of an additional context switch is
  // comparable with the overhead of Session::Run().
  int32 inter_op_thread_pool = 3;

  // Whether the partition graph(s) executed by the executor(s) should be
  // outputted via RunMetadata.
  bool output_partition_graphs = 5;

  // EXPERIMENTAL.  Options used to initialize DebuggerState, if enabled.
  DebugOptions debug_options = 6;

  // When enabled, causes tensor allocation information to be included in
  // the error message when the Run() call fails because the allocator ran
  // out of memory (OOM).
  //
  // Enabling this option can slow down the Run() call.
  bool report_tensor_allocations_upon_oom = 7;

  // Everything inside Experimental is subject to change and is not subject
  // to API stability guarantees in
  // https://www.tensorflow.org/guide/version_compat.
  message Experimental {
    // If non-zero, declares that this graph is going to use collective
    // ops and must synchronize step_ids with any other graph with this
    // same group_key value (in a distributed computation where tasks
    // run disjoint graphs).
    int64 collective_graph_key = 1;
    // If true, then operations (using the inter-op pool) across all
    // session::run() calls will be centrally scheduled, optimizing for (median
    // and tail) latency.
    // Consider using this option for CPU-bound workloads like inference.
    bool use_run_handler_pool = 2;
  };

  Experimental experimental = 8;

  reserved 4;
}

// Metadata output (i.e., non-Tensor) for a single Run() call.
message RunMetadata {
  // Statistics traced for this step. Populated if tracing is turned on via the
  // "RunOptions" proto.
  // EXPERIMENTAL: The format and set of events may change in future versions.
  StepStats step_stats = 1;

  // The cost graph for the computation defined by the run call.
  CostGraphDef cost_graph = 2;

  // Graphs of the partitions executed by executors.
  repeated GraphDef partition_graphs = 3;

  message FunctionGraphs {
    // TODO(nareshmodi): Include some sort of function/cache-key identifier?
    repeated GraphDef partition_graphs = 1;

    GraphDef pre_optimization_graph = 2;
    GraphDef post_optimization_graph = 3;
  }
  // This is only populated for graphs that are run as functions in TensorFlow
  // V2. There will be an entry below for each function that is traced.
  // The main use cases of the post_optimization_graph and the partition_graphs
  // is to give the caller insight into the graphs that were actually run by the
  // runtime. Additional information (such as those in step_stats) will match
  // these graphs.
  // We also include the pre_optimization_graph since it is usually easier to
  // read, and is helpful in situations where the caller wants to get a high
  // level idea of what the built graph looks like (since the various graph
  // optimization passes might change the structure of the graph significantly).
  repeated FunctionGraphs function_graphs = 4;
}

// Defines a connection between two tensors in a `GraphDef`.
message TensorConnection {
  // A tensor name. The value of this tensor will be substituted for
  // the tensor named in `to_tensor`.
  string from_tensor = 1;

  // A tensor name. The value of this tensor will be bound to the
  // value of the tensor named in `from_tensor`.
  string to_tensor = 2;
}

// Defines a subgraph in another `GraphDef` as a set of feed points and nodes
// to be fetched or executed.
//
// Compare with the arguments to `Session::Run()`.
message CallableOptions {
  // Tensors to be fed in the callable. Each feed is the name of a tensor.
  repeated string feed = 1;

  // Fetches. A list of tensor names. The caller of the callable expects a
  // tensor to be returned for each fetch[i] (see RunStepResponse.tensor). The
  // order of specified fetches does not change the execution order.
  repeated string fetch = 2;

  // Target Nodes. A list of node names. The named nodes will be run by the
  // callable but their outputs will not be returned.
  repeated string target = 3;

  // Options that will be applied to each run.
  RunOptions run_options = 4;

  // Tensors to be connected in the callable. Each TensorConnection denotes
  // a pair of tensors in the graph, between which an edge will be created
  // in the callable.
  repeated TensorConnection tensor_connection = 5;

  // The Tensor objects fed in the callable and fetched from the callable
  // are expected to be backed by host (CPU) memory by default.
  //
  // The options below allow changing that - feeding tensors backed by
  // device memory, or returning tensors that are backed by device memory.
  //
  // The maps below map the name of a feed/fetch tensor (which appears in
  // 'feed' or 'fetch' fields above), to the fully qualified name of the device
  // owning the memory backing the contents of the tensor.
  //
  // For example, creating a callable with the following options:
  //
  // CallableOptions {
  //   feed: "a:0"
  //   feed: "b:0"
  //
  //   fetch: "x:0"
  //   fetch: "y:0"
  //
  //   feed_devices: {
  //     "a:0": "/job:localhost/replica:0/task:0/device:GPU:0"
  //   }
  //
  //   fetch_devices: {
  //     "y:0": "/job:localhost/replica:0/task:0/device:GPU:0"
  //  }
  // }
  //
  // means that the Callable expects:
  // - The first argument ("a:0") is a Tensor backed by GPU memory.
  // - The second argument ("b:0") is a Tensor backed by host memory.
  // and of its return values:
  // - The first output ("x:0") will be backed by host memory.
  // - The second output ("y:0") will be backed by GPU memory.
  //
  // FEEDS:
  // It is the responsibility of the caller to ensure that the memory of the fed
  // tensors will be correctly initialized and synchronized before it is
  // accessed by operations executed during the call to Session::RunCallable().
  //
  // This is typically ensured by using the TensorFlow memory allocators
  // (Device::GetAllocator()) to create the Tensor to be fed.
  //
  // Alternatively, for CUDA-enabled GPU devices, this typically means that the
  // operation that produced the contents of the tensor has completed, i.e., the
  // CUDA stream has been synchronized (e.g., via cuCtxSynchronize() or
  // cuStreamSynchronize()).
  map<string, string> feed_devices = 6;
  map<string, string> fetch_devices = 7;

  // By default, RunCallable() will synchronize the GPU stream before returning
  // fetched tensors on a GPU device, to ensure that the values in those tensors
  // have been produced. This simplifies interacting with the tensors, but
  // potentially incurs a performance hit.
  //
  // If this options is set to true, the caller is responsible for ensuring
  // that the values in the fetched tensors have been produced before they are
  // used. The caller can do this by invoking `Device::Sync()` on the underlying
  // device(s), or by feeding the tensors back to the same Session using
  // `feed_devices` with the same corresponding device name.
  bool fetch_skip_sync = 8;

  // Next: 9
}