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";

import "tensorflow/core/framework/cost_graph.proto";
import "tensorflow/core/framework/graph.proto";
import "tensorflow/core/framework/step_stats.proto";
import "tensorflow/core/protobuf/debug.proto";
import "tensorflow/core/protobuf/cluster.proto";
import "tensorflow/core/protobuf/rewriter_config.proto";

message GPUOptions {
  // A value between 0 and 1 that indicates what fraction of the
  // available GPU memory to pre-allocate for each process.  1 means
  // to pre-allocate all of the GPU memory, 0.5 means the process
  // allocates ~50% of the available GPU memory.
  double per_process_gpu_memory_fraction = 1;

  // 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;

  // 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;

  // 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 "/gpu:0", and "/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: 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.
  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;

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

// 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;

  // 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;
  }

  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;
};

// 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.
  //
  // 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;

  // Next: 15
};

// 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.
  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;

  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;
}