auto_shard.cc

/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "tensorflow/core/grappler/optimizers/data/auto_shard.h"

#include "absl/container/flat_hash_map.h"
#include "absl/container/flat_hash_set.h"
#include "absl/strings/match.h"
#include "tensorflow/core/data/dataset_utils.h"
#include "tensorflow/core/framework/attr_value.pb.h"
#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/framework/function.pb.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/grappler/clusters/cluster.h"
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/mutable_graph_view.h"
#include "tensorflow/core/grappler/op_types.h"
#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h"
#include "tensorflow/core/grappler/optimizers/data/graph_utils.h"
#include "tensorflow/core/grappler/utils/functions.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/platform/errors.h"

namespace tensorflow {
namespace grappler {
namespace {

using tensorflow::data::AutoShardPolicy;

constexpr char kAssertCardinalityDatasetOpName[] = "AssertCardinalityDataset";
constexpr char kShardDatasetOpName[] = "ShardDataset";
constexpr char kShuffleDatasetOpName[] = "ShuffleDataset";
constexpr char kShuffleDatasetV2OpName[] = "ShuffleDatasetV2";
constexpr char kShuffleDatasetV3OpName[] = "ShuffleDatasetV3";
constexpr char kPrefetchDatasetOpName[] = "PrefetchDataset";
constexpr char kFinalizeDatasetOpName[] = "FinalizeDataset";
constexpr char kOptionsDatasetOpName[] = "OptionsDataset";
constexpr char kRebatchDatasetOpName[] = "RebatchDataset";
constexpr char kRebatchDatasetV2OpName[] = "RebatchDatasetV2";
constexpr char kTensorDatasetOpName[] = "TensorDataset";
constexpr char kTensorSliceDatasetOpName[] = "TensorSliceDataset";
constexpr char kPlaceholderOpName[] = "Placeholder";
constexpr char kConstOpName[] = "Const";

constexpr char kNumWorkersAttrName[] = "num_workers";
constexpr char kNumReplicasAttrName[] = "num_replicas";
constexpr char kIndexAttrName[] = "index";
constexpr char kAutoShardPolicyAttrName[] = "auto_shard_policy";
constexpr char kReshuffleEachIteration[] = "reshuffle_each_iteration";
constexpr char kOutputShapes[] = "output_shapes";
constexpr char kOutputTypes[] = "output_types";

// clang-format off
constexpr std::array<const char*, 5> kReaderDatasetOps = {
    "FixedLengthRecordDataset",
    "RecordIODataset",
    "SSTableDataset",
    "TextLineDataset",
    "TFRecordDataset"
};

constexpr std::array<const char*, 2> kMultipleInputsDatasetOps = {
    "ConcatenateDataset",
    "ZipDataset"
};

constexpr std::array<const char*, 30> kPassThroughOps = {
    "_Retval",
    "AssertNextDataset",
    "BatchDataset",
    "CacheDataset",
    "ExperimentalMapAndBatchDataset",
    "ExperimentalParseExampleDataset",
    "ExperimentalRebatchDataset",
    "FilterDataset",
    "FinalizeDataset",
    "Identity",
    "MapAndBatchDataset",
    "MapDataset",
    "MaxIntraOpParallelismDataset",
    "ModelDataset",
    "OptimizeDataset",
    "OptionsDataset",
    "PaddedBatchDataset",
    "ParallelMapDataset",
    "ParseExampleDataset",
    "PrefetchDataset",
    "PrivateThreadPoolDataset",
    "ReduceDataset",
    "RebatchDataset",
    "RepeatDataset",
    "ShardDataset",
    "ShuffleAndRepeatDataset",
    "ShuffleDataset",
    "SkipDataset",
    "TakeDataset",
    "WindowDataset",
};

// TODO(frankchn): Process functions within kFuncDatasetOps as well.
constexpr std::array<const char*, 5> kFuncDatasetOps = {
    "ExperimentalParallelInterleaveDataset",
    "FlatMapDataset",
    "InterleaveDataset",
    "LegacyParallelInterleaveDataset",
    "ParallelInterleaveDataset",
};

constexpr std::array<const char*, 5> kUnshardableSourceDatasetOps = {
    "GeneratorDataset",
    "RangeDataset",
    "SparseTensorsSliceDataset",
    "TensorDataset",
    "TensorSliceDataset",
};
// clang-format on

Status OptimizeGraph(const GrapplerItem& item, int64 num_workers, int64 index,
                     AutoShardPolicy policy, int64 num_replicas,
                     GraphDef* output);

template <std::size_t SIZE>
bool IsDatasetNodeOfType(const NodeDef& node,
                         const std::array<const char*, SIZE>& arr) {
  for (const auto& dataset_op_name : arr) {
    if (tensorflow::data::MatchesAnyVersion(/*op_prefix=*/dataset_op_name,
                                            /*op_to_match=*/node.op())) {
      return true;
    }
  }
  return false;
}

// Adds a ShardDataset node before `add_before`.
Status AddShardNode(MutableGraphView* graph, const NodeDef& add_before,
                    int64 num_workers, int64 index) {
  NodeDef new_node;
  new_node.set_op(kShardDatasetOpName);
  graph_utils::SetUniqueGraphNodeName(kShardDatasetOpName, graph->graph(),
                                      &new_node);

  // Construct argument nodes
  NodeDef* num_shards_node =
      graph_utils::AddScalarConstNode<int64>(num_workers, graph);
  NodeDef* index_node = graph_utils::AddScalarConstNode<int64>(index, graph);

  // Add inputs to new node
  new_node.add_input(add_before.input(0));
  new_node.add_input(num_shards_node->name());
  new_node.add_input(index_node->name());

  // Ensure that each shard will have at least one element.
  (*(new_node.mutable_attr()))["require_non_empty"].set_b(true);

  // Add shapes and other attributes
  NodeDef* add_after = graph->GetNode(add_before.input(0));

  if (absl::StrContains(add_after->op(), "Dataset")) {
    // We still may or may not have the right attributes because Datasets like
    // TFRecordDataset doesn't have a output type or shape, and by default we
    // set them to DT_STRING and an unknown shape.
    if (add_after->attr().count(kOutputShapes) > 0) {
      graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
    } else {
      tensorflow::TensorShapeProto* shape =
          (*(new_node.mutable_attr()))[kOutputShapes]
              .mutable_list()
              ->add_shape();
      shape->set_unknown_rank(true);
    }

    if (add_after->attr().count(kOutputTypes) > 0) {
      graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);
    } else if (add_after->attr().count("Toutput_types") > 0) {
      (*(new_node.mutable_attr()))[kOutputTypes] =
          add_after->attr().at("Toutput_types");
    } else {
      (*(new_node.mutable_attr()))[kOutputTypes].mutable_list()->add_type(
          tensorflow::DataType::DT_STRING);
    }
  } else {
    // TODO(frankchn): Make this work for datasets where input(0) is a Const,
    // and we need to shard the Const.
    // This is probably not a dataset, so we bail because we can't infer the
    // output types and shape.
    return errors::NotFound(
        "Unable to shard this input. You may need to wrap the inputs to your "
        "reader dataset in a TensorSliceDataset. Input node is ",
        add_after->DebugString());
  }

  // Add new node into graph and update edges
  NodeDef* new_node_graph = graph->AddNode(std::move(new_node));
  TF_RETURN_IF_ERROR(
      graph->UpdateFanouts(add_after->name(), new_node_graph->name()));

  return Status::OK();
}

Status AddShuffleDataset(MutableGraphView* graph, const NodeDef& add_before,
                         const string& buffer_size_node,
                         const string& seed_node, const string& seed2_node,
                         bool reshuffle_each_iteration) {
  NodeDef* add_after = graph->GetNode(add_before.input(0));
  NodeDef new_node;
  new_node.set_op(kShuffleDatasetOpName);
  graph_utils::SetUniqueGraphNodeName(kShuffleDatasetOpName, graph->graph(),
                                      &new_node);

  new_node.add_input(add_before.input(0));
  new_node.add_input(buffer_size_node);
  new_node.add_input(seed_node);
  new_node.add_input(seed2_node);

  graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
  graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);

  AttrValue reshuffle_attr;
  reshuffle_attr.set_b(reshuffle_each_iteration);
  (*new_node.mutable_attr())[kReshuffleEachIteration] = reshuffle_attr;

  NodeDef* new_node_graph = graph->AddNode(std::move(new_node));

  TF_RETURN_IF_ERROR(
      graph->UpdateFanouts(add_after->name(), new_node_graph->name()));
  return Status::OK();
}

Status AddShuffleDatasetV2(MutableGraphView* graph, const NodeDef& add_before,
                           const string& buffer_size_node,
                           const string& seed_generator_node) {
  NodeDef* add_after = graph->GetNode(add_before.input(0));
  NodeDef new_node;
  new_node.set_op(kShuffleDatasetV2OpName);
  graph_utils::SetUniqueGraphNodeName(kShuffleDatasetV2OpName, graph->graph(),
                                      &new_node);

  new_node.add_input(add_before.input(0));
  new_node.add_input(buffer_size_node);
  new_node.add_input(seed_generator_node);

  graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
  graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);

  NodeDef* new_node_graph = graph->AddNode(std::move(new_node));

  TF_RETURN_IF_ERROR(
      graph->UpdateFanouts(add_after->name(), new_node_graph->name()));
  return Status::OK();
}

Status AddShuffleDatasetV3(MutableGraphView* graph, const NodeDef& add_before,
                           const string& buffer_size_node,
                           const string& seed_node, const string& seed2_node,
                           const string& seed_generator_node,
                           bool reshuffle_each_iteration) {
  NodeDef* add_after = graph->GetNode(add_before.input(0));
  NodeDef new_node;
  new_node.set_op(kShuffleDatasetV3OpName);
  graph_utils::SetUniqueGraphNodeName(kShuffleDatasetV3OpName, graph->graph(),
                                      &new_node);

  new_node.add_input(add_before.input(0));
  new_node.add_input(buffer_size_node);
  new_node.add_input(seed_node);
  new_node.add_input(seed2_node);
  new_node.add_input(seed_generator_node);

  graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
  graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);

  AttrValue reshuffle_attr;
  reshuffle_attr.set_b(reshuffle_each_iteration);
  (*new_node.mutable_attr())[kReshuffleEachIteration] = reshuffle_attr;

  NodeDef* new_node_graph = graph->AddNode(std::move(new_node));

  TF_RETURN_IF_ERROR(
      graph->UpdateFanouts(add_after->name(), new_node_graph->name()));
  return Status::OK();
}

bool ReaderOpInFunction(const NodeDef& node,
                        const FunctionLibraryDefinition& flib) {
  const FunctionDef* func = flib.Find(node.attr().at("f").func().name());
  for (int i = 0; i < func->node_def_size(); i++) {
    NodeDef node_in_func = func->node_def(i);
    if (IsDatasetNodeOfType(node_in_func, kReaderDatasetOps) &&
        node_in_func.input_size() > 0 &&
        absl::StartsWith(node_in_func.input(0), "args_0")) {
      return true;
    }
    if (IsDatasetNodeOfType(func->node_def(i), kFuncDatasetOps) &&
        ReaderOpInFunction(func->node_def(i), flib)) {
      return true;
    }
  }
  return false;
}

Status RemoveShuffleDataset(MutableGraphView* graph, const NodeDef& node,
                            absl::flat_hash_set<string>* nodes_to_delete,
                            string* op_name, string* buffer_size_node,
                            string* seed_node, string* seed2_node,
                            bool* reshuffle_each_iteration) {
  if (node.op() == kShuffleDatasetOpName) {
    *op_name = node.op();
    *buffer_size_node = node.input(1);
    *seed_node = node.input(2);
    *seed2_node = node.input(3);
    *reshuffle_each_iteration = node.attr().at(kReshuffleEachIteration).b();
    TF_RETURN_IF_ERROR(graph->UpdateFanouts(node.name(), node.input(0)));
    nodes_to_delete->insert(node.name());
  }

  for (const auto& fanin : graph->GetFanins(node, true)) {
    TF_RETURN_IF_ERROR(RemoveShuffleDataset(
        graph, *fanin.node, nodes_to_delete, op_name, buffer_size_node,
        seed_node, seed2_node, reshuffle_each_iteration));
  }

  // TODO(frankchn): Traverse functions too.
  return Status::OK();
}

Status RemoveShuffleDatasetV2(MutableGraphView* graph, const NodeDef& node,
                              absl::flat_hash_set<string>* nodes_to_delete,
                              string* op_name, string* buffer_size_node,
                              string* seed_generator_node) {
  if (node.op() == kShuffleDatasetV2OpName) {
    *op_name = node.op();
    *buffer_size_node = node.input(1);
    *seed_generator_node = node.input(2);
    TF_RETURN_IF_ERROR(graph->UpdateFanouts(node.name(), node.input(0)));
    nodes_to_delete->insert(node.name());
  }

  for (const auto& fanin : graph->GetFanins(node, true)) {
    TF_RETURN_IF_ERROR(
        RemoveShuffleDatasetV2(graph, *fanin.node, nodes_to_delete, op_name,
                               buffer_size_node, seed_generator_node));
  }

  // TODO(frankchn): Traverse functions too.
  return Status::OK();
}

Status RemoveShuffleDatasetV3(MutableGraphView* graph, const NodeDef& node,
                              absl::flat_hash_set<string>* nodes_to_delete,
                              string* op_name, string* buffer_size_node,
                              string* seed_node, string* seed2_node,
                              string* seed_generator_node,
                              bool* reshuffle_each_iteration) {
  if (node.op() == kShuffleDatasetV3OpName) {
    *op_name = node.op();
    *buffer_size_node = node.input(1);
    *seed_node = node.input(2);
    *seed2_node = node.input(3);
    *seed_generator_node = node.input(4);
    *reshuffle_each_iteration = node.attr().at(kReshuffleEachIteration).b();
    TF_RETURN_IF_ERROR(graph->UpdateFanouts(node.name(), node.input(0)));
    nodes_to_delete->insert(node.name());
  }

  for (const auto& fanin : graph->GetFanins(node, true)) {
    TF_RETURN_IF_ERROR(RemoveShuffleDatasetV3(
        graph, *fanin.node, nodes_to_delete, op_name, buffer_size_node,
        seed_node, seed2_node, seed_generator_node, reshuffle_each_iteration));
  }

  // TODO(frankchn): Traverse functions too.
  return Status::OK();
}

Status ProcessDatasetSourceNode(MutableGraphView* graph, const NodeDef& node,
                                absl::flat_hash_set<string>* nodes_to_delete,
                                int64 num_workers, int64 index) {
  string shuffle_op_name = "";
  string buffer_size_node = "";
  string seed_node = "";
  string seed2_node = "";
  string seed_generator_node = "";
  bool reshuffle_each_iteration;

  TF_RETURN_IF_ERROR(AddShardNode(graph, node, num_workers, index));
  TF_RETURN_IF_ERROR(RemoveShuffleDataset(
      graph, node, nodes_to_delete, &shuffle_op_name, &buffer_size_node,
      &seed_node, &seed2_node, &reshuffle_each_iteration));
  if (shuffle_op_name.empty()) {
    TF_RETURN_IF_ERROR(
        RemoveShuffleDatasetV2(graph, node, nodes_to_delete, &shuffle_op_name,
                               &buffer_size_node, &seed_generator_node));
  }
  if (shuffle_op_name.empty()) {
    TF_RETURN_IF_ERROR(RemoveShuffleDatasetV3(
        graph, node, nodes_to_delete, &shuffle_op_name, &buffer_size_node,
        &seed_node, &seed2_node, &seed_generator_node,
        &reshuffle_each_iteration));
  }

  if (shuffle_op_name == kShuffleDatasetOpName) {
    TF_RETURN_IF_ERROR(AddShuffleDataset(graph, node, buffer_size_node,
                                         seed_node, seed2_node,
                                         reshuffle_each_iteration));
  } else if (shuffle_op_name == kShuffleDatasetV2OpName) {
    TF_RETURN_IF_ERROR(AddShuffleDatasetV2(graph, node, buffer_size_node,
                                           seed_generator_node));
  } else if (shuffle_op_name == kShuffleDatasetV3OpName) {
    TF_RETURN_IF_ERROR(AddShuffleDatasetV3(
        graph, node, buffer_size_node, seed_node, seed2_node,
        seed_generator_node, reshuffle_each_iteration));
  }

  return Status::OK();
}

const NodeDef* FindFuncAndTensorSliceDataset(
    const NodeDef* node, int64 num_workers, int64 index,
    FunctionLibraryDefinition* flib, MutableGraphView* graph,
    absl::flat_hash_set<string>* nodes_to_delete) {
  if (IsDatasetNodeOfType(*node, kFuncDatasetOps)) {
    const NodeDef* input_node = graph_utils::GetInputNode(*node, *graph, 0);
    if (input_node->op() == kTensorSliceDatasetOpName ||
        input_node->op() == kTensorDatasetOpName) {
      const NodeDef* next_input_node =
          graph_utils::GetInputNode(*input_node, *graph, 0);
      if (next_input_node->op() == kPlaceholderOpName) {
        return node;
      }
    }
  }

  if (!IsDatasetNodeOfType(*node, kPassThroughOps)) {
    return nullptr;
  }

  // Sometimes there are other nodes between the last InterleaveDataset and the
  // second to last FlatMapDataset, so we need to skip over those.
  const NodeDef* input_node = graph_utils::GetInputNode(*node, *graph, 0);
  return FindFuncAndTensorSliceDataset(input_node, num_workers, index, flib,
                                       graph, nodes_to_delete);
}

Status RecursivelyHandleOp(const NodeDef& node, int64 num_workers, int64 index,
                           FunctionLibraryDefinition* flib,
                           MutableGraphView* graph,
                           absl::flat_hash_set<string>* nodes_to_delete) {
  if (node.op() == kAssertCardinalityDatasetOpName) {
    LOG(WARNING) << "The `assert_cardinality` transformation is currently not "
                    "handled by the auto-shard rewrite and will be removed.";
    nodes_to_delete->insert(node.name());
    TF_RETURN_IF_ERROR(graph->UpdateFanouts(node.name(), node.input(0)));
    const NodeDef* input_node = graph_utils::GetInputNode(node, *graph, 0);
    return RecursivelyHandleOp(*input_node, num_workers, index, flib, graph,
                               nodes_to_delete);
  }

  if (IsDatasetNodeOfType(node, kUnshardableSourceDatasetOps)) {
    return errors::NotFound("Found an unshardable source dataset: ",
                            node.DebugString());
  }

  if (IsDatasetNodeOfType(node, kMultipleInputsDatasetOps)) {
    for (int i = 0; i < node.input_size(); ++i) {
      const NodeDef* input_node = graph_utils::GetInputNode(node, *graph, i);
      TF_RETURN_IF_ERROR(RecursivelyHandleOp(*input_node, num_workers, index,
                                             flib, graph, nodes_to_delete));
    }
    return Status::OK();
  }

  // This handles the case for the following subgraph:
  //   Placeholder -> TensorSliceDataset -> FlatMapDataset -x->
  //   (other preprocessing datasets) -> InterleaveDataset
  // and then inserting the shard node immediately after the FlatMapDataset.
  //
  // This is used for some training pipelines where a dataset is created with
  // the following code:
  //
  // def make_dataset_pipeline():
  //   file_globs = [...]
  //   datasets = []
  //   for file_glob in file_globs:
  //     datasets.append(Dataset.list_files(file_glob).map(TFRecordReader))
  //   dataset = Dataset.from_tensor_slices(datasets)
  //   dataset = dataset.flat_map(lambda x: x)
  //   dataset = ...  # additional preprocessing
  //   dataset = dataset.interleave(lambda x: x, cycle_length=...)
  //   return dataset
  if (IsDatasetNodeOfType(node, kFuncDatasetOps)) {
    const NodeDef* input_node = graph_utils::GetInputNode(node, *graph, 0);
    const NodeDef* flat_map_node = FindFuncAndTensorSliceDataset(
        input_node, num_workers, index, flib, graph, nodes_to_delete);

    if (flat_map_node != nullptr) {
      auto fanouts = graph->GetFanouts(*flat_map_node, false);
      // FlatMapDataset should only be the input to one other dataset.
      if (fanouts.size() == 1) {
        return ProcessDatasetSourceNode(graph, *fanouts.begin()->node,
                                        nodes_to_delete, num_workers, index);
      }
    }
  }

  // This handles the case where a reader Dataset is contained within a
  // FuncDataset (e.g. FlatMap, ParallelInterleave, etc...). For example:
  //
  // dataset = Dataset.list_files("/path/to/data")
  // dataset = dataset.flat_map(core_readers.TFRecordDataset)
  //
  // where the list of files is passed in one-by-one as an argument to the
  // function in flat_map.
  if (IsDatasetNodeOfType(node, kFuncDatasetOps) &&
      ReaderOpInFunction(node, *flib)) {
    return ProcessDatasetSourceNode(graph, node, nodes_to_delete, num_workers,
                                    index);
  }

  if (IsDatasetNodeOfType(node, kReaderDatasetOps)) {
    // We reached a reader dataset directly and we try to shard input 0.
    return ProcessDatasetSourceNode(graph, node, nodes_to_delete, num_workers,
                                    index);
  }

  if (!IsDatasetNodeOfType(node, kPassThroughOps)) {
    return errors::NotFound(
        "Did not find a shardable source, walked to ",
        "a node which is not a dataset: ", node.DebugString(),
        ". Consider either turning off auto-sharding or switching the "
        "auto_shard_policy to DATA to shard this dataset. You can do this by "
        "creating a new `tf.data.Options()` object then setting "
        "`options.experimental_distribute.auto_shard_policy = "
        "AutoShardPolicy.DATA` before applying the options object to the "
        "dataset via `dataset.with_options(options)`.");
  }

  const NodeDef* input_node = graph_utils::GetInputNode(node, *graph, 0);
  return RecursivelyHandleOp(*input_node, num_workers, index, flib, graph,
                             nodes_to_delete);
}

// Recursively walk the dataset graph from sink to source, searching for
// the first (i.e. closest to the sink) occurence of a ReaderDataset, such as
// CSVDataset, TFRecordDataset, etc. We then insert a ShardDataset op before
// that nodes input, so that each worker only reads a subset of files.
// Additionally, we remove sources of randomness (e.g. ShuffleDataset) that
// occur upstream of the ShardDataset transformation to ensure that sharding
// returns a sensible result.
Status ShardByFile(const NodeDef& sink_node, int64 num_workers, int64 index,
                   FunctionLibraryDefinition* flib, MutableGraphView* graph) {
  absl::flat_hash_set<string> nodes_to_delete;
  TF_RETURN_IF_ERROR(RecursivelyHandleOp(sink_node, num_workers, index, flib,
                                         graph, &nodes_to_delete));
  return graph->DeleteNodes(nodes_to_delete);
}

Status RewriteRebatchV2ToV1(const NodeDef& sink_node, int64 num_replicas,
                            MutableGraphView* graph) {
  // The final node before AutoShardDataset is RebatchDataset.
  // This is always the case as RebatchDataset and AutoShardDataset are internal
  // APIs used directly by tf.distribute's input_lib. As such, instead of
  // walking the entire dataset graph, we can walk up directly from the
  // sink_node to get the RebatchDataset.
  NodeDef* input_node = graph_utils::GetInputNode(sink_node, *graph);
  if (input_node->op() != kRebatchDatasetV2OpName) {
    return Status::OK();
  }

  NodeDef* rebatch_node = input_node;
  // Update RebatchDatasetV2 in place. Since Rebatch is an internal API, no
  // other nodes should have it as an input.
  rebatch_node->set_op(kRebatchDatasetOpName);
  // Delete the `batch_sizes` and `drop_remainder` input.
  rebatch_node->mutable_input()->DeleteSubrange(/*start=*/1, /*num=*/2);
  // Add the `num_replicas` input.
  if (num_replicas < 1) {
    return errors::InvalidArgument(
        "Cannot rewrite RebatchDatasetV2 to legacy RebatchDataset with invalid "
        "num_replicas argument. `num_replicas` is ",
        num_replicas, ", but expected to be >= 1.");
  }
  auto num_replicas_node = graph_utils::AddScalarConstNode(num_replicas, graph);
  rebatch_node->add_input(num_replicas_node->name());

  // Set `use_fallback` attr. This attr is not used anywhere, so its value
  // does not matter
  (*rebatch_node->mutable_attr())["use_fallback"].set_b(true);

  // Update the output_shapes attr to set all its batch dimensions to -1
  // (unknown).
  auto* shapes_attr =
      gtl::FindOrNull(*rebatch_node->mutable_attr(), "output_shapes");
  if (shapes_attr == nullptr) {
    return errors::InvalidArgument(
        "Cannot rewrite RebatchDatasetV2 with missing `output_shapes` attr.");
  }
  for (int i = 0; i < shapes_attr->list().shape_size(); ++i) {
    auto* shape = shapes_attr->mutable_list()->mutable_shape(i);
    if (shape->unknown_rank()) continue;
    shape->mutable_dim(0)->set_size(-1);
  }

  return Status::OK();
}

Status ShardByData(const NodeDef& sink_node, int64 num_workers, int64 index,
                   int64 num_replicas, MutableGraphView* graph) {
  const NodeDef* shard_before = &sink_node;
  // We sometimes insert a PrefetchDataset, OptionsDataset, and FinalizeDataset
  // at the end of the input pipeline before autosharding. When sharding by
  // data, we should insert the shard before the these datasets so that the
  // right number of elements is prefetched.
  NodeDef* input_node = graph_utils::GetInputNode(sink_node, *graph);
  while (input_node->op() == kPrefetchDatasetOpName ||
         input_node->op() == kOptionsDatasetOpName ||
         input_node->op() == kFinalizeDatasetOpName) {
    shard_before = input_node;
    input_node = graph_utils::GetInputNode(*input_node, *graph);
  }
  // Sharding by data only works with legacy RebatchDataset. As such, we rewrite
  // all instances of RebatchDatasetV2 to RebatchDataset.
  TF_RETURN_IF_ERROR(RewriteRebatchV2ToV1(*shard_before, num_replicas, graph));
  return AddShardNode(graph, *shard_before, num_workers, index);
}

// Searches the dataset graph replacing any occurence of `shard(1, 0)` with
// `shard(num_workers, index)`.
Status ShardByHint(const NodeDef& sink_node, int64 num_workers, int64 index,
                   int64 num_replicas, MutableGraphView* graph) {
  auto get_shard_node = [graph](const NodeDef& node) -> const NodeDef* {
    if (node.op() != kShardDatasetOpName) return nullptr;
    auto num_workers_node = graph->GetNode(node.input(1));
    if (num_workers_node->op() != kConstOpName) return nullptr;
    if (num_workers_node->attr().at("value").tensor().int64_val(0) !=
        tensorflow::data::kShardHint)
      return nullptr;
    return &node;
  };

  auto* num_workers_node =
      graph_utils::AddScalarConstNode(static_cast<int64>(num_workers), graph);
  auto* worker_index_node =
      graph_utils::AddScalarConstNode(static_cast<int64>(index), graph);

  for (const NodeDef& node : graph->graph()->node()) {
    const NodeDef* shard_node = get_shard_node(node);
    if (!shard_node) continue;
    auto mutable_node = graph->GetNode(shard_node->name());
    *mutable_node->mutable_input(1) = num_workers_node->name();
    *mutable_node->mutable_input(2) = worker_index_node->name();
  }
  return Status::OK();
}

Status OptimizeGraph(const GrapplerItem& item, int64 num_workers, int64 index,
                     AutoShardPolicy policy, int64 num_replicas,
                     GraphDef* output) {
  if (policy == AutoShardPolicy::OFF ||
      (policy == AutoShardPolicy::FILE && num_workers == 1 && index == 0)) {
    return Status::OK();
  }

  *output = item.graph;
  MutableGraphView graph(output);
  FunctionLibraryDefinition flib(OpRegistry::Global(), item.graph.library());

  NodeDef* sink_node;
  TF_RETURN_IF_ERROR(graph_utils::GetFetchNode(graph, item, &sink_node));

  switch (policy) {
    case AutoShardPolicy::OFF:
      return Status::OK();
    case AutoShardPolicy::FILE:
      return ShardByFile(*sink_node, num_workers, index, &flib, &graph);
    case AutoShardPolicy::DATA:
      return ShardByData(*sink_node, num_workers, index, num_replicas, &graph);
    case AutoShardPolicy::HINT:
      return ShardByHint(*sink_node, num_workers, index, num_replicas, &graph);
    case AutoShardPolicy::AUTO:
    default:
      Status s = ShardByFile(*sink_node, num_workers, index, &flib, &graph);
      if (errors::IsNotFound(s)) {
        LOG(WARNING) << "AUTO sharding policy will apply DATA sharding policy "
                        "as it failed to apply FILE sharding policy because of "
                        "the following reason: "
                     << s.error_message();
        return ShardByData(*sink_node, num_workers, index, num_replicas,
                           &graph);
      }
      return s;
  }
}

}  // anonymous namespace

Status AutoShard::Init(
    const tensorflow::RewriterConfig_CustomGraphOptimizer* config) {
  if (!config) return errors::InvalidArgument("RewriterConfig not found.");

  if ((config->parameter_map().find(kNumWorkersAttrName) ==
       config->parameter_map().end())) {
    return errors::InvalidArgument(kNumWorkersAttrName, " parameter missing.");
  }

  if ((config->parameter_map().find(kIndexAttrName) ==
       config->parameter_map().end())) {
    return errors::InvalidArgument(kIndexAttrName, " parameter missing.");
  }

  num_workers_ = config->parameter_map().at(kNumWorkersAttrName).i();
  index_ = config->parameter_map().at(kIndexAttrName).i();
  auto_shard_policy_ =
      AutoShardPolicy(config->parameter_map().at(kAutoShardPolicyAttrName).i());
  num_replicas_ = config->parameter_map().at(kNumReplicasAttrName).i();

  if (auto_shard_policy_ != AutoShardPolicy::OFF &&
      auto_shard_policy_ != AutoShardPolicy::AUTO &&
      auto_shard_policy_ != AutoShardPolicy::DATA &&
      auto_shard_policy_ != AutoShardPolicy::FILE &&
      auto_shard_policy_ != AutoShardPolicy::HINT) {
    return errors::InvalidArgument(kAutoShardPolicyAttrName, " is invalid.");
  }

  if (num_workers_ < 1) {
    return errors::InvalidArgument(kNumWorkersAttrName,
                                   " should be >= 1, currently ", num_workers_);
  }

  if (index_ < 0 || index_ >= num_workers_) {
    return errors::InvalidArgument(kIndexAttrName, " should be >= 0 and < ",
                                   num_workers_, ", currently ", index_);
  }

  if (num_replicas_ < 0) {
    return errors::InvalidArgument(kNumReplicasAttrName, " should be >= 0");
  }

  return Status::OK();
}

Status AutoShard::OptimizeAndCollectStats(Cluster* /* cluster */,
                                          const GrapplerItem& item,
                                          GraphDef* output,
                                          OptimizationStats* stats) {
  *output = item.graph;
  TF_RETURN_IF_ERROR(OptimizeGraph(item, num_workers_, index_,
                                   auto_shard_policy_, num_replicas_, output));
  stats->num_changes++;
  return Status::OK();
}

REGISTER_GRAPH_OPTIMIZER_AS(AutoShard, "tf_auto_shard");

}  // namespace grappler
}  // namespace tensorflow