* omitted} or {@linkplain #withPadChar(char) replaced}.
   *
   * <p>No line feeds are added by default, as per <a
   * href="http://tools.ietf.org/html/rfc4648#section-3.1">RFC 4648 section 3.1</a>, Line Feeds in
   * Encoded Data. Line feeds may be added using {@link #withSeparator(String, int)}.
   */
  public static BaseEncoding base64() {
    return BASE64;
  }

   * omitted} or {@linkplain #withPadChar(char) replaced}.
   *
   * <p>No line feeds are added by default, as per <a
   * href="http://tools.ietf.org/html/rfc4648#section-3.1">RFC 4648 section 3.1</a>, Line Feeds in
   * Encoded Data. Line feeds may be added using {@link #withSeparator(String, int)}.
   */
  public static BaseEncoding base64() {
    return BASE64;
  }

http://code.google.com/p/rietveld/issues/detail?id=155

The server side of the protocol is trivial:
  1. add a <link rel="hub" href="hub-server"> tag to all
     feeds that will be pubsubhubbubbed.
  2. every time one of those feeds changes, tell the hub
     with a simple POST request.

I have tested this by adding debug prints to a local hub
server and checking that the server got the right publish
requests.

  // inference.
  pm.addPass(mlir::TF::CreateGuaranteeAllFuncsOneUsePass());
  pm.addPass(mlir::TF::CreateTFShapeInferencePass());
  // For V1 compatibility, we process a module where the graph does not have
  // feeds and fetched. We extract first the TPU computation in a submodule,
  // where it'll be in a function with args and returned values, much more
  // like a TF v2 module. We can then run the usual pipeline on this nested

}

// Returns a TPUPartitionedInput op connected to a `tf_device.cluster_func`
// operand value if it has an XLA sharding. If value is a resource type then
// TPUPartitionedInput op will be connected to a ReadVariable op that feeds into
// a `tf_device.cluster_func`.
mlir::Operation* GetXlaShardingFromOperand(Value value) {
  Value value_to_visit = value;
  if (auto read_var = value_to_visit.getDefiningOp<mlir::TF::ReadVariableOp>())

      related config.
    representative_dataset: a generator that returns a dictionary in {input_key:
      input_value} format or a tuple with signature key and a dictionary in
      {input_key: input_value} format that feeds calibration data for quantizing
      model. This should be provided when the model is not a QAT model.

  Returns:
    A SavedModel object with TF quantization applied.

  Raises:

     * Adapt a {@link MavenExecutionRequest} to a {@link Settings} object for use in the Maven core.
     * We want to make sure that what is ask for in the execution request overrides what is in the settings.
     * The CLI feeds into an execution request so if a particular value is present in the execution request
     * then we will take that over the value coming from the user settings.
     */

// constant arguments? Even constant arguments get an _Arg node in the graph
// instantiated for Function compilation. The tf2xla kernel for constant _Arg
// nodes takes the constant value, converts it to XlaLiteral, and feeds it
// to xla::ComputationBuilder.ConstantLiteral, which returns the handle. This
// constant XlaLiteral is included in the HLO graph, and subsequently, in
// the actual executable, which is copied to the device before being

// leaving behind non-GC'ed state.
//
// Detailed motivation behind this approach, from ashankar@:
//
// Each call to Session::Run() that identifies a new subgraph (based on feeds
// and fetches) creates some datastructures that live as long as the session
// (the partitioned graph, associated executors etc.).
//
// A pathological case of this would be if say the initialization op

be wrapped in the `tf_executor.island` operation.

The `tf_executor.graph` operation does not accept any operands, inputs are
implicitly captured by the region, representing the feeds to the graph.

The region attached to `tf_executor.graph` is terminated by a
`tf_executor.fetch` operation. The non-control operands of the terminator
correspond to the result values (or fetches) of the `tf_executor.graph`