}

// CHECK-LABEL: @float_conv_3
// CHECK:  "tf.PartitionedCall"
// CHECK-SAME: f = @composite_conv2d_with_bias_and_relu6_fn_6
// CHECK:  "tf.PartitionedCall"
// CHECK-SAME: f = @composite_conv2d_with_bias_and_relu6_fn_5

// CHECK-LABEL: @float_conv_1
// CHECK:  "tf.PartitionedCall"
// CHECK-SAME: f = @composite_conv2d_with_bias_and_relu6_fn_2
// CHECK:  "tf.PartitionedCall"

module {
  func.func @wrap_composite_func(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
    %0 = "tf.PartitionedCall"(%arg0, %arg1) <{f = @composite_conv2d_with_relu6_fn}> {_tfl_quant_trait = "fully_quantizable"}
          : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

  // CHECK: return %[[SUBGRAPH_1]] : tensor<1024x3xf32>
  // CHECK: }
}

// -----

// main function contains PartitionedCall and StatefulPartitionedCall ops which
// is used to preserve aliased functions. This test make sure stablehlo ops in
// each PartitionedCall functions are lifted.

module attributes {tf.versions = {bad_consumers = [], min_consumer = 12 : i32, producer = 1629 : i32}, tf_saved_model.semantics} {

func.func @serving_default(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x2x2x2xf32>) attributes {tf.entry_function = {control_outputs = "", inputs = "serving_default_input_tensor:0", outputs = "PartitionedCall:0"}} {

// CHECK-NEXT: [[q_bias:%.+]] = "quantfork.qcast"([[bias]]) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>

    up creating nested XLA launch ops.

    For example, the `tf.PartitionedCall` operation in the following code

    ```mlir
    func.func @convert_partitioned_call_with_resources(%arg0: tensor<!tf_type.resource>, %arg1: tensor<i32>) -> tensor<i32> {

// -----

// This test makes sure functions without tf._noinline=true is inlined.

module {
  func.func @partitioned_call(%arg0: tensor<1x2x2x3xf32>) -> (tensor<1x2x2x3xf32>) {
    %0 = "tf.PartitionedCall"(%arg0) <{
      config = "", config_proto = "", executor_type = "", f = @some_func
    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>

  // Collect all the ops that needs to have token input names attributes. These
  // ops are communication ops and all their parent ops via nesting or function
  // calls. For example, IfRegion op and PartitionedCall op.
  std::vector<Operation*> worklist;
  absl::flat_hash_set<Operation*> ops_with_tokens;
  module.walk([&](Operation* op) {
    if (IsCommunicationOp(op)) {
      ops_with_tokens.insert(op);

  // Guarantee all functions have one use, which enables more exact shape
  // inference.
  pm.addPass(mlir::TF::CreateGuaranteeAllFuncsOneUsePass());
  pm.addPass(mlir::TF::CreateTFShapeInferencePass());
  // Encapsulate PartitionedCall ops within a cluster so that the composite
  // resource ops can be decomposed.
  pm.addPass(tensorflow::tf2xla::internal::CreateXlaClusterFormationPass());

  for (int i = 0; i < num_results; ++i) {
    main_func.setResultAttr(
        i, kTfSavedModelIndexPathAttr,
        ArrayAttr::get(context, {StringAttr::get(context, output_names[i])}));
  }

  // Creates PartitionedCall ops to call exported functions.
  auto guard = OpBuilder::InsertionGuard(builder);
  int arg_idx = 0;
  int result_idx = 0;
  llvm::SmallVector<Value> call_op_returns;