}
    Value broadcast_multiplier = rewriter.create<mhlo::BroadcastInDimOp>(
        mul_op.getLoc(), filter.getType(), multiplier, broadcast_dims);
    Value new_filter = rewriter.create<mhlo::MulOp>(
        mul_op.getLoc(), filter.getType(), filter, broadcast_multiplier);
    Value new_conv = rewriter.create<mhlo::ConvolutionOp>(
        mul_op.getLoc(), conv_op.getType(), conv_op.getLhs(), new_filter,

  OpBuilder builder(func_.getBody());
  auto op = builder.create<mlir::TFL::NonMaxSuppressionV4Op>(
      func_.getLoc(), output_type0, output_type1, boxes, scores,
      max_output_size, iou_threshold, score_threshold);

  builder.create<mlir::func::ReturnOp>(func_.getLoc(), op.getResults());
}

LogicalResult ConvertNMSPaddedFunc::VerifySignature() {
  // Verify high-level function signature.

            RequirementMatcher matcher = provides.get(key);

            if (matcher == null) {
                getLog().debug("Toolchain {} is missing required property: {}", this, key);
                return false;
            }
            if (!matcher.matches(requirement.getValue())) {
                getLog().debug("Toolchain {} doesn't match required property: {}", this, key);
                return false;
            }
        }

            const_op->getLoc(),
            DenseElementsAttr::get(
                tensorflow::GetTypeFromTFTensorShape(
                    {}, splat_elements_attr.getType().getElementType()),
                splat_elements_attr.getSplatValue<Attribute>()));
    TFL::FillOp fill = op_builder->create<TFL::FillOp>(
        const_op->getLoc(), splat_elements_attr.getType(), fill_shape,
        fill_value);

        CreateI8F32UniformQuantizedType(
            uniform_quantize_call_op.getLoc(), *rewriter.getContext(),
            input_scale_value, input_zero_point_value);
    auto input_uniform_quantize_op =
        rewriter.create<stablehlo::UniformQuantizeOp>(
            uniform_quantize_call_op.getLoc(),
            /*result=*/
            mlir::cast<TensorType>(input_value.getType())

    } else {
      initial_value = mlir::Attribute();
    }
    opToName[globalTensor] = name;
    auto variableOp = globalBuilder.create<ml_program::GlobalOp>(
        globalTensor.getLoc(), name, globalTensor.getType(),
        globalTensor.getIsMutable(), initial_value,
        /*visibility=*/globalBuilder.getStringAttr("private"));
    variableOp.setPrivate();
  }

  SymbolTable syms(module);

  builder.create<tf_executor::YieldOp>(terminator.getLoc(),
                                       terminator.getOperands());
  terminator.erase();

  builder.setInsertionPoint(island_op);
  IRMapping mapping;
  for (int i : llvm::seq<int>(0, num_replicas)) {
    // Create new island for replica.
    auto replica = builder.create<tf_executor::IslandOp>(
        island_op.getLoc(), output_types, control_type, replica_inputs);

  tf_executor::YieldOp yield = island.GetYield();
  if (yield.getNumOperands() == 0) {
    builder.create<TF::NoOp>(island.getLoc(), TypeRange{}, ValueRange{});
  } else if (yield.getNumOperands() == 1) {
    Value operand = yield.getOperand(0);
    auto identity = builder.create<TF::IdentityOp>(island.getLoc(),
                                                   operand.getType(), operand);
    yield.setOperand(0, identity.getOutput());

  SendTPUEmbeddingGradientsOp send_op;
  if (failed(GetOp(region, &send_op))) return failure();

  // No TPU embedding ops.
  if (!recv_op && !send_op) return success();

  Location loc = recv_op ? recv_op.getLoc() : send_op.getLoc();
  StringRef config = recv_op ? recv_op.getConfig() : send_op.getConfig();

  // Create XlaRecvTPUEmbeddingDeduplicationData op.
  OpBuilder builder(region);
  auto output_ty =

  void runOnOperation() override;
};

void ReplaceClusterReturnWithReturn(tf_device::ReturnOp cluster_return_op,
                                    OpBuilder* builder) {
  builder->create<func::ReturnOp>(cluster_return_op.getLoc(),
                                  cluster_return_op.getOperands());
  cluster_return_op.erase();
}

// Builds a function that outlines region attached to cluster_op or launch_op,