if (auto call_op = dyn_cast<TF::PartitionedCallOp>(op)) {
    StringRef function_name =
        mlir::cast<FlatSymbolRefAttr>(call_op.getFAttr()).getValue();
    if (!function_name.starts_with("composite_")) {
      return spec;
    }
    if (function_name.contains("depthwise_conv2d")) {
      spec->coeff_op_quant_dim[1] = 3;
      if (function_name.contains("with_bias")) {

    for (auto user : users) {
      if (!llvm::isa<mlir::stablehlo::UniformDequantizeOp>(user)) {
        return failure();
      }
    }
    auto func_name = call_op.getCallee();
    if (!func_name.starts_with("quantized_")) return failure();
    if (call_op->getNumResults() != 1) return failure();
    if (!mlir::isa<UniformQuantizedType>(
            getElementTypeOrSelf(call_op->getResult(0).getType())))
      return failure();

// Returns the associated option name for the given op definition.
static inline std::string GetOperatorOptionName(const Record &def) {
  assert(def.getName().starts_with("TFL_") && "unexpected op prefix");
  assert(def.getName().ends_with("Op") && "unexpected op suffix");

  auto *custom_option = dyn_cast<StringInit>(def.getValueInit("customOption"));
  std::ostringstream oss;

                                PatternRewriter& rewriter) const override {
    const auto f_attr = mlir::dyn_cast<FlatSymbolRefAttr>(op.getFAttr());
    StringRef function_name = f_attr.getValue();
    if (!function_name.starts_with(kDequantizeFunctionName)) return failure();

    llvm::SmallVector<Operation*> users(op->getUsers().begin(),
                                        op->getUsers().end());

    bool changed = false;

  }

  // Our MLIR might be the result of a conversion from a previously generated
  // flatbuffer file.
  if (auto custom_op = dyn_cast<mlir::TFL::CustomOp>(*op)) {
    return custom_op.getCustomCode().starts_with("Flex");
  }

  // We never see TFL::IfOps in the IR -- it is flatbuffer_export that rewrites
  // them from TF::IfOps.
  if (isa<TF::IfOp>(op)) {
    return false;
  }

      return StringRef();
    }
  }

  StringRef result;
  for (Attribute class_attr : classes_attr) {
    StringRef node_name = mlir::cast<StringAttr>(class_attr).getValue();
    if (!node_name.starts_with(kLocationPrefix)) {
      continue;
    }
    if (!result.empty()) {
      // Invalid case since there are multiple loc:@ attributes.
      op->emitOpError()

    auto entry_function =
        call_op->getAttrOfType<FlatSymbolRefAttr>("_entry_function");
    StringRef function_name = entry_function.getValue();
    if (!function_name.starts_with("composite_")) {
      return spec;
    }

    if (function_name.contains("conv")) {
      // Looks up `Method` to see if it should be per-channel quantized and
      // populates the spec accordingly.

    new_func.setPrivate();
    symbol_table.insert(new_func);

    // For consistency, we require all quantized composite function to have
    // the "tf_quant.quantized_ops" attribute.
    if (!new_func.getSymName().starts_with("quantized_")) continue;
    if (!new_func->hasAttrOfType<ArrayAttr>("tf_quant.quantized_ops")) {
      new_func->emitError() << "Missing \"tf_quant.quantized_ops\" "

 public:
  using OpRewritePattern<TFL::CustomOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(TFL::CustomOp op,
                                PatternRewriter& rewriter) const override {
    if (!op.getCustomCode().starts_with(kFlexOpNamePrefix)) {
      return failure();
    }

    llvm::StringRef tf_op_name =
        op.getCustomCode().substr(kFlexOpNamePrefix.size());

    StringRef function_name = f_attr.getValue();
    // TODO(b/228928859): Improve the getter function to match attributes rather
    // than function name.
    if (!function_name.starts_with("composite_")) {
      return failure();
    }

    if (function_name.contains("depthwise_conv2d")) {
      // Uniform Quantized op requires weights of tf.DepthwiseConv2dNative to