// unquantized_tensor_data.pb as it is used by unquantized dump model.
    // After saving unquantized dump model, the file name will be changed to
    // quantized_tensor_data.pb.
    // Since this process doesn't happen for per layer, we need to set file_name
    // as quantized_tensor_data.pb here.
    // TODO: b/296933893 - Refactor the debugger code when no quantize option
    // is added
    std::string file_name =

  "32-bit quantized integer">;
def TF_Quint8  : AnyTypeOf<
  [TF_TensorFlowType<"Quint8", "quint8">, TF_Quint8Ref],
  "8-bit quantized unsigned integer">;
def TF_Quint16 : AnyTypeOf<
  [TF_TensorFlowType<"Quint16", "quint16">, TF_Quint16Ref],
  "16-bit quantized unsigned integer">;

// Any quantized type
def TF_Quantized : AnyTypeOf<
  [TF_Qint8, TF_Qint16, TF_Qint32, TF_Quint8, TF_Quint16], "quantized">;

include "tensorflow/compiler/mlir/lite/ir/tfl_ops.td"

// Quantize attribute $0 by using quantization parameter from %1.
def QuantizeByQuantizedType : NativeCodeCall<"quant::Quantize($0, $1.getValue())">;
def F32ElementsAttr : ElementsAttrBase<
  CPred<"$_self.cast<ElementsAttr>().getShapedType().getElementType().isF32()">, "float constant tensor">;

// Squash tfl.dequantize and tfl.quantize pairs.

namespace mlir {
namespace TFL {

bool IsQuantized(const tflite::TensorT& tensor);

absl::StatusOr<mlir::quant::QuantizedType> GetQuantizedType(
    const tflite::TensorT& tensor, mlir::Builder builder,
    bool is_constant = false, mlir::Type storage_type = {});

// Imports float tensor with calibration value into calibrated quantized type.
absl::StatusOr<mlir::quant::QuantizedType> GetCalibratedQuantizedType(

      != _PresetMethod.METHOD_STATIC_RANGE_WEIGHT_ONLY_INT8
  ):
    raise ValueError(
        'StableHLO quantized opset currently only supports static range'
        ' quantization and weight-only quantizationvia TF Quantizer.'
    )

  # Set `force_graph_mode_calibration` to True to avoid skipping op execution,
  # which are not connected to return ops, during calibration execution.

// RUN: tf-quant-opt %s -quant-insert-quantized-functions='quantization-method=weight_only target-opset=XLA' | FileCheck %s

// Empty module
module {
  func.func @simple_fn(%arg0: tensor<*xf32>) -> tensor<*xf32> {
    func.return %arg0 : tensor<*xf32>
  }
}

// CHECK-NOT: func private @internal_dequantize_f32
// CHECK-NOT: func private @internal_conv3d_fn
// CHECK-NOT: func private @internal_batch_matmul_fn

    "tf-custom-opdefs", llvm::cl::desc("List of custom opdefs when importing "
                                       "graphdef"));

// Quantize and Dequantize ops pair can be optionally emitted before and after
// the quantized model as the adaptors to receive and produce floating point
// type data with the quantized model. Set this to `false` if the model input is
// integer types.
// NOLINTNEXTLINE
opt<bool> emit_quant_adaptor_ops(

// Stores information about how to quantize a user-specified custom operation.
// CustomOpInfo contains info of its corresponding CustomOp registered in the
// CustomOpMap. 'quantizable_input_indices' is used to determine which indices
// of the CustomOp are quantizable. 'is_weight_only' is used specify whether the
// custom op is quantized only for storage and dequantized at runtime.

  // The following two passes find specific uniform quantization patterns in
  // StableHLO and converts them to TFLite ops that accept or produce uniform
  // quantized types. They only target a specific set of models that contain
  // "decomposed" quantized ops produced from the framework level. This is why
  // they are placed right after the `LegalizeTFXlaCallModuleToStablehloPass`

          clEnumValN(OpSet::XLA, "XLA", "Uses TF XLA ops"),
          clEnumValN(OpSet::UNIFORM_QUANTIZED, "UNIFORM_QUANTIZED",
                     "Uses TF Uniform Quantized ops"))};

  Option<QuantMethod> quantization_method_{
      *this, "quantization-method",
      llvm::cl::init(tensorflow::quantization::QuantizationMethod::
                         METHOD_STATIC_RANGE_INT8),