llvm::SmallVector<Value, 4> quantized;
      for (auto user : returned.getUsers()) {
        if (auto q = Quantized(user)) {
          quantized.push_back(q);
        }
      }
      if (quantized.size() == 1) {
        ret.setOperand(i, quantized.front());
      }
      i++;
    }
  });

  // We prefer to placing quantization emulation ops on the results of the

// Returns element type from attribute Type 'type_attr'.
mlir::Type GetShapeStrippedType(mlir::TypeAttr type_attr);

// Returns true if 'val' is not from Quantize op or
// from Quantize Op with same quant type as 'qtype_attr'
bool NotFromQuantOpOrSameQuantType(mlir::Value val, mlir::TypeAttr qtype_attr);

}  // namespace tflite

                            << quantized_per_axis_type << ".\n");
    return false;
  }

  return true;
}

// Determines whether the storage type of a quantized type is supported by
// `tfl.quantize` or `tfl.dequantize` ops. ui8, i8 and i16 are supported.
bool IsSupportedByTfliteQuantizeOrDequantizeOps(IntegerType storage_type) {
  if (storage_type.getWidth() == 8 ||

    // the textual format (on the commandline for example).
    return "quant-propagate-quantize-type";
  }
  StringRef getDescription() const final {
    // This is a brief description of the pass.
    return "Propagate quantized type through allowed ops.";
  }

  void runOnOperation() override;
};

// Propagate dequantize op if the next op supports the data type.
// Given the below graph,

  auto func_op = module_op->lookupSymbol<func::FuncOp>("quantize");
  ASSERT_THAT(func_op, NotNull());

  auto uniform_quantize_op_itr =
      func_op.getBody().op_begin<mlir::stablehlo::UniformQuantizeOp>();
  ASSERT_THAT(
      uniform_quantize_op_itr,
      Ne(func_op.getBody().op_end<mlir::stablehlo::UniformQuantizeOp>()));

  // `uniform_quantize` is considered partially quantized because its output is

        /*value=*/filter_i8_value_attr);

    // Replace filter uses with uniform quantized filter.
    rewriter.replaceAllUsesWith(filter_op->getResult(0),
                                quantized_filter_constant_op.getResult());

    // Replace conv op with a new convolution op that has quantized output type.
    // Quantize -> Dequantize following r3.
    auto output_uniform_quantize_call_op = cast<func::CallOp>(

                  const tensorflow::quantization::PyFunctionLibrary*
                      quantization_py_function_library = nullptr);

// Quantize the model with calibration data. Throw errors if `fully_quantize`
// is specified by the calibration data are not sufficient to quantize the
// model.
PyObject* MlirQuantizeModel(PyObject* data, bool disable_per_channel,
                            bool fully_quantize, int inference_type,

    // execution to be safe. Although normally they should support float
    // execution. Not Quantized ops.
    if (!int8_type_observed && !uint8_type_observed) return;

    // Insert dequantize ops for every quantized input.
    SmallVector<Value, 4> dequantized_inputs;
    for (auto& input : op->getOpOperands()) {
      auto input_type = input.get().getType();

  %zps = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>

  // CHECK: %[[QUANTIZE:.*]] = mhlo.uniform_quantize %arg0 : (tensor<2xf32>) -> tensor<2x!quant.uniform<i8:f32, 1.000000e+00:3>>
  // CHECK: %[[CONVERT_1:.*]] = mhlo.bitcast_convert %[[QUANTIZE]] : (tensor<2x!quant.uniform<i8:f32, 1.000000e+00:3>>) -> tensor<2xi8>

        }
        assign_variable_op.erase();
        dq_op.erase();
      } else {
        // Add quantize op.
        builder.setInsertionPoint(assign_variable_op);
        auto new_q_op = builder.create<QuantizeOp>(
            assign_variable_op.getLoc(), ref_qtype,
            assign_variable_op.getValue(), TypeAttr::get(ref_qtype));