// inject min/max to the input and output tensors
        auto& input_tensor = sub_graph->tensors[op->inputs[0]];
        input_tensor->quantization->scale.clear();
        input_tensor->quantization->zero_point.clear();
        input_tensor->quantization->min.push_back(-1.0);
        input_tensor->quantization->max.push_back(1.0);

        auto& output_tensor = sub_graph->tensors[op->outputs[0]];

              tensor_spec.TensorSpec(
                  shape=[None], dtype=input_type, name='input_tensor'
              )
          ]
      )
      def model(self, input_tensor: core.Tensor) -> Mapping[str, core.Tensor]:
        """Performs a gather and a 2D convolution operation.

        Args:
          input_tensor: Input tensor to perform operation on.

        Returns:

absl::StatusOr<mlir::ElementsAttr> ConvertTensorProto(
    const TensorProto& input_tensor, mlir::Builder* builder);

// Converts an TensorFlow tensor into an MLIR elements attribute.
absl::StatusOr<mlir::ElementsAttr> ConvertTensor(const Tensor& input_tensor,
                                                 mlir::Builder* builder);

                             absl::string_view tensor_name, Device* device,
                             Tensor* cpu_tensor, StatusCallback done) override;
  void CopyTensorInSameDevice(const Tensor* input_tensor, Device* device,
                              Tensor* output_tensor,
                              StatusCallback done) const override;

  bool use_pjrt_tensor_buffer() const { return use_pjrt_tensor_buffer_; }

  // CHECK: %[[CONSTANT_1:.*]] = stablehlo.constant dense<1.000000e+03> : tensor<1x64xf32>
  // CHECK: return
  // CHECK: }

    module attributes {tf_saved_model.semantics} {
      func.func @main(%arg: tensor<1x2xf32> {tf_saved_model.index_path = ["input_tensor:0"]}) -> (tensor<1x2xf32> {tf_saved_model.index_path = ["output_tensor:0"]}) attributes {tf.entry_function = {inputs = "input_tensor:0", outputs = "output_tensor:0"}, tf_saved_model.exported_names = ["main"]} {
        %0 = tf_executor.graph {
          tf_executor.fetch %arg : tensor<1x2xf32>

      return %2 : tensor<1x3xf32>
    }
  }
)mlir";

constexpr absl::string_view kModuleXlaCallModule = R"mlir(
  module {
    func.func @main(%arg0: tensor<?x2xf32> {tf_saved_model.index_path = ["input_tensor"]}) -> (tensor<?x2xf32>) {
      %0 = stablehlo.constant dense<[-0.211145893, -0.708605706]> : tensor<2xf32>
      %1 = stablehlo.constant dense<[[-0.630731344, 0.54962182], [0.180364341, -0.764542698]]> : tensor<2x2xf32>

  ASSERT_THAT(weights_zero_points, SizeIs(out_channel_size));
  ASSERT_THAT(input_tensor->quantization->scale, SizeIs(1));
  ASSERT_THAT(output_tensor->quantization->scale, SizeIs(1));

  const float eps = 1e-7;

  // Bias scale should be input * per_channel_weight_scale.
  for (size_t i = 0; i < out_channel_size; i++) {
    EXPECT_THAT(bias_scales[i], FloatNear(input_tensor->quantization->scale[0] *

}
constexpr absl::string_view kModuleDotWeightOnlyPtq = R"mlir(
  module {
    func.func @main(%arg0: tensor<?x2xf32> {tf_saved_model.index_path = ["input_tensor"]}) -> (tensor<?x2xf32>) {
      %0 = stablehlo.constant dense<[-0.211145893, -0.708605706]> : tensor<2xf32>
      %1 = stablehlo.constant dense<[[-0.630731344, 0.54962182], [0.180364341, -0.764542698]]> : tensor<2x2xf32>

    for (int i = 0; i < num_inputs; ++i) {
      if (input_tensors[i] == nullptr) continue;
      all_input_tensors.emplace_back();
      Tensor& input_tensor = all_input_tensors.back();
      status->status = TF_TensorToTensor(input_tensors[i], &input_tensor);
      if (!status->status.ok()) return;
      input_tensors_vector[i] = &input_tensor;
    }
  }