func.return %24 : tensor<1x4xf32>

    auto get_num_elements = [](RankedTensorType tensor) {
      int num_elements = 1;
      for (int i = 0; i < tensor.getRank(); ++i) {
        // Assume dynamic dim size as the dim size one.
        if (!tensor.isDynamicDim(i)) {
          num_elements *= tensor.getDimSize(i);
        }
      }
      return num_elements;
    };

// been converted to constants, all the assign ops are gone, etc.). However,
// TFLite has these variable tensors semantics. So the variable mapping from TF
// to TFLite is actually broken here, we sort of hard-code the variable tensors
// based on the actual ops using them, such as unidirectional_sequence_lstm.
//
// MLIRConverter also benefits from lots of typical compiler optimization like

  for (auto arg_num : input_mapping) {
    const Tensor* tensor;
    if (auto it = variable_snapshots.find(arg_num);
        it != variable_snapshots.end()) {
      tensor = it->second;
    } else {
      tensor = inputs[arg_num - num_missing_prefix_ctx_inputs];
    }

    // The input tensor can have the following cases.
    // 1. Tensor with PjRtTensorBuffer, containing a PjRtBuffer. This case

func.func @main(tensor<4x4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4x4xf32> {
// CHECK: {
// CHECK-NEXT:   version: 3,

// required to collect tensor statistics.
void AddPreCalibrationPasses(
    OpPassManager& pm,
    const ::stablehlo::quantization::CalibrationOptions& calibration_options,
    const ::stablehlo::quantization::QuantizationSpecs& specs,
    const ::stablehlo::quantization::DebuggerConfig& debugger_config);

// Adds passes for static-range quantization post-calibration. Utilizes tensor

  // functionality will be supported in the future.

  // Create a host tensor and copy data into it.
  //
  // Raises an error if shape or dtype are incompatible with T.
  // TODO(b/189458441): Update this when we decide on the representation of
  // shape and dtype in this API.
  // Disclaimer: This API is subject to change as we add support for creating
  // device tensors b/187222691 and enable buffer re-use b/187223179.

opt<std::string> input_shapes(
    "tf-input-shapes",
    llvm::cl::desc(
        "Input tensor shapes. Shapes for different tensors are separated by "
        "':', and dimension sizes for the same tensor are separated by ','"),
    llvm::cl::init(""));

// NOLINTNEXTLINE
opt<std::string> output_arrays(
    "tf-output-arrays", llvm::cl::desc("Output tensor names, separated by ','"),
    llvm::cl::init(""));

// NOLINTNEXTLINE

    base_offset *= dimension;
  }

  return contiguous_offset;
}

// Performs transposition of a tensor represented as a contiguous element array.
// Assumes row-major order. The shape of the input tensor and the desired
// permutation is registered during construction, and calling `TransposeValues`
// returns the transposed tensor values.
class DenseElementsTransposer {
 public:

  // quantization aware training or calibration, for the remaining tensors.
  std::vector<std::pair<std::optional<double>, std::optional<double>>>
      input_ranges;

  // Whether to disable setting the quantization parameters of the input nodes
  // using input ranges.
  bool disable_set_input_nodes_quantization_params = false;

  // The default ranges can be used when a tensor doesn't have quantization