void AddPostQuantizationStableHloToTfPasses(
    const mlir::TFL::PassConfig& pass_config,
    mlir::OpPassManager& pass_manager);

// This is the early part of the conversion in isolation. This enables a caller
// to inject more information in the middle of the conversion before resuming it
// (like freezing variables for example).
void AddPreVariableFreezingTFToTFLConversionPasses(
    const mlir::TFL::PassConfig& pass_config,

#include <optional>
#include <string>

#include "mlir/IR/OperationSupport.h"  // from @llvm-project
#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/graph/graph.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

struct MlirDumpConfig;

// Dumps 'graph_def' to a file, as textual IR. Returns the file name chosen.
//

// This applies both the PassManagerCLOptions provided by MLIR along with any
// tensorflow specific options.
//
// Note that this function should be in a more appropriate file, but it is
// unclear what a proper file would be as no other functions would currently be
// in the file also.
void applyTensorflowAndCLOptions(mlir::PassManager& pm,

// Inference type.
constexpr char kInferenceType[] = "tac.inference_type";

// Inference type.
constexpr char kSkipTargetAnnotation[] = "tac.skip_target_annotation";

// TODO(renjieliu): Add more inference types.
enum InferenceType {
  UNKNOWN = 0,
  FLOAT = 1,
  QUANTIZED_INT8 = 2,
  QUANTIZED_UINT8 = 3,
  HYBRID = 4
};

inline InferenceType GetInferenceTypeEnum(llvm::StringRef inference_type_str) {

#include <memory>

#include "tensorflow/core/framework/full_type.pb.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/platform/refcount.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Abstract interface to a Tensor handle in either tracing or immediate
// execution mode.
class AbstractTensorHandle : public core::RefCounted {

#include <memory>
#include <string>

#include "tensorflow/c/tf_status_helper.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/platform/status.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/util/tensor_bundle/tensor_bundle.h"
#include "tensorflow/core/util/tensor_slice_reader.h"

namespace tensorflow {
namespace checkpoint {

class TensorSliceReader;

// TODO(jpienaar): This file and class are confusingly named. This seems to be
// a SavedModel only import options file that exposes a subset of the
// GraphImportConfig options, but the naming would make one think it is more
// general.
struct MLIRImportOptions {
  // If true, functionalize the input graph before importing it into MLIR.
  bool upgrade_legacy = false;

                             absl::Span<AbstractTensorHandle* const> inputs,
                             std::vector<AbstractTensorHandle*>* params);

// A callable that takes tensor inputs and returns zero or more tensor outputs.
using Model = std::function<Status(AbstractContext*,
                                   absl::Span<AbstractTensorHandle* const>,
                                   absl::Span<AbstractTensorHandle*>)>;

#include "tensorflow/compiler/mlir/quantization/tensorflow/exported_model.pb.h"
#include "tensorflow/compiler/mlir/quantization/tensorflow/quantization_options.pb.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/protobuf/meta_graph.pb.h"
#include "tensorflow/python/lib/core/pybind11_lib.h"
#include "tsl/platform/protobuf.h"  // IWYU pragma: keep

namespace pybind11::detail {
namespace internal {

// Loads a TFLite flatbuffer, convert to TOSA for backend compilation and
// produce an MLIR bytecode file as output.
// TODO(jpienaar): Refactor this when we use more implicit module passing
// between calls to avoid serialization overhead.
void ExperimentalTFLiteToTosaBytecode(
    const std::string &flatbuffer_file, const std::string &tosa_bytecode_file,
    bool use_external_constant,