%3 = "tfl.dequantize"(%arg2) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100x!quant.uniform<i8:f32, 2.000000e-01:-3>>) -> tensor<100xf32>
    %4 = tfl.mul %2, %3 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<100xf32>

    // 1. Collect quantizable ops.
    QuantizationUnits quantizable_ops = GetQuantizableOps(op);
    if (quantizable_ops.empty()) {
      return failure();
    }

    // 2. Quantize collected ops.
    if (!QuantizeOps(rewriter, op, quantizable_ops)) {
      return failure();
    }

    // 3. Complete the Q-DQ pair for each inference type.
    if (!ConvertToFloat16Constant(rewriter, op)) {

// 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.

  bool reduce_type_precision = false;
  // Whether to consider this model a quantized model with quantize/dequantize
  // ops and to convert kernels to quantized kernels wherever appropriate.
  quant::QDQConversionMode qdq_conversion_mode =
      quant::QDQConversionMode::kQDQNone;

  // When set to true, StableHLO Quantizer is run. The full configuration for
  // the quantizer is at `TocoFlags::quantization_config`.

    enable_whole_model_verify,
    denylisted_ops,
    denylisted_nodes,
    enable_variable_quantization,
    disable_per_channel_for_dense_layers,
    debug_options_str,
):
  """Wraps experimental mlir quantize model."""
  return _pywrap_converter_api.ExperimentalMlirQuantizeModel(
      input_data_str,
      disable_per_channel,
      fully_quantize,
      inference_type,
      input_data_type,
      output_data_type,

};

void QuantizeCompositeFunctionsPass::runOnOperation() {
  MLIRContext& ctx = getContext();

  PassManager pm(&ctx);
  // Intermediate output from QuantizePass will have quantized ops
  // (XlaCallModuleOps) with quantized input and output types, which are not
  // allowed in the TF dialect.
  pm.enableVerifier(false);

  PrepareQuantizePassOptions options;
  options.enable_per_channel_quantized_weight_ =

        per_axis_type.getStorageTypeMin(), per_axis_type.getStorageTypeMax());
  }

  auto quantize = builder.create<quantfork::QuantizeCastOp>(
      q_op.getLoc(), new_value_type.clone(new_qtype), new_value);
  auto dequantize = builder.create<quantfork::DequantizeCastOp>(
      dq_op.getLoc(), new_value_type, quantize.getResult());
  return dequantize.getResult();
}

      func_name, rewriter, quant_type, val_to_dequantize, result_type,
      LogicsForUniformDequanization);

  return dequant_op;
}
}  // namespace

// Generate quantize and dequantize functions with uniform quantization.
std::optional<TF::PartitionedCallOp> ApplyUniformQuantization(
    PatternRewriter& rewriter, TF::ConstOp op,
    tensorflow::quantization::QuantizationComponentSpec& weight_spec) {

)

filegroup(
    name = "quantize_header",
    srcs = ["quantize.h"],
    visibility = ["//visibility:public"],
)

cc_library(
    name = "tfcompile_lib",
    srcs = [
        "codegen.cc",
        "compile.cc",
        "flags.cc",
    ],
    hdrs = [
        "codegen.h",
        "compile.h",
        "flags.h",
        "quantize.h",
    ],
    compatible_with = [],

    lines.push_back(absl::StrFormat(
        "Number of quantized layers with quantized outputs: %d/%d",
        total_quantized_func_count - float_output_func_count,
        total_quantized_func_count));
    lines.push_back(absl::StrFormat("Number of quantize layers added: %d",
                                    quantize_func_count));
    lines.push_back(absl::StrFormat("Number of dequantize layers added: %d",