llvm::DenseSet<Operation*> redundant_stats_ops;

  // Step 0: remove the quantfork::StatisticsOp which are used by the
  // quant.qcast op in case it overrides the information from training FakeQuant
  // ops.
  func.walk([&](quantfork::QuantizeCastOp q) {
    auto input_op = q.getArg().getDefiningOp();
    if (auto stats = dyn_cast_or_null<quantfork::StatisticsOp>(input_op)) {
      q.setOperand(stats.getArg());

  // outputs.
  llvm::DenseMap<SignedInteger, QuantParamsForResults> restricted_output_params;

  // Coefficient operand index and whether supporting per-channel quantization.
  // For QAT, this information is carried by the FakeQuant*/Quantize/Dequantize
  // ops, but post-training quantization, the quantization parameters need to be
  // inferred from the tensor content and op property. A "-1" value indicates

        values[output_index] = item;
      }
    }

    rewriter.replaceOpWithNewOp<PackOp>(op, op.getType(), values);
    return success();
  }
};

// This pass performs a manual conversion with FakeQuant, converting between
// floating point and quantized space. It is designed to reproduce TF's
// implementation, mirroring the previous XLA implementation.
//

    loader = saved_model_loader.SavedModelLoader(self._output_saved_model_path)
    graphdef = loader.get_meta_graph_def_from_tags(tags).graph_def
    # The Conv2D op shouldn't be quantized as it has no FakeQuant on input.
    self.assertTrue(self._contains_op(graphdef, 'Conv2D'))
    # If the Gather op is quantized, input_model_size / output_model_size > 2.
    self.assertLess(
        testing.get_size_ratio(

  let results = (outs TFL_FpTensor:$output);
}

def TFL_FakeQuantOp : TFL_Op<"fake_quant", [
    Pure,
    QuantizableResult]> {
  let summary = "FakeQuant operator";

  let description = [{
    Fake-quantize the 'inputs' tensor of type float via float scalars min and
    max to 'outputs' tensor of same shape as inputs.
  }];

  let arguments = (

        tensorflow is running in.
    *   It is now safe to call any of the C API's TF_Delete\* functions on
        nullptr.
    *   Log some errors on Android to logcat.
    *   Match FakeQuant numerics in TFLite to improve accuracy of TFLite
        quantized inference models.
    *   Optional bucket location check for the GCS Filesystem.
    *   Performance enhancements for StringSplitOp & StringSplitV2Op.