return false;
      });
}

// Return the half_range value that is used by DequantizeOp. half_range is used
// to offset the quantized representation before it gets scaled. In the case
// of negative quantize types, this offset is half the type's range.
static DenseElementsAttr DequantizeHalfRange(OpBuilder *builder, Value input) {
  auto input_type = mlir::dyn_cast_or_null<ShapedType>(input.getType());

  %4 = "tfl.add"(%2, %3) {tac.device = "GPU", tac.inference_type = "FLOAT", fused_activation_function = "NONE"} : (tensor<1x384x384xf32>, tensor<1x384x384xf32>) -> tensor<1x384x384xf32>
  %5 = "tfl.quantize"(%4) {qtype = tensor<1x384x1x!quant.uniform<i8:f32, 0.003:-128>>} : (tensor<1x384x384xf32>) -> tensor<1x384x384x!quant.uniform<i8:f32, 0.003:-128>>

// RUN: stablehlo-quant-opt %s -split-input-file -verify-diagnostics \
// RUN:     -stablehlo-quantize-composite-functions | FileCheck %s
// RUN: stablehlo-quant-opt %s -split-input-file -verify-diagnostics \
// RUN:     -stablehlo-quantize-composite-functions=enable-per-channel-quantized-weight=false | FileCheck --check-prefix=CHECK-PER-TENSOR %s

// Tests that basic dot_general is properly quantized.

module attributes {tf_saved_model.semantics} {

  // before converting TF_Conv to TFL_Conv
  (void)applyPatternsAndFoldGreedily(func, std::move(patterns));

  // Remove the wrapper of the tf.FakeQuant* ops and also insert the
  // tfl.quantize and tfl.dequantize to preserve the quantization parameters.
  // This is done after the first round of optimization to make sure all the
  // min/max operands of the tf.FakeQuant* are constants to be matched. The