auto out_count = rewriter.create<MulOp>(
        loc, tensorflow::GetTypeFromTFTensorShape({}, out_size_element_ty),
        out_y, out_x);

    // Generate what the final output shape will look like.
    auto out_shape = rewriter.create<PackOp>(
        loc, tensorflow::GetTypeFromTFTensorShape({4}, out_size_element_ty),
        ValueRange({batch, out_y, out_x, channels}));

    // Compute the indices along the vertical dimension.

// TF-DAG: %[[conv1_dequantized_0:.*]] = "tf.PartitionedCall"(%[[conv1_quantized]], %[[out_scale]], %[[in_out_zp]]) <{config = "", config_proto = "", executor_type = "", f = @dequantize_i8}>

// quantize_conv_with_bias_dynamic_fn, omitting stablehlo.maximum.
// This is because activation clipping which includes 0.0f can be simply
// omitted from the graph as the lifted function's out_scale and out_zp are
// already calculated based on the clipped distribution.
// Note that the resulting scale and zero point should be calculated based on
// clipped range [0, r_max].

        print('TODO: use "node_type"')
      if node.attr == 'shape' and tensor_type == TFRTypes.TENSOR:
        ssa_value = self._ssa_name('shape')
        self._emit_with_loc(
            '\n{} = tfr.get_shape {} -> !shape.shape'.format(ssa_value, value),
            node)
        return (ssa_value, TFRTypes.SHAPE)

    if isinstance(node.value, ast.Attribute):
      if isinstance(node.value.value, ast.Name):