def optimizeConsecutiveConv#OpsTuple[0]#OpsTuple[1] : Pat<
    (OpsTuple[1]
      (OpsTuple[0] $input, $zp_offset, $broadcast_dims_1),
      $bias, $broadcast_dims_2),
    (OpsTuple[0]
      $input,
      (OpsTuple[2] $zp_offset, $bias, $broadcast_dims_2), $broadcast_dims_1),
    [
      (IsNull $broadcast_dims_1),
      (IsNull $broadcast_dims_2),
      (TensorOf<[AnyInteger]> $input),

// CHECK: return %[[TRANSPOSE_1]]

// -----

// Tests that a `add(convolution(%activation, %weight), %bias)` pattern with the
// activation tensor of NCHW format and non-constant bias is converted to NHWC
// convolution, but without the deferred transpose for `stablehlo.add`.
// Transpose ops are inserted to the activation and output of

// Specify this trait if the bias-th input of the op is a bias input, which
// needs a scale based on the scales of op1 and op2.
class AccumulatorUniformScale<int bias, int op1, int op2> : NativeOpTrait<
  !strconcat("quant::AccumulatorUniformScale<",
             !interleave([bias, op1, op2], ", "),
             ">::Impl")>;

def FoldQuantWeightsIntoTposeConv : Pat<
  (TFL_TransposeConvOp
    $output_shape,
    (TFL_DequantizeOp $quant_weights),
    $quant_input,
    $bias, $padding, $stride_h, $stride_w, $faf),
  (TFL_TransposeConvOp $output_shape, $quant_weights,
    $quant_input, $bias, $padding, $stride_h, $stride_w, $faf), 

  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>
  %bias = arith.constant dense<[7.11401462, 7.05456924]> : tensor<2xf32>

  %q_input= "quantfork.qcast"(%input) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3x!quant.uniform<i8:f32, 0.58810077742034317:-128>>

  //     constant is used by multiple ops as a bias, duplicate constants and
  //     let each op assign its own quantization parameter for bias.
  //   - Adds all the non-bias constants (weights) to a set for looking up
  //     later.
  //   - Adds all per-channel weights to a set for looking up later.
  void PreprocessConstantOps();

      bias_fn: Optional[ops.Operation] = None,
      activation_fn: Optional[ops.Operation] = None,
  ) -> module.Module:
    class MatmulModel(module.Module):
      """A simple model with a single matmul.

      Bias and activation function are optional.
      """

      def __init__(
          self,
          weight_shape: Sequence[int],
      ) -> None:
        """Initializes a MatmulModel.

        Args:

                    "\""# func_name #"\", $0...)", returns>;

// Add the second argument to the first argument, which is expected to be an
// argument list.
// bias(einsum(inputs), bias) --> einsum_with_bias(AppendToVector(inputs, bias))
// Since inputs is a vector in case of einsum, we cannot use ArgumentList here.
def AppendToVector : NativeCodeCall<"AppendToVector($0, $1)">;

    A.reset(A_raw);
  }
  // Bias
  float Bias_vals[] = {2.0f, 3.0f};
  int64_t Bias_dims[] = {2};
  AbstractTensorHandlePtr Bias;
  {
    AbstractTensorHandle* Bias_raw;
    status_ = TestTensorHandleWithDims<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), Bias_vals, Bias_dims, 1, &Bias_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    Bias.reset(Bias_raw);
  }

                         PatternRewriter& rewriter) {
  auto bias = op->getOperand(bias_idx);

  if (!mlir::isa<NoneType>(bias.getType())) return failure();

  // Proceed to create a zero bias.
  auto output = op->getResult(0);
  auto output_type = mlir::dyn_cast_or_null<RankedTensorType>(output.getType());
  if (!output_type) return failure();

  // bias should be a vector sized of the last output dim.