%arg4: tensor<10xf32> {tf._user_specified_name = "b2"}) -> tensor<10xf32>
          attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<32x28x28x1>, #tf_type.shape<3x3x1x5>, #tf_type.shape<5>, #tf_type.shape<3920x10>, #tf_type.shape<10>]} {
  return %arg4 : tensor<10xf32>
}

// -----

// CHECK-LABEL: strips_attributes_with_tf_values
// CHECK-NOT: tf

    %7 = "quantfork.dcast"(%6) : (tensor<2x3x1x1024x!quant.uniform<i8:f32, 5.000000e-01:2>>) -> tensor<2x3x1x1024xf32>

//
// This is needed to support Conv+BatchNorm pattern from Jax models converted
// using jax2tf w/o native serialization. Note that Jax2tf patterns always
// extend bias shapes to a rank of 4, e.g. 1x1x1x5.
def ConvertSubMulToMulAdd : Pat<
  (TF_MulOp
    (TF_SubOp
      (SupportedAffineOpMatcher $conv_out, $input, $weight),
      (TF_ConstOp:$sub_rhs IsFloatElementsAttr:$sub_rhs_value)),

  %0 = stablehlo.constant() {value = dense<3> : tensor<3x3x1x4xi8>} : () -> tensor<3x3x1x4x!quant.uniform<i8:f32:3, {2.000000e+2, 3.000000e+3, 2.000000e+2, 3.000000e+3}>>

  func.func @conv2d_unranked_input_and_dynamic_channel(%arg0: tensor<*xf32>, %arg1: tensor<3x3x3x?xf32>) -> tensor<*xf32> {
    // CHECK: "tf.Conv2D"
    // CHECK-SAME: -> tensor<?x?x?x?xf32>
    %0 = "tf.Conv2D"(%arg0, %arg1) {padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<*xf32>, tensor<3x3x3x?xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

  // CHECK-LABEL: func @conv2d_dynamic_batch