%handle = "tf.VarHandleOp"() {device = "/device:CPU:0", container = "", shared_name = "x"} : () -> tensor<!tf_type.resource<tensor<3xf32>>>
  %0 = "tf.ReadVariableOp"(%handle) {_output_shapes = ["tfshape$dim { size: 3 }"], device = "/device:CPU:0", dtype = f32} : (tensor<!tf_type.resource<tensor<3xf32>>>) -> tensor<3xf32>

// CHECK:           %[[VAL_4:.*]]:2 = call @func_0_GPU_FLOAT(%[[VAL_0]], %[[VAL_1]], %[[VAL_2]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>, tensor<1xf32>)

      %outputs_59, %control_60 = tf_executor.island wraps "tf.Reshape"(%outputs_57, %outputs_23) {_tpu_replicate = "cluster_sample_sequence", device = ""} : (tensor<*xf32>, tensor<4xi32>) -> tensor<*xf32>

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>

    %0 = tf_executor.graph {
      %outputs, %control = tf_executor.island wraps "tf.Const"() {device = "", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>
      %outputs_0, %control_1 = tf_executor.island wraps "tf.Const"() {device = "", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>

^bb0(%arg0: tensor<40x37xf32>, %arg1: tensor<40x37xf32>):
  %cst = arith.constant dense<1.0> : tensor<40xf32>
  %0:2 = "tfl.fully_connected"(%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<40x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> (tensor<40x40xf32>, tensor<40x40xf32>)
  func.return %0 : tensor<40x40xf32>

  %dense_default_0 = "tf.Const"() {device = "/device:CPU:0", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>
  %dense_default_1 = "tf.Const"() {device = "/device:CPU:0", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>
  %dense_keys = "tf.Const"() {device = "/device:CPU:0", dtype = !tf_type.string, value = dense<""> : tensor<2x!tf_type.string>} : () -> tensor<2x!tf_type.string>

  // CHECK: %[[dq2:.*]] = "quantfork.dcast"(%[[q2]])
  // CHECK-SAME: quant.uniform<i4:f32, 0.13329366842905679:-1>
  %0 = "quantfork.stats"(%arg0) {bitsNum = 8 : i64, layerStats = dense<[-0.999415695, 0.99998933]> : tensor<2xf32>, narrowRange = false} : (tensor<?x3xf32>) -> tensor<?x3xf32>
  // CHECK: %[[dot:.*]] = stablehlo.dot %[[dq2]], %[[dq1]]
  %1 = stablehlo.dot %0, %cst : (tensor<?x3xf32>, tensor<3x2xf32>) -> tensor<?x2xf32>

// CHECK:         }
func.func @tfl_wrapped_jax_random_uniform(%arg0: tensor<2xui32>) -> tuple<tensor<1x2xf32>> {
  // This is a fake jax random uniform body.
  %0 = stablehlo.constant dense<0.0> : tensor<2xf32>
  %1 = "stablehlo.reshape"(%0) : (tensor<2xf32>) -> tensor<1x2xf32>
  %2 = "stablehlo.tuple"(%1) : (tensor<1x2xf32>) -> tuple<tensor<1x2xf32>>
  func.return %2 : tuple<tensor<1x2xf32>>

    %pred = arith.constant false
    cf.cond_br %pred, ^bb1(%v : tensor<!tf_type.resource<tensor<?xf32>>>), ^bb1(%v1 : tensor<!tf_type.resource<tensor<?xf32>>>)
  ^bb1(%either: tensor<!tf_type.resource<tensor<?xf32>>>):
    %ret = "tf.ReadVariableOp"(%either) : (tensor<!tf_type.resource<tensor<?xf32>>>) -> tensor<?xf32>