%0 = "tf.GatherV2"(%cst_0, %arg0, %cst) {batch_dims = 0 : i64, device = ""} : (tensor<128x32xf32>, tensor<6xi64>, tensor<i32>) -> tensor<6x32xf32>
  return %0 : tensor<6x32xf32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<i32>
// CHECK-DAG: %[[CST_1:.*]] = "tf.Const"() {{.*}} : () -> tensor<128x32xf32>

  func.return %1 : tensor<128x32xf32>
  // CHECK: %[[FULLY_CONNECTED:.*]] = "tfl.fully_connected"(%arg0, %arg1, %arg2) <{fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"}> : (tensor<128x64xf32>, tensor<32x64xf32>, tensor<32xf32>) -> tensor<128x32xf32>
  // CHECK: return %[[FULLY_CONNECTED]] : tensor<128x32xf32>
}

  %0 = "tf.CollectivePermute"(%arg0, %source_target_pairs) {
  } : (tensor<128x32xf32>, tensor<3x2xi32>) -> tensor<128x32xf32>

  func.return %0 : tensor<128x32xf32>
}

// -----

// CHECK-LABEL: @cross_replica_sum
func.func @cross_replica_sum(%input: tensor<10xf32>) -> tensor<10xf32> {
  %replica_groups = "tf.Const" () {

}

// -----

module {
func.func private @func_20_GPU_FLOAT(%arg0: tensor<128x128xf32>, %arg1: tensor<3xi32>) -> tensor<1x128x128xf32> attributes {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_20"} {
  %0 = "tfl.reshape"(%arg0, %arg1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x128xf32>, tensor<3xi32>) -> tensor<1x128x128xf32>
  func.return %0 : tensor<1x128x128xf32>
}

func.func @Int64SliceBeginSize(%arg0: tensor<4x128x32xf32>) -> tensor<1x128x32xf32> {
  %0 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi64>} : () -> tensor<3xi64>
  %1 = "tfl.pseudo_const"() {value = dense<[1, 128, 32]> : tensor<3xi64>} : () -> tensor<3xi64>
  %2 = "tfl.slice"(%arg0, %0, %1) : (tensor<4x128x32xf32>, tensor<3xi64>, tensor<3xi64>) -> tensor<1x128x32xf32>
  func.return %2 : tensor<1x128x32xf32>

// CHECK-NEXT:  %[[cst:.*]] = arith.constant dense<1> : tensor<i32>
  %cst = arith.constant dense<1> : tensor<i32>
// CHECK-NEXT:  %[[softmax:.*]] = "tfl.softmax"(%arg0) <{beta = 1.000000e+00 : f32}> : (tensor<128x16xf32>) -> tensor<128x16xf32>
  %0 = "tfl.softmax"(%arg0) {beta = 1.000000e+00 : f32} : (tensor<128x16xf32>) -> tensor<128x16xf32>

  func.func @main(%arg0: tensor<128x10xf32>, %arg1: tensor<10x1024xf32>, %arg2: tensor<128x1024xf32>) -> (tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, tensor<128x1024xf32> {mhlo.sharding = ""}) {
    func.return %arg0, %arg1, %arg2 : tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>
  }
}

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

// CHECK-LABEL: matmulBiasAdd
func.func @matmulBiasAdd(%arg0: tensor<64xf32>, %arg1: tensor<8x32xf32>, %arg2: tensor<32x64xf32>) -> (tensor<*xf32>) {
  // CHECK: %[[VAL_3:.*]] = "tf._FusedMatMul"(%arg1, %arg2, %arg0) <{epsilon = 0.000000e+00 : f32, fused_ops = ["BiasAdd"], transpose_a = false, transpose_b = false}> : (tensor<8x32xf32>, tensor<32x64xf32>, tensor<64xf32>) -> tensor<*xf32>

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>