%values, %indices = "tf.ApproxTopK"(%arg0) {aggregate_to_topk = true, is_max_k = true, k = 4 : i64, recall_target = 8.500000e-01 : f32, reduction_dimension = 1 : i64, reduction_input_size_override = -1 : i64} : (tensor<1x4xf32>) -> (tensor<1x4xf32>, tensor<1x4xi32>)
  func.return %values, %indices: tensor<1x4xf32>, tensor<1x4xi32>

  // CHECK-LABEL: approx_top_k_with_max_k_last_reduction_dimension

      auto name = attr.getName();
      if (!(name == "top_k" || name == "reduction_dim" ||
            name == "recall_target" || name == "aggregate_to_topk" ||
            name == "reduction_input_size_override" || name == "is_fallback")) {
        return op.emitOpError()
               << name.getValue() << " is not a supported backend_config"
               << " attribute for ApproxTopK";
      }
    }

    // CHECK: mhlo.compare  GT
    %values, %indices = "tf.ApproxTopK"(%0) {aggregate_to_topk = true, device = "", is_max_k = true, k = 10 : i64, recall_target = 0.949999988 : f32, reduction_dimension = -1 : i64, reduction_input_size_override = -1 : i64} : (tensor<10x500xbf16>) -> (tensor<10x10xbf16>, tensor<10x10xi32>)
    return %values : tensor<10x10xbf16>
  }

  // CHECK-LABEL: fusedBatchNormV3_noTraining

// CHECK:          %[[VALUES:.*]], %[[INDICES:.*]] = "tf.ApproxTopK"(%[[ARG_0]]) <{aggregate_to_topk = true, is_max_k = true, k = 4 : i64, recall_target = 8.500000e-01 : f32, reduction_dimension = 1 : i64, reduction_input_size_override = -1 : i64}> : (tensor<1x4xf32>) -> (tensor<1x4xf32>, tensor<1x4xi32>)
// CHECK:          return %[[VALUES]], %[[INDICES]] : tensor<1x4xf32>, tensor<1x4xi32>
// CHECK:        }

    DefaultValuedOptionalAttr<F32Attr, "0.95f">:$recall_target,
    DefaultValuedOptionalAttr<BoolAttr, "true">:$is_max_k,
    DefaultValuedOptionalAttr<I64Attr, "-1">:$reduction_input_size_override,
    DefaultValuedOptionalAttr<BoolAttr, "true">:$aggregate_to_topk
  );

  let results = (outs