# See the License for the specific language governing permissions and
# limitations under the License.

set -o errexit
set -o nounset
set -o pipefail

red=$(tput setaf 1)
reset=$(tput sgr0)
readonly red reset

KUBE_ROOT=$(dirname "${BASH_SOURCE[0]}")/../..
ALL_TARGETS=$(make -C "${KUBE_ROOT}" PRINT_HELP=y -rpn | sed -n -e '/^$/ { n ; /^[^ .#][^ ]*:/ { s/:.*$// ; p ; } ; }' | sort)

    %0:4 = tf_executor.graph {
      %outputs:4, %control = tf_executor.island {
       "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "cluster", device = "device", num_replicas = 1, topology = "topology"} : () -> ()
        %1 = "tf.opA"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>

# bazel test invocation as normal.
test:nonpip_filters --test_tag_filters=-no_oss,-oss_serial,-gpu,-tpu,-benchmark-test,-v1only,-no_aarch64,-no_oss_py38,-no_oss_py39,-no_oss_py310
test:nonpip_filters --build_tag_filters=-no_oss,-oss_serial,-gpu,-tpu,-benchmark-test,-v1only,-no_aarch64,-no_oss_py38,-no_oss_py39,-no_oss_py310

  // symbol_table caches callees in the CallGraph.
  SymbolTableCollection symbol_table;
  // List pending nodes to traverse with their root TPU cluster.
  llvm::SmallVector<std::pair<CallGraphNode*, tf_device::ClusterOp>>
      pending_call_nodes;
  // Cache the host device for each TPU cluster.
  std::unordered_map<Operation*, std::optional<std::string>> cluster_to_host;

  auto insert_pending_op = [&](Operation* op,

      "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "cluster", device = "device", num_replicas = 1 : i64, topology = "topology"} : () -> ()
      %0 = "tf.opA"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>

  });
  return failure(walk_result.wasInterrupted());
}

// Updates the operand of TPU embedding enqueue ops depending on whether
// the graph is in training mode or in non-training mode.
// If SendTPUEmbeddingGradients op is present, this means that graph is in
// training mode. As so, correctly feed in `then` branch value of SelectV2
// operand as inputs to the TPU embedding enqueue ops.
LogicalResult UpdateEmbeddingEnqueueOpInput(

}

// "x's type V is not a named type and T is a type parameter, and x is assignable to each specific type in T's type set."
func _[
	TP0 any,
	TP1 ~_Chan,
	TP2 ~chan int | ~chan byte,
]() {
	var (
		_ TP0 = c // ERRORx `cannot use c .* as TP0 value`
		_ TP0 = C // ERRORx `cannot use C .* as TP0 value`
		_ TP1 = c
		_ TP1 = C // ERRORx `cannot use C .* as TP1 value`

#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/protobuf/config.pb.h"
#include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h"
#include "tensorflow/core/tpu/kernels/tpu_compile_op_support.h"
#include "tsl/lib/core/status_test_util.h"

namespace mlir::quant::stablehlo {
namespace {

using ::testing::Test;

#include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h"
#include "tsl/platform/errors.h"

namespace tensorflow {
namespace tf2xla {
namespace internal {
namespace {

constexpr char kEntryFuncName[] = "main";

absl::Status SetupArguments(mlir::ModuleOp module,
                            std::vector<TensorShape>& arg_shapes,
                            tpu::TPUCompileMetadataProto& metadata_proto) {

// RUN: tf-tfrt-opt -tfrt-lower-cluster-to-runtime-ops-tpu -split-input-file -verify-diagnostics %s | FileCheck %s

module attributes {tf.versions = {producer = 888 : i32}, tf.devices = ["/job:worker/replica:0/task:0/device:CPU:0", "/job:worker/replica:0/task:0/device:TPU_SYSTEM:0", "/job:worker/replica:0/task:0/device:TPU:0"]} {

  // CHECK-LABEL: @converts_cluster
  func.func @converts_cluster() {
    // CHECK: %0:2 = "tf_device.launch"() <{{.*}}> ({