%6 = "tf.TPUExecute"(%3#1, %1, %program#1) {_producer_name = "UNKNOWN", device = "/TPU:1"} : (tensor<1x1xf32>, tensor<1x1xf32>, tensor<3x!tf_type.string>) -> tensor<1x1xf32>

# 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)

  if (!op->hasAttr(TF::kTpuReplicateAttr)) {
    op->emitOpError("TF2XLA TPU bridge input check: " + errormsg() +
                    "missing _tpu_replicate attr");
    return false;
  }
  auto opattr = op->getAttr(TF::kTpuReplicateAttr);
  if (opattr != attr) {
    op->emitOpError("TF2XLA TPU bridge input check: " + errormsg() +
                    "invalid _tpu_replicate attr.")

        if (!exported_model.ok()) return exported_model.status();

        // Remove the `tpu` tag from the debug quantized saved model as it is
        // for CPU. Note the 'tpu' value should be the same as `TPU` defined in
        // tensorflow/python/saved_model/tag_constants.py.
        if (quantization_options.has_debugger_config()) {
          tags.erase("tpu");
        }
        py_function_library.SaveExportedModel(

# 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

    mlir::ArrayAttr device_assignment_attr);

// Finds the TPU compilation device and execution devices from `devices` for a
// TPU computation subgraph. Compilation device is determined from looking up
// all TPU_SYSTEM:0 devices and choosing the CPU device associated to the first
// TPU_SYSTEM device sorted lexicographically by replica and task. Execution
// devices are determined by looking up all TPU devices associated with each

  // 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,

  });
  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;