TFE_DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
} TFE_ContextDevicePlacementPolicy;
// LINT.ThenChange(//tensorflow/c/eager/immediate_execution_context.h)

// Sets the default execution mode (sync/async). Note that this can be
// overridden per thread using TFE_ContextSetExecutorForThread.
TF_CAPI_EXPORT extern void TFE_ContextOptionsSetAsync(TFE_ContextOptions*,

namespace {

void TestRemoteExecuteSilentCopiesFunc(bool async, bool remote,
                                       bool heavy_load_on_streaming_rpc,
                                       bool remote_func_outputs = false,
                                       bool has_packed_input = false) {
  return TestRemoteExecuteSilentCopies(async, remote, /*func=*/true,
                                       heavy_load_on_streaming_rpc,

// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.
// The effect is bounding the memory held by inflight TensorHandles that are

  virtual Status InputLength(const char* input_name, int* length) = 0;
  virtual Status OutputLength(const char* output_name, int* length) = 0;

  // Set stack trace to be used for potential async error reporting.
  virtual void SetStackTrace(ManagedStackTrace stack_trace) = 0;

  virtual const tensorflow::AbstractOpAttrs* GetOpAttrs() const = 0;
  virtual void AddAttrs(const AbstractOpAttrs* op_attrs) = 0;

#include <memory>

#include "tensorflow/core/common_runtime/eager/eager_executor.h"

struct TFE_Executor {
  explicit TFE_Executor(bool async, bool enable_streaming_enqueue,
                        int in_flight_nodes_limit)
      : owned_executor(new tensorflow::EagerExecutor(
            async, enable_streaming_enqueue, in_flight_nodes_limit)) {}

  explicit TFE_Executor(tensorflow::EagerExecutor* executor)

  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  // TODO(gjn): Add support for waiting on async local mirrors
  if (!remote && !async && !remote_func_outputs) {
    auto remote_arg =
        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(h1_task2));
    // The input handles should never change since they have been mirrored.

      tensorflow::unwrap(cancellation_manager));
  status->status = absl::OkStatus();
}

TFE_Executor* TFE_NewExecutor(bool is_async, bool enable_streaming_enqueue,
                              int in_flight_nodes_limit) {
  return new TFE_Executor(is_async, enable_streaming_enqueue,
                          in_flight_nodes_limit);
}

void TFE_DeleteExecutor(TFE_Executor* executor) { delete executor; }

}

// Note: this function looks up a thread local policy. So it should be called in
// the appropriate client thread. In particular, in async mode, it may not be
// safe to call this function from the async EagerExecutor threads.
extern TFE_ContextDevicePlacementPolicy TFE_ContextGetDevicePlacementPolicy(
    TFE_Context* ctx) {
  return static_cast<TFE_ContextDevicePlacementPolicy>(

    "if_google",
    "if_oss",
    "if_xla_available",
    "tf_cc_shared_object",
    "tf_custom_op_library_additional_deps_impl",
    "tf_monitoring_python_deps",
    "tf_native_cc_binary",
    "tsl_async_value_deps",
)
load(
    "//tensorflow:tensorflow.default.bzl",
    "ADDITIONAL_API_INDEXABLE_SETTINGS",
    "tf_cc_shared_library",
)
load(
    "@local_xla//xla/tsl/mkl:build_defs.bzl",
    "if_mkl_ml",
)

*   `tf.distribute`:
    *   Support added for global sync `BatchNormalization` by using the newly
        added `tf.keras.layers.experimental.SyncBatchNormalization` layer. This
        layer will sync `BatchNormalization` statistics every step across all
        replicas taking part in sync training.
    *   Performance improvements for GPU multi-worker distributed training using