#ifndef TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_
#define TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_

// Run a function containing a MatMul op and check its output.
// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.

# Use a cache directory mounted to /tf/cache. Very useful!
build:sigbuild_local_cache --disk_cache=/tf/cache
# Use the public-access TF DevInfra cache (read only)
build:sigbuild_remote_cache --remote_cache="https://storage.googleapis.com/tensorflow-devinfra-bazel-cache/manylinux2014" --remote_upload_local_results=false
# Change the value of CACHEBUSTER when upgrading the toolchain, or when testing
# different compilation methods. E.g. for a PR to test a new CUDA version, set

        "//tensorflow/core:protos_all_cc",
        "//tensorflow/core:test",
        "//tensorflow/core:test_main",
    ],
)

tf_cc_test(
    name = "parallel_device_remote_test",
    srcs = ["parallel_device_remote_test.cc"],
    # TODO(b/136478427): Enable global heap checking when servers shut down
    # cleanly.
    args = ["--heap_check="],
    deps = [
        ":parallel_device",

  virtual const char* DeviceType(Status* status) const = 0;
  // Returns the device ID which created the handle.
  virtual int DeviceId(Status* status) const = 0;
  // Returns a tensor for the handle. If tensor is remote, it will be copied.
  virtual AbstractTensorInterface* Resolve(Status* status) = 0;

  std::string DebugString() const override;

  // Returns a Boolean hint indicating whether callers should prefer

  TestRemoteExecuteSilentCopiesFunc(/*async=*/true, /*remote=*/false,
                                    /*heavy_load_on_streaming_rpc=*/false,
                                    /*remote_func_outputs=*/true);
}
TEST(CAPI, RemoteExecuteSilentCopiesLocalAsyncFuncOrdering) {
  // A remote input may be not ready when we start running a function. Test that
  // the function execution should wait until the remote input is ready.

A. Unique TensorFlower <******@****.***> 1682633218 -0700

  // Set up distributed execution environment on local and remote tasks.
  // When `reset_context` is true, initialize new cluster context state based
  // on cluster configurations provided in `server_def`; otherwise, update
  // existing context state with the provided `server_def`. Contexts created
  // on remote tasks will be considered stale and garbage collected after
  // `keep_alive_secs` of inactivity.

  // 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.
    ASSERT_TRUE(remote_arg->HasLocalMirror(nullptr));
  }

  if (remote_func_outputs) {
    const string backing_device =

# Use a cache directory mounted to /tf/cache. Very useful!
build:sigbuild_local_cache --disk_cache=/tf/cache
# Use the public-access TF DevInfra cache (read only)
build:sigbuild_remote_cache --remote_cache="https://storage.googleapis.com/tensorflow-devinfra-bazel-cache/manylinux2014" --remote_upload_local_results=false
# Change the value of CACHEBUSTER when upgrading the toolchain, or when testing
# different compilation methods. E.g. for a PR to test a new CUDA version, set

#      Ex. nightly_upload -- for TF nightly official builds; changes version numbers
#      Ex. no_upload      -- Disable all uploads, usually for temporary CI issues

# Recommended: use a local+remote cache.
#
#   Bazel will cache your builds in tensorflow/build_output/cache,
#   and will also try using public build cache results to speed up
#   your builds. This usually saves a lot of time, especially when