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

# 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

TF_CAPI_EXPORT extern TFE_ContextDevicePlacementPolicy
TFE_ContextGetDevicePlacementPolicy(TFE_Context* ctx);

// A tensorflow.ServerDef specifies remote workers (in addition to the current
// workers name). Operations created in this context can then be executed on
// any of these remote workers by setting an appropriate device.
//
// If the following is set, all servers identified by the
// ServerDef must be up when the context is created.

  // 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 =

  // 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.

# CuDNN: https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#ubuntu-network-installation
libcudnn8=8.9.6.50-1+cuda12.2
libcudnn8-dev=8.9.6.50-1+cuda12.2

# This can be removed once NVIDIA publishes a cuda-12.3.2 Docker image.
# For now it ensures that we install at least version 12.3.107 of PTXAS,
# since 12.3.103 has a bug.

// Dynamic cluster API.

// Update an existing context with a new set of servers defined in a ServerDef
// proto. Servers can be added to and removed from the list of remote workers
// in the context. A New set of servers identified by the ServerDef must be up
// when the context is updated.
//
// This API is for experimental usage and may be subject to change.

     *
     *    dA = U
     *    dbias = reduceSum(U, dims = channel_dim)
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

    // Recover data format from forward pass for gradient.
    std::string data_format;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("data_format", &data_format));

    // Grad for A
    grad_inputs[0] = upstream_grad;

void TF_InitGraph(TP_OptimizerRegistrationParams* params, TF_Status* status);

// Get a set of node names that must be preserved. They can not be transformed
// or removed during the graph transformation. This includes feed and fetch
// nodes, keep_ops, init_ops. Fills in `num_values` and `storage_size`, they
// will be used in `TF_GetNodesToPreserveList`.