// 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.
void TestRemoteExecuteSilentCopies(bool async, bool remote, bool func,
                                   bool heavy_load_on_streaming_rpc,

#define TENSORFLOW_C_EXPERIMENTAL_FILESYSTEM_PLUGINS_GCS_GCS_HELPER_H_

#include <fstream>
#include <string>

class TempFile : public std::fstream {
 public:
  // We should specify openmode each time we call TempFile.
  TempFile(const std::string& temp_file_name, std::ios::openmode mode);
  TempFile(TempFile&& rhs);
  ~TempFile() override;
  const std::string getName() const;
  bool truncate();

// Core APIs
// -----------------------------------------------------------------------------

// A TF_ExecutionContext stores knowledge about how to execute an operation.
// E.g. it could know whether we're in eager mode or graph mode, keeps track
// of gradient tapes, etc.
typedef struct TF_ExecutionContext TF_ExecutionContext;

// A TF_AbstractTensor is an input to an operation. E.g. it could be a union

                       const ComputeFunc& compute_func, TF_Status* status) {
    if (max_age_ == 0) {
      return compute_func(key, value, status);
    }

    // Note: we hold onto mu_ for the rest of this function. In practice, this
    // is okay, as stat requests are typically fast, and concurrent requests are
    // often for the same file. Future work can split this up into one lock per

//       `TP_Optimizer`, `TP_OptimizerRegistrationParams`
//     * Struct that should be filled by the proper: `TF_GrapplerItem`,
//       `TF_GraphProperties`, `TF_FunctionLibraryDefinition`
//   * We use `struct_size` for version checking. It should be set both by
//     core and the plugin.
//     * For example, `TF_InitGraph` function receives
//       `TP_OptimizerRegistrationParams*` as input with `struct_size`

// APIs for generically dealing with op attributes (e.g. when forwarding them
// through custom device implementations).
//
// TODO(allenl): Currently these are black boxes, but we should have some way to
// inspect values. This would let people e.g. copy over most attributes and then
// modify some based on their values.

// A reference to an op's name -> attribute mapping

//
// Below here we do the gradient algorithm. It works as follows:
//
// First we filter the tape to just the subset of operations we want to
// differentiate. In the process of doing so we count how many times each Tensor
// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output

// that provides the filesystem functionality. However, the POSIX filesystem
// needs to be statically registered in some tests and utilities for building
// the API files at the time of creating the pip package. Hence, we need to
// expose this function so that this filesystem can be statically registered
// when needed.
void TF_InitPlugin(TF_FilesystemPluginInfo* info);

//   like shadowing a numpy array.
// * We do not provide TF_OK since it is not strictly necessary and we
//   are not optimizing for convenience.
// * We make assumption that one session has one graph.  This should be
//   fine since we have the ability to run sub-graphs.
// * We could allow NULL for some arguments (e.g., NULL options arg).
//   However since convenience is not a primary goal, we don't do this.

// not be deleted. The returned error status represents the reason that any one
// of these entries could not be deleted.
//
// Typical status codes:
//  * TF_OK - dirname exists and we were able to delete everything underneath
//  * TF_NOT_FOUND - dirname doesn't exist
//  * TF_PERMISSION_DENIED - dirname or some descendant is not writable
//  * TF_UNIMPLEMENTED - some underlying functions (like Delete) are not