TF_Status* status);

// Fetches the current number of inputs attached to `op`.
//
// Does not use the operation's definition to determine how many inputs should
// be attached. It is intended for use with TFE_OpGetFlatInput to inspect an
// already-finalized operation.
//
// Note that TFE_OpGetFlatInputCount and TFE_OpGetFlatInput operate on a flat

            "issue at https://github.com/tensorflow/tensorflow/issues.");
        return;
      }
      // We have a convoluted scheme here: Using the C++ graph construction API
      // to add potentially many nodes to the graph without running the checks
      // (such as uniqueness of the names of nodes) we run with other functions
      // that add a node to the graph (like TF_FinishOperation).

  // Method to delete a device.
  void (*delete_device)(void* device_info);

  // Implements TFE_CreatePackedTensorHandle when one of `handles` is on this
  // custom device.
  //
  // Many devices will want to simply return an "unimplemented" status
  // here. This is the default behavior if `pack` is null when passed to
  // TFE_RegisterCustomDevice.

*   Various performance improvements, especially for RNNs and convolutions
*   Many bugfixes
*   Nightly builds, tutorial tests, many test improvements
*   New examples: transfer learning and deepdream ipython notebook
*   Added tutorials, many documentation fixes.

## Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

### Resource allocation

A denial of service caused by one model could bring down the entire server, but
we don't consider this as a vulnerability, given that models can exhaust
resources in many different ways and solutions exist to prevent this from
happening (e.g., rate limits, ACLs, monitors to restart broken servers).

### Model sharing

Status RegisterFilesystemPluginImpl(const TF_FilesystemPluginInfo* info) {
  TF_RETURN_IF_ERROR(ValidatePluginMemoryRoutines(info));

  // Validate and register all filesystems
  // Try to register as many filesystems as possible.
  // Free memory once we no longer need it
  Status status;
  for (int i = 0; i < info->num_schemes; i++) {
    status.Update(ValidateAndRegisterFilesystems(info, i));

namespace tf_writable_file {
typedef struct GCSFile {
  const std::string bucket;
  const std::string object;
  gcs::Client* gcs_client;  // not owned
  TempFile outfile;
  bool sync_need;
  // `offset` tells us how many bytes of this file are already uploaded to
  // server. If `offset == -1`, we always upload the entire temporary file.
  int64_t offset;
} GCSFile;

        "//tensorflow/cc:client_session",
        "//tensorflow/cc:scope",
        "//tensorflow/core:tensorflow",
        "//tensorflow/core/data:standalone",
        # Exports for pywrap_tensorflow_internal. Many of these are transitive
        # depedencies of the above, but must be explicitly listed for
        # cc_shared_library to work.
        "//tensorflow/c/eager:c_api_experimental",

machine learning system that operates at large scale and in heterogeneous environments. TensorFlow uses dataflow graphs to represent computation, shared state, and the operations that mutate that state. It maps the nodes of a dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, general purpose GPUs, and custom-designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility to the application...

//  - target: dy in dx/dy
//  - source: dx in dx/dy
//  - tensor: one of the many inputs or outputs of an operation
//
// 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