}

    // 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
    // key if this proves to be a significant performance bottleneck.
    absl::MutexLock lock(&mu_);
    if (LookupLocked(key, value)) {
      return TF_SetStatus(status, TF_OK, "");
    }

//
// Returns an error into `status` if:
//   * `output` is not in `graph`.
TF_CAPI_EXPORT extern int TF_GraphGetTensorNumDims(TF_Graph* graph,
                                                   TF_Output output,
                                                   TF_Status* status);

// Returns the shape of the Tensor referenced by `output` in `graph`

// parallel device creates a parallel tensor `x` with `a` on the first of
// `underlying_devices` and `b` on the second. Running `a_unpacked, b_unpacked =
// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//
// The filled `device` struct and the allocated `device_info` struct may be
// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.
void AllocateParallelDevice(const char* device_name,

// Disables only graph collection in RunMetadata on the functions executed from
// this context.
TF_CAPI_EXPORT extern void TFE_ContextDisableGraphCollection(TFE_Context* ctx);

// TODO(fishx): Move these monitoring APIs into a separate file.
// -----------------------------------------------------------------------------
// Monitoring Counter APIs.
// These APIs de-templated monitoring Counter for swig.

void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,
    std::array<TensorHandlePtr, num_replicas>* components, TF_Status* status);

// Helper to pack `num_replicas` TFE_TensorHandles into one parallel handle.
template <std::size_t num_replicas>
TensorHandlePtr CreatePerDeviceValues(
    TFE_Context* context,
    const std::array<TFE_TensorHandle*, num_replicas>& components,

  // ParallelDevice, and in turn runs one operation per component device with
  // its corresponding inputs from the input ParallelTensors. Wraps the
  // resulting per-device and per-output TFE_TensorHandles into one
  // ParallelTensor per output of the original operation.
  //
  // Attributes are forwarded to executed operations unmodified.
  //
  // The returned optional has a value if and only if `status` evaluates to

                                                       int tensor_id,
                                                       TF_Status* status);

// On success, enqueues `tensor` into a TF-managed FifoQueue given by
// `tensor_id`, associated with `session`. There must be a graph node named
// "fifo_queue_enqueue_<tensor_id>", to be executed by this API call. It reads

// upstream gradients.
//
// TODO(apassos) provide concrete template instantiations for TFE_TensorHandle
// specialization, which is blocked by quite a few things needing to loop back
// into python now.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class VSpace {
 public:
  virtual ~VSpace() {}

  // Returns the number of elements in the gradient tensor.

  // code that is layered on top of the runtime.
  virtual bool UsesTFRT() = 0;

  // List attributes of available devices
  virtual void ListDevices(std::vector<DeviceAttributes>* devices) = 0;

  // Add `devices` into context's device manager. Context's device manager
  // will take ownership and maintain devices' lifetime.
  virtual Status AddDevices(std::vector<std::unique_ptr<Device>> devices) = 0;

  // Returns a map from variable names to their shapes.  Slices of a partitioned
  // tensor are combined into a single entry.
  const TensorSliceReader::VarToShapeMap& GetVariableToShapeMap() const;

  // Returns a map from variable names to their data types.  Slices of a
  // partitioned tensor are combined into a single entry.
  const TensorSliceReader::VarToDataTypeMap& GetVariableToDataTypeMap() const;