// in the memory of the device name 'device_name'.
// If source and destination are the same device, then this creates a new handle
// that shares the underlying buffer. Otherwise, it currently requires at least
// one of the source or destination devices to be CPU (i.e., for the source or
// destination tensor to be placed in host memory).
// If async execution is enabled, the copy may be enqueued and the call will

// nodes, keep_ops, init_ops. Fills in `values` and `lengths`, each of which
// must point to an array of length at least `num_values`.
//
// The elements of values will point to addresses in `storage` which must be at
// least `storage_size` bytes in length.  `num_values` and `storage` can be
// obtained from TF_GetNodesToPreserveSize
//

// `values` and `lengths`, each of which must point to an array of length at
// least `max_values`.
//
// The elements of values will point to addresses in `storage` which must be at
// least `storage_size` bytes in length.  Ideally, max_values would be set to
// TF_AttrMetadata.list_size and `storage` would be at least
// TF_AttrMetadata.total_size, obtained from TF_OperationGetAttrMetadata(oper,
// attr_name).
//

//
// This serves as a factory for creating `AbstractOperation`s and for
// registering traced functions.
// Operations creation within a context can only be executed in that context
// (for now at least).
// Implementations of the context may contain some state e.g. an execution
// environment, a traced representation etc.
class AbstractContext {
 protected:

  // watched inputs.
  void Watch(int64_t tensor_id);

  // Records an operation with inputs `input_tensor_id` and outputs
  // `output_tensors` on the tape and marks all its outputs as watched if at
  // least one input of the op is watched and has trainable dtype.
  //
  // op_type is used to decide which of the incoming gradients can be left as
  // nullptr instead of building zeros when build_default_zeros_grads == true.

  ///    the file contents do not extend past `offset` and thus nothing was
  ///    placed in `out`.
  /// 4) `TF_OK` otherwise (i.e. the read succeeded, and at least one byte was
  /// placed
  ///    in `buffer`).
  ///
  /// Caller is responsible for allocating memory for `buffer`.
  /// `buffer` will be left unchanged in case of errors.

  // watched inputs.
  void Watch(const AbstractTensorHandle*);
  // Records an operation with given inputs and outputs
  // on the tape and marks all its outputs as watched if at
  // least one input of the op is watched and has a trainable dtype.
  // op_name is optional and is used for debugging only.
  void RecordOperation(absl::Span<AbstractTensorHandle* const> inputs,

  void (*cleanup)(TF_RandomAccessFile* file);

  /// Reads up to `n` bytes from `*file` starting at `offset`.
  ///
  /// The output is in `buffer`, core TensorFlow owns the buffer and guarantees
  /// that at least `n` bytes are available.
  ///
  /// Returns number of bytes read or -1 in case of error. Because of this
  /// constraint and the fact that `ssize_t` is not defined in `stdint.h`/C++