//   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.
// * Devices are not in this API.  Instead, they are created/used internally

// `enable_xla_compilation` is non-zero, and OFF otherwise.
// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.
// c) ConfigProto.device_count is set to `num_cpu_devices`.
TF_CAPI_EXPORT extern TF_Buffer* TF_CreateConfig(
    unsigned char enable_xla_compilation, unsigned char gpu_memory_allow_growth,
    unsigned int num_cpu_devices);

//
// Gradient is the type returned by gradient functions. In Python TF it's either
// Tensor or IndexedSlices or None, which here we map to nullptr. Gradients need
// to allow their size to be computed and they need to be passable to a backward
// function and deleted (as the backprop code creates lots of gradients the user
// is not interested in).
//

// maintains the map from `op_id` to a `OpTapeEntry` which stores the `op_type`,
// inputs and outputs and the gradient function These data structures combined
// allow us to trace the data dependencies between operations and hence compute
// gradients.
//
// `ZerosLike` is not expected to be called and returns a nullptr. The creation

///   * any other change to these tables is only allowed on a major TensorFlow
///     version change (e.g., from 2.x to 3.0). This is provided as an escape
///     hatch to allow cleaning up these tables. Since any of these changes
///     break ABI compatibility and cause all plugins to be recompiled, these
///     type of changes should be extremely rare.
///