Right now all headers are in https://github.com/pytorch-labs/tokenizers/tree/main/include/pytorch/tokenizers. That probably includes internal ones. Users doesn't know what headers they are expected to use ("exported headers") in their own project. Say in linux, they put internal headers in src/ like https://github.com/torvalds/linux/blob/master/drivers/accel/qaic/mhi_controller.h, or have some subdir like https://github.com/torvalds/linux/tree/master/include/crypto/internal, whereas exported headers are in https://github.com/torvalds/linux/tree/master/include/crypto