1. Created `LongNetTokenizer` class: This class handles tokenization of both text and images for a multimodal model. It relies on the CLIPProcessor for image tokenization and GPT-NeoX's tokenizer for text tokenization.
2. Refactored `LongNet` class: The original `LongNet` class was refactored to use the Perceiver Resampler, a model that resamples the outputs of the CLIP model to a lower dimension, and combines it with the output of a Transformer-based decoder.
3. Created `LongNetLanguageOnly` class: This is a new class for a language-only model that doesn't consider image input. It leverages the same Transformer-based decoder as the multimodal model but only takes text input.
4. Created `LongNetSelector` class: This class is essentially a factory that allows the user to select and initialize a LongNet model of a specific type ('multimodal' or 'language_only') easily.
This way, the original code is better structured, and the usage of different versions of LongNet is more intuitive. Users can now choose which LongNet model they want to use and easily switch between them. These changes enhance the flexibility and adaptability of the code.