Breaking Out of the 2048 Token Context Limit in Oobabooga
Blog post from RunPod
Oobabooga, a text-generation tool, has increased its context limit from 2048 to 8192 tokens, allowing users to maintain longer and more immersive interactions without losing significant entries. To utilize this expanded capacity, users must ensure they have the latest version installed, which requires enabling the trust-remote-code flag. Updating involves removing the existing installation and reinstalling the software, which may necessitate backing up text logs, models, and plugins. Users are advised to experiment with different models, as some can handle more than the previous limit of 2048 tokens, with examples like the WizardLM-7B-Landmark and Minotaur-13b-Landmark models supporting over 10,000 tokens using landmark attention. However, each model's capacity varies, and exceeding a model's token limit may lead to output issues. Combining this increased context with plugins like Long Term Memory, which use some tokens for invoking memories, can further enhance creative writing by offering additional context space.