I suggest using 'form' instead of 'shape'; the latter is mainly concerned with external form. In context of LLMs, form would be the internal mapping, and shape

I suggest using 'form' instead of 'shape'; the latter is mainly concerned with external form. In context of LLMs, form would be the internal mapping, and shape the decoded text that is emitted.