Currently switching models makes the TPU cry and we do not want to upset
our dearest hardware friend.
In the future it'd probably be best to teach the TPU how to not be
afraid of switching models, provided we have some limitations that
prevent loading itty bitty or way too big models on the TPU.
This helps people who clone the repository without knowing submodules need to be cloned, on the dependency installation it will now automatically check the submodules.
The maximum memory length was hardcoded to 512 tokens independent of
the overall context token budget. This was restrictive when the token
budget was increased, and excessive when the budget was decreased.
Based on feedback, this change modifies the max memory length to be 50%
of the context length. I tested with a memory consisting of the first
five paragraphs of The Three Musketeers, and the change seemed
reasonable with context lengths of 512, 1024 and 2048.
In order to minimize changes, this commit only modifies
max_memory_length; max_prompt_length and max_authornote_length are still
hardcoded as 512 tokens. We might want to make the max prompt length
flexible in the future, but since the author's note is injected into
the story, it's probably fine to restrict its length since a long
author's note will likely derail the AI.
There was an exception loading or importing a V1 story json into UI2 if
the story had zero actions. Do to incorrect indendation, the V1 load
process would attempt to use an uninitialized temp_story_class variable.
However, even after fixing this, it was impossible to continue playing
the saved story, because it was not possible to submit a blank action
when a story only had a prompt and zero actions. Modify the logic around
the first submission to handle a non-empty initial prompt. This allows
the player to begin playing the loaded game.
The .github folder was causing workflow errors for everyone who didn't configure this correctly. I have updated the docker with the latest ideas for flexibility. If a quick build is desirable people can build upon the official KoboldAI United image a seperate base image should not be needed.
The art guide now optionally supports the sequence <|>
If this exists in the art guide, <|> is replaced by the story summary.
Otherwise, the art guide is appended to the summary as before.
Update the default art guide to place the medium before the summary.
Also update the art guide tooltip to include the new default.
