proud cliff
heavily based on the ChatGPT node. Asks a local oobabooga server to generate a more complex (hopefully better) prompt based on what was entered in the text box.
Cave-At : oobabooga must be started with the --api argument, and you need to make sure both InvokeAI's and Oobabooga's models have enough VRAM. This was successfully tested on a RTX3090/24GB using a 13B-GPTQ model.
proud cliff
Fixed the default context prompt and temperature.
just as an example :
photography of a beautiful woman
with TheBlock_Nous-Hermes-Llama2-GPTQ resulted in the generated prompt :
AI: Here is an image of a stunning woman posing gracefully against a serene landscape backdrop. Her features are delicate and her smile radiant, capturing the essence of beauty in its purest form. The light gently dances across her skin, accentuating every curve and contour, while the vibrant colors of nature surround her, creating a breathtaking contrast that will leave you mesmerized.
And that's the result with protogenNovaExperimental_X80 (steps: 50 , CFG:7, 512x512)
proud cliff
The default p and k values were negating the effect of the temperature. Also slightly better default system prompt.
wide halo
Would I be able to load/utilize a 30b local model with a 4090 (single) system and use this node feature?
proud cliff
Would I be able to load/utilize a 30b local model with a 4090 (single) system an...
not too sure to be honest. So far I only tried 13B models (with a 3090 and 24GB VRAM)
wide halo
Where am I supposed to place this Oobabooga.py file and the oobabooga-with-HDFix.json file at?
strange echo
X:\example_install_dir\invokeai.venv\Lib\site-packages\invokeai\app\invocations\ for the .py and the user interface to load the graph with the .json.
fast inlet
This is great, you should do the GGML models on kobold.cpp so people can use them and use stable diffusion/invoke at the same time
@proud cliff If you haven't tried them yet, you've got to, they're the fastest thing in LLMs with the least resources required, and the results are astoundingly good. https://github.com/LostRuins/koboldcpp
