Replace t5xxl with other LLM ?

#64
by thesby - opened

Is it possible to replace t5xxl model with other LLM or MLLM, such as gemma or qwen-vl ?

yes but costly
maybe after v50 i'll experiment on it

What about that flan t5, and unchauned t5? They are uncensored compared to original t5xxl

Gemma and Qwen support multilingual and multi-modal.

What about that flan t5, and unchauned t5? They are uncensored compared to original t5xxl

t5xxl_flan_latest.safetensors works perfectly in Forge

yeah.. replace then encoder mid training ...great idea πŸ˜‚

t5 based replacements work flawlessly, flan, uncensored, abliterated. Other llms might need a bit more work as the header is different... might be something for later as Lodestone said. Llama works fine on HiDream and has nice variants just for the TE, just as a suggestion. Complete llms might be a bit overkill when only the text encoding is needed...

Sign up or log in to comment