Open Pelochus opened 2 months ago
Initial point of reference:
https://gist.github.com/av1d/ce58ef738902ec0365fe828720be31e5
Important link to this issue:
https://www.reddit.com/r/RockchipNPU/comments/1cpngku/rknnllm_v101_lets_talk_about_converting_and/
I could also add an option to not use files, use parameters or something. Perhaps reconverting everything with a modified tokenizer.json could be also a good option, for having the same template for every converted model
We anticipated better with an update. Instead we get useless.
Essentially changing the LLM template dynamically at program execution. Add a parameter that reads a file and changes the corresponding variables. Especially useful if using various different LLMs. This needs some research first, as I know pretty much nothing about the templates.