* add api/static/* to .githignore
* add init_prompt to params so that we can set n_ctx to len(init_prompt) + params.n_ctx
* change get_prompt so it properly builds a instruction prompt from the recent history
- keeps the prompt below n_ctx
- goes backwards and adds only the most recent sentences
- always prepends the init_prompt
* ignore eventSource.onerror
I don't know why it does this, but this gets called locally for me and messes up the history, just doing nothing here works fine for me
* fix split on ! not *
* run black formatter on stream.py
* revert previous black format, just do the one change it wants manually
* third times the charm
---------
Co-authored-by: Juan Calderon-Perez <835733+gaby@users.noreply.github.com>
- Added convert script, thanks to @eiz, will catch unconverted .bin files and convert them on startup.
- Switched back to main branch of llama.cpp
- Got rid of code to handle magic.dat
