[main] fix infinite generation (-n == -1)

[anzz1]

Fixes the command line option -n / --n_predict value -1 (generate indefinitely) in interactive mode.

[Green-Sky]

just wanted to test it and just throw a large file at it:
main: error: prompt is too long (335687 tokens, max 2044)

... i suppose thats ok, gotta pipe it to stdin then

edit: nvm, i did not think, i can just let it dream

[Green-Sky]

I let it dream, but it ended itself after 516 tokens. I think it works.

[anzz1]

I let it dream, but it ended itself after 516 tokens. I think it works.

if you want to continue even when it think it should stop, u also need the '--ignore-eos' option to force it continue indefinitely. this is just a fix for 79b2b26 and n==-1 makes it not stop when running out of tokens. end of stream is different thing. with -n -1 --ignore-eos it should continue forever unless something goes wrong.

[rabidcopy]

I let it dream, but it ended itself after 516 tokens. I think it works.

Did you do this with interactive or a reverse prompt? Currently EOS is only ignored with --ignore-eos or by being in interactive mode without a reverse prompt. (EOS is still technically ignored with a reverse prompt but it'll hand back the reverse prompt and give control on EOS) Otherwise just passing a prompt will still hit EOS and break.

[Green-Sky]

hehe, i know guys. I just forgot to enable it. 😄