Add a CLI option in main.cpp in order to stop generation at newline token

[WilliamTambellini]

Prerequisites

Please answer the following questions for yourself before submitting an issue.

• [X ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
• [X ] I carefully followed the README.md.
• [X ] I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• [X ] I reviewed the Discussions, and have a new bug or useful enhancement to share.

Feature Description

Just add a CLI option in main.cpp in order to stop generation at newline token

Motivation

For some users like us, the goal is not to chat/have a discussion with a LLM, just to get a single line of output.

Possible Implementation

in main.cpp: eg:

        if (params.stop_at_nl && output_tokens.size() > 0 && !embd.empty() && embd.back() == llama_token_nl(model)) {
            break; // only generate 1 single output line
        }

[Jeximo]

-e -r "\n" on CLI should stop generation at a newline.

-e, --escape ... process prompt escapes sequences (\n, \r, \t, \', \", \\)

llama.cpp/examples/main/README.md

Line 104 in f87f7b8

- `-r PROMPT, --reverse-prompt PROMPT`: Specify one or multiple reverse prompts to pause text generation and switch to interactive mode. For example, `-r "User:"` can be used to jump back into the conversation whenever it's the user's turn to speak. This helps create a more interactive and conversational experience. However, the reverse prompt doesn't work when it ends with a space.

[HanClinto]

I think that @Jeximo has the best solution, but an alternate approach would be to utilize grammars.

--grammar '[^\n]+'

This should force the LLM to generate anything except for newline characters, keeping everything on a single line.

[WilliamTambellini]

Tks @Jeximo but sadly that does not do exactly what I need: a single line non-empty output:
-e -r "\n" does stop at the newline char but does not enforce the output to be non empty. That s why I added "output_tokens.size() > 0" in my implementation.
Now, perhaps "-e -r \n" could do the job if there is a way to force main to generate a non empty output. Is there ?

Tks @HanClinto but "--grammar '[^\n]+'" is not the desired feature: I still want to authorize the llm to 'softmax' a newline char but main to stop generating when the LLM advices a newline, not to continue.

[HanClinto]

Now, perhaps "-e -r \n" could do the job if there is a way to force main to generate a non empty output. Is there ?

For this, yes -- you can use grammars to require at least one non-newline character before generating a newline.

This worked for me:

./main -m ./models/llama-2-7b.Q4_0.gguf -e -r "\n" --grammar "root ::= [^\n][^\x00]*" -p "My favorite flavor is "

This grammar requires at least one non-newline character, followed one or more of any other character. The reason why I used [^\x00] instead of . is that it seems like this symbol might not be implemented in the grammar parser -- I intend to look into this -- it might simply be my own issue. The intended grammar (if I could get "." to work) would be:
root ::= [^\n].*

Hopefully that helps you visualize what the chosen grammar is doing.

I realize this isn't as clean as your suggested solution -- it's always a tough balance of too much clutter in the command-line options, vs. making it easier to do common tasks. I think if enough other people found this to be a common use-case, then the project might want to support it as a first-class command-line option.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.