feat(help): list all accepted llama.cpp options in --help

[RichardoC]

Description

Why this PR?

Currently, it's not possible to use just the cli to run a model from a gguf file. Yuo have to consult the upstream docs to work out the correct arguments to pass.

What is this PR?

lamafile --help (and --chat/--cli --help) printed a small, hand-written option list that was missing many flags llamafile actually accepts -- e.g. --top-k, --top-p, --temperature, --ctx-size, --repeat-penalty, -t, --port. Only --server --help was complete, because it fell through to llama.cpp's own usage printer.

Rather than maintain a second copy of the option list (which is what drifted), delegate the full catalogue to the same code path --server --help already uses: common_params_parse() with the llama_example the target mode feeds it at runtime (LLAMA_EXAMPLE_SERVER for server/auto, LLAMA_EXAMPLE_CLI for chat/cli).

Changes:

• main.cpp: add print_llama_cpp_usage() which calls common_params_parse() to print the categorized llama.cpp option list (common/sampling/spec/ example-specific) and exit(0) on --help. The print_*_help() functions now take a show_full_list_hint bool: the --help path prints the llamafile intro (modes + llamafile-specific flags) and appends the full llama.cpp list; the missing--m error path prints the intro plus a pointer to --help. --server --help is unchanged (still falls through to server_main). The duplicate SERVER/AUTO missing--m arms are collapsed (they printed identical text).
• tests/integration/tests/test_help.py: new model-free regression tests that invoke every --help variant and assert exit 0, the llamafile identity, the llamafile-specific flags, and that the previously-missing llama.cpp options (--top-k, --top-p, --temperature, --ctx-size, --repeat-penalty, -ngl) now appear. Also covers the missing--m error path.
• tests/integration/pyproject.toml: register the "help" test marker.

Because the full option list comes from llama.cpp's own parser, --help stays in sync with what llamafile accepts as upstream evolves -- no hand-maintained copy to drift again.

Verified: make check passes (extract_data_uris 19/19, fa_helpers 3/3, gpu_backend 23/23); new pytest tests/integration/tests/test_help.py --executable o//llamafile/llamafile passes 17/17; manual check of all four --help variants confirms --top-k/--top-p/--temperature/--ctx-size/ --repeat-penalty/-ngl now appear alongside the llamafile-specific flags.

Generated with pi 0.80.2 and GLM 5.2

PR Type

• 🆕 New Feature
• 📚 Documentation

Relevant issues

Checklist

• I understand the code I am submitting.
• I have run this code locally and verified the change.
• New and existing tests pass locally, or I have explained why tests were not run.
• Documentation was updated where necessary.
• If I changed code in llama.cpp/, whisper.cpp/, or stable-diffusion.cpp/, I also updated the matching *.patches/ files.
• I have read and followed the contribution guidelines.
• AI Usage:
  • No AI was used.
  • AI was used in an assistive capacity.
  • This PR includes substantial AI-generated content.

AI Usage Information

• AI Model used: GLM 5.2
• AI Developer Tool used: pi 0.80.2
• Any other info you'd like to share: Built in a scratch macos VM

When answering reviewer questions, please respond yourself rather than pasting reviewer comments into an AI system and posting the reply back unchanged.

• I am an AI Agent filling out this form (check box if true)

[RichardoC]

Turns out my actual issue was that the options aren't exposed by --chat, only --server and yes this PR fixes that

[aittalam]

Hi @RichardoC , and thank you for making the help system more clear!

I ran the tests and they seem to work both standalone (without provinding any model) and when the models are provided (see below). I think that's great, and I am not too worried about repeating the help tests as one can turn it off (-m "not help") for pre-built models and just run it once with the plain executable, or run it for each model as it is quite fast and does not add a big overhead.

Just a minor UX inconsistency: when one runs llamafile --server the help provides examples for all three "run llamafile --server --help", "run llamafile --cli --help", etc rather than just the --server example. I already have a fix for that (adding a server-specific help and activating it in the case section), would you mind testing it and letting me know what you think of it?

[aittalam]

Re-tested and merged, thanks again @RichardoC !

[RichardoC]

Thanks for sorting all this folks! Sorry about the delay in responses, got busy with work. I'm happy with the changes you folks have made