llama : fix --reverse-prompt crashing issue

[MollySophia]

Currently running llama-cli with --reverse-prompt option makes it crash in common_sampler_last() due to calling gsmpl->prev.rat(0) on an empty ring buffer, causing ring buffer: index out of bounds exception.

This fixes #14513 .

[MollySophia]

BTW I noticed that the --reverse-prompt option isn't enabled for LLAMA_EXAMPLE_SERVER. It's also not able to set in the webui. Is that an intended behavior? @ggerganov

[ggerganov]

I think we have "stop" strings that do the same in llama-server?

[MollySophia]

I think we have "stop" strings that do the same in llama-server?

Yeah I saw that it can be set in the completion api. I was wondering if the default value can be set on server startup.
After further reading, I found that the default value can be set with --props argument.
Thanks!

[MollySophia]

Hmmm nope, --props option enables properties modifying with API, rather than overrides certain properties' default values

[ggerganov]

I guess we can enable LLAMA_EXAMPLE_SERVER for the reverse prompt and propagate it to the parameter parsing logic to use as a default:

llama.cpp/tools/server/server.cpp

Lines 481 to 494 in 9220426

	{
	params.antiprompt.clear();
	const auto & stop = data.find("stop");
	if (stop != data.end() && stop->is_array()) {
	for (const auto & word : *stop) {
	if (!word.empty()) {
	params.antiprompt.push_back(word);
	}
	}
	}
	}