llama : fix segfault from unknown model arch name

[compilade]

As described in #5795 (comment), when comparing the model arch name, if llm_arch_from_string gets to LLM_ARCH_UNKNOWN in the LLM_ARCH_NAMES map, a NULL pointer for the name is obtained. This case needs to be handled to avoid a segmentation fault.

To test that this works, load a model with a name that is not in LLM_ARCH_NAMES (an easy way is to do this is to change the names in LLM_ARCH_NAMES (e.g. changing "llama" to "lllama")).

With master, it segfaults. With this PR, it correctly errors with

llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'whatever-unknown-name-you-tried'

[slaren]

I think we need to stop using the operator[] here to avoid adding unknown architectures to the list:

llama.cpp/llama.cpp

Lines 364 to 366 in c2224f0

	std::string operator()(llm_kv kv) const {
	return ::format(LLM_KV_NAMES[kv], LLM_ARCH_NAMES[arch]);
	}

[cebtenzzre]

I think we need to stop using the operator[] here to avoid adding unknown architectures to the list:

Yep, this should use std::map::at instead.

[compilade]

Hmm, using std::map::at there instead errors for every model type (even known ones) with

llama_model_load: error loading model: map::at
llama_load_model_from_file: failed to load model

This happens because the operator() of llm_kv is first called with LLM_ARCH_UNKNOWN to get the architecture name, but it's not in LLM_ARCH_NAMES.

Maybe the way to fix this would be to define LLM_ARCH_UNKNOWN to have an empty string as a name?

[ggerganov]

And also make all LLM maps const

[cebtenzzre]

Maybe the way to fix this would be to define LLM_ARCH_UNKNOWN to have an empty string as a name?

I think the only acceptable name for LLM_ARCH_UNKNOWN is something like unknown or (unknown) - in case it ever gets printed. It shouldn't be a problem to use a non-empty string to substitute a placeholder that doesn't exist.