model: don't crash on unsupported architecture

[giladgd]

Overview

Before #22004, when trying to load a model with an unsupported architecture using llama_model_load_from_file it used to just fail silently and return nullptr instead of crashing.
This PR restores the old behavior.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: No

[giladgd]

@ggml-bot I'm not a new contributor

[am17an]

@giladgd the bot looks at the last 90 days for this check for performance reasons. Sorry about the noise

[CISC]

Before #22004, when trying to load a model with an unsupported architecture using llama_model_load_from_file it used to just fail silently and return nullptr instead of crashing. This PR restores the old behavior.

You lost me?

[giladgd]

@CISC In the past, when trying to load a model with an unsupported architecture (for example, loading an mmproj using llama_model_load_from_file) it’ll just fail - calling llama_model_load_from_file with such a model will log an error, and return nullptr.
PR #22004 (inadvertently) changed that behavior, so now trying to load an unsupported model using llama_model_load_from_file reaches an GGML_ABORT and so the process crashes.

This PR fix that, so that trying to load an unsupported model will just fail (like before) and won’t crash the process.

[CISC]

Ah, I see, because llama_model_load catches the runtime error...