llama: max ctx by default, fix fit magic number

[JohannesGaessler]

Fixes #18376 .

The intent with llama_params_fit is to only change those parameters that a user is not setting explicitly themself. However, if a user is explicitly setting -c 0 this is currently being changed by llama_params_fit. I would propose the following changes to fix this:

• Change the data type of llama_context_params::n_ctx from uint32_t to int32_t with all values <= 0 being interpreted to mean that the full context should be used.
• Change the default value set llama_context_default_params from 512 to -1. This would mean with the llama C API by default the full context of the model would now be used, consistent with the common API.
• In llama_params_fit, check the context size against the new default value instead of vs. 0.
• For the CLI, add strings "auto" and "full" for --ctx-size which by default both use the full context but differ w.r.t. whether they should be changed by llama_params_fit.

An alternative approach would be keep the changes entirely in the common API.

[CISC]

Hmmm, what's with the sign disparity between common and llama params?

[JohannesGaessler]

I think the biggest reason is simply that no one invested the effort to make things consistent. I would suggest that going forward, if we have non-negative quantities like the context size we store them internally as unsigned values but use signed values in llama_model_params and llama_context_params in order to encode special values.

[ggerganov]

An alternative approach would be keep the changes entirely in the common API.

This seems like the better approach.

[ggerganov]

Hm, I don't see why this change is needed. Looks like it allows negative values that have the same functionality as n_ctx == 0. So the user code can simply used params.n_ctx = max(0, n_ctx); to achieve the same.

[JohannesGaessler]

Superseded by #19070 .