ggml_status introduction

[Xarbirus]

I would like to add the ability to explicitly specify the execution result for the ggml_backend_graph_plan_compute and ggml_backend_graph_compute functions.

This will allow us to determine whether the execution was aborted by abort_callback or by some other reasons (and allow us to use different reasons). I used the additional ggml_compute_result_t type because when executing ggml_*_compute multi-threaded, each thread may return its own error/warning, and in order not to lose them, I decided to concatenate exit_code's using OR.

Also I didn't remove GGML_EXIT_SUCCESS and GGML_EXIT_ABORTED because they can be used by other party libraries that use ggml.

[slaren]

I think this is a good change. Some comments:

• ggml_compute_exit_code does not need to be a bit field, in practice only one error can happen, and even if in the future there is some bizarre condition that may cause different errors in different threads, I think it is still better to report one error only to keep the API simpler
• I would prefer a more generic name that can be reused in other ggml APIs in the future, such as ggml_error or ggml_status
• Add a error_to_string function, otherwise every app will start implementing their own

[ggerganov]

+1 for ggml_status

[Xarbirus]

Thank you for your review, made changes:

• renamed new type from ggml_compute_exit_code to ggml_status
• added ggml_status_to_string (to cast a status to a string)
• updated statuses, now negative values ​​mean errors and positive values ​​mean warnings

And I get an important question. The llama_decode function has its own error codes, for example -1 is used here and 1 is used here. Do I need to add all possible errors from llama.cpp to ggml_status or do I need to create a separate enum in llama.cpp that will expand and complement ggml_status later (after syncing ggml with llama.cpp)?

[slaren]

My opinion is that llama.cpp needs its own llama_status with its own error codes, since llama.cpp has its own error conditions that are completely independent and separate of ggml.

[Xarbirus]

@slaren , I completely agree with you!