GGMM is an extension of the original GGML format, by prepending meta information to the Config structure.
The meta field is just a JSON string allowing non ASCII characters (json.dumps(..., , ensure_ascii=False)).
Ideally, with this meta information, Config structure can be eliminated. But at present, meta information is
not used yet. A use case will be customizing model name shown in banner:
{
"model_name": "...",
"model_native_name": "...",
}For example, python chatllm.py -m :aqualichat2:7b.
M is Succ('L') and the first letter of meta.
-
I don't like the idea of defining a kv data structure again with JSON already in-hand.
-
The "Universal" concept does not make sense to me.
GGMM is tightly couple with this inference: whenever this is a GGMM file, it will just works.