ModelSpecs are described in a `model_registry.json``
for local prototyping additional model specs can be defined in a model_registry_custom.json (not version controlled)
A Model is mainly retrieved by backends.get_model_for(model_spec) method which performs a unification operation with existing ModelSpecs from the model registry. The first unifying model spec is returned or the one given retained.
A ModelSpec must define a "backend" (name) to properly load a Model. The backend name must match a backend file like <name>_api.py.
Aside:
generation arguments (temp, max_tokens) are directly attached to Model (use model.set_gen_args(arg0=1,arg1=2) for prototyping)
Backends are now lazy loaded and the one to be used must be specified in the ModelSpec
adjusted all pre-defined bechmark games to newly introduced classes
Implements #26:
Backends now provideModelsModels are defined byModelSpecsModelSpecs are described in a `model_registry.json``model_registry_custom.json(not version controlled)A
Modelis mainly retrieved bybackends.get_model_for(model_spec)method which performs a unification operation with existingModelSpecs from the model registry. The first unifying model spec is returned or the one given retained.A
ModelSpecmust define a "backend" (name) to properly load aModel. The backend name must match a backend file like<name>_api.py.Aside:
Model(usemodel.set_gen_args(arg0=1,arg1=2)for prototyping)Backends are now lazy loaded and the one to be used must be specified in theModelSpec