We currently have a complicated initialization process, with a lot happening in hydra.py The issue is that we need the number of channels to be able to instantiate the model. This is the historical reason for making a lot of instantiation steps manually (and using model_partial, ...) I see a way to get rid of these manual steps by first instantiating the tokenization/embedding pipelines and passing a cached sample, then updating the config to perform an automatic instantiation. This would result in: * simpler config files * simpler hydra.py * no config_partial, model_partial, ...