espnet2.speechlm.model.speechlm.lm.parallel.configure_moe_model

source

espnet2.speechlm.model.speechlm.lm.parallel.configure_moe_model(model)

Use Transformers’ grouped MM experts and SpeechLM’s router loss.

Configure both pretrained models and PP replicas before wrapping them. The native experts read this shared config on every forward call.