Method initialize

verl/models/mcore/model_initializer.py:141–147

(self, **kwargs)

        return transformer_layer_spec

    def initialize(self, **kwargs):
        model = super().initialize(**kwargs)
        freeze_moe_router = kwargs.get("freeze_moe_router", False)
        if freeze_moe_router:
            for layer in model.decoder.layers:
                layer.mlp.router.weight.requires_grad = False
        return model


class Qwen3MoEModel(BaseModelInitializer):