Are DSGEs always linearised for forecasting?

Are DSGE models always (or almost always) log-linearised when generating forecasts?

Yes, most of them are.