This sounds like an inference time optimization so we can use it without retraining the model? Any idea when this might hit Forge?