Normal QAD
Deployment-informed
Identical token prefix
Frozen teacher
Original BF16 model
Quantizers disabled
Original weights
Training actor
QDQ student
Quantized forward
High-precision backward
Rollout engine
Native FP4 student
Used after training
Preserved scales
Sync weights
Distillation loss
Teacher ↔ training actor
Backprop through actor
Teacher probs
Student probs
Student probs