None defined yet.
Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss