TL;DR of the paper, generated automatically.

Distillation tolerates substantial rollout mismatch, and FP8 helps efficiency.

• Tolerance: a 7B student keeps quality at over 200x BF16 mismatch.
• Memory: FP8 KV cache holds at least 1.93x BF16's tokens.
• Speed: FP8 with async rollouts hits the same quality 1.62x sooner.