training-instability

Proposes mechanism-driven monitors for preemptive detection of LLM training instability by deriving internal signals from low-precision flash attention and MoE routers, enabling detection thousands of steps before loss divergence.