Keeping large training runs stable: AdamW and its memory cost, warmup and cosine schedules, bf16 and fp8 precision, loss-spike guardrails, ZeRO and FSDP.