Imported from LJX2017/personal-wiki (inbox/AGENTS.md). Install upstream with npx skills add LJX2017/personal-wiki --skill inbox. Copyright stays with the author.

Working in the inbox

The inbox optimizes for low-friction capture, not tidiness.

• Preserve the user's original wording. Fix obvious transcription errors only when the intended wording is unambiguous.
• Do not force frontmatter, tags, or a formal template onto rough captures.
• When asked to process a capture, identify the smallest useful destination: journal/ for a dated record, notes/ for a durable authored idea, or wiki/ for derived synthesis.
• Preserve the capture's substance when moving it. Do not delete uncertain fragments; leave them in place or mark them for the user's review.
• Add links only when they are genuinely useful.

谁的通讯发起的慢？谁的通讯结束的慢？发现慢节点 + hang 节点

checkpoint -> 1T model 14T optimizer 存一次checkpoint压力很大，频繁存会有网络带宽+容量压力

显存->内存->异步写入HDFS，不阻塞训练 恢复：每DP组一个

存储压力？按需把checkpoint删掉，有些为了快速恢复，可以直接删除

ByteScale: 长度变长->attn占比 + DP负载不均 DP CP结合，动态划分CP？