https://cybersecurityerial.github.io/echo_blog/tags/communication/Recent content in Communication on Echo的技术博客Hugozh-cnWed, 17 Jun 2026 00:00:00 +0800https://cybersecurityerial.github.io/echo_blog/posts/llm-system-training-framework-notes-04-megatron-communicator-deadlock/Wed, 17 Jun 2026 00:00:00 +0800https://cybersecurityerial.github.io/echo_blog/posts/llm-system-training-framework-notes-04-megatron-communicator-deadlock/<blockquote> <p>本篇目标：</p> </blockquote> <h2 id="为什么通信器会死锁">为什么通信器会死锁</h2> <h2 id="megatron-的-p2p-通信抽象">Megatron 的 P2P 通信抽象</h2> <h2 id="batch-p2p-comm">batch p2p comm</h2> <h2 id="overlap-p2p-comm">overlap p2p comm</h2> <h2 id="warmup--steady--cooldown-里的通信顺序">warmup / steady / cooldown 里的通信顺序</h2> <h2 id="如何设计防死锁的通信接口">如何设计防死锁的通信接口</h2> <h2 id="todo">TODO</h2>