https://lilinji.github.io/tags/cudastream/Recent content in CUDAStream on Ringi's LogHugo -- 0.152.2zh-cnMon, 31 Aug 2026 00:00:00 +0800https://lilinji.github.io/2026/08/%E7%AC%AC07%E8%AE%B2pytorch-%E5%BC%82%E6%AD%A5%E6%89%A7%E8%A1%8C%E6%B5%81cuda-stream-%E4%B8%8E-cuda-graph-%E5%8E%9F%E7%90%86/Mon, 31 Aug 2026 00:00:00 +0800https://lilinji.github.io/2026/08/%E7%AC%AC07%E8%AE%B2pytorch-%E5%BC%82%E6%AD%A5%E6%89%A7%E8%A1%8C%E6%B5%81cuda-stream-%E4%B8%8E-cuda-graph-%E5%8E%9F%E7%90%86/深入剖析 GPU 异步执行第一性原理：Host-Device 生产者消费者队列、Default Stream 与多流并发调度、Event 流间同步屏障、record_stream 显存踩踏暗礁、CPU Launch Overhead 瓶颈剖析，以及 CUDA Graph 录制回放与 vLLM 多桶图优化工业级系统实战。