AI 导读
第一条 DSpark trace 有 47% 处于空闲：在 10.62 ms 的一步中有 4.99 ms 空闲。 FlashInfer plan() 为获取 CPU 早已知道的元数据，进行了阻塞式的设备到主机读取。改用主机数据构建 plan 后，去掉了同步，让 CPU 能在 GPU 需要之前就排队 draft 和 verify。
正文
The first DSpark trace was 47% idle: 4.99 ms of idle in a 10.62 ms step.
FlashInfer plan() made blocking device to host reads for metadata the CPU already knew. Building plans from host data removed the sync and let the CPU queue draft and verify before the GPU needed them.
来源：@AntLingAGI · x.com