Pipeline parallelism, a crucial technique for large-model training, is hindered by runtime variability in computation and communication. Existing systems rely on static or adaptively generated schedules, which can lead to inefficiencies when task readiness diverges from the pre-committed order. A new approach introduces a readiness-driven runtime that adapts to runtime variability, enabling more efficient pipeline-parallel training. This runtime dynamically adjusts the execution order based on task readiness, reducing waiting times and improving overall performance. By doing so, it addresses the limitations of traditional pipeline systems, which can be hindered by static schedules1. The implications of this research extend beyond the realm of machine learning, as efficient pipeline parallelism can have significant impacts on various fields that rely on large-scale computing. This breakthrough matters to practitioners because it can significantly improve the scalability and efficiency of large-model training, ultimately leading to faster and more accurate model development.