CFP: Low-overhead Profiling-based Intra-operator Parallelism Generation by Preserving Communication-Free Structures

This paper introduces CFP, a system that search intra-operator parallelism configurations by leveraging runtime profiles of actual parallel programs. The key idea is to profile a limited space by identifying a new structure named ParallelBlock, which is a group of operators with the property of communication-free tensor partition propagation: the partition of its input tensor can propagate through all operators to its output tensor without introducing communication or synchronization. Based on this property, an optimal tensor partition of operators within a ParallelBlock should be inferred from the partition of input tensor through partition propagation to prevent the avoidable communication. Thus, the search space can be reduced by only profiling each ParallelBlock with different input tensor partitions at its entry, instead of enumerating all combinations among operators within the ParallelBlock. Moreover, the search space is further reduced by identifying ParallelBlock sequences (segments) with similar parallel behavior. CFP computes the overall performance of the model based on the profiles of all segments. On GPT, LLAMA, and MoE models, CFP achieves up to a 1.51x, 1.31x, and 3.43x speedup over the state-of-the-art framework, Alpa.