Accelerating Heterogeneous Tensor Parallelism via Flexible Workload Control

Abstract: Transformer-based models are becoming deeper and larger recently. For better scalability, an underlying training solution in industry is to split billions of parameters (tensors) into many tasks and then run them across homogeneous accelerators (e.g., GPUs). However, such dedicated compute cluster is prohibitively expensive in academia and moderate companies. An economic replacement is to aggregate existing heterogeneous devices and share resources among multi-tenants. Nevertheless, static hardware configurations and dynamic resource contention definitely cause straggling tasks, which heavily slows down the overall training efficiency. Existing works feature contributions mainly tailored for traditional data parallelism. They cannot work well for the new tensor parallelism due to strict communication and correctness constraints. In this paper we first present ZERO-resizing, a novel dynamic workload balancing technique without any data migration. We tune workloads in real-time by temporarily resizing matrices involved in core tensor-related computations. We particularly design data imputation and priority selection policies to respectively satisfy consistency constraint required by normal training and reduce the accuracy loss. We also give a lightweight data migration technique without loss of accuracy, to cope with heavy heterogeneity. Our final SEMI-migration solution is built on top of these two techniques and can adaptively distinguish their respective balancing missions, to achieve an overall success in efficiency and accuracy. Extensive experiments on the representative Colossal-AI platform validate the effectiveness of our proposals.