An Efficient 2D Method for Training Super-Large Deep Learning Models
This work addresses the problem of efficiently training large-scale models for AI researchers and practitioners, representing an incremental improvement over existing model parallelism methods.
The paper tackles the memory constraints in training super-large deep learning models by proposing Optimus, a 2D-partition paradigm for model parallelism that partitions activations to reduce redundancy, resulting in up to 1.78X speedup in inference and 8X increase in maximum batch size compared to Megatron.
Huge neural network models have shown unprecedented performance in real-world applications. However, due to memory constraints, model parallelism must be utilized to host large models that would otherwise not fit into the memory of a single device. Previous methods like Megatron partition the parameters of the entire model among multiple devices, while each device has to accommodate the redundant activations in forward and backward pass. In this work, we propose Optimus, a highly efficient and scalable 2D-partition paradigm of model parallelism that would facilitate the training of infinitely large language models. In Optimus, activations are partitioned and distributed among devices, further reducing redundancy. In terms of isoefficiency, Optimus significantly outperforms Megatron. On 64 GPUs of TACC Frontera, Optimus achieves 1.48X speedup for training, 1.78X speedup for inference, and 8X increase in maximum batch size over Megatron. Optimus surpasses Megatron in scaling efficiency by a great margin. The code is available at https://github.com/xuqifan897/Optimus.