Dynamic sequences with varying lengths have been widely used in the training of Transformer-based large language models (LLMs). However, current training frameworks adopt a pre-defined static parallel strategy for these sequences, causing neither communication-parallelization cancellation on short sequences nor out-of-memory on long sequences. To mitigate these issues, we propose ParaDySe, a novel adaptive Parallel strategy switching framework for Dynamic Sequences. ParaDySe enables on-the-fly optimal strategy adoption according to the immediate input sequence. It first implements the modular function libraries for parallel strategies with unified tensor layout specifications, and then builds sequence-aware memory and time cost models with hybrid methods. Guided by cost models, ParaDySe selects optimal layer-wise strategies for dynamic sequences via an efficient heuristic algorithm. By integrating these techniques together, ParaDySe achieves seamless hot-switching of optimal strategies through its well-designed function libraries. We compare ParaDySe with baselines on representative LLMs under datasets with sequence lengths up to 624K. Experimental results indicate that ParaDySe addresses OOM and CPC bottlenecks in LLM training by systematically integrating long-sequence optimizations with existing frameworks.

Transformer-based models like large language models (LLMs) have attracted significant attention in recent years due to their superior performance. A long sequence of input tokens is essential for industrial LLMs to provide better user services. However, memory consumption increases quadratically with the increase of sequence length, posing challenges for scaling up long-sequence training. Current parallelism methods produce duplicated tensors during execution, leaving space for improving memory efficiency. Additionally, tensor parallelism (TP) cannot achieve effective overlap between computation and communication. To solve these weaknesses, we propose a general parallelism method called memory-efficient tensor parallelism (METP), designed for the computation of two consecutive matrix multiplications and a possible function between them ( O = f ( AB ) C ), which is the kernel computation component in Transformer training. METP distributes subtasks of computing O to multiple devices and uses send/recv instead of collective communication to exchange submatrices for finishing the computation, avoiding producing duplicated tensors. We also apply the double buffering technique to achieve better overlap between computation and communication. We present the theoretical condition of full overlap to help instruct the long-sequence training of Transformers. Suppose the parallel degree is p ; through theoretical analysis, we prove that METP provides O (1/ p 3 ) memory overhead when not using FlashAttention to compute attention and could save at least 41.7% memory compared to TP when using FlashAttention to compute multi-head self-attention. Our experimental results demonstrate that METP can increase the sequence length by 2.38–2.99 times compared to other methods when using eight A100 graphics processing units (GPUs).

Large-scale deep learning models are trained distributedly due to memory and computing resource limitations. Few existing strategy generation approaches take optimal memory minimization as the objective. To fill in this gap, we propose a novel algorithm that generates optimal parallelism strategies with the constraint of minimal memory redundancy. We propose a novel redundant memory cost model to calculate the memory overhead of each operator in a given parallel strategy. To generate the optimal parallelism strategy, we formulate the parallelism strategy search problem into an integer linear programming problem and use an efficient solver to find minimal-memory intra-operator parallelism strategies. Furthermore, the proposed algorithm has been extended and implemented in a multi-dimensional parallel training framework and is characterized by high throughput and minimal memory redundancy. Experimental results demonstrate that our approach achieves memory savings of up to 67% compared to the latest Megatron-LM strategies; in contrast, the gap between the throughput of our approach and its counterparts is not large.

Large-scale models have gained significant attention in a wide range of fields, such as computer vision and natural language processing, due to their effectiveness across various applications. However, a notable hurdle in training these large-scale models is the limited memory capacity of graphics processing units (GPUs). In this paper, we present a comprehensive survey focused on training large-scale models with limited GPU memory. The exploration commences by scrutinizing the factors that contribute to the consumption of GPU memory during the training process, namely model parameters, model states, and model activations. Following this analysis, we present an in-depth overview of the relevant research work that addresses these aspects individually. Finally, the paper concludes by presenting an outlook on the future of memory optimization in training large-scale language models, emphasizing the necessity for continued research and innovation in this area. This survey serves as a valuable resource for researchers and practitioners keen on comprehending the challenges and advancements in training large-scale language models with limited GPU memory.

Transformer models, such as BERT, GPT, and ViT, have been applied to a wide range of areas in recent years, due to their efficacy. In order to improve the training efficiency of Transformer models, different distributed training approaches have been proposed, like Megatron-LM [8]. However, when multi-dimensional parallelism strategies are considered, due to the complexity, existing works can not harmonize the different strategies well enough to obtain a globally optimal solution. In this paper, we propose a parallelism strategy searching algorithm PTIP, which generates operator-level parallelism strategies consisting of three schemes: data parallelism, tensor parallelism, and pipeline parallelism. PTIP abstracts these three parallelism schemes simultaneously into an auxiliary graph, reformulates the searching problem into a mixed-integer programming (MIP) problem, and uses a MIP solver to obtain a high-quality multi-dimensional strategy. Experiments conducted on Transformers demonstrate that PTIP obtains 13.9% − 24.7% performance improvement compared to Megatron-LM [8].

Deep learning (DL) has gained great success in recent years, leading to state-of-the-art performance in research community and industrial fields like computer vision and natural language processing. One of the reasons for this success is the huge amount parameters adopted in DL models. However, it is impractical to train a moderately large model with a large number of parameters on a typical single device. Thus, It is necessary to train DL models in clusters with distributed training algorithms. However, traditional distributed training algorithms are usually sub-optimal and highly customized, which owns the drawbacks to train large-scale DL models in varying computing clusters. To handle the above problem, researchers propose auto-parallelism, which is promising to train large-scale DL models efficiently and practically in various computing clusters. In this survey, we perform a broad and thorough investigation on challenges, basis, and strategy searching methods of auto-parallelism in DL training. First, we abstract basic parallelism schemes with their communication cost and memory consumption in DL training. Further, we analyze and compare a series of current auto-parallelism works and investigate strategies and searching methods which are commonly used in practice. At last, we discuss several trends in auto-parallelism which are promising in further research.

As the deep learning model grows larger, training model with a single computational resource becomes impractical. To solve this, hybrid parallelism, which combines data and pipeline parallelism emerges to train large models with multiple GPUs. In practice, using heterogeneous GPU clusters to train large models is a common need due to the upgrade of a part of hardware. However, existing hybrid parallelism approaches in the heterogeneous environment do not work well in communication efficacy, workload balance among GPUs and utilizing the memory constrained GPU. To address these problems, we present a parallel DNN training approach, Hybrid Parallelism on Heterogeneous clusters (HPH). In HPH, we propose a topology designer that minimizes the communication time cost. Furthermore, HPH uses a partition algorithm that automatically partitions DNN layers among workers to maximize throughput. Besides, HPH adopts recomputation-aware scheduling to reduce memory consumption and further reschedule the pipeline to eliminate the extra time overhead of recomputation. Our experimental results on a 32-GPU heterogeneous cluster show that HPH achieves up to 1.42x training speed-ups compared with the state-of-the-art approach.