Daniel Barley

Daniel Barley, Jonathan Leis, Benjamin Klenk et al.

Pipeline parallelism is a key technique for distributed training of large language models because it reduces per-device parameter and activation memory. However, comparing pipeline schedules is difficult: analytical models expose structural quantities such as bubble ratios, while end-to-end hardware experiments are costly and system-specific. In this work, we introduce a tabular schedule abstraction and a unified multi-abstraction methodology that connects formula-based reasoning, idealized schedule tables, and communication-aware execution simulation. Using this framework, we compare GPipe, 1F1B, Chimera, and Hanayo in its restricted regime across multiple modeled system configurations. Our results show that schedule rankings are not abstraction-invariant: communication can negate structural advantages suggested by bubble analysis alone. Under the assumptions considered here, GPipe and 1F1B are runtime-equivalent, but 1F1B achieves a lower activation-memory peak. Chimera is advantageous mainly at low microbatch counts and in communication-favorable regimes, while Hanayo is effective in its intended restricted operating point but remains sensitive to network bottlenecks. We further study an asymmetric Chimera-style placement, which does not reduce the global peak memory requirement but reveals limited runtime gains in shallow pipelines. Overall, pipeline schedule quality is meaningful only in the context of the modeled execution environment.

Daniel Barley, Holger Fröning

The rise of Deep Neural Networks (DNNs) has led to an increase in model size and complexity, straining the memory capacity of GPUs. Sparsity in DNNs, characterized as structural or ephemeral, has gained attention as a solution. This work focuses on ephemeral sparsity, aiming to reduce memory consumption during training. It emphasizes the significance of activations, an often overlooked component, and their role in memory usage. This work employs structured pruning in Block Sparse Compressed Row (BSR) format in combination with a magnitude-based criterion to efficiently prune activations. We furthermore introduce efficient block-sparse operators for GPUs and showcase their effectiveness, as well as the superior compression offered by block sparsity. We report the effectiveness of activation pruning by evaluating training speed, accuracy, and memory usage of large-scale neural architectures on the example of ResMLP on image classification tasks. As a result, we observe a memory reduction of up to 32% while maintaining accuracy. Ultimately, our approach aims to democratize large-scale model training, reduce GPU requirements, and address ecological concerns.

Daniel Barley, Holger Fröning

The ever-growing scale of deep neural networks (DNNs) has lead to an equally rapid growth in computational resource requirements. Many recent architectures, most prominently Large Language Models, have to be trained using supercomputers with thousands of accelerators, such as GPUs or TPUs. Next to the vast number of floating point operations the memory footprint of DNNs is also exploding. In contrast, GPU architectures are notoriously short on memory. Even comparatively small architectures like some EfficientNet variants cannot be trained on a single consumer-grade GPU at reasonable mini-batch sizes. During training, intermediate input activations have to be stored until backpropagation for gradient calculation. These make up the vast majority of the memory footprint. In this work we therefore consider compressing activation maps for the backward pass using pooling, which can reduce both the memory footprint and amount of data movement. The forward computation remains uncompressed. We empirically show convergence and study effects on feature detection at the example of the common vision architecture ResNet. With this approach we are able to reduce the peak memory consumption by 29% at the cost of a longer training schedule, while maintaining prediction accuracy compared to an uncompressed baseline.