IoT devices have limited hardware capabilities and are often deployed in remote areas. Consequently, advanced vision models surpass such devices' processing and storage capabilities, requiring offloading of such tasks to the cloud. However, remote areas often rely on LPWANs technology with limited bandwidth, high packet loss rates, and extremely low duty cycles, which makes fast offloading for time-sensitive inference challenging. Today's approaches, which are deployable on weak devices, generate a non-progressive bit stream, and therefore, their decoding quality suffers strongly when data is only partially available on the cloud at a deadline due to limited bandwidth or packet losses.

The architecture of Vision Transformers (ViTs), particularly the Multi-head Attention (MHA) mechanism, imposes substantial hardware demands. Deploying ViTs on devices with varying constraints, such as mobile phones, requires multiple models of different sizes. However, this approach has limitations, such as training and storing each required model separately. This paper introduces HydraViT, a novel approach that addresses these limitations by stacking attention heads to achieve a scalable ViT. By repeatedly changing the size of the embedded dimensions throughout each layer and their corresponding number of attention heads in MHA during training, HydraViT induces multiple subnetworks. Thereby, HydraViT achieves adaptability across a wide spectrum of hardware environments while maintaining performance. Our experimental results demonstrate the efficacy of HydraViT in achieving a scalable ViT with up to 10 subnetworks, covering a wide range of resource constraints. HydraViT achieves up to 5 p.p. more accuracy with the same GMACs and up to 7 p.p. more accuracy with the same throughput on ImageNet-1K compared to the baselines, making it an effective solution for scenarios where hardware availability is diverse or varies over time. The source code is available at https://github.com/ds-kiel/HydraViT.