BridgeTower fine-tuning on Habana Gaudi2: 2.5x A100, 1.4x H100 speedups with Optimum Habana v1.7

AI Impact Summary

BridgeTower fine-tuning on Habana Gaudi2 using Optimum Habana v1.7 achieves substantial throughput gains versus A100 and H100, signaling a strong hardware-software fit for vision-language model workloads. The performance uplift relies on hardware-accelerated data loading and Habana's media pipeline to shift decoding and augmentations onto the device, with dataloader_num_workers tuning providing clear CPU-parallelism gains across Gaudi2, H100, and A100. For teams retraining BridgeTower or similar VL models, this implies faster experiments and lower compute costs on Gaudi2 when adopting Optimum Habana tooling and GaudiTrainingArguments.