CUDA OOM in DA3 for 1000+ Image Datasets

When processing large-scale datasets (e.g., 1000+ images), DA3 encounters CUDA out-of-memory errors due to its current single-GPU execution pipeline. Is there an existing mechanism—or planned support—for distributing computation across multiple GPUs to handle high-resolution, long-sequence inputs more efficiently? Additionally, are there recommended strategies for memory-aware batching or multi-GPU parallelization to mitigate OOM issues during training and inference?