How can I use PertData with PyTorch DDP DistributedSampler?

[l024258]

I’m training a perturbation‑prediction model using datasets managed via GEARS PertData, and I need to run multi‑GPU training with PyTorch Distributed Data Parallel (DDP). What’s the recommended way to connect a PertData dataset to a DataLoader that uses torch.utils.data.distributed.DistributedSampler so that each rank gets a disjoint shard, preserves split integrity (train/val/test), and supports epoch‑level shuffling?

I’m specifically looking for guidance or example code for:

• Converting or wrapping PertData into a PyTorch‑style dataset that DistributedSampler understands
• Handling split selection (train, val, test) so that all ranks see consistent subsets
• Ensuring deterministic shuffling across epochs via sampler.set_epoch(epoch)
• Best practices around rank/world size, seed control, and worker initialization