Data loader fails with multi-gpu

[moradza]

My model gets trained on a single gpu but when I want to scale it up to multi-gpu it gets killed with no obvious error message. I am feeling it has to do with data loader memory use. Anybody has similar experience? Strategy = DDP