At inference time the same input gives slightly different outputs on every call. The model has dropout layers. What was missed?
Answer: B. model.eval(): dropout (and batch norm) behave differently in training mode
In training mode, dropout randomly zeroes activations. Call model.eval() (and use torch.no_grad()) for inference.