ddp fail with new pl_lightning / pytorch

[ragier]

@thibo73800 To reproduce, you can launch a train (detr for example) with this modification if you have only one GPU :

Replace :
accelerator=None if torch.cuda.device_count() <= 1 else "ddp",
by
accelerator="ddp",

[anhtu293]

It's fixed in the latest version of aloception-oss (v0.5.1).
Since v1.8 of lightning, it doesnt accept ddp as value for accelerator anymore. We have strategy for that instead.