https://yestino.com/entities/nccl2-a54a29Every event involving NCCL2enhttps://yestino.com/events/trunk-05b558d4e79ea7a049dbc351a9067f0bf2a63faf-c10d-bind-loc-e61bf3https://yestino.com/events/trunk-05b558d4e79ea7a049dbc351a9067f0bf2a63faf-c10d-bind-loc-e61bf3Tue, 25 Aug 2026 06:43:08 GMTNCCL2 initializes eagerly, so selecting torch.cuda.current_device() before honoring LOCAL_RANK can bind forked workers to the same GPU. Sources: PyTorch Releases