MCPcopy
hub / github.com/microsoft/Cream / init_distributed_mode

Function init_distributed_mode

EfficientViT/classification/utils.py:206–228  ·  view source on GitHub ↗
(args)

Source from the content-addressed store, hash-verified

204 torch.save(*args, **kwargs)
205
206def init_distributed_mode(args):
207 if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:
208 args.rank = int(os.environ["RANK"])
209 args.world_size = int(os.environ['WORLD_SIZE'])
210 args.gpu = int(os.environ['LOCAL_RANK'])
211 elif 'SLURM_PROCID' in os.environ:
212 args.rank = int(os.environ['SLURM_PROCID'])
213 args.gpu = args.rank % torch.cuda.device_count()
214 else:
215 print('Not using distributed mode')
216 args.distributed = False
217 return
218
219 args.distributed = True
220
221 torch.cuda.set_device(args.gpu)
222 args.dist_backend = 'nccl'
223 print('| distributed init (rank {}): {}'.format(
224 args.rank, args.dist_url), flush=True)
225 torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
226 world_size=args.world_size, rank=args.rank)
227 torch.distributed.barrier()
228 setup_for_distributed(args.rank == 0)
229
230
231def replace_batchnorm(net):

Callers

nothing calls this directly

Calls 3

formatMethod · 0.80
printFunction · 0.70
setup_for_distributedFunction · 0.70

Tested by

no test coverage detected