MCPcopy Create free account
hub / github.com/Standard-Intelligence/hertz-dev / init_dist

Function init_dist

utils/dist.py:56–76  ·  view source on GitHub ↗
()

Source from the content-addressed store, hash-verified

54 printonce(*args, **kwargs)
55
56def init_dist():
57 if T.distributed.is_initialized():
58 print0('Distributed already initialized')
59 rank = T.distributed.get_rank()
60 local_rank = int(os.environ.get('LOCAL_RANK', 0))
61 world_size = T.distributed.get_world_size()
62 else:
63 try:
64 rank = int(os.environ['RANK'])
65 local_rank = int(os.environ['LOCAL_RANK'])
66 world_size = int(os.environ['WORLD_SIZE'])
67 device = f'cuda:{local_rank}'
68 T.cuda.set_device(device)
69 T.distributed.init_process_group(backend='nccl', timeout=timedelta(minutes=30), rank=rank, world_size=world_size, device_id=T.device(device))
70 print(f'Rank {rank} of {world_size}.')
71 except Exception as e:
72 print0once(f'Not initializing distributed env: {e}')
73 rank = 0
74 local_rank = 0
75 world_size = 1
76 return rank, local_rank, world_size
77
78def load_ckpt(load_from_location, expected_hash=None):
79 os.environ['HF_HUB_ENABLE_HF_TRANSFER'] = '1' #Disable this to speed up debugging errors with downloading from the hub

Callers

nothing calls this directly

Calls 2

print0Function · 0.85
print0onceFunction · 0.85

Tested by

no test coverage detected