[Core][Test] move local_rank to the last arg with default value(#3711)

[Core][Test] move local_rank to the last arg with default value to keep api compatible (#3711)
This commit is contained in:
youkaichao
2024-03-28 21:19:45 -07:00
committed by GitHub
parent 395aa823ea
commit 756b30a5f3
7 changed files with 24 additions and 14 deletions
+6 -1
View File
@@ -202,11 +202,11 @@ class NCCLCommunicator:
init_method=None,
timeout=datetime.timedelta(seconds=10),
world_size: int = -1,
local_rank: int = -1,
rank: int = -1,
store=None,
group_name: str = "",
pg_options=None,
local_rank: int = -1,
):
if not dist.is_initialized():
backend = backend or "nccl"
@@ -220,6 +220,11 @@ class NCCLCommunicator:
store=store,
group_name=group_name,
pg_options=pg_options)
self.rank = dist.get_rank()
self.world_size = dist.get_world_size()
if local_rank == -1:
local_rank = self.rank
self.local_rank = local_rank
torch.cuda.set_device(local_rank)
if rank == 0:
self.unique_id = ncclGetUniqueId()
@@ -35,8 +35,10 @@ def set_pynccl_stream(stream: torch.cuda.Stream):
pass
def init_process_group(world_size: int, local_rank: int, rank: int,
init_method: str) -> None:
def init_process_group(world_size: int,
rank: int,
init_method: str,
local_rank: int = -1) -> None:
assert not is_initialized()
global comm
logger.info(f"vLLM is using nccl=={ncclGetVersion()}")