A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 07:15 UTC
[BugFix] fix data parallel construct ipv6 url addres (#18991)
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
rongfu.leng committed
Jun 1, 2025 at 11:42 UTC
aa54a7bf7bae7e1db43693470ebe93e3dcd30f9d
1 file changed
+3
-3
vllm/distributed/parallel_state.py
+3
-3
@@ -41,8 +41,8 @@ from vllm.distributed.device_communicators.base_device_communicator import (
41
DeviceCommunicatorBase)
42
from vllm.distributed.utils import StatelessProcessGroup
43
from vllm.logger import init_logger
44
-from vllm.utils import (direct_register_custom_op, resolve_obj_by_qualname,
45
- supports_custom_op)
44
+from vllm.utils import (direct_register_custom_op, get_distributed_init_method,
45
+ resolve_obj_by_qualname, supports_custom_op)
46
47
48
@dataclass
@@ -929,7 +929,7 @@ def init_distributed_environment(
929
world_size = parallel_config.world_size_across_dp
930
ip = parallel_config.data_parallel_master_ip
931
port = parallel_config.get_next_dp_init_port()
932
- distributed_init_method = f"tcp://{ip}:{port}" # noqa
932
+ distributed_init_method = get_distributed_init_method(ip, port)
933
logger.info(
934
"Adjusting world_size=%d rank=%d distributed_init_method=%s for DP",
935
world_size, rank, distributed_init_method)