xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-13 12:46:07 +08:00
Code Issues Packages Projects Releases Wiki Activity

enable multi-node in external launcher mode (#29833)

Browse Source
This commit is contained in:
Xieyang Xu 2025-12-03 17:02:02 -08:00 committed by GitHub
parent 1109f98288
commit ad32e3e19c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 27 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/config/parallel.py
View File

@ -593,10 +593,14 @@ class ParallelConfig:
"max_parallel_loading_workers is currently " "max_parallel_loading_workers is currently "
"not supported and will be ignored." "not supported and will be ignored."
) )
if self.distributed_executor_backend not in ("mp", "uni") and self.nnodes > 1: allowed_backends = ("mp", "uni", "external_launcher")
if (
self.distributed_executor_backend not in allowed_backends
and self.nnodes > 1
):
raise ValueError( raise ValueError(
"nnodes > 1 can only be set when distributed executor " "nnodes > 1 can only be set when distributed executor "
"backend is mp or uni." "backend is mp, uni or external_launcher."
) )
@property @property

39
vllm/distributed/parallel_state.py
View File

@ -1169,17 +1169,13 @@ def init_distributed_environment(
from vllm.config import get_current_vllm_config from vllm.config import get_current_vllm_config
config = get_current_vllm_config() config = get_current_vllm_config()
if config is not None and config.parallel_config.nnodes > 1: if (
parallel_config = config.parallel_config
ip = parallel_config.master_addr
rank = parallel_config.data_parallel_rank * world_size + rank
world_size = parallel_config.world_size_across_dp
port = parallel_config.master_port
distributed_init_method = get_distributed_init_method(ip, port)
elif (
config is not None config is not None
and config.parallel_config.data_parallel_size > 1
and config.parallel_config.distributed_executor_backend != "external_launcher" and config.parallel_config.distributed_executor_backend != "external_launcher"
and (
config.parallel_config.nnodes > 1
or config.parallel_config.data_parallel_size > 1
)
): ):
parallel_config = config.parallel_config parallel_config = config.parallel_config
# adjust to take into account data parallelism # adjust to take into account data parallelism
@ -1187,15 +1183,22 @@ def init_distributed_environment(
rank = parallel_config.data_parallel_rank * world_size + rank rank = parallel_config.data_parallel_rank * world_size + rank
# adjust the world size to take into account data parallelism # adjust the world size to take into account data parallelism
world_size = parallel_config.world_size_across_dp world_size = parallel_config.world_size_across_dp
ip = parallel_config.data_parallel_master_ip
port = parallel_config.get_next_dp_init_port() # Use appropriate IP and port based on configuration
distributed_init_method = get_distributed_init_method(ip, port) if parallel_config.nnodes > 1:
logger.debug( ip = parallel_config.master_addr
"Adjusting world_size=%d rank=%d distributed_init_method=%s for DP", port = parallel_config.master_port
world_size, distributed_init_method = get_distributed_init_method(ip, port)
rank, else:
distributed_init_method, ip = parallel_config.data_parallel_master_ip
) port = parallel_config.get_next_dp_init_port()
distributed_init_method = get_distributed_init_method(ip, port)
logger.debug(
"Adjusting world_size=%d rank=%d distributed_init_method=%s for DP",
world_size,
rank,
distributed_init_method,
)
if not torch.distributed.is_initialized(): if not torch.distributed.is_initialized():
logger.info( logger.info(
"world_size=%d rank=%d local_rank=%d distributed_init_method=%s backend=%s", "world_size=%d rank=%d local_rank=%d distributed_init_method=%s backend=%s",