[UX] Suppress gloo log spam (#29250)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-11-25 17:19:35 -08:00
committed by GitHub
parent 56531b79cc
commit 8d6a89dffd
3 changed files with 63 additions and 26 deletions
+3 -1
View File
@@ -51,6 +51,7 @@ from vllm.distributed.utils import StatelessProcessGroup
from vllm.logger import init_logger
from vllm.utils.import_utils import resolve_obj_by_qualname
from vllm.utils.network_utils import get_distributed_init_method
from vllm.utils.system_utils import suppress_stdout
from vllm.utils.torch_utils import (
direct_register_custom_op,
supports_custom_op,
@@ -329,7 +330,8 @@ class GroupCoordinator:
)
# a group with `gloo` backend, to allow direct coordination between
# processes through the CPU.
cpu_group = torch.distributed.new_group(ranks, backend="gloo")
with suppress_stdout():
cpu_group = torch.distributed.new_group(ranks, backend="gloo")
if self.rank in ranks:
self.ranks = ranks
self.world_size = len(ranks)
+27 -25
View File
@@ -30,6 +30,7 @@ from torch.distributed.rendezvous import rendezvous
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.utils.network_utils import get_tcp_uri
from vllm.utils.system_utils import suppress_stdout
from vllm.utils.torch_utils import is_torch_equal_or_newer
logger = init_logger(__name__)
@@ -427,33 +428,34 @@ def init_gloo_process_group(
Stateless init ProcessGroup with gloo backend compatible with
different torch versions.
"""
if is_torch_equal_or_newer("2.6"):
pg = ProcessGroup(
prefix_store,
group_rank,
group_size,
)
else:
options = ProcessGroup.Options(backend="gloo")
pg = ProcessGroup(
prefix_store,
group_rank,
group_size,
options,
)
from torch.distributed.distributed_c10d import ProcessGroupGloo
with suppress_stdout():
if is_torch_equal_or_newer("2.6"):
pg = ProcessGroup(
prefix_store,
group_rank,
group_size,
)
else:
options = ProcessGroup.Options(backend="gloo")
pg = ProcessGroup(
prefix_store,
group_rank,
group_size,
options,
)
from torch.distributed.distributed_c10d import ProcessGroupGloo
backend_class = ProcessGroupGloo(
prefix_store, group_rank, group_size, timeout=timeout
)
backend_type = ProcessGroup.BackendType.GLOO
device = torch.device("cpu")
if is_torch_equal_or_newer("2.6"):
# _set_default_backend is supported in torch >= 2.6
pg._set_default_backend(backend_type)
backend_class._set_sequence_number_for_group()
backend_class = ProcessGroupGloo(
prefix_store, group_rank, group_size, timeout=timeout
)
backend_type = ProcessGroup.BackendType.GLOO
device = torch.device("cpu")
if is_torch_equal_or_newer("2.6"):
# _set_default_backend is supported in torch >= 2.6
pg._set_default_backend(backend_type)
backend_class._set_sequence_number_for_group()
pg._register_backend(device, backend_type, backend_class)
pg._register_backend(device, backend_type, backend_class)
return pg