[UX] Suppress gloo log spam (#29250)

Signed-off-by: mgoin <mgoin64@gmail.com>

vllm/distributed/parallel_state.py

@@ -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,
@@ -329,6 +330,7 @@ class GroupCoordinator:
             )
             # a group with `gloo` backend, to allow direct coordination between
             # processes through the CPU.
+            with suppress_stdout():
                 cpu_group = torch.distributed.new_group(ranks, backend="gloo")
             if self.rank in ranks:
                 self.ranks = ranks

vllm/distributed/utils.py

@@ -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__)
@@ -427,6 +428,7 @@ def init_gloo_process_group(
     Stateless init ProcessGroup with gloo backend compatible with
     different torch versions.
     """
+    with suppress_stdout():
         if is_torch_equal_or_newer("2.6"):
             pg = ProcessGroup(
                 prefix_store,

vllm/utils/system_utils.py

@@ -56,6 +56,39 @@ def set_env_var(key: str, value: str) -> Iterator[None]:
             os.environ[key] = old
 
 
+@contextlib.contextmanager
+def suppress_stdout():
+    """
+    Suppress stdout from C libraries at the file descriptor level.
+
+    Only suppresses stdout, not stderr, to preserve error messages.
+    Suppression is disabled when VLLM_LOGGING_LEVEL is set to DEBUG.
+
+    Example:
+        with suppress_stdout():
+            # C library calls that would normally print to stdout
+            torch.distributed.new_group(ranks, backend="gloo")
+    """
+    # Don't suppress if logging level is DEBUG
+    if envs.VLLM_LOGGING_LEVEL == "DEBUG":
+        yield
+        return
+
+    stdout_fd = sys.stdout.fileno()
+    stdout_dup = os.dup(stdout_fd)
+    devnull_fd = os.open(os.devnull, os.O_WRONLY)
+
+    try:
+        sys.stdout.flush()
+        os.dup2(devnull_fd, stdout_fd)
+        yield
+    finally:
+        sys.stdout.flush()
+        os.dup2(stdout_dup, stdout_fd)
+        os.close(stdout_dup)
+        os.close(devnull_fd)
+
+
 # File path utilities