[Perf] Low-latency CPU all-reduce for DP metadata syncs (#61767)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: hjjq <50634613+hjjq@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Nick Hill
2026-10-15 15:29:13 -07:00
committed by GitHub
co-authored by hjjq Claude Opus 5.5
parent 784d1fff0c
commit de7d70676d
7 changed files with 159 additions and 8 deletions
+5
View File
@@ -50,6 +50,7 @@ from vllm.distributed.device_communicators.base_device_communicator import (
)
from vllm.distributed.utils import (
StatelessProcessGroup,
cpu_all_reduce_,
create_tcp_store,
get_cached_tcp_store_client,
)
@@ -1376,6 +1377,10 @@ class GroupCoordinator:
"""
torch.distributed.barrier(group=self.cpu_group)
def cpu_all_reduce_(self, tensor: torch.Tensor) -> torch.Tensor:
"""In-place SUM all-reduce of a small CPU tensor over `cpu_group`."""
return cpu_all_reduce_(tensor, self.cpu_group)
def send(self, tensor: torch.Tensor, dst: int | None = None) -> None:
"""Sends a tensor to the destination rank in a blocking way."""
"""NOTE: `dst` is the local rank of the destination rank."""