biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Log] Optimize Startup Log (#26601)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2025-10-13 14:06:57 -04:00
committed by GitHub
parent afc47e4de7
commit e251e457c5
6 changed files with 17 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/distributed/device_communicators/cuda_communicator.py
View File

@@ -13,6 +13,7 @@ from vllm.distributed.device_communicators.pynccl import register_nccl_symmetric
from vllm.distributed.device_communicators.pynccl_allocator import (
is_symmetric_memory_enabled,
)
from vllm.distributed.parallel_state import is_global_first_rank
from vllm.logger import init_logger
from vllm.platforms import current_platform
@@ -95,35 +96,35 @@ class CudaCommunicator(DeviceCommunicatorBase):
from .all2all import NaiveAll2AllManager
self.all2all_manager = NaiveAll2AllManager(self.cpu_group)
logger.info("Using naive all2all manager.")
elif all2all_backend == "allgather_reducescatter":
from .all2all import AgRsAll2AllManager
self.all2all_manager = AgRsAll2AllManager(self.cpu_group)
logger.info("Using AllGather-ReduceScatter all2all manager.")
elif all2all_backend == "pplx":
from .all2all import PPLXAll2AllManager
self.all2all_manager = PPLXAll2AllManager(self.cpu_group)
logger.info("Using PPLX all2all manager.")
elif all2all_backend == "deepep_high_throughput":
from .all2all import DeepEPHTAll2AllManager
self.all2all_manager = DeepEPHTAll2AllManager(self.cpu_group)
logger.info("Using DeepEP High-Throughput all2all manager.")
elif all2all_backend == "deepep_low_latency":
from .all2all import DeepEPLLAll2AllManager
self.all2all_manager = DeepEPLLAll2AllManager(self.cpu_group)
logger.info("Using DeepEP Low-Latency all2all manager.")
elif all2all_backend == "flashinfer_all2allv":
from .all2all import FlashInferAllToAllManager
self.all2all_manager = FlashInferAllToAllManager(self.cpu_group)
logger.info("Using Flashinfer all2allv manager.")
else:
raise ValueError(f"Unknown all2all backend: {all2all_backend}")
if is_global_first_rank():
logger.info(
"Using %s all2all manager.",
self.all2all_manager.__class__.__name__,
)
def all_reduce(self, input_):
# since currently we perform copy input -> symm_input -> out-of-place AR
# return symm_output, we don't need to check if input is symmetric

3
vllm/distributed/device_communicators/pynccl.py
View File

@@ -105,11 +105,10 @@ class PyNcclCommunicator:
self.disabled = False
self.nccl_version = self.nccl.ncclGetRawVersion()
logger.info("vLLM is using nccl==%s", self.nccl.ncclGetVersion())
if self.rank == 0:
# get the unique id from NCCL
self.unique_id = self.nccl.ncclGetUniqueId()
logger.info("vLLM is using nccl==%s", self.nccl.ncclGetVersion())
else:
# construct an empty unique id
self.unique_id = ncclUniqueId()