[Misc] Remove use of CUDA_VISIBLE_DEVICES for device selection (fix DP slow startup time &c) (#26709)

Signed-off-by: ilmarkov <markovilya197@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
This commit is contained in:
Ilya Markov
2025-10-23 20:58:39 +08:00
committed by GitHub
co-authored by Tyler Michael Smith
parent faee3ccdc2
commit 237cf6d32a
4 changed files with 35 additions and 7 deletions
+2 -2
View File
@@ -8,7 +8,6 @@ import torch.distributed as dist
from vllm.config import ParallelConfig
from vllm.distributed.parallel_state import get_dp_group, is_global_first_rank
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.v1.worker.ubatch_utils import (
UBatchSlices,
check_ubatch_thresholds,
@@ -20,7 +19,8 @@ logger = init_logger(__name__)
def _get_device_and_group(parallel_config: ParallelConfig):
device = current_platform.device_type
# Use the actual device assigned to the DP group, not just the device type
device = get_dp_group().device
group = get_dp_group().device_group
# Transfering this tensor from GPU to CPU will introduce a GPU sync
+21
View File
@@ -169,6 +169,27 @@ class Worker(WorkerBase):
if self.device_config.device.type == "cuda":
# This env var set by Ray causes exceptions with graph building.
os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
if (
self.parallel_config.data_parallel_size > 1
and self.parallel_config.data_parallel_size_local > 0
and self.parallel_config.data_parallel_backend != "ray"
):
# Use local DP rank if available, otherwise use global DP rank.
dp_local_rank = self.parallel_config.data_parallel_rank_local
if dp_local_rank is None:
dp_local_rank = self.parallel_config.data_parallel_rank
tp_pp_world_size = (
self.parallel_config.pipeline_parallel_size
* self.parallel_config.tensor_parallel_size
)
# DP_LOCAL_RANK * TP_PP_WORLD_SIZE + TP_LOCAL_RANK
self.local_rank += dp_local_rank * tp_pp_world_size
assert self.local_rank <= torch.cuda.device_count(), (
f"DP adjusted local rank {self.local_rank} is out of bounds. "
)
self.device = torch.device(f"cuda:{self.local_rank}")
current_platform.set_device(self.device)