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

[V1] AsyncLLM data parallel (#13923)

Browse Source 
Signed-off-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
Nick Hill
2025-03-27 16:14:41 -07:00
committed by GitHub
parent 112b3e5b3b
commit 15dac210f0
18 changed files with 722 additions and 156 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm/config.py
View File

@@ -40,7 +40,8 @@ from vllm.transformers_utils.config import (
from vllm.transformers_utils.s3_utils import S3Model
from vllm.transformers_utils.utils import is_s3, maybe_model_redirect
from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
get_cpu_memory, random_uuid, resolve_obj_by_qualname)
get_cpu_memory, get_open_port, random_uuid,
resolve_obj_by_qualname)
if TYPE_CHECKING:
from ray.util.placement_group import PlacementGroup
@@ -1389,6 +1390,8 @@ class ParallelConfig:
tensor_parallel_size: int = 1 # Number of tensor parallel groups.
data_parallel_size: int = 1 # Number of data parallel groups.
data_parallel_rank: int = 0 # Rank of the data parallel group.
# Local rank of the data parallel group, defaults to global rank.
data_parallel_rank_local: Optional[int] = None
# IP of the data parallel master.
data_parallel_master_ip: str = "127.0.0.1"
data_parallel_master_port: int = 29500 # Port of the data parallel master.
@@ -1493,10 +1496,18 @@ class ParallelConfig:
self.world_size = self.pipeline_parallel_size * \
self.tensor_parallel_size
self.data_parallel_size = envs.VLLM_DP_SIZE
self.data_parallel_rank = envs.VLLM_DP_RANK
self.data_parallel_master_ip = envs.VLLM_DP_MASTER_IP
self.data_parallel_master_port = envs.VLLM_DP_MASTER_PORT
if self.data_parallel_size > 1:
# Data parallel was specified in the engine args.
self.data_parallel_master_port = get_open_port()
# TODO multi-node
else:
# Otherwise fall back to env vars (e.g. for offline SPMD case).
self.data_parallel_size = envs.VLLM_DP_SIZE
self.data_parallel_rank = envs.VLLM_DP_RANK
self.data_parallel_rank_local = envs.VLLM_DP_RANK_LOCAL
self.data_parallel_master_ip = envs.VLLM_DP_MASTER_IP
self.data_parallel_master_port = envs.VLLM_DP_MASTER_PORT
self.world_size_across_dp = self.world_size * self.data_parallel_size
if self.distributed_executor_backend == "external_launcher":