xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-09 23:55:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor ux

Browse Source 
Signed-off-by: Robert Shaw <robshaw@redhat.com>
This commit is contained in:
Robert Shaw 2025-07-20 20:01:57 +00:00
parent 6206a06a84
commit c22990470f
3 changed files with 36 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
vllm/engine/arg_utils.py
View File

@ -295,6 +295,7 @@ class EngineArgs:
tensor_parallel_size: int = ParallelConfig.tensor_parallel_size tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
data_parallel_size: int = ParallelConfig.data_parallel_size data_parallel_size: int = ParallelConfig.data_parallel_size
data_parallel_rank: Optional[int] = None data_parallel_rank: Optional[int] = None
data_parallel_start_rank: Optional[int] = None
data_parallel_size_local: Optional[int] = None data_parallel_size_local: Optional[int] = None
data_parallel_address: Optional[str] = None data_parallel_address: Optional[str] = None
data_parallel_rpc_port: Optional[int] = None data_parallel_rpc_port: Optional[int] = None
@ -606,6 +607,11 @@ class EngineArgs:
type=int, type=int,
help='Data parallel rank of this instance. ' help='Data parallel rank of this instance. '
'When set, enables external load balancer mode.') 'When set, enables external load balancer mode.')
parallel_group.add_argument('--data-parallel-start-rank',
'-dpr',
type=int,
help='Starting data parallel rank '
'for secondary nodes.')
parallel_group.add_argument('--data-parallel-size-local', parallel_group.add_argument('--data-parallel-size-local',
'-dpl', '-dpl',
type=int, type=int,
@ -1091,19 +1097,36 @@ class EngineArgs:
# but we should not do this here. # but we should not do this here.
placement_group = ray.util.get_current_placement_group() placement_group = ray.util.get_current_placement_group()
# data_parallel_external_lb = self.data_parallel_rank is not None # Organize --data-parallel-start-rank and --data-parallel-rank.
# if data_parallel_external_lb: if self.data_parallel_start_rank is not None:
# assert self.data_parallel_size_local in (1, None), ( if self.data_parallel_rank is not None:
# "data_parallel_size_local must be 1 when data_parallel_rank " raise ValueError(
# "is set") "Found --data-parallel-rank and --data-parallel-start-rank."
# data_parallel_size_local = 1 "Only one should be set (use --data-parallel-start-rank).")
# elif self.data_parallel_size_local is not None: else:
data_parallel_external_lb = False self.data_parallel_rank = self.data_parallel_start_rank
if self.data_parallel_size_local is not None:
# Validate External LB.
data_parallel_external_lb = True
if data_parallel_external_lb:
if self.data_parallel_size_local is None:
raise ValueError(
"With external LB, --data-parallel-size-local must be set."
)
if self.data_parallel_size_local >= self.data_parallel_size:
raise ValueError(
"With external LB, --data-parallel-size-local must be less "
"than --data-parallel-size.")
if (self.data_parallel_rank is not None
and self.data_parallel_size_local > 1):
raise ValueError(
"With --data-parallel-size-local > 1, use --data-parall"
"--data-parallel-rank")
data_parallel_size_local = self.data_parallel_size_local data_parallel_size_local = self.data_parallel_size_local
else:
# Local DP size defaults to global DP size if not set. # Local DP size defaults to global DP size if not set.
data_parallel_size_local = self.data_parallel_size data_parallel_size_local = (self.data_parallel_size_local
or self.data_parallel_size)
# DP address, used in multi-node case for torch distributed group # DP address, used in multi-node case for torch distributed group
# and ZMQ sockets. # and ZMQ sockets.

7
vllm/entrypoints/cli/serve.py
View File

@ -45,11 +45,6 @@ class ServeSubcommand(CLISubcommand):
if args.headless or args.api_server_count < 1: if args.headless or args.api_server_count < 1:
run_headless(args) run_headless(args)
else: else:
if args.data_parallel_start_rank:
raise ValueError(
"data_parallel_start_rank is only applicable "
"in headless mode. "
"Add --headless flag to enable headless mode.")
if args.api_server_count > 1: if args.api_server_count > 1:
run_multi_api_server(args) run_multi_api_server(args)
else: else:
@ -122,7 +117,7 @@ def run_headless(args: argparse.Namespace):
engine_manager = CoreEngineProcManager( engine_manager = CoreEngineProcManager(
target_fn=EngineCoreProc.run_engine_core, target_fn=EngineCoreProc.run_engine_core,
local_engine_count=local_engine_count, local_engine_count=local_engine_count,
start_index=args.data_parallel_start_rank, start_index=vllm_config.parallel_config.data_parallel_rank,
local_start_index=0, local_start_index=0,
vllm_config=vllm_config, vllm_config=vllm_config,
local_client=False, local_client=False,

7
vllm/entrypoints/openai/cli_args.py
View File

@ -253,13 +253,6 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
default=False, default=False,
help="Run in headless mode. See multi-node data parallel " help="Run in headless mode. See multi-node data parallel "
"documentation for more details.") "documentation for more details.")
parser.add_argument(
"--data-parallel-start-rank",
"-dpr",
type=int,
default=0,
help="Starting data parallel rank for secondary nodes. "
"Requires --headless.")
parser.add_argument("--api-server-count", parser.add_argument("--api-server-count",
"-asc", "-asc",
type=int, type=int,