A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 13:30 UTC
Use `@property` and private field for `data_parallel_rank_local` (#17053)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Apr 23, 2025 at 16:50 UTC
bdb366031252cd8d99270a3f81dfe76021a760c1
2 files changed
+18
-5
vllm/config.py
+15
-2
@@ -1593,8 +1593,21 @@ class ParallelConfig:
1593
the product of the tensor parallel size and data parallel size."""
1594
data_parallel_rank: int = 0
1595
"""Rank of the data parallel group."""
1596
- data_parallel_rank_local: Optional[int] = None
1597
- """Local rank of the data parallel group, defaults to global rank."""
1596
+ _data_parallel_rank_local: Optional[int] = field(default=None, init=False)
1597
+ """Private field to store the local rank of the data parallel group."""
1598
+
1599
+ @property
1600
+ def data_parallel_rank_local(self) -> int:
1601
+ """Local rank of the data parallel group, defaults to global rank."""
1602
+ if self._data_parallel_rank_local is None:
1603
+ return self.data_parallel_rank
1604
+ return self._data_parallel_rank_local
1605
+
1606
+ @data_parallel_rank_local.setter
1607
+ def data_parallel_rank_local(self, value: int) -> None:
1608
+ """Set the local rank of the data parallel group."""
1609
+ self._data_parallel_rank_local = value
1610
+
1611
data_parallel_master_ip: str = "127.0.0.1"
1612
"""IP of the data parallel master."""
1613
data_parallel_master_port: int = 29500
vllm/v1/engine/core_client.py
+3
-3
@@ -439,10 +439,10 @@ class MPClient(EngineCoreClient):
439
) -> None:
440
441
# Default case - single core engine.
442
- dp_rank = vllm_config.parallel_config.data_parallel_rank
443
- local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local
442
core_engine = new_core_engine(
445
- dp_rank, local_dp_rank if local_dp_rank is not None else dp_rank)
443
+ vllm_config.parallel_config.data_parallel_rank,
444
+ vllm_config.parallel_config.data_parallel_rank_local,
445
+ )
446
core_engines.append(core_engine)
447
self.core_engine = core_engine
448