[DP][ray] Support different VLLM_RAY_DP_PACK_STRATEGY (#23849)

Signed-off-by: Rui Qiao <ruisearch42@gmail.com>
This commit is contained in:
Rui Qiao
2025-10-09 19:53:43 -07:00
committed by GitHub
parent c6187f55f7
commit 757fa4a4da
2 changed files with 86 additions and 33 deletions
+12
View File
@@ -133,6 +133,7 @@ if TYPE_CHECKING:
VLLM_DP_MASTER_PORT: int = 0
VLLM_MOE_DP_CHUNK_SIZE: int = 256
VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
VLLM_RAY_DP_PACK_STRATEGY: str = "strict"
VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
VLLM_MXFP4_USE_MARLIN: Optional[bool] = None
VLLM_V0_USE_OUTLINES_CACHE: bool = False
@@ -1000,6 +1001,17 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
)
== "1",
# Strategy to pack the data parallel ranks for Ray.
# Available options:
# - "fill":
# for DP master node, allocate exactly data-parallel-size-local DP ranks,
# for non-master nodes, allocate as many DP ranks as can fit;
# - "strict":
# allocate exactly data-parallel-size-local DP ranks to each picked node;
# This environment variable is ignored if data-parallel-backend is not Ray.
"VLLM_RAY_DP_PACK_STRATEGY": lambda: os.getenv(
"VLLM_RAY_DP_PACK_STRATEGY", "strict"
),
# Whether to use S3 path for model loading in CI via RunAI Streamer
"VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
# Use model_redirect to redirect the model name to a local folder.