Update num_tokens_across_dp to use nccl instead of gloo (#24105)

Signed-off-by: Sage Moore <sage@neuralmagic.com>
This commit is contained in:
Sage Moore
2025-09-15 19:05:48 +00:00
committed by GitHub
parent a0b26701c9
commit 49bfc538e4
2 changed files with 25 additions and 4 deletions
+8
View File
@@ -92,6 +92,7 @@ if TYPE_CHECKING:
VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
VLLM_SKIP_P2P_CHECK: bool = False
VLLM_DISABLED_KERNELS: list[str] = []
VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION: bool = False
VLLM_USE_V1: bool = True
VLLM_ROCM_USE_AITER: bool = False
VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
@@ -745,6 +746,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
lambda: [] if "VLLM_DISABLED_KERNELS" not in os.environ else os.environ[
"VLLM_DISABLED_KERNELS"].split(","),
# Swaps the all reduce backend that we use to coordinate the DP padding
# information from NCCL to gloo.
"VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION":
lambda:
(os.getenv("VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION", "False").lower() in
("true", "1")),
# If set, use the V1 code path.
"VLLM_USE_V1":
lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),