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 14, 05:30 UTC
[Bug] Fix deepep low latency use nvlink by default (#27677)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 28, 2025 at 19:53 UTC
d3ab240f39219df0175ec662416f630d7bf273d8
1 file changed
+2
-2
vllm/envs.py
+2
-2
@@ -205,7 +205,7 @@ if TYPE_CHECKING:
205
VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
206
VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
207
VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
208
- VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK: bool = False
208
+ VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK: bool = True
209
VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
210
VLLM_DBO_COMM_SMS: int = 20
211
GPT_OSS_SYSTEM_TOOL_MCP_LABELS: list[str] = []
@@ -1362,7 +1362,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
1362
# Allow DeepEP to use nvlink for internode_ll kernel, turn this on for
1363
# better latency on GB200 like system
1364
"VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK": lambda: bool(
1365
- int(os.getenv("VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK", "0"))
1365
+ int(os.getenv("VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK", "1"))
1366
),
1367
# Allow DeepEP to use MNNVL (multi-node nvlink) for internode_ll kernel,
1368
# turn this for better latency on GB200 like system