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
[env] default nixl side port conflicts with kv-event zmq port (#25056)
Signed-off-by: Peter Pan <Peter.Pan@daocloud.io>
Peter Pan committed
Sep 27, 2025 at 23:02 UTC
49996cd59789c36d0dc18bcb9dc9454b4d94dde7
1 file changed
+2
-2
vllm/envs.py
+2
-2
@@ -155,7 +155,7 @@ if TYPE_CHECKING:
155
VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
156
VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
157
VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
158
- VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
158
+ VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
159
VLLM_ALL2ALL_BACKEND: Literal["naive", "pplx",
160
"deepep_high_throughput",
161
"deepep_low_latency",
@@ -1220,7 +1220,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
1220
1221
# Port used for NIXL handshake between remote agents.
1222
"VLLM_NIXL_SIDE_CHANNEL_PORT":
1223
- lambda: int(os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5557")),
1223
+ lambda: int(os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")),
1224
1225
# all2all backend for vllm's expert parallel communication
1226
# Available options: