@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 22e718ff1a
Fork 0

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 16, 19:30 UTC

[Misc] Revive to use loopback address for driver IP (#7091)

Signed-off-by: Rui Qiao <ruisearch42@gmail.com>

Rui Qiao committed Aug 2, 2024 at 15:50 UTC 22e718ff1a51930231d87c89d6c43676af59860b
1 file changed +10
vllm/executor/ray_gpu_executor.py
+10
@@ -216,6 +216,16 @@ class RayGPUExecutor(DistributedGPUExecutor):
216 self._run_workers("update_environment_variables",
217 all_args=all_args_to_update_environment_variables)
218
219 + if len(node_gpus) == 1:
220 + # in single node case, we don't need to get the IP address.
221 + # the loopback address is sufficient
222 + # NOTE: a node may have several IP addresses, one for each
223 + # network interface. `get_ip()` might return any of them,
224 + # while they might not work for communication inside the node
225 + # if the network setup is complicated. Using the loopback address
226 + # solves this issue, as it always works for communication inside
227 + # the node.
228 + driver_ip = "127.0.0.1"
229 distributed_init_method = get_distributed_init_method(
230 driver_ip, get_open_port())
231