Is it possible to bind SGLang Scheduler to 0.0.0.0 or 127.0.0.1 for CRIU C/R Purpose

[ZeroExistence]

Problem Description

My current concern is that it seems that SGLang does not have any way to adjust scheduler IP address bind to first network interface on container or machine.

When I run SGLang, only the SGLang inference endpoint is configurable by --host flag as shown below.

## Run command
NCCL_P2P_DISABLE=1 http_proxy='' https_proxy='' python3 -m sglang.launch_server --model-path meta-llama/Llama-3.1-8B --host 0.0.0.0 --log-level debug

root@dgtest-sglang:/sgl-workspace/sglang# ss -tanlup
Netid            State             Recv-Q            Send-Q                       Local Address:Port                         Peer Address:Port            Process
tcp              LISTEN            0                 2048                               0.0.0.0:30000                             0.0.0.0:*                users:(("python3",pid=1520,fd=54))
tcp              LISTEN            0                 4096                            10.8.1.167:45475                             0.0.0.0:*                users:(("sglang::schedul",pid=1659,fd=83))
tcp              LISTEN            0                 4096                            10.8.1.167:42277                             0.0.0.0:*                users:(("sglang::schedul",pid=1659,fd=95))
tcp              LISTEN            0                 4096                            10.8.1.167:45223                             0.0.0.0:*                users:(("sglang::schedul",pid=1659,fd=91))
tcp              LISTEN            0                 4096                            10.8.1.167:39335                             0.0.0.0:*                users:(("sglang::schedul",pid=1659,fd=87))
tcp              LISTEN            0                 4096                                     *:30142                                   *:*                users:(("sglang::schedul",pid=1659,fd=77))

## Process List
root@dgtest-sglang:/sgl-workspace/sglang# ps -ef | grep 1659
root        1659    1520 97 07:04 pts/0    00:00:51 sglang::scheduler
root        2037    1659 12 07:04 pts/0    00:00:04 /usr/bin/python3 /usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_worker/__main__.py --pickler=torch._inductor.compile_worker.subproc_pool.SubprocPickler --kind=fork --workers=32 --parent=1659 --read-fd=96 --write-fd=99 --torch-key=xxx
root        2147    2037  0 07:04 pts/0    00:00:00 /usr/bin/python3 /usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_worker/__main__.py --pickler=torch._inductor.compile_worker.subproc_pool.SubprocPickler --kind=fork --workers=32 --parent=1659 --read-fd=96 --write-fd=99 --torch-key=xxx
root        2149    2037  0 07:04 pts/0    00:00:00 /usr/bin/python3 /usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_worker/__main__.py --pickler=torch._inductor.compile_worker.subproc_pool.SubprocPickler --kind=fork --workers=32 --parent=1659 --read-fd=96 --write-fd=99 --torch-key=xxx

I would like to set SGLang scheduler to bind on same address configured on --host or atleast bind to 127.0.0.1. I am currently not sure if the open ports by scheduler is required for multi-node/pod setup.

Is it possible to do this or any way to configure it?

My use-case is basically like this.

1. Run SGLang with 192.0.0.1.
2. Checkpoint container with criu.
3. Restore SGLang with new IP address (ex. 192.0.0.2).
4. Restore process succeed.

Right now, I am failing to restore process and forced to re-use the original address during checkpointing. It is not ideal since the idea is to allow SGLang to scale-out using the same set of checkpoint.

Thank you very much!

[ZeroExistence]

Actually, I have found the configuration variable that I am looking for to bind these processes to loopback address instead of eth0. I will consider this discussion as resolved. Thank you.