[Question] Deploy multi node inference on wsl2, master is trying to connect worker.

[o0olele]

env

• sglang version: 0.4.4.post3
• python version: 3.10
• wsl version: wsl2, network mode-nat

master node

• command

  hostname -I # eg aa.aa.aa.aa
  TP_SOCKET_IFNAME=th0 GLOO_SOCKET_IFNAME=th0 NCCL_SOCKET_IFNAME=th0 python -m sglang.launch_server 
  --model-path path/to/qwq --tp 2 --dist-init-addr aa.aa.aa.aa:20000 --nnodes 2 --node-rank 0 

• add portproxy follow here
  master host ip is: bb.bb.bb.bb

worker node

• command

  hostname -I # eg cc.cc.cc.cc
  TP_SOCKET_IFNAME=th0 GLOO_SOCKET_IFNAME=th0 NCCL_SOCKET_IFNAME=th0 python -m sglang.launch_server 
  --model-path path/to/qwq --tp 2 --dist-init-addr bb.bb.bb.bb:20000 --nnodes 2 --node-rank 1

result

• the worker node connected to the master, use netstat and show tcp connection is connected with port 20000.
• but master node subprocess exited after a while:

  RuntimeError: Gloo connectFullMesh failed with [/pytorch/third_party/gloo/gloo/transport/tcp/pair.cc:144]

• then i used tcpdump capture the tcp packets, and found that the master was triying to establish connection to the worker node.

  aa.aa.aa.aa > cc.cc.cc.cc: Flags [S], seq 34745689, win 64240, options [mss 1460, sackOk, TS val...]

• cc.cc.cc.cc is obviously unreachable on master.
• Could someone explain the reason for connecting to the worker? Also, where can I find the relevant code section?