vllm - 💡(How to fix) Fix [Bug]: P2pNcclConnector ZMQ port conflict when restarting disaggregated_prefill.py

Your current environment

Environment:

• vLLM version: 0.22.0
• Hardware: 2x NVIDIA L40 46GB
• OS: Ubuntu, CUDA 13.0
• Model: Qwen/Qwen3-8B

🐛 Describe the bug

Description: Running examples/disaggregated/disaggregated_prefill.py, the script fails immediately with ZMQError because both the prefill and decode processes attempt to bind the same port:

zmq.error.ZMQError: Address already in use (addr='tcp://<YOUR_IP>:14579')

Root cause: In p2p_nccl_engine.py line 91, the port is calculated as kv_port + port_offset where port_offset=0 for both processes. Both prefill (rank=0) and decode (rank=1) end up binding the same port 14579.

Expected behavior: Each process should bind a different port. The fix should use kv_rank as port_offset to differentiate the two instances.

Steps to reproduce:

1. Run python examples/disaggregated/disaggregated_prefill.py with any model
2. Script fails immediately with ZMQError on the prefill process

disaggregated_prefill_error2.log

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.