运行时报错端口被占用，但是用命令检查时并没有被占用，换其他空闲端口也没用

[lumo-s]

[lumo-s]

RuntimeError: Address already in use

[AsafCodes]

This is a common issue with distributed PyTorch training.

Based on the U2PL repository structure, this bug is likely related to the distributed training setup used in their shell scripts.

Root Causes & Solutions:
1. Process Cleanup Issue (Most Likely)
The previous training run may have left zombie processes that still hold the port, even though netstat or similar commands don't show them.

# Kill any remaining Python processes
pkill -f python
# Or more specifically for PyTorch distributed
pkill -f "torch.distributed"
# Wait a moment, then try again

2. Port Range Conflict
PyTorch distributed training doesn't just use the specified port - it uses a range of ports starting from your specified port.
Solutions:

Try ports with larger gaps (e.g., if using 12345, try 12400 or 13000)
Use ports above 20000 to avoid common service ranges

3. Script Configuration Issues
Looking at U2PL's training scripts, they likely use torch.distributed.launch or similar. The issue might be:
Check the training scripts:

# Set environment variables explicitly
export MASTER_ADDR=localhost
export MASTER_PORT=12345
export WORLD_SIZE=1
export RANK=0