LocalAI P2P connection failure between Linux server and Windows 11 WSL2 instances

[berry-13]

LocalAI version:
localai/localai:v2.20.1-cublas-cuda12-ffmpeg-core

Environment, CPU architecture, OS, and Version:

1. Linux server: x86 GPU 2060 12 GB, Docker Ubuntu 22.04
2. Windows 11 PC: x86 WSL2, GPU 4070 12 GB, Docker Desktop (with WSL2) Ubuntu image

Describe the bug
Unable to establish P2P connection between LocalAI instances running on a Linux server and a Windows 11 PC with WSL2. The server UI shows 0/0 nodes connected, despite both instances being on the same LAN and able to communicate otherwise

To Reproduce

1. Start the LocalAI container on the Windows 11 PC with WSL2 using the provided docker-compose configuration
2. Start the LocalAI container on the Linux server using the provided docker-compose configuration
3. Open the LocalAI server UI
4. Observe that the UI shows 0/0 nodes connected

Expected behavior
The LocalAI instances should establish a P2P connection, and the server UI should show at least one connected node

Logs
I don't think it's usefull to attach all logs but they mainly say"found node" or "successfully announced"

Additional context

1. Both machines are on the same LAN and can communicate with each other for other applications
2. CUDA and NVCC are running correctly on both systems
3. The llama-cpp-args parameter is not recognized when trying to set a specific port
4. There might be issues with WSL network port forwarding, but it's unclear as the port can't be set manually
5. Both instances are using the same token for authentication
6. The PC instance is using the worker p2p-llama-cpp-rpc command, while the server instance is using whisper-base with the --p2p flag

[JackBekket]

It is also does not work between NAT and also do not work with vast.ai enviroment. We are currently trying our own implementation of libp2p+pubsub instead of edgevpn, but not yet build a solution

[mintyleaf]

@berry-13

#4200 allows you to specify listen and remapped maddress
you can also use https://github.com/mintyleaf/p2boot to create a custom local dht node instead of using IPFS defaults for faster p2p address resolving

can you confirm that this working now?

[berry-13]

@berry-13

#4200 allows you to specify listen and remapped maddress

you can also use https://github.com/mintyleaf/p2boot to create a custom local dht node instead of using IPFS defaults for faster p2p address resolving

can you confirm that this working now?

sorry for the late response but I didn't had much time to look at this in the last months. yes, confirming it now correctly. thanks!

[Podolyakofs]

Good day, I've the same problem - can you share your config pls, oh how you solve this?