Description:
Recently, Gradient released a new decentralized inference system called Parallax, which could serve as a comparison target for our paper. I have been benchmarking Parallax and Petals for performance evaluation.
When running Parallax on the same machine (no cross-network communication), everything works fine. However, when attempting to test across devices, the system fails with the following network error:
relay does not work on this type of NAT (Symmetric NAT)
This appears to be caused by Lattica, the underlying networking layer used by Parallax, which cannot operate under Symmetric NAT.
I have reproduced this issue on two setups:
Yotta AI platform (GPU 4090 server)
Chameleon Cloud (RTX 6000 node)
Both environments encounter the same NAT issue when running distributed inference across devices.
Expected behavior:
Comprehensive evaluation for parallax.