Question about environment variable

[PhoenixGS]

Hello, I couldn't find any documentation regarding the environment variable QWEN_URL in the code repository. Do we need to deploy Qwen models on our own machines, such as using the vLLM framework? Also, which specific Qwen model should we deploy? Thank you!

[wangyu-ustc]

Yes you need to deploy your own Qwen model and set QWEN_URL as something like "localhost:8001/v1" using vLLM framework. I deployed Qwen3-32B for training.

[PhoenixGS]

Thank you! If we use four nodes for training, should the Qwen model be deployed on the other nodes or on the master node directly?

[wangyu-ustc]

The Qwen model should be deployed on a separate node, so I actually used five nodes for training, one is used to deploy Qwen3-32B model and it outputs a api url. Using this url we can run the training on the remaining four nodes.

Otherwise you can use the API from Qwen or Openrouter as the fifth node, but that might be too expensive for training.