[Reproduction] GPU nodes blocked from internet on Slurm cluster

[Fallforgotten]

I’m reproducing InternAgent on a Slurm HPC cluster using a locally deployed gpt-oss gateway. The main blocker is that GPU nodes have no outbound internet. At the moment, I’m unable to obtain an HTTP(S) proxy for outbound access. I was wondering whether you’ve encountered a similar situation. I’d really appreciate any advice you could share.

[JiakangYuan]

@Fallforgotten In the current version of InternAgent, it is necessary to retrieve papers from Semantic Scholar. This operation happens before the idea generation phase. Therefore, the two processes can be decoupled. The paper retrieval can be performed on a machine with internet access, while the idea generation and subsequent experiments can be conducted on the server where gpt-oss is deployed.

To implement this:

• First, run the code in launch_dolphin.py before Line 202. This will save the retrieved papers to a json file.
• Afterwards, on the server, you can comment out Line 196-Line 202 and use the deployed model to execute the subsequent code. Note that you need to modify Line 212 to pass in the file path of the saved json file.

[Fallforgotten]

@JiakangYuan Thanks for your suggestion. Much appreciated!