about llm endpoint

[kyk666123]

If I want to use VERL for training, and within the agent pipeline, some nodes use local models while others call commercial APIs—while only training the open-source model—do I need to pass an agl.LLM object (which wraps the actual endpoint and model name of the commercial LLM) via the resources parameter when initializing the agent class?
Additionally, for the local model, are its endpoint and model name directly specified in VERL’s configuration file, and will there automatically be a main_llm field in resources that I can access?
Finally, do both the local model and the commercial API model require calling llm.get_base_url() to obtain their final endpoints?
Thank you very much for your time and guidance—I really appreciate any clarification you can provide!

[kyk666123]

I’ve also previously encountered issues where certain spans showed warnings like “no parent” or “cannot log.” I suspect this happened because, when calling the commercial API, I didn’t wrap it in an agl.LLM object but instead invoked it directly.

[ultmaster]

If I want to use VERL for training, and within the agent pipeline, some nodes use local models while others call commercial APIs—while only training the open-source model—do I need to pass an agl.LLM object (which wraps the actual endpoint and model name of the commercial LLM) via the resources parameter when initializing the agent class?

The agl.LLM object is the communication object between algorithm and agent, that specifies the LLM that is going to be tuned. If you are also using some 3rd-party LLMs simultaneously that is not being tuned, you don't need to put it into an agl.LLM class.

Additionally, for the local model, are its endpoint and model name directly specified in VERL’s configuration file, and will there automatically be a main_llm field in resources that I can access?

The endpoint and model is generated by VERL on-the-fly because it's in the process of being tuned by VERL.

Finally, do both the local model and the commercial API model require calling llm.get_base_url() to obtain their final endpoints?
Thank you very much for your time and guidance—I really appreciate any clarification you can provide!

For the model being tuned, yes. Otherwise, no.

I’ve also previously encountered issues where certain spans showed warnings like “no parent” or “cannot log.” I suspect this happened because, when calling the commercial API, I didn’t wrap it in an agl.LLM object but instead invoked it directly.

I think that's because you didn't use get_base_url?