aboue "Cannot log to store."

[kyk666123]

When I was training, I encountered the following message:
"No headers found in {len(subtree_spans)} subtree spans of root {root_span_id}. Cannot log to store."
After that, all rollout.final_reward values became None. What could be the possible causes?

[ultmaster]

How did you write your rollout function?

This issue is normally caused by using the get_base_url in the wrong way so that the LLM endpoint used doesn't contain rollout_id information to tell which rollout the LLM request comes from.

You also mentioned that there is no final reward. That's a different issue.

I would suggest going through the debugging tutorial first: https://microsoft.github.io/agent-lightning/stable/tutorials/debug/

[kyk666123]

I'm running an example of sql_agent, which includes both SqlAgent and LitsqlAgent. In this setup, I replaced the check component of SqlAgent with the Qwen Bailian large model API, while keeping all other components using the open-source model qwen2.5-coder-3b. I did not modify LitsqlAgent at all. During training, the final_reward shows as None, even though rewards are indeed being generated—this is evident because I can observe the average reward. The only changes compared to the previous version are that final_reward became None, and the warning mentioned in my initial question started appearing.
Additionally, is it feasible to replace certain components of the agent with a commercial API (like Qwen Bailian), keep the rest of the components as open-source models (e.g., qwen2.5-coder-3b), and then use agent-match to specify this hybrid configuration—training only the open-source model parts while leaving the commercial API components frozen?

[ultmaster]

During training, the final_reward shows as None, even though rewards are indeed being generated—this is evident because I can observe the average reward.

I'm not sure I fully get the situation. Could you at least check the dashboard and see what the traces look like?

Additionally, is it feasible to replace certain components of the agent with a commercial API (like Qwen Bailian), keep the rest of the components as open-source models (e.g., qwen2.5-coder-3b), and then use agent-match to specify this hybrid configuration—training only the open-source model parts while leaving the commercial API components frozen?

Yes. That's exactly an intended use case.

[ligang-cs]

When I was training, I encountered the following message: "No headers found in {len(subtree_spans)} subtree spans of root {root_span_id}. Cannot log to store." After that, all rollout.final_reward values became None. What could be the possible causes?

I met the same issue ("cannot log to store"). Increasing the span attribute limit addressed it: export OTEL_SPAN_ATTRIBUTE_COUNT_LIMIT=1024.

[ultmaster]

@ligang-cs I was not aware of this environment variable. Thank you for pointing this out!