Question about reproduce #3
Description
Very interesting work! I am currently trying to reproduce Router-R1 following the official training setup. However, during the first training step, I encountered an issue where the variable queries becomes an empty list at the following line:
ret = access_routing_pool(queries=queries, api_base=self.config.api_base, api_key=self.config.api_key) (line 470 in /router_r1/llm_agent/generation.py)
This usually occurs during the third progress bar update within a single step, and it leads to a training failure. I would like to ask whether the authors have encountered a similar issue, or if there are any configuration details that might prevent queries from becoming empty.
Environment
- Model: Qwen2.5-3B-Instruct
- Batch size: 64
- Api base: https://integrate.api.nvidia.com/v1