In rollout phase, input data is evenly split among VLLM workers, but response lengths vary, causing load imbalance and idle workers. Also, early training may produce repetitive, useless outputs, wasting resources. （like kimi1.5 Partial Rollouts for Long CoT RL ）

Proposal:

Use a shared queue for all VLLM workers to ensure better load balancing.

Add a feature to stop repetitive responses.

Would the community be interested in implementing this?