Yang

### Title: Feat: Implement Streaming Asynchronous Sampling for Zero-Waste Generation ### Description This pull request introduces a streaming asynchronous sampling mechanism designed to significantly improve the efficiency of sample generation...

Enable logits extraction from vLLM for training

4

This PR adds support for directly extracting logits from vLLM during training. By enabling this feature, we avoid redundant computation of sequence logits during reinforcement learning (RL) fine-tuning, which previously...