model training

vLLM V1 Skews RL Results: Why Inference Correctness Beats Speed

Upgrading vLLM from V0 to V1 causes output inconsistencies in RL. If inference frameworks trade accuracy for speed, dependent models silently drift.