vLLM 官方博客T1·Infra
78 pts
Google's ASA and Ranking AI Research teams optimized the open-source TPU stack (vLLM, Tunix, MaxText, Pallas) for RL training of a 0.6B dense model on TPU v7, cutting the full rollout-to-update cycle from roughly 100 minutes to about 56 seconds (~183K tokens/s), a 107x speedup while retaining the full 8,192-token response budget. All changes were upstreamed, so developers can reproduce the path via the merged tpu-recipes recipe without redoing bring-up work.