vLLM achieves bitwise-consistent RL training and inference, open-sources agent skill

vLLM's new blog addresses how inference optimizations alter recomputed logprobs in RL training, causing KL divergence and performance drops. The bitwise-consistent solution achieved exactly zero KL over 100 RL steps on DeepSeek-V4-Flash across 32 GB200s, retained 87.5%-91.3% throughput under quantization, and shipped as an open-source agent skill (PR #61551).

2026-10-15 ~ 2026-10-15 · 4 related posts

• vLLM Achieves Bitwise-Consistent RL Training, Fixing KL Blowups and Score Collapse — vllm_project · 2026-10-15
• vLLM achieves bitwise-consistent RL training: DeepSeek-V4-Flash hits 0 diff over 100 steps — vllm_project · 2026-10-15
• vLLM's bitwise-consistent RL: zero KL on Nemotron NVFP4, 91.3% throughput on DeepSeek-V4-Flash — DynamicWebPaige · 2026-10-15
• vLLM open-sources agent skill for bitwise-identical train-inference logprobs in RL — vllm_project · 2026-10-15