dirl-diffusion-rl

【部署与运维】Enable effective RL for diffusion language models via DiPO (unbiased GRPO for dLLMs) and framework optimizations. FlexAttention accelerates blockwise training, LMDeploy optimizes inference, achieving training-inference consistency—improving dLLM math performance to rival larger autoregressive models.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准