Blog

Release notes and engineering writeups for RL-Kernel.

• RL-Kernel v0.1.0: Achieving Bitwise Consistency Between Training and Rollout

  ---

  How vime and RL-Kernel align numerical execution between training and rollout, with operator ablations and 200-step ROCm and CUDA results.

• DSV4 Flash MoE Train-Inference Consistency Roadmap

  ---

  A roadmap for DeepSeek-V4-Flash MoE consistency, covering expert routing, tensor layouts, and reduction order across training and inference.

• Announcing RL-Kernel for vime: Faster and Leaner linear_logp for Full RL Rollouts

  ---

  Introducing the fused tensor-parallel linear_logp integration for vime to reduce CUDA time and memory use in full RL rollout training.