Summary
This is a thin slot: two curated reposts, both from @ContentCase and both resurfacing June material, not new news. The real signal is RL-Kernel, an open-source GPU kernel library for RL post-training (GRPO and PPO). It is built on FlashInfer and ships custom sampling, prefix-shared attention and TMA kernels. The authors claim up to 163x speedups on hot training components, which matters because rollout and loss computation dominate the cost of RL post-training. The wiki already covered it in June, so treat this as a reminder, not a fresh release. The second post is an opinion piece on RL interview questions and PhD-versus-industry careers. It is career content, not research.
Posts
- RL-Kernel: GPU kernels for GRPO/PPO post-training (@ContentCase · repo · Zhihu write-up). This is a FlashInfer-integrated kernel library that targets the RL training loop rather than inference. It aims for bitwise-matching numerics between training and inference engines, and claims up to 163x speedups plus lower memory on high-frequency components. The 163x is a per-op figure, not end-to-end. See the June summary and the RL for LLMs concept page.
- RL Interview Questions 2026 (@ContentCase). An X Article on RL interview prep and whether to skip the PhD for industry. Click through to read. It is career advice with no technical claim. Skip unless you are hiring.