Audited against SGLang main: 1ebe1c57eddd0ea7915b408e35a1b9b33cd10c41 (2026-04-19) Audited against HF transformers main: 631d45082cbd23f4146f1e79c37b3875a3dbc4f4 (2026-04-20) Model: Qwen/Qwen3-30B-...

In Part 1, I explored the fundamental challenge of training-inference context mismatch in reasoning models and prototyped three solutions. While those initial experiments on a single conversation d...

After implementing correct and scalable tokenization and masking1 for multi-turn rollout, there remains a critical challenge to achieve full consistency between training and inference: the context ...

After completing our multi-turn tokenization and masking refactoring, we eliminated a critical bottleneck that was preventing us from building a more consistent and flexible rollout system for our ...

I recently spent two weeks refactoring multiturn tokenization and masking for VeRL. While VeRL already had a functional implementation, what initially seemed like a straightforward refactor turned ...