verl — Reinforcement Learning Infrastructure
I contribute to verl, focusing on compatibility and reliability in reinforcement learning training and rollout workflows.
Selected merged contribution
PR #7625 — Fix TRL value-head model imports · Merged August 31, 2026.
Fixed the remaining value-head model imports in verl/utils/model.py after TRL moved AutoModelForCausalLMWithValueHead into trl.experimental.ppo. This restores FSDP value-head critic loading with newer TRL releases while preserving compatibility with older releases through an import fallback.
Bug reporting and investigation
Issue #7624 — vLLM startup failure with chunked prefill disabled. Reported and reproduced a rollout startup regression that appeared when the token budget was smaller than the model context length, with a minimal reproducer and root-cause analysis.