AsyncGRPOTrainer now stable in TRL v1.18.0: generation and training on separate GPUs
SergioPaniego · x · 2026-11-13
Hugging Face's TRL v1.18.0 marks AsyncGRPOTrainer as stable. It runs GRPO's generation and training phases simultaneously on separate GPUs instead of taking turns, improving RL training throughput.
The team's write-up covers how to run it, what you can train with it, and how it compares against the synchronous GRPOTrainer.
Related event: Hugging Face TRL v1.18 makes AsyncGRPOTrainer stable(2 posts)→
More from coding & agent
- Developer downgrades Claude Ultra to Pro and moves all work to Codex over 5-hour rate limits — umeshonai · 2026-11-14
- DocETL: an open-source Python library for agentic document ETL and AI data processing — mdancho84 · 2026-11-14
- Harness Engineering and Graph Engineering: two practices for reliable AI agents — Pavan_Belagatti · 2026-11-14
- CodeRabbit launches Secure: AI-powered code security scans with free first scan — heyshrutimishra · 2026-11-14
- Dev observes Fable 5.2 is making design decisions closer to his own than Opus 5.5 — adonis_singh · 2026-11-14
- LangChain raises $300M Series C at $3B valuation, CapitalG leads — hwchase17 · 2026-11-14