prime-rl 0.9.0 ships adaptive concurrency, online agentic evals during SFT, CPU optimizer offload
samsja19 · x · 2026-08-26
prime-rl 0.9.0 is out, a release its author calls the one he's proudest of — favorite features added while removing a lot of code:
- Adaptive Concurrency
- Online agentic evals during SFT: an inference server runs next to the SFT trainer to execute agentic evals, reusing the weight-broadcast infra to pick up fresh weight updates
- Local run dashboard: panes for logs/metrics, rollout and config viewing
- Full CPU optimizer offload: gradients, optimizer stats and the optimizer step fully offloaded to CPU, with all communication and computation hidden during backward; at long sequence lengths where attention dominates step time, the offloading can be fully hidden
- Custom kernels shipped
Related event: prime-rl 0.9.0 Released with Adaptive Concurrency and Local Dashboard(3 posts)→
More from coding & agent
- Dev bills $20,000 for an app built with $200 of Claude Code — IgorCarron · 2026-10-15
- Dan Shipper: run product and labs in parallel, even with a labs team of one — danshipper · 2026-10-15
- Gemini 4 Argon wrote 5,000+ lines of pure Python to build its own ray-tracing 3D engine — aitrendz_xyz · 2026-10-15
- Minecraft fully ported from Java to Rust, file by file, server and client included — Vjeux · 2026-10-15
- mitsuhiko builds a full Python IDE for iOS in an afternoon with new AI models — mitsuhiko · 2026-10-15
- How to get Gemini 4 Argon to build games that actually look decent — chongdashu · 2026-10-15