TRL v1.18.0
AsyncGRPOTrainer is now a stable trainer in TRL v1.18.0, moved to the top-level package with a public rollout loop.
The Deploy Log | Products | Free, no signup
TRL changelog
5 versions of TRL from Hugging Face in The Deploy Log. From to . On the median a new version every 7 days, publisher's dates only; the longest wait was 29 days. The newest is v1.18.0. 1 version in the last 30 days. 1 row with no publisher's date. Each links the publisher's own page.
| Version | Publisher's date | Days after the one before |
|---|---|---|
| v1.18.0 | 29 | |
| v1.16.0 | 7 | |
| v1.15.0 | 2 | |
| v1.14.2 | ||
| v1.0 | No date |
TRL v1.16.0
TRL v1.16.0 adds AWS Trainium training support with pad_to_multiple_of defaulting to auto.
TRL v1.15.0
Fused LM head on by default, up to 6.9x longer sequences on the same GPU.
TRL v1.14.2
Patch release fixes two cases of silently wrong training and three crashes, including end-of-turn handling for GRPO, RLOO, and Distillation.
TRL v1.0
TRL v1.0 is released with a stable core following semantic versioning and an experimental layer, and the library is downloaded 3 million times a month.
One line every time AI actually ships, around the clock, free on Threads and Bluesky. Every one is in The Deploy Log with its source.