Issue evidence · Run · huggingface/trl

GRPO/RLOO on Trainium: each new prompt length bucket recompiles generation for ~12 minutes

GitHub issue: https://github.com/huggingface/trl/issues/7733

Not run

On AWS Trainium, each new prompt-length bucket triggers a roughly 12-minute recompile of generation.

What Badgr ran

Nothing was run: it needs Trainium hardware, which Badgr's providers do not offer.

What came back

  • The report gives a GRPO example based on the Trainium guide.

Not run.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence