Issue evidence · Run · huggingface/trl
GitHub issue: https://github.com/huggingface/trl/issues/7674
Reproduced, workaround proven
With log_completions=True, GRPOTrainer crashes at the first logging step whose rollouts include a tool call, a regression in TRL 1.15.0.
Ran the reporter's repro.py unchanged on an RTX 4090 through Badgr (pytorch 2.8.0 cu126 image, transformers 5.19.0, pyarrow 26.0.0), both modes (content_blocks, arg_types), on TRL 1.14.2 and then 1.15.0.
Pinning trl==1.14.2 is the workaround that passed here. main was not tested. The same script hit unrelated errors on a Mac CPU, so only the GPU run is reported.
Checked 2026-10-14 in Badgr’s local development environment.
← All issue evidence