[sergiopaniego]

Each GRPO step opens 64 sessions, and openenv/wordle allows 8 (MAX_CONCURRENT_ENVS). With the official Space as the default, a run without --env-url fails partway into the first step. Now grpo.py requires --env-url, so it fails at startup, and the README and tutorial already say to pass the URL of your copy.

evaluate.py keeps its default, since it plays one game at a time. Both scripts now call get_token() once.

Same change as in the TRL Wordle example (huggingface/trl#7932).

---

Note

Low Risk
Example-script CLI and token-handling only; no changes to training logic, auth flows, or production paths.

Overview
GRPO training now requires --env-url instead of defaulting to the official Wordle Space. A GRPO step needs many concurrent environment sessions (64), which exceeds the public Space limit, so runs used to fail mid-step; missing the flag now fails immediately at startup.

Evaluation still defaults --env-url to the public Space, since it runs one game at a time.

Both grpo.py and evaluate.py call get_token() once and reuse the result when building optional auth headers for private Spaces.

^{Reviewed by Cursor Bugbot for commit 3158933. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 3158933. Configure here.}

[cursor]

Required flag breaks GRPO tests

High Severity

Making --env-url required causes test_grpo_rollout_budget to fail at startup. run_training still invokes grpo.py with only --model and --hub-model-id, so parse_args() exits before any rollout logic runs.

 

^{Reviewed by Cursor Bugbot for commit 3158933. Configure here.}