[albertvillanova]

This PR declares huggingface_hub, numpy, pandas, psutil, pyarrow and pyyaml as dependencies, with a floor for each.

Fix #7653.

Motivation

TRL imports these six packages directly but only gets them through accelerate, datasets and transformers. If one of those drops a package or makes it optional, import trl breaks.

Solution

Under the criterion in CONTRIBUTING.md (#7638), a first floor drops support for nothing and only records what we test.

Each floor is the higher of two versions: the floor accelerate, datasets or transformers already declare, and the oldest release with Python 3.12 wheels. None of the APIs TRL uses requires anything newer. The full suite passes with exactly these floors on Python 3.12: https://github.com/huggingface/trl/actions/runs/38434120299

Package	Floor	Released	Reason
huggingface_hub	0.34.0	Jul 2025	transformers 4.56.2 floor
numpy	1.26.0	Sep 2023	first release with Python 3.12 wheels
pandas	2.1.1	Sep 2023	first release with Python 3.12 wheels
psutil	5.9.4	Nov 2022	first release with Python 3.12 wheels
pyarrow	21.0.0	Jul 2025	datasets 4.7.0 floor
pyyaml	6.0.1	Jul 2023	first release with Python 3.12 wheels

numpy, pandas, psutil and pyyaml are all more than 2 years old. Their floors only exclude releases that cannot be installed on Python 3.12, plus, on Python 3.11, numpy 1.23 to 1.25 and pandas 1.5 to 2.1.0. huggingface_hub and pyarrow match floors the stack already enforces.

pyyaml is imported lazily, but only by the CLI config parsing, which is still a direct use. torch is declared by #7642.

No version guard, shim or warning filter is retired, since nothing in the code depended on a version older than these floors.

Changes

• Add huggingface_hub>=0.34.0, numpy>=1.26.0, pandas>=2.1.1, psutil>=5.9.4, pyarrow>=21.0.0 and pyyaml>=6.0.1 to the project dependencies

---

Note

Low Risk
Packaging-only change that documents and pins already-used libraries; no runtime logic changes in this diff.

Overview
Adds explicit minimum-version dependencies for six packages TRL imports directly but previously only received transitively via accelerate, datasets, and transformers: huggingface_hub>=0.34.0, numpy>=1.26.0, pandas>=2.1.1, psutil>=5.9.4, pyarrow>=21.0.0, and pyyaml>=6.0.1.

This hardens installs so import trl does not break if an upstream package drops or optionalizes one of these libraries. The chosen floors align with existing stack constraints and Python 3.12 wheel availability rather than new API requirements.

^{Reviewed by Cursor Bugbot for commit a4faf33. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[qgallouedec]

Thanks! #7648 moved the minimum-versions job to Python 3.13 after this ran on 3.12; merging main to this branch right now so that it's green (hopefully) tomorrow with 3.13

[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 713143b. Configure here.}