Require peft>=0.17.0 and drop the target_parameters guards

[albertvillanova]

This PR raises the peft floor from 0.13.0 (September 2024) to 0.17.0 (August 2025) and removes the guards that floor made necessary.

Part of #7644.

Depends on

• Reject prompt learning in distillation before wrapping the model #7718

Motivation

peft 0.13.0 is 24 months old, so under the criterion in #7116 it can be dropped without further justification. 0.17.0 is the newest release the same criterion allows today, since it is 14 months old and the rule does not require anything younger than 1 year.

LoraConfig.target_parameters was introduced in peft 0.17.0, so everything #7119 added to keep 0.13 through 0.16 working becomes dead at this floor.

Changes

• Raise peft>=0.13.0 to >=0.17.0 in the peft extra and in dev
• Drop the Version("0.17.0") <= lower bound from the reference-adapter guard in the DPO, GRPO, KTO and RLOO trainers, and the comment sentence explaining it
• Pass target_parameters to LoraConfig directly in get_peft_config again, instead of only when set
• Remove the require_peft_target_parameters test mark and restore @require_peft on the six tests that used it
• Pin peft==0.17.0 in the minimum-versions CI job

The < Version("0.20.0") upper bound stays: it encodes peft#3340, fixed by peft#3350 in 0.20.0, and 0.17 to 0.19 are still supported.

---

Note

Medium Risk
Raises the supported PEFT floor (breaking for 0.13–0.16 installs) while RL trainer reference-adapter logic for 0.17–0.19 is unchanged aside from simplified version checks.

Overview
Raises the minimum PEFT dependency from >=0.13.0 to >=0.17.0 in pyproject.toml (peft extra and dev) and pins peft==0.17.0 in the minimum-versions CI job.

Because LoraConfig.target_parameters exists from 0.17 onward, the PR removes older-version shims: the require_peft_target_parameters pytest marker (MoE tests use @require_peft again), unconditional target_parameters in get_peft_config, and the lower bound on the reference-adapter warning in DPO, GRPO, KTO, and RLOO (still warns for PEFT < 0.20.0 when target_parameters is set).

^{Reviewed by Cursor Bugbot for commit 12027e1. 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.

[albertvillanova]

CI with minimum versions is red for TestAddFusedLMHeadLoss::test_forward_matches_reference_with_peft[prefix_tuning].

This PR requires marking that test as xfail:

• Mark the fused LM head prefix-tuning test xfail on peft 0.17.1 and below #7645

#7646 is independent of the merge order and turns the resulting crash into a clear error across the trainers.

[qgallouedec]

nice, thanks!

[qgallouedec]

👍

[albertvillanova]

The minimum-versions job is red on tests/test_distillation_trainer.py::TestDistillationTrainer::test_peft_prompt_learning_raises:

E       AssertionError: Regex pattern did not match.
E         Expected regex: 'Prompt-learning'
E         Actual message: 'Prefix tuning requires `peft>=0.18.0` when the model is loaded with a device map. Upgrade peft, or load the model with `device_map=None`. See https://github.com/huggingface/peft/issues/2821.'

The rest of the job is green (1 failed, 1997 passed), and the collection error from the previous run is gone.

https://github.com/huggingface/trl/actions/runs/38547934903/job/115798519219

Investigating.

[albertvillanova]

Investigated: the failure is a bug on main, tracked in #7717 and fixed by

• Reject prompt learning in distillation before wrapping the model #7718

I will merge main and push here once that lands.