[albertvillanova]

This PR fixes AsyncGRPOTrainer raising AttributeError: 'AsyncGRPOTrainer' object has no attribute 'args' when it is built without a train_dataset.

FAILED tests/experimental/test_async_grpo_trainer.py::TestAsyncGRPOTrainer::test_environment_owned_data_requires_max_steps - AttributeError: 'AsyncGRPOTrainer' object has no attribute 'args'

Bug introduced by PR:

• Environment-owned dataset: make train_dataset optional #6349

The feature #6349 added has never worked.

Motivation

AsyncGRPOTrainer.__init__ reads self.args while building the placeholder dataset, but self.args is only set by Trainer.__init__, which runs a few lines later. Every environment-owned run, meaning train_dataset=None together with an environment_factory, fails there before it can reach either the intended max_steps ValueError or the placeholder dataset construction.

The existing regression test, TestAsyncGRPOTrainer::test_environment_owned_data_requires_max_steps, does cover this, but it never ran: tests/experimental is excluded from the regular suite through norecursedirs, and in tests-experimental.yml the whole class is gated behind is_ampere_or_newer(), which was false on the T4 runner. Moving that job to an L40S in #6741 makes the test execute for the first time and surfaces the failure.

Solution

Use the local args instead of self.args in the pre-super().__init__() region, consistent with every other config read in that block. AsyncDistillationTrainer is unaffected: it assigns self.args = args before its own super().__init__().

Changes

• Read max_steps, per_device_train_batch_size and gradient_accumulation_steps from the local args when building the placeholder dataset in AsyncGRPOTrainer.__init__

---

Note

^{Cursor Bugbot is generating a summary for commit cfa4913. 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]

I am merging this as a hotfix.