fix: pass missing gts argument to _dump_generations call

[Alexi5000]

Summary

• Fix a runtime TypeError in _dump_generations() that occurs when rollout_data_dir is configured, in both AgentLightningTrainer._train_step and EnvAgentLightningTrainer._train_step
• Make the fix version-robust across verl 0.5.0 and 0.6.0 (see Problem below)
• Remove a leftover print(batch.batch.keys()) debug statement from both call sites

Note: This PR supersedes the duplicate PRs #493 and #522, which addressed the same bug. This version additionally handles the verl 0.5.0 / 0.6.0 signature difference so it does not break the CI-pinned verl 0.5.0.

Problem

RayPPOTrainer._dump_generations() is called from both trainer subclasses, but its signature changed between verl releases:

• verl 0.5.0 (CI-pinned in examples-compat.yml): _dump_generations(self, inputs, outputs, scores, reward_extra_infos_dict, dump_path) — no gts parameter
• verl >=0.6.0: _dump_generations(self, inputs, outputs, gts, scores, reward_extra_infos_dict, dump_path) — adds a required gts (ground truths) positional parameter

The original call sites omitted gts entirely, which raises a TypeError on verl 0.6.0:

TypeError: RayPPOTrainer._dump_generations() missing 1 required positional argument: 'gts'

Conversely, unconditionally passing gts=None breaks on verl 0.5.0:

TypeError: _dump_generations() got an unexpected keyword argument 'gts'

Ground truth is not available in agent-mode training, so None is the correct value when the parameter exists.

Fix

Feature-detect the _dump_generations signature with inspect.signature(...) and only pass gts=None when the gts parameter is present. This works on both verl 0.5.0 and 0.6.0:

dump_kwargs = dict(
    inputs=inputs,
    outputs=outputs,
    scores=scores,
    reward_extra_infos_dict=reward_extra_infos_dict,
    dump_path=rollout_data_dir,
)
if "gts" in inspect.signature(self._dump_generations).parameters:
    dump_kwargs["gts"] = None
self._dump_generations(**dump_kwargs)

Fixes #492

Files Changed

• agentlightning/verl/trainer.py — AgentLightningTrainer._train_step
• contrib/agentlightning/contrib/algorithm/env_verl/trainer.py — EnvAgentLightningTrainer._train_step

Test plan

• Run training with rollout_data_dir configured on verl 0.5.0 — no TypeError
• Run training with rollout_data_dir configured on verl 0.6.0 — no TypeError, gts=None passed
• Run training without rollout_data_dir — no behavior change

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR aligns rollout generation dumping behavior across the two VERL trainer implementations by removing a stray debug print and passing an explicit gts argument into _dump_generations.

Changes:

• Removed print(batch.batch.keys()) debug output during rollout dumping.
• Added gts=None to _dump_generations(...) calls in both trainer implementations.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
contrib/agentlightning/contrib/algorithm/env_verl/trainer.py	Removes debug print and adds gts=None when dumping generations.
agentlightning/verl/trainer.py	Mirrors the same rollout dumping change for consistency.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Alexi5000]

@microsoft-github-policy-service agree

[Alexi5000]

@microsoft-github-policy-service agree