[JiahaoTanXX (JiahaoTanXX)]

Require VERL's default data.val_batch_size=null at the start of run_ppo, with an actionable error before importing the VERL runtime helper or accessing Ray. Document the full-dataset validation requirement and correct YAML/CLI syntax.

Implements the maintainer's suggestion in #663. The dedicated assertion tests have been removed as requested in review.

Validation on macOS, Python 3.12.13, VERL 0.8.0 and PyTorch 2.13.0, using the frozen CPU dependencies:

• .venv/bin/pytest -v --durations=20 tests: 190 passed, 4 skipped (CUDA/vLLM-dependent cases).
• .venv/bin/pyright: 0 errors, 0 warnings.
• .venv/bin/ruff check ., .venv/bin/ruff format --check ., .venv/bin/python scripts/check_headers.py, .venv/bin/pre-commit run --all-files --show-diff-on-failure, and git diff --check: passed.
• .venv/bin/mkdocs build --strict: passed.
• uv build --no-sources: source distribution and wheel built.

No live GPU training was run.

AI assistance: prepared and validated with OpenAI Codex.

[Copilot]

🟢 Approval recommended

No active review issues are reported for the change enforcing full-dataset validation.

No open issues 🎉

What changed in this PR

This PR makes Agent Lightning’s full-dataset validation requirement explicit, following the maintainer recommendation in #663.

Changes:

• Add an early assertion requiring data.val_batch_size to be None, before importing the VERL runtime helper or accessing Ray.
• Add six regression cases covering the accepted default and rejected non-null values.
• Document the requirement and correct YAML and CLI syntax.

File	Description
tests/​verl/​test_entrypoint.py	Tests valid and invalid batch-size values at the Ray startup boundary.
docs/​20-trainer-configuration.md	Explains full-dataset validation and null configuration syntax.
agentlightning/​verl/​entrypoint.py	Rejects non-null validation batch sizes with an actionable message.

_{Review effort: Balanced}

[Zhiyuan He (hzy46)]

Thanks. Please remove the tests. We do not need a specific test for an assert.