Proposal: module-aware fast local validation for contributors

Description

AutoGluon is a large multi-module project. Contributors often change one area (for example, Tabular) but the full test suite can take several hours. This proposal adds a fast, module-aware local validation command to catch common problems before a pull request reaches CI.

Initial scope: tabular only. This does not replace full CI, platform tests, benchmark tests, or GPU tests.

Proposed files

• CI/quick_validate.py — CLI entry point, invoked as uv run python CI/quick_validate.py --module tabular.
• CI/smoke_tests/test_tabular_smoke.py — a tiny end-to-end test that trains TabularPredictor on a small in-memory dataset and verifies prediction succeeds.
• CONTRIBUTING.md — documents the command, expected duration, and the checks intentionally left to CI.

Proposed behavior

For --module tabular, the command would:

1. Run Ruff formatting and lint checks on the relevant paths.
2. Run the existing fast, targeted Tabular unit tests.
3. Run the small train/predict smoke test.
4. Print a summary of passed checks and explicitly list expensive checks not run locally.

Example:

uv run python CI/quick_validate.py --module tabular

Example summary:

✓ Ruff format and lint passed
✓ Targeted Tabular tests passed
✓ Tabular train/predict smoke test passed

Not run: full cross-platform CI, benchmarks, GPU tests, and the complete test suite.

Acceptance criteria

• Works with the documented supported Python versions and does not add a new runtime dependency.
• Uses the repository's existing uv, Ruff, and pytest workflow.
• Has a small, deterministic smoke test with a short time limit and cleanup of temporary model artifacts.
• Fails with actionable messages if linting, targeted tests, or the smoke test fails.
• Does not alter the existing full-CI or benchmark workflows.

Why this helps

This creates a quick feedback loop for first-time and experienced contributors, reduces avoidable CI failures, and makes the expected pre-PR validation steps explicit. Additional modules such as TimeSeries and MultiModal can be considered later, after the Tabular implementation is reviewed.