tests/test_evaluate.py

# Copyright Lightning AI. Licensed under the Apache License 2.0, see LICENSE file.

import subprocess
from contextlib import redirect_stdout
from dataclasses import asdict
from io import StringIO
from unittest import mock

import pytest
import torch
import yaml

import litgpt.eval.evaluate as module
from litgpt import GPT, Config
from litgpt.scripts.download import download_from_hub


@pytest.mark.flaky(reruns=3)
def test_evaluate_script(tmp_path):
    ours_config = Config.from_name("pythia-14m")
    download_from_hub(repo_id="EleutherAI/pythia-14m", tokenizer_only=True, checkpoint_dir=tmp_path)
    checkpoint_dir = tmp_path / "EleutherAI" / "pythia-14m"
    ours_model = GPT(ours_config)
    torch.save(ours_model.state_dict(), checkpoint_dir / "lit_model.pth")
    with open(checkpoint_dir / "model_config.yaml", "w", encoding="utf-8") as fp:
        yaml.dump(asdict(ours_config), fp)

    stdout = StringIO()
    with redirect_stdout(stdout), mock.patch("sys.argv", ["eval/evaluate.py"]):
        with pytest.raises(ValueError) as excinfo:
            module.convert_and_evaluate(
                checkpoint_dir,
                out_dir=tmp_path / "out_dir",
                device=None,
                dtype=torch.float32,
                limit=5,
                tasks="logiqa",
                batch_size=0,  # Test for non-positive integer
            )
        assert "batch_size must be a positive integer, 'auto', or in the format 'auto:N'." in str(excinfo.value)

        with pytest.raises(ValueError) as excinfo:
            module.convert_and_evaluate(
                checkpoint_dir,
                out_dir=tmp_path / "out_dir",
                device=None,
                dtype=torch.float32,
                limit=5,
                tasks="logiqa",
                batch_size="invalid",  # Test for invalid string
            )
        assert "batch_size must be a positive integer, 'auto', or in the format 'auto:N'." in str(excinfo.value)

    stdout = StringIO()
    with redirect_stdout(stdout), mock.patch("sys.argv", ["eval/evaluate.py"]):
        module.convert_and_evaluate(
            checkpoint_dir,
            out_dir=tmp_path / "out_dir",
            device=None,
            dtype=torch.float32,
            limit=5,
            tasks="logiqa",
            batch_size=1,  # Valid case
        )
    stdout = stdout.getvalue()
    assert (tmp_path / "out_dir" / "results.json").is_file()
    assert "logiqa" in stdout
    assert "Metric" in stdout
    assert "Loading checkpoint shards" not in stdout


def test_cli():
    args = ["litgpt", "evaluate", "-h"]
    output = subprocess.check_output(args)
    output = str(output.decode())
    assert "Evaluate a model with the LM Evaluation Harness" in output