|
| 1 | +import functools |
| 2 | +import subprocess |
| 3 | +import sys |
| 4 | +from contextlib import redirect_stdout |
| 5 | +from io import StringIO |
| 6 | +from pathlib import Path |
| 7 | +from unittest import mock |
| 8 | +from unittest.mock import Mock, PropertyMock, call, ANY |
| 9 | + |
| 10 | +import pytest |
| 11 | +import torch |
| 12 | + |
| 13 | +wd = Path(__file__).parent.parent.absolute() |
| 14 | + |
| 15 | + |
| 16 | +@functools.lru_cache(maxsize=1) |
| 17 | +def load_generate_script(): |
| 18 | + sys.path.append(str(wd)) |
| 19 | + |
| 20 | + import generate |
| 21 | + |
| 22 | + return generate |
| 23 | + |
| 24 | + |
| 25 | +@pytest.mark.parametrize("B", (1, 2)) |
| 26 | +def test_generate(B): |
| 27 | + generate = load_generate_script() |
| 28 | + |
| 29 | + T, C = 5, 3 |
| 30 | + logits = torch.randn(B, T, C) |
| 31 | + input_idx = torch.randint(10, size=(B, T)) |
| 32 | + |
| 33 | + model = Mock(return_value=logits) |
| 34 | + max_new_tokens = 20 |
| 35 | + |
| 36 | + multinomial_results = [] |
| 37 | + original_multinomial = torch.multinomial |
| 38 | + |
| 39 | + def multinomial(*args, **kwargs): |
| 40 | + out = original_multinomial(*args, **kwargs) |
| 41 | + multinomial_results.append(out) |
| 42 | + return out |
| 43 | + |
| 44 | + with mock.patch("torch.multinomial", multinomial): |
| 45 | + out = generate.generate(model, input_idx, max_new_tokens, max_seq_length=10) |
| 46 | + |
| 47 | + assert out.shape == (B, T + max_new_tokens) |
| 48 | + multinomial_results = torch.hstack(multinomial_results) |
| 49 | + expected = torch.cat((input_idx, multinomial_results), dim=1) |
| 50 | + assert out.shape == expected.shape |
| 51 | + torch.testing.assert_close(out, expected) |
| 52 | + |
| 53 | + |
| 54 | +def test_main(tmp_path, monkeypatch): |
| 55 | + generate = load_generate_script() |
| 56 | + |
| 57 | + checkpoint_path = tmp_path / "ckpt" |
| 58 | + checkpoint_path.touch() |
| 59 | + tokenizer_path = tmp_path / "tokenizer" |
| 60 | + tokenizer_path.touch() |
| 61 | + |
| 62 | + class FabricMock(PropertyMock): |
| 63 | + @property |
| 64 | + def device(self): |
| 65 | + return torch.device("cpu") |
| 66 | + |
| 67 | + fabric_mock = FabricMock() |
| 68 | + monkeypatch.setattr(generate.L, "Fabric", fabric_mock) |
| 69 | + model_mock = Mock() |
| 70 | + monkeypatch.setattr(generate.LLaMA, "from_name", model_mock) |
| 71 | + load_mock = Mock() |
| 72 | + monkeypatch.setattr(generate.torch, "load", load_mock) |
| 73 | + tokenizer_mock = Mock() |
| 74 | + tokenizer_mock.return_value.encode.return_value = torch.tensor([[1, 2, 3]]) |
| 75 | + tokenizer_mock.return_value.decode.return_value = "foo bar baz" |
| 76 | + monkeypatch.setattr(generate, "Tokenizer", tokenizer_mock) |
| 77 | + generate_mock = Mock() |
| 78 | + generate_mock.return_value = torch.tensor([[3, 2, 1]]) |
| 79 | + monkeypatch.setattr(generate, "generate", generate_mock) |
| 80 | + |
| 81 | + num_samples = 2 |
| 82 | + out = StringIO() |
| 83 | + with redirect_stdout(out): |
| 84 | + generate.main( |
| 85 | + checkpoint_path=checkpoint_path, |
| 86 | + tokenizer_path=tokenizer_path, |
| 87 | + model_size="1T", |
| 88 | + accelerator="litpu", |
| 89 | + temperature=2.0, |
| 90 | + top_k=2, |
| 91 | + num_samples=num_samples, |
| 92 | + ) |
| 93 | + |
| 94 | + model_mock.assert_called_once_with("1T") |
| 95 | + load_mock.assert_called_once_with(checkpoint_path) |
| 96 | + tokenizer_mock.assert_called_once_with(tokenizer_path) |
| 97 | + assert len(tokenizer_mock.return_value.decode.mock_calls) == num_samples |
| 98 | + assert torch.allclose(tokenizer_mock.return_value.decode.call_args[0][0], generate_mock.return_value) |
| 99 | + model = model_mock.return_value |
| 100 | + assert fabric_mock.mock_calls == [ |
| 101 | + call(accelerator="litpu", devices=1), |
| 102 | + call().device.__enter__(), |
| 103 | + call().device.__exit__(None, None, None), |
| 104 | + call().setup_module(model), |
| 105 | + ] |
| 106 | + model = fabric_mock.return_value.setup_module.return_value |
| 107 | + assert ( |
| 108 | + generate_mock.mock_calls |
| 109 | + == [call(model, ANY, 50, model.config.block_size, temperature=2.0, top_k=2)] * num_samples |
| 110 | + ) |
| 111 | + # only the generated result is printed to stdout |
| 112 | + assert out.getvalue() == "foo bar baz\n" * num_samples |
| 113 | + |
| 114 | + |
| 115 | +def test_cli(): |
| 116 | + cli_path = wd / "generate.py" |
| 117 | + output = subprocess.check_output([sys.executable, cli_path, "-h"]) |
| 118 | + output = str(output.decode()) |
| 119 | + assert "Generates text samples" in output |
0 commit comments