| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import argparse |
| 4 | import json |
| 5 | from pathlib import Path |
| 6 | |
| 7 | import pytest |
| 8 | from transformers import AutoTokenizer, PreTrainedTokenizerBase |
| 9 | |
| 10 | from vllm.benchmarks.datasets import get_samples |
| 11 | |
| 12 | |
| 13 | @pytest.fixture(scope="session") |
| 14 | def hf_tokenizer() -> PreTrainedTokenizerBase: |
| 15 | return AutoTokenizer.from_pretrained("openai-community/gpt2") |
| 16 | |
| 17 | |
| 18 | def _write_jsonl(path: Path, n_rows: int) -> None: |
| 19 | with path.open("w") as f: |
| 20 | for i in range(n_rows): |
| 21 | f.write(json.dumps({"prompt": f"row {i}: unique prompt content."}) + "\n") |
| 22 | |
| 23 | |
| 24 | def _args_for_custom(dataset_path: str, seed: int) -> argparse.Namespace: |
| 25 | return argparse.Namespace( |
| 26 | dataset_name="custom", |
| 27 | dataset_path=dataset_path, |
| 28 | disable_shuffle=False, |
| 29 | num_prompts=30, |
| 30 | custom_output_len=32, |
| 31 | skip_chat_template=True, |
| 32 | no_oversample=False, |
| 33 | seed=seed, |
| 34 | request_id_prefix="", |
| 35 | ) |
| 36 | |
| 37 | |
| 38 | @pytest.mark.benchmark |
| 39 | def test_custom_dataset_seed_propagates( |
| 40 | hf_tokenizer: PreTrainedTokenizerBase, tmp_path: Path |
| 41 | ) -> None: |
| 42 | """--seed must control the CustomDataset shuffle used by get_samples. |
| 43 | |
| 44 | Without the fix, CustomDataset was instantiated without random_seed, |
| 45 | so its load-time shuffle always used DEFAULT_SEED=0 regardless of |
| 46 | args.seed, causing every run with --dataset-name custom to pick the |
| 47 | same subset of rows from a larger file. |
| 48 | """ |
| 49 | jsonl = tmp_path / "data.jsonl" |
| 50 | _write_jsonl(jsonl, n_rows=60) |
| 51 | |
| 52 | samples_a = get_samples(_args_for_custom(str(jsonl), seed=0), hf_tokenizer) |
| 53 | samples_b = get_samples(_args_for_custom(str(jsonl), seed=42), hf_tokenizer) |
| 54 | |
| 55 | prompts_a = {s.prompt for s in samples_a} |
| 56 | prompts_b = {s.prompt for s in samples_b} |
| 57 | |
| 58 | assert len(prompts_a) == 30 |
| 59 | assert len(prompts_b) == 30 |
| 60 | assert prompts_a != prompts_b |
| 61 | |
| 62 | |
| 63 | @pytest.mark.benchmark |
| 64 | def test_custom_dataset_same_seed_is_deterministic( |
| 65 | hf_tokenizer: PreTrainedTokenizerBase, tmp_path: Path |
| 66 | ) -> None: |
| 67 | """Same --seed must yield the same CustomDataset subset.""" |
| 68 | jsonl = tmp_path / "data.jsonl" |
| 69 | _write_jsonl(jsonl, n_rows=60) |
| 70 | |
| 71 | samples_a = get_samples(_args_for_custom(str(jsonl), seed=7), hf_tokenizer) |
| 72 | samples_b = get_samples(_args_for_custom(str(jsonl), seed=7), hf_tokenizer) |
| 73 | |
| 74 | prompts_a = [s.prompt for s in samples_a] |
| 75 | prompts_b = [s.prompt for s in samples_b] |
| 76 | |
| 77 | assert prompts_a == prompts_b |