| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | |
| 5 | from vllm.exceptions import VLLMValidationError |
| 6 | |
| 7 | MODEL_NAME = "sentence-transformers/all-MiniLM-L12-v2" |
| 8 | max_model_len = 128 |
| 9 | |
| 10 | input_str = """Immerse yourself in the enchanting chronicle of calculus, a |
| 11 | mathematical domain that has radically transformed our comprehension of |
| 12 | change and motion. Despite its roots in ancient civilizations, the |
| 13 | formal birth of calculus predominantly occurred in the 17th century, |
| 14 | primarily under the influential guidance of Sir Isaac Newton and Gottfried |
| 15 | Wilhelm Leibniz. The earliest traces of calculus concepts are found in |
| 16 | ancient Greek mathematics,most notably in the works of Eudoxus and |
| 17 | Archimedes, around 300 BCE. They utilized the 'method of exhaustion'—a |
| 18 | technique for computing areas and volumes through the use of finite sums. |
| 19 | This methodology laid crucial foundational work for integral calculus. |
| 20 | In the 17th century, both Newton and Leibniz independently pioneered |
| 21 | calculus, each contributing unique perspectives that would shape this new |
| 22 | field.""" |
| 23 | |
| 24 | |
| 25 | @pytest.fixture(scope="module") |
| 26 | def vllm_model(vllm_runner): |
| 27 | with vllm_runner( |
| 28 | MODEL_NAME, runner="pooling", max_model_len=max_model_len |
| 29 | ) as model: |
| 30 | yield model |
| 31 | |
| 32 | |
| 33 | def test_smaller_truncation_size( |
| 34 | vllm_model, |
| 35 | ): |
| 36 | truncate_prompt_tokens = 10 |
| 37 | |
| 38 | vllm_output = vllm_model.llm.embed( |
| 39 | input_str, |
| 40 | tokenization_kwargs=dict(truncate_prompt_tokens=truncate_prompt_tokens), |
| 41 | ) |
| 42 | |
| 43 | prompt_tokens = vllm_output[0].prompt_token_ids |
| 44 | |
| 45 | assert len(prompt_tokens) == truncate_prompt_tokens |
| 46 | |
| 47 | |
| 48 | def test_max_truncation_size(vllm_model): |
| 49 | truncate_prompt_tokens = -1 |
| 50 | |
| 51 | vllm_output = vllm_model.llm.embed( |
| 52 | input_str, |
| 53 | tokenization_kwargs=dict(truncate_prompt_tokens=truncate_prompt_tokens), |
| 54 | ) |
| 55 | |
| 56 | prompt_tokens = vllm_output[0].prompt_token_ids |
| 57 | |
| 58 | assert len(prompt_tokens) == max_model_len |
| 59 | |
| 60 | |
| 61 | def test_bigger_truncation_size(vllm_model): |
| 62 | truncate_prompt_tokens = max_model_len + 1 |
| 63 | |
| 64 | with pytest.raises(VLLMValidationError): |
| 65 | vllm_model.llm.embed( |
| 66 | input_str, |
| 67 | tokenization_kwargs=dict(truncate_prompt_tokens=truncate_prompt_tokens), |
| 68 | ) |