| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import weakref |
| 4 | |
| 5 | import pytest |
| 6 | |
| 7 | from vllm import LLM, PoolingRequestOutput |
| 8 | from vllm.config import PoolerConfig |
| 9 | from vllm.tasks import PoolingTask |
| 10 | |
| 11 | MODEL_NAME = "intfloat/multilingual-e5-small" |
| 12 | |
| 13 | prompt = "The chef prepared a delicious meal." |
| 14 | prompt_token_ids = [0, 581, 21861, 133888, 10, 8, 150, 60744, 109911, 5, 2] |
| 15 | embedding_size = 384 |
| 16 | |
| 17 | |
| 18 | @pytest.fixture(scope="module") |
| 19 | def llm(vllm_runner): |
| 20 | with vllm_runner( |
| 21 | MODEL_NAME, |
| 22 | max_model_len=None, |
| 23 | pooler_config=PoolerConfig(task="token_embed"), |
| 24 | max_num_batched_tokens=32768, |
| 25 | tensor_parallel_size=1, |
| 26 | gpu_memory_utilization=0.75, |
| 27 | enforce_eager=True, |
| 28 | seed=0, |
| 29 | enable_chunked_prefill=None, |
| 30 | ) as runner: |
| 31 | assert embedding_size == runner.llm.model_config.embedding_size |
| 32 | # pytest caches yielded fixtures until after teardown, so use a proxy to |
| 33 | # avoid retaining the LLM while VllmRunner.__exit__ releases ROCm memory. |
| 34 | yield weakref.proxy(runner.llm) |
| 35 | |
| 36 | |
| 37 | @pytest.mark.skip_global_cleanup |
| 38 | def test_str_prompts(llm: LLM): |
| 39 | outputs = llm.encode(prompt, pooling_task="token_embed", use_tqdm=False) |
| 40 | assert len(outputs) == 1 |
| 41 | assert isinstance(outputs[0], PoolingRequestOutput) |
| 42 | assert outputs[0].outputs.data.shape == (11, 384) |
| 43 | |
| 44 | |
| 45 | @pytest.mark.skip_global_cleanup |
| 46 | def test_token_ids_prompts(llm: LLM): |
| 47 | outputs = llm.encode([prompt_token_ids], pooling_task="token_embed", use_tqdm=False) |
| 48 | assert len(outputs) == 1 |
| 49 | assert isinstance(outputs[0], PoolingRequestOutput) |
| 50 | assert outputs[0].outputs.data.shape == (11, 384) |
| 51 | |
| 52 | |
| 53 | @pytest.mark.parametrize("task", ["embed", "classify", "token_classify", "plugin"]) |
| 54 | def test_unsupported_tasks(llm: LLM, task: PoolingTask, caplog_vllm): |
| 55 | if task == "plugin": |
| 56 | err_msg = "No IOProcessor plugin installed." |
| 57 | elif task == "embed": |
| 58 | err_msg = "Try switching the model's pooling_task via.+" |
| 59 | else: |
| 60 | err_msg = "Classification API is not supported by this model.+" |
| 61 | |
| 62 | with pytest.raises(ValueError, match=err_msg): |
| 63 | llm.encode(prompt, pooling_task=task, use_tqdm=False) |