| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | |
| 6 | from tests.utils import create_new_process_for_each_test |
| 7 | from vllm import LLM, SamplingParams |
| 8 | from vllm.assets.image import ImageAsset |
| 9 | from vllm.multimodal.image import convert_image_mode |
| 10 | |
| 11 | |
| 12 | @create_new_process_for_each_test() |
| 13 | def test_plugin( |
| 14 | monkeypatch: pytest.MonkeyPatch, |
| 15 | dummy_opt_path: str, |
| 16 | ): |
| 17 | with monkeypatch.context() as m: |
| 18 | m.setenv("VLLM_PLUGINS", "") |
| 19 | |
| 20 | with pytest.raises(ValueError, match="are not supported for now"): |
| 21 | LLM(model=dummy_opt_path, load_format="dummy") |
| 22 | |
| 23 | |
| 24 | @create_new_process_for_each_test() |
| 25 | def test_oot_registration_text_generation( |
| 26 | monkeypatch: pytest.MonkeyPatch, |
| 27 | dummy_opt_path: str, |
| 28 | ): |
| 29 | with monkeypatch.context() as m: |
| 30 | m.setenv("VLLM_PLUGINS", "register_dummy_model") |
| 31 | prompts = ["Hello, my name is", "The text does not matter"] |
| 32 | sampling_params = SamplingParams(temperature=0) |
| 33 | llm = LLM(model=dummy_opt_path, load_format="dummy") |
| 34 | first_token = llm.get_tokenizer().decode(0) |
| 35 | outputs = llm.generate(prompts, sampling_params) |
| 36 | |
| 37 | for output in outputs: |
| 38 | generated_text = output.outputs[0].text |
| 39 | # make sure only the first token is generated |
| 40 | rest = generated_text.replace(first_token, "") |
| 41 | assert rest == "" |
| 42 | |
| 43 | |
| 44 | @create_new_process_for_each_test() |
| 45 | def test_oot_registration_embedding( |
| 46 | monkeypatch: pytest.MonkeyPatch, |
| 47 | dummy_gemma2_embedding_path: str, |
| 48 | ): |
| 49 | with monkeypatch.context() as m: |
| 50 | m.setenv("VLLM_PLUGINS", "register_dummy_model") |
| 51 | prompts = ["Hello, my name is", "The text does not matter"] |
| 52 | llm = LLM( |
| 53 | model=dummy_gemma2_embedding_path, load_format="dummy", max_model_len=512 |
| 54 | ) |
| 55 | outputs = llm.embed(prompts) |
| 56 | |
| 57 | for output in outputs: |
| 58 | assert all(v == 0 for v in output.outputs.embedding) |
| 59 | |
| 60 | |
| 61 | image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB") |
| 62 | |
| 63 | |
| 64 | @create_new_process_for_each_test() |
| 65 | def test_oot_registration_multimodal( |
| 66 | monkeypatch: pytest.MonkeyPatch, |
| 67 | dummy_llava_path: str, |
| 68 | ): |
| 69 | with monkeypatch.context() as m: |
| 70 | m.setenv("VLLM_PLUGINS", "register_dummy_model") |
| 71 | prompts = [ |
| 72 | { |
| 73 | "prompt": "What's in the image?<image>", |
| 74 | "multi_modal_data": {"image": image}, |
| 75 | }, |
| 76 | { |
| 77 | "prompt": "Describe the image<image>", |
| 78 | "multi_modal_data": {"image": image}, |
| 79 | }, |
| 80 | ] |
| 81 | |
| 82 | sampling_params = SamplingParams(temperature=0) |
| 83 | llm = LLM( |
| 84 | model=dummy_llava_path, |
| 85 | load_format="dummy", |
| 86 | max_num_seqs=1, |
| 87 | trust_remote_code=True, |
| 88 | gpu_memory_utilization=0.98, |
| 89 | max_model_len=4096, |
| 90 | enforce_eager=True, |
| 91 | limit_mm_per_prompt={"image": 1}, |
| 92 | ) |
| 93 | |
| 94 | first_token = llm.get_tokenizer().decode(0) |
| 95 | outputs = llm.generate(prompts, sampling_params) |
| 96 | |
| 97 | for output in outputs: |
| 98 | generated_text = output.outputs[0].text |
| 99 | # make sure only the first token is generated |
| 100 | rest = generated_text.replace(first_token, "") |
| 101 | assert rest == "" |