| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | |
| 6 | from vllm.assets.image import ImageAsset |
| 7 | from vllm.assets.video import VideoAsset |
| 8 | from vllm.multimodal.image import convert_image_mode |
| 9 | |
| 10 | models = ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"] |
| 11 | |
| 12 | |
| 13 | def base_prompt(modalities_str: str) -> str: |
| 14 | return f"<|im_start|>user {modalities_str}\nDescribe what you see from these items.<|im_end|><|im_start|>assistant\n" # noqa: E501 |
| 15 | |
| 16 | |
| 17 | INTERLEAVED_PROMPT = base_prompt("<image><video><image>\n") |
| 18 | NONINTERLEAVED_PROMPT = base_prompt("<image><image><video>\n") |
| 19 | |
| 20 | |
| 21 | @pytest.mark.core_model |
| 22 | @pytest.mark.parametrize("model", models) |
| 23 | @pytest.mark.parametrize("dtype", ["float16"]) |
| 24 | @pytest.mark.parametrize("max_tokens", [128]) |
| 25 | def test_models(vllm_runner, model, dtype: str, max_tokens: int) -> None: |
| 26 | """This is a simple test to check if interleaved and non-interleaved prompts |
| 27 | give the same result. |
| 28 | """ |
| 29 | image_cherry = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB") |
| 30 | image_stop = convert_image_mode(ImageAsset("stop_sign").pil_image, "RGB") |
| 31 | images = [image_cherry, image_stop] |
| 32 | video = VideoAsset(name="baby_reading", num_frames=16).np_ndarrays |
| 33 | |
| 34 | inputs = [ |
| 35 | ( |
| 36 | [INTERLEAVED_PROMPT], |
| 37 | [images], |
| 38 | [video], |
| 39 | ), |
| 40 | ( |
| 41 | [NONINTERLEAVED_PROMPT], |
| 42 | [images], |
| 43 | [video], |
| 44 | ), |
| 45 | ] |
| 46 | |
| 47 | with vllm_runner( |
| 48 | model, |
| 49 | runner="generate", |
| 50 | dtype=dtype, |
| 51 | limit_mm_per_prompt={"image": 2}, |
| 52 | max_model_len=32768, |
| 53 | max_num_seqs=2, |
| 54 | tensor_parallel_size=1, |
| 55 | enforce_eager=True, |
| 56 | ) as vllm_model: |
| 57 | vllm_outputs_per_case = [ |
| 58 | vllm_model.generate_greedy( |
| 59 | prompts, max_tokens, images=images, videos=videos |
| 60 | ) |
| 61 | for prompts, images, videos in inputs |
| 62 | ] |
| 63 | |
| 64 | all_results = [output[0][1] for output in vllm_outputs_per_case] |
| 65 | outputs = [ |
| 66 | (total_str, total_str.find("assistant\n") + len("assistant\n")) |
| 67 | for total_str in all_results |
| 68 | ] |
| 69 | prompt_lengths = [prompt_len for _, prompt_len in outputs] |
| 70 | generated_strs = [total_str[prompt_len:] for total_str, prompt_len in outputs] |
| 71 | interleaved_prompt_len, noninterleaved_prompt_len = prompt_lengths |
| 72 | interleaved_output_str, noninterleaved_output_str = generated_strs |
| 73 | |
| 74 | # The two prompts are identical except for the order of modality tokens. |
| 75 | assert interleaved_prompt_len == noninterleaved_prompt_len |
| 76 | |
| 77 | # The two generated strings should be different because of the |
| 78 | # interleaved modality tokens. |
| 79 | assert interleaved_output_str != noninterleaved_output_str |