| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for smolvlm's multimodal preprocessing kwargs.""" |
| 4 | |
| 5 | import pytest |
| 6 | from transformers import SmolVLMConfig |
| 7 | |
| 8 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 9 | |
| 10 | from ....conftest import ImageTestAssets |
| 11 | from ...utils import build_model_context |
| 12 | |
| 13 | |
| 14 | @pytest.mark.parametrize("model_id", ["HuggingFaceTB/SmolVLM2-2.2B-Instruct"]) |
| 15 | @pytest.mark.parametrize( |
| 16 | ("mm_processor_kwargs", "expected_toks_per_img"), |
| 17 | [ |
| 18 | ({"max_image_size": {"longest_edge": 384}}, 1377), |
| 19 | ({"max_image_size": {"longest_edge": 768}}, 405), |
| 20 | ({"do_image_splitting": False}, 81), |
| 21 | ], |
| 22 | ) |
| 23 | @pytest.mark.parametrize("num_imgs", [1, 2]) |
| 24 | @pytest.mark.parametrize("kwargs_on_init", [True, False]) |
| 25 | def test_processor_override( |
| 26 | image_assets: ImageTestAssets, |
| 27 | model_id: str, |
| 28 | mm_processor_kwargs: dict[str, object], |
| 29 | expected_toks_per_img: int, |
| 30 | num_imgs: int, |
| 31 | kwargs_on_init: bool, |
| 32 | ): |
| 33 | """Ensure Idefics3MultiModalProcessor handles num_crops properly.""" |
| 34 | # Same as the previous test - don't initialize mm_processor_kwargs |
| 35 | # in this test and assume that the kwargs will be correctly expanded by |
| 36 | # the partial when calling the custom input processor. |
| 37 | ctx = build_model_context( |
| 38 | model_id, |
| 39 | mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None, |
| 40 | limit_mm_per_prompt={"image": num_imgs}, |
| 41 | ) |
| 42 | processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config) |
| 43 | hf_processor_mm_kwargs = {} if kwargs_on_init else mm_processor_kwargs |
| 44 | |
| 45 | # Build the image str / prompt based on the number of images we pass |
| 46 | placeholders = ( |
| 47 | "<image>" |
| 48 | if num_imgs == 1 |
| 49 | else "\n".join(f"Image-{i}: <image>\n" for i in range(1, num_imgs + 1)) |
| 50 | ) |
| 51 | prompt = f"<|im_start|>User:{placeholders}\n<end_of_utterance>\nAssistant:" # noqa: E501 |
| 52 | |
| 53 | # Build mm_data |
| 54 | image_size = ctx.get_hf_config(SmolVLMConfig).vision_config.image_size |
| 55 | dummy_image_size = (image_size * 4, image_size * 4) |
| 56 | dummy_image = image_assets[0].pil_image.resize(dummy_image_size) |
| 57 | mm_data = {"image": [dummy_image] * num_imgs} |
| 58 | |
| 59 | processed_inputs = processor( |
| 60 | prompt, |
| 61 | mm_items=processor.info.parse_mm_data(mm_data), |
| 62 | hf_processor_mm_kwargs=hf_processor_mm_kwargs, |
| 63 | ) |
| 64 | |
| 65 | # Ensure the placeholders format are correct |
| 66 | hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 67 | num_patches = processor.info.get_num_patches( |
| 68 | image_width=dummy_image.width, |
| 69 | image_height=dummy_image.height, |
| 70 | processor=hf_processor, |
| 71 | mm_kwargs=hf_processor_mm_kwargs, |
| 72 | ) |
| 73 | assert num_patches == expected_toks_per_img // hf_processor.image_seq_len |
| 74 | |
| 75 | hf_processed_inputs = hf_processor( |
| 76 | text=prompt, |
| 77 | images=mm_data["image"], |
| 78 | **processor.info.ctx.get_merged_mm_kwargs(hf_processor_mm_kwargs), |
| 79 | ) |
| 80 | assert processed_inputs["prompt_token_ids"] == hf_processed_inputs["input_ids"][0] |
| 81 | |
| 82 | # Ensure we have the right number of placeholders per num_crops size |
| 83 | image_token_id = ctx.get_hf_config().image_token_id |
| 84 | img_tok_count = processed_inputs["prompt_token_ids"].count(image_token_id) |
| 85 | assert img_tok_count == expected_toks_per_img * num_imgs |