test_phi3v.py 2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
"""Tests for phi3v's multimodal preprocessing kwargs."""
zhuwenwen's avatar
zhuwenwen committed
3
import os
4
5
import pytest

6
from vllm.multimodal import MULTIMODAL_REGISTRY
7

8
9
from ....conftest import _ImageAssets
from ...utils import build_model_context
zhuwenwen's avatar
zhuwenwen committed
10
from ....utils import models_path_prefix
11
12


zhuwenwen's avatar
zhuwenwen committed
13
@pytest.mark.parametrize("model_id", [os.path.join(models_path_prefix, "microsoft/Phi-3.5-vision-instruct")])
14
# yapf: disable
15
@pytest.mark.parametrize(
16
    ("mm_processor_kwargs", "expected_toks_per_img"),
17
    [
18
19
        ({"num_crops": 4}, 757),
        ({"num_crops": 16}, 1921),
20
        # the default num_crops of phi-3.5-vision is 4
21
        ({}, 757),
22
    ])
23
# yapf: enable
24
@pytest.mark.parametrize("num_imgs", [1, 2])
25
@pytest.mark.parametrize("kwargs_on_init", [True, False])
26
27
28
29
30
31
def test_processor_override(
    image_assets: _ImageAssets,
    model_id: str,
    mm_processor_kwargs: dict[str, int],
    expected_toks_per_img: int,
    num_imgs: int,
32
    kwargs_on_init: bool,
33
):
34
    """Ensure Phi3VMultiModalProcessor handles num_crops properly."""
35
36
37
    # Avoid initializing CUDA early
    from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID

38
    ctx = build_model_context(
39
        model_id,
40
        mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
41
        limit_mm_per_prompt={"image": num_imgs},
42
    )
43
    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
44
    hf_processor_mm_kwargs = {} if kwargs_on_init else mm_processor_kwargs
45

46
47
48
    # Build the image str / prompt based on the number of images we pass
    img_str = "".join([f"<|image_{idx}|>\n" for idx in range(1, num_imgs + 1)])
    prompt = f"<|user|>\n{img_str}<|end|>\n<|assistant|>\n"
49
    mm_data = {"image": [image_assets[0].pil_image] * num_imgs}
50

51
    processed_inputs = processor.apply(prompt, mm_data, hf_processor_mm_kwargs)
52
53
54

    # Ensure we have the right number of placeholders per num_crops size
    img_tok_count = processed_inputs["prompt_token_ids"].count(_IMAGE_TOKEN_ID)
zhuwenwen's avatar
zhuwenwen committed
55
    assert img_tok_count == expected_toks_per_img * num_imgs