biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc] Standardize handling of mm_processor_kwargs.size (#35284)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-02-26 21:05:46 +08:00
committed by GitHub
parent ec13e549d3
commit 845ee348ef
9 changed files with 135 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
tests/models/multimodal/processing/test_qwen2_vl.py
View File

@@ -2,6 +2,8 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from packaging.version import Version
from transformers import __version__ as TRANSFORMERS_VERSION
from vllm.multimodal import MULTIMODAL_REGISTRY
@@ -15,6 +17,16 @@ from ...utils import build_model_context
[
({}, 1426, (5704, 1176)),
({"min_pixels": 64**2, "max_pixels": 512**2}, 330, (1320, 1176)),
(
{
"size": {
"shortest_edge": 64**2,
"longest_edge": 512**2,
},
},
330,
(1320, 1176),
),
],
)
@pytest.mark.parametrize("num_imgs", [1, 2])
@@ -29,6 +41,12 @@ def test_processor_override(
kwargs_on_init: bool,
):
"""Ensure Qwen2VLMultiModalProcessor handles min/max pixels properly."""
if (
Version(TRANSFORMERS_VERSION) < Version("5.2.0")
and "size" in mm_processor_kwargs
):
pytest.skip("`size` ignored by `Qwen2VLProcessor.__call__`")
ctx = build_model_context(
model_id,
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
@@ -60,21 +78,34 @@ def test_processor_override(
@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])
@pytest.mark.parametrize("max_pixels", [1280 * 28 * 28, 1283 * 28 * 28])
@pytest.mark.parametrize(
"mm_processor_kwargs",
[
{"min_pixels": 28 * 28, "max_pixels": 1280 * 28 * 28},
{"min_pixels": 28 * 28, "max_pixels": 1283 * 28 * 28},
{"size": {"shortest_edge": 28 * 28, "longest_edge": 1280 * 28 * 28}},
{"size": {"shortest_edge": 28 * 28, "longest_edge": 1283 * 28 * 28}},
],
)
def test_get_image_size_with_most_features(
image_assets: ImageTestAssets,
model_id: str,
max_pixels: int,
mm_processor_kwargs: dict[str, object],
):
if (
Version(TRANSFORMERS_VERSION) < Version("5.2.0")
and "size" in mm_processor_kwargs
):
pytest.skip("`size` ignored by `Qwen2VLProcessor.__call__`")
ctx = build_model_context(
model_id,
mm_processor_kwargs={"max_pixels": max_pixels},
mm_processor_kwargs=mm_processor_kwargs,
limit_mm_per_prompt={"image": 1},
)
processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
hf_processor_mm_kwargs: dict[str, object] = {}
hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs)
hf_processor = processor.info.get_hf_processor(**mm_processor_kwargs)
merge_size = processor.info.get_hf_config().vision_config.spatial_merge_size
max_image_size = processor.info.get_image_size_with_most_features()
@@ -82,7 +113,7 @@ def test_get_image_size_with_most_features(
image_width=max_image_size.width,
image_height=max_image_size.height,
image_processor=hf_processor.image_processor,
mm_kwargs=hf_processor_mm_kwargs,
mm_kwargs=mm_processor_kwargs,
)
prompt = "<|vision_start|><|image_pad|><|vision_end|>"
@@ -91,7 +122,7 @@ def test_get_image_size_with_most_features(
processed_inputs = processor(
prompt,
mm_items=processor.info.parse_mm_data(mm_data),
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
hf_processor_mm_kwargs=mm_processor_kwargs,
)
grid_thw = processed_inputs["mm_kwargs"].get_data()["image_grid_thw"].tolist()
t, h, w = grid_thw[0]