A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI] Move testing image from remote URL to S3 (#23980)
Signed-off-by: Roger Wang <hey@rogerw.io>
Roger Wang committed
Aug 29, 2025 at 21:41 UTC
ee52a32705988bcac1833feab34af977addf5cca
2 files changed
+5
-8
tests/models/multimodal/generation/vlm_utils/custom_inputs.py
+4
-7
@@ -1,12 +1,9 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""Custom input builders for edge-cases in different models."""
4
-from io import BytesIO
4
from typing import Callable
5
7
-import requests
8
-from PIL import Image
9
-
6
+from vllm.assets.image import ImageAsset
7
from vllm.multimodal.image import rescale_image_size
8
from vllm.multimodal.video import (rescale_video_size, resize_video,
9
sample_frames_from_video)
@@ -118,9 +115,9 @@ def different_patch_input_cases_internvl():
115
116
117
def windows_attention_image_qwen2_5_vl():
121
- # image from regression issue: https://github.com/vllm-project/vllm/issues/15122
122
- image_url = "https://github.com/AOMediaCodec/av1-avif/blob/main/testFiles/Link-U/hato.jpg?raw=true"
123
- image = Image.open(BytesIO(requests.get(image_url).content))
118
+
119
+ # image from regression issue: https://github.com/vllm-project/vllm/issues/15122 # noqa: E501
120
+ image = ImageAsset("hato").pil_image
121
122
question = "Describe the image."
123
img_prompt = "<|vision_start|><|image_pad|><|vision_end|>"
vllm/assets/image.py
+1
-1
@@ -11,7 +11,7 @@ from .base import get_vllm_public_assets
11
12
VLM_IMAGES_DIR = "vision_model_images"
13
14
-ImageAssetName = Literal["stop_sign", "cherry_blossom"]
14
+ImageAssetName = Literal["stop_sign", "cherry_blossom", "hato"]
15
16
17
@dataclass(frozen=True)