[CI/Build] Serve images used by multimodal tests through local HTTP Server (#23907)

Signed-off-by: Divyansh Singhvi <divyanshsinghvi@gmail.com>
Signed-off-by: dsinghvi <divyanshsinghvi@gmail.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>

tests/conftest.py

@@ -1,9 +1,14 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import http.server
 import json
 import math
+import mimetypes
 import os
+import socket
 import tempfile
+import threading
+from collections.abc import Generator
 from enum import Enum
 from typing import Any, Callable, Optional, TypedDict, TypeVar, Union, cast
 
@@ -32,6 +37,7 @@ from vllm.distributed import (cleanup_dist_env_and_memory,
 from vllm.inputs import (ExplicitEncoderDecoderPrompt, TextPrompt,
                          to_enc_dec_tuple_list, zip_enc_dec_prompts)
 from vllm.logger import init_logger
+from vllm.multimodal.utils import fetch_image
 from vllm.outputs import RequestOutput
 from vllm.sampling_params import BeamSearchParams
 from vllm.sequence import Logprob
@@ -1253,3 +1259,119 @@ def cli_config_file():
 def cli_config_file_with_model():
     """Return the path to the CLI config file with model."""
     return os.path.join(_TEST_DIR, "config", "test_config_with_model.yaml")
+
+
+class AssetHandler(http.server.BaseHTTPRequestHandler):
+    # _IMAGE_CACHE : Dict[str, bytes] = {}
+
+    def log_message(self, *args, **kwargs):
+        pass
+
+    def do_GET(self):
+        # Accepts paths like: /1280px-Venn_diagram_rgb.jpg
+        filename = self.path.lstrip("/")
+        if not filename or "." not in filename:
+            self.send_error(404, "Missing filename (expected /<name>.<ext>)")
+            return
+
+        base, ext = filename.rsplit(".", 1)
+        ext = ext.lower()
+
+        if ext not in ["jpg", "png"]:
+            self.send_error(404, f"Unsupported extension: .{ext}")
+            return
+
+        try:
+            data = ImageAsset(base).read_bytes(ext=ext)
+        except Exception as e:
+            self.send_error(500, f"Failed to load asset: {ext} {base} {e} ")
+            return
+
+        ctype, _ = mimetypes.guess_type(filename)
+        if ctype is None:
+            ctype = {"jpg": "image/jpg", "png": "image/png"}[ext]
+        self.send_response(200)
+        self.send_header("Content-Type", ctype)
+        self.send_header("Content-Length", str(len(data)))
+        self.end_headers()
+        self.wfile.write(data)
+
+
+def _find_free_port() -> int:
+    with socket.socket() as s:
+        s.bind(("127.0.0.1", 0))
+        return s.getsockname()[1]
+
+
+class LocalAssetServer:
+
+    address: str
+    port: int
+    server: Optional[http.server.ThreadingHTTPServer]
+    thread: Optional[threading.Thread]
+
+    def __init__(self, address: str = "127.0.0.1") -> None:
+        self.address = address
+        self.port = -1
+        self.server = None
+        self.thread = None
+
+    def __enter__(self):
+        self.port = _find_free_port()
+        self.server = http.server.ThreadingHTTPServer(
+            (self.address, self.port), AssetHandler)
+        self.thread = threading.Thread(target=self.server.serve_forever,
+                                       daemon=True)
+        self.thread.start()
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        if self.server:
+            self.server.shutdown()
+            del self.server
+
+        if self.thread:
+            self.thread.join()
+            del self.thread
+
+        if exc_type is None:
+            return None
+
+        return False
+
+    @property
+    def base_url(self) -> str:
+        assert self.port is not None
+        return f"http://{self.address}:{self.port}"
+
+    def url_for(self, name: str) -> str:
+        """e.g., name='RGBA_comp.png' -> 'http://127.0.0.1:PORT/RGBA_comp.png'"""
+        return f"{self.base_url}/{name}"
+
+    def get_image_asset(self, name: str) -> Image.Image:
+        return fetch_image(self.url_for(name))
+
+
+@pytest.fixture(scope="session")
+def local_asset_server() -> Generator[LocalAssetServer, None, None]:
+    """
+    Starts a thread based HTTP server bound to 127.0.0.1 on a random free port. 
+    The server currently servers images at:
+    http://127.0.0.1:<port>/<name>.<ext>
+    """
+    with LocalAssetServer() as srv:
+        yield srv
+
+
+@pytest.fixture
+def image_url(request, local_asset_server) -> str:
+    # request.param is one of the IMAGE_ASSETS filenames
+    name = request.param
+    return local_asset_server.url_for(name)
+
+
+@pytest.fixture
+def image_urls(request, local_asset_server) -> list[str]:
+    """Indirect fixture: takes a list of names, returns list of full URLs."""
+    names: list[str] = request.param
+    return [local_asset_server.url_for(name) for name in names]

tests/entrypoints/llm/test_chat.py

@@ -7,7 +7,7 @@ import pytest
 from vllm import LLM
 from vllm.distributed import cleanup_dist_env_and_memory
 
-from ..openai.test_vision import TEST_IMAGE_URLS
+from ..openai.test_vision import TEST_IMAGE_ASSETS
 
 
 @pytest.fixture(scope="function")
@@ -95,7 +95,8 @@ def vision_llm():
 
 
 @pytest.mark.parametrize("image_urls",
-                         [[TEST_IMAGE_URLS[0], TEST_IMAGE_URLS[1]]])
+                         [[TEST_IMAGE_ASSETS[0], TEST_IMAGE_ASSETS[1]]],
+                         indirect=True)
 def test_chat_multi_image(vision_llm, image_urls: list[str]):
     messages = [{
         "role":

tests/entrypoints/openai/test_vision.py

@@ -16,11 +16,11 @@ MODEL_NAME = "microsoft/Phi-3.5-vision-instruct"
 MAXIMUM_IMAGES = 2
 
 # Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
-TEST_IMAGE_URLS = [
+TEST_IMAGE_ASSETS = [
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",
+    "2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
-    "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
+    "Grayscale_8bits_palette_sample_image.png",  # "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
+    "1280px-Venn_diagram_rgb.svg.png",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
-    "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
+    "RGBA_comp.png",  # "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
 ]
 
 EXPECTED_MM_BEAM_SEARCH_RES = [
@@ -69,10 +69,11 @@ async def client(server):
 
 
 @pytest.fixture(scope="session")
-def base64_encoded_image() -> dict[str, str]:
+def base64_encoded_image(local_asset_server) -> dict[str, str]:
     return {
-        image_url: encode_image_base64(fetch_image(image_url))
+        image_asset:
-        for image_url in TEST_IMAGE_URLS
+        encode_image_base64(local_asset_server.get_image_asset(image_asset))
+        for image_asset in TEST_IMAGE_ASSETS
     }
 
 
@@ -97,7 +98,7 @@ def get_hf_prompt_tokens(model_name, content, image_url):
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_single_chat_session_image(client: openai.AsyncOpenAI,
                                          model_name: str, image_url: str):
     content_text = "What's in this image?"
@@ -157,7 +158,7 @@ async def test_single_chat_session_image(client: openai.AsyncOpenAI,
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_error_on_invalid_image_url_type(client: openai.AsyncOpenAI,
                                                model_name: str,
                                                image_url: str):
@@ -187,7 +188,7 @@ async def test_error_on_invalid_image_url_type(client: openai.AsyncOpenAI,
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_single_chat_session_image_beamsearch(client: openai.AsyncOpenAI,
                                                     model_name: str,
                                                     image_url: str):
@@ -223,10 +224,11 @@ async def test_single_chat_session_image_beamsearch(client: openai.AsyncOpenAI,
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("raw_image_url", TEST_IMAGE_ASSETS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_single_chat_session_image_base64encoded(
-        client: openai.AsyncOpenAI, model_name: str, image_url: str,
+        client: openai.AsyncOpenAI, model_name: str, raw_image_url: str,
-        base64_encoded_image: dict[str, str]):
+        image_url: str, base64_encoded_image: dict[str, str]):
 
     content_text = "What's in this image?"
     messages = [{
@@ -237,7 +239,7 @@ async def test_single_chat_session_image_base64encoded(
                 "type": "image_url",
                 "image_url": {
                     "url":
-                    f"data:image/jpeg;base64,{base64_encoded_image[image_url]}"
+                    f"data:image/jpeg;base64,{base64_encoded_image[raw_image_url]}"
                 }
             },
             {
@@ -287,12 +289,12 @@ async def test_single_chat_session_image_base64encoded(
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_idx", list(range(len(TEST_IMAGE_URLS))))
+@pytest.mark.parametrize("image_idx", list(range(len(TEST_IMAGE_ASSETS))))
 async def test_single_chat_session_image_base64encoded_beamsearch(
         client: openai.AsyncOpenAI, model_name: str, image_idx: int,
         base64_encoded_image: dict[str, str]):
     # NOTE: This test also validates that we pass MM data through beam search
-    image_url = TEST_IMAGE_URLS[image_idx]
+    raw_image_url = TEST_IMAGE_ASSETS[image_idx]
     expected_res = EXPECTED_MM_BEAM_SEARCH_RES[image_idx]
 
     messages = [{
@@ -303,7 +305,7 @@ async def test_single_chat_session_image_base64encoded_beamsearch(
                 "type": "image_url",
                 "image_url": {
                     "url":
-                    f"data:image/jpeg;base64,{base64_encoded_image[image_url]}"
+                    f"data:image/jpeg;base64,{base64_encoded_image[raw_image_url]}"
                 }
             },
             {
@@ -326,7 +328,7 @@ async def test_single_chat_session_image_base64encoded_beamsearch(
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_chat_streaming_image(client: openai.AsyncOpenAI,
                                     model_name: str, image_url: str):
     messages = [{
@@ -385,7 +387,8 @@ async def test_chat_streaming_image(client: openai.AsyncOpenAI,
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 @pytest.mark.parametrize(
     "image_urls",
-    [TEST_IMAGE_URLS[:i] for i in range(2, len(TEST_IMAGE_URLS))])
+    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
+    indirect=True)
 async def test_multi_image_input(client: openai.AsyncOpenAI, model_name: str,
                                  image_urls: list[str]):
 

tests/entrypoints/openai/test_vision_embedding.py

@@ -19,11 +19,11 @@ vlm2vec_jinja_path = VLLM_PATH / "examples/template_vlm2vec.jinja"
 assert vlm2vec_jinja_path.exists()
 
 # Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
-TEST_IMAGE_URLS = [
+TEST_IMAGE_ASSETS = [
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",
+    "2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
-    "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
+    "Grayscale_8bits_palette_sample_image.png",  # "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
+    "1280px-Venn_diagram_rgb.svg.png",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
-    "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
+    "RGBA_comp.png",  # "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
 ]
 
 
@@ -49,10 +49,11 @@ def server():
 
 
 @pytest.fixture(scope="session")
-def base64_encoded_image() -> dict[str, str]:
+def base64_encoded_image(local_asset_server) -> dict[str, str]:
     return {
-        image_url: encode_image_base64(fetch_image(image_url))
+        image_url:
-        for image_url in TEST_IMAGE_URLS
+        encode_image_base64(local_asset_server.get_image_asset(image_url))
+        for image_url in TEST_IMAGE_ASSETS
     }
 
 
@@ -70,7 +71,7 @@ def get_hf_prompt_tokens(model_name, content, image_url):
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_image_embedding(server: RemoteOpenAIServer, model_name: str,
                                image_url: str):
     content_text = "Represent the given image."

tests/models/multimodal/generation/test_pixtral.py

@@ -29,10 +29,10 @@ MISTRAL_SMALL_3_1_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
 MODELS = [PIXTRAL_ID, MISTRAL_SMALL_3_1_ID]
 
 IMG_URLS = [
-    "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/237-400x300.jpg",
+    "237-400x300.jpg",  # "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/237-400x300.jpg",
-    "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/231-200x300.jpg",
+    "231-200x300.jpg",  # "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/237-400x300.jpg",
-    "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/27-500x500.jpg",
+    "27-500x500.jpg",  # "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/237-400x300.jpg",
-    "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/17-150x600.jpg",
+    "17-150x600.jpg",  # "https://huggingface.co/datasets/Isotr0py/mistral-test-images/resolve/main/237-400x300.jpg",
 ]
 PROMPT = "Describe each image in one short sentence."
 
@@ -105,12 +105,6 @@ def _create_engine_inputs_hf(urls: list[str]) -> TextPrompt:
     return engine_inputs
 
 
-MSGS = [
-    _create_msg_format(IMG_URLS[:1]),
-    _create_msg_format(IMG_URLS[:2]),
-    _create_msg_format(IMG_URLS),
-]
-
 SAMPLING_PARAMS = SamplingParams(max_tokens=512, temperature=0.0, logprobs=5)
 LIMIT_MM_PER_PROMPT = dict(image=4)
 
@@ -156,12 +150,8 @@ def load_outputs_w_logprobs(filename: "StrPath") -> OutputsLogprobs:
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("max_model_len", MAX_MODEL_LEN)
 @pytest.mark.parametrize("dtype", ["bfloat16"])
-def test_chat(
+def test_chat(vllm_runner, max_model_len: int, model: str, dtype: str,
-    vllm_runner,
+              local_asset_server) -> None:
-    max_model_len: int,
-    model: str,
-    dtype: str,
-) -> None:
     EXPECTED_CHAT_LOGPROBS = load_outputs_w_logprobs(
         FIXTURE_LOGPROBS_CHAT[model])
     with vllm_runner(
@@ -174,7 +164,14 @@ def test_chat(
             limit_mm_per_prompt=LIMIT_MM_PER_PROMPT,
     ) as vllm_model:
         outputs = []
-        for msg in MSGS:
+
+        urls_all = [local_asset_server.url_for(u) for u in IMG_URLS]
+        msgs = [
+            _create_msg_format(urls_all[:1]),
+            _create_msg_format(urls_all[:2]),
+            _create_msg_format(urls_all),
+        ]
+        for msg in msgs:
             output = vllm_model.llm.chat(msg, sampling_params=SAMPLING_PARAMS)
 
             outputs.extend(output)
@@ -190,14 +187,24 @@ def test_chat(
                          name_1="output")
 
 
-@pytest.mark.parametrize("prompt,expected_ranges",
+@pytest.fixture
-                         [(_create_engine_inputs_hf(IMG_URLS[:1]),
+def prompt(request, local_asset_server) -> TextPrompt:
-                           [PlaceholderRange(offset=11, length=494)]),
+    names = request.param
-                          (_create_engine_inputs_hf(IMG_URLS[1:4]), [
+    urls = [local_asset_server.url_for(n) for n in names]
-                              PlaceholderRange(offset=11, length=266),
+    return _create_engine_inputs_hf(urls)
-                              PlaceholderRange(offset=277, length=1056),
+
-                              PlaceholderRange(offset=1333, length=418)
+
-                          ])])
+@pytest.mark.parametrize(
+    "prompt,expected_ranges",
+    [
+        pytest.param(IMG_URLS[:1], [PlaceholderRange(offset=11, length=494)]),
+        pytest.param(IMG_URLS[1:4], [
+            PlaceholderRange(offset=11, length=266),
+            PlaceholderRange(offset=277, length=1056),
+            PlaceholderRange(offset=1333, length=418)
+        ])
+    ],
+)
 def test_multi_modal_placeholders(vllm_runner, prompt: TextPrompt,
                                   expected_ranges: list[PlaceholderRange],
                                   monkeypatch) -> None:

tests/multimodal/test_utils.py

@@ -31,11 +31,11 @@ if TYPE_CHECKING:
     from vllm.multimodal.inputs import MultiModalPlaceholderDict
 
 # Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
-TEST_IMAGE_URLS = [
+TEST_IMAGE_ASSETS = [
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",
+    "2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
-    "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
+    "Grayscale_8bits_palette_sample_image.png",  # "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
+    "1280px-Venn_diagram_rgb.svg.png",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
-    "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
+    "RGBA_comp.png",  # "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
 ]
 
 TEST_VIDEO_URLS = [
@@ -45,12 +45,11 @@ TEST_VIDEO_URLS = [
 
 
 @pytest.fixture(scope="module")
-def url_images() -> dict[str, Image.Image]:
+def url_images(local_asset_server) -> dict[str, Image.Image]:
-    connector = MediaConnector()
 
     return {
-        image_url: connector.fetch_image(image_url)
+        image_url: local_asset_server.get_image_asset(image_url)
-        for image_url in TEST_IMAGE_URLS
+        for image_url in TEST_IMAGE_ASSETS
     }
 
 
@@ -69,7 +68,7 @@ def _image_equals(a: Image.Image, b: Image.Image) -> bool:
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_fetch_image_http(image_url: str):
     connector = MediaConnector()
 
@@ -79,12 +78,12 @@ async def test_fetch_image_http(image_url: str):
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("raw_image_url", TEST_IMAGE_ASSETS)
 @pytest.mark.parametrize("suffix", get_supported_suffixes())
 async def test_fetch_image_base64(url_images: dict[str, Image.Image],
-                                  image_url: str, suffix: str):
+                                  raw_image_url: str, suffix: str):
     connector = MediaConnector()
-    url_image = url_images[image_url]
+    url_image = url_images[raw_image_url]
 
     try:
         mime_type = Image.MIME[Image.registered_extensions()[suffix]]
@@ -117,7 +116,7 @@ async def test_fetch_image_base64(url_images: dict[str, Image.Image],
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_fetch_image_local_files(image_url: str):
     connector = MediaConnector()
 
@@ -152,8 +151,8 @@ async def test_fetch_image_local_files(image_url: str):
 
 
 @pytest.mark.asyncio
-async def test_fetch_image_local_files_with_space_in_name():
+@pytest.mark.parametrize("image_url", [TEST_IMAGE_ASSETS[0]], indirect=True)
-    image_url = TEST_IMAGE_URLS[0]
+async def test_fetch_image_local_files_with_space_in_name(image_url: str):
     connector = MediaConnector()
 
     with TemporaryDirectory() as temp_dir:

tests/v1/entrypoints/openai/responses/test_image.py

@@ -8,17 +8,17 @@ import pytest
 import pytest_asyncio
 
 from tests.utils import RemoteOpenAIServer
-from vllm.multimodal.utils import encode_image_base64, fetch_image
+from vllm.multimodal.utils import encode_image_base64
 
 # Use a small vision model for testing
 MODEL_NAME = "Qwen/Qwen2.5-VL-3B-Instruct"
 MAXIMUM_IMAGES = 2
 # Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
-TEST_IMAGE_URLS = [
+TEST_IMAGE_ASSETS = [
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",
+    "2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
-    "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
+    "Grayscale_8bits_palette_sample_image.png",  # "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
+    "1280px-Venn_diagram_rgb.svg.png",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
-    "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
+    "RGBA_comp.png",  # "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
 ]
 
 
@@ -52,16 +52,17 @@ async def client(image_server):
 
 
 @pytest.fixture(scope="session")
-def base64_encoded_image() -> dict[str, str]:
+def base64_encoded_image(local_asset_server) -> dict[str, str]:
     return {
-        image_url: encode_image_base64(fetch_image(image_url))
+        image_url:
-        for image_url in TEST_IMAGE_URLS
+        encode_image_base64(local_asset_server.get_image_asset(image_url))
+        for image_url in TEST_IMAGE_ASSETS
     }
 
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_single_chat_session_image(client: openai.AsyncOpenAI,
                                          model_name: str, image_url: str):
     content_text = "What's in this image?"
@@ -91,11 +92,11 @@ async def test_single_chat_session_image(client: openai.AsyncOpenAI,
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+@pytest.mark.parametrize("raw_image_url", TEST_IMAGE_ASSETS)
 async def test_single_chat_session_image_base64encoded(
     client: openai.AsyncOpenAI,
     model_name: str,
-    image_url: str,
+    raw_image_url: str,
     base64_encoded_image: dict[str, str],
 ):
     content_text = "What's in this image?"
@@ -106,7 +107,7 @@ async def test_single_chat_session_image_base64encoded(
             {
                 "type": "input_image",
                 "image_url":
-                f"data:image/jpeg;base64,{base64_encoded_image[image_url]}",
+                f"data:image/jpeg;base64,{base64_encoded_image[raw_image_url]}",
                 "detail": "auto",
             },
             {
@@ -127,7 +128,8 @@ async def test_single_chat_session_image_base64encoded(
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 @pytest.mark.parametrize(
     "image_urls",
-    [TEST_IMAGE_URLS[:i] for i in range(2, len(TEST_IMAGE_URLS))])
+    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
+    indirect=True)
 async def test_multi_image_input(client: openai.AsyncOpenAI, model_name: str,
                                  image_urls: list[str]):
     messages = [{

tests/v1/tpu/test_multimodal.py

@@ -4,18 +4,19 @@
 import openai
 import pytest
 
-from vllm.multimodal.utils import encode_image_base64, fetch_image
+from vllm.multimodal.utils import encode_image_base64
 from vllm.platforms import current_platform
 
-from ...entrypoints.openai.test_vision import TEST_IMAGE_URLS
+from ...entrypoints.openai.test_vision import TEST_IMAGE_ASSETS
 from ...utils import RemoteOpenAIServer
 
 
 @pytest.fixture(scope="session")
-def base64_encoded_image() -> dict[str, str]:
+def base64_encoded_image(local_asset_server) -> dict[str, str]:
     return {
-        image_url: encode_image_base64(fetch_image(image_url))
+        image_asset:
-        for image_url in TEST_IMAGE_URLS
+        encode_image_base64(local_asset_server.get_image_asset(image_asset))
+        for image_asset in TEST_IMAGE_ASSETS
     }
 
 
@@ -66,7 +67,7 @@ async def test_basic_vision(model_name: str, base64_encoded_image: dict[str,
         client: openai.AsyncOpenAI = remote_server.get_async_client()
 
         # Other requests now should be much faster
-        for image_url in TEST_IMAGE_URLS:
+        for image_url in TEST_IMAGE_ASSETS:
             image_base64 = base64_encoded_image[image_url]
             chat_completion_from_base64 = await client.chat.completions\
                 .create(

vllm/assets/image.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from dataclasses import dataclass
+from pathlib import Path
 from typing import Literal
 
 import torch
@@ -11,17 +12,29 @@ from .base import get_vllm_public_assets
 
 VLM_IMAGES_DIR = "vision_model_images"
 
-ImageAssetName = Literal["stop_sign", "cherry_blossom", "hato"]
+ImageAssetName = Literal["stop_sign", "cherry_blossom", "hato",
+                         "2560px-Gfp-wisconsin-madison-the-nature-boardwalk",
+                         "Grayscale_8bits_palette_sample_image",
+                         "1280px-Venn_diagram_rgb", "RGBA_comp", "237-400x300",
+                         "231-200x300", "27-500x500", "17-150x600",
+                         "handelsblatt-preview", "paper-11"]
 
 
 @dataclass(frozen=True)
 class ImageAsset:
     name: ImageAssetName
 
+    def get_path(self, ext: str) -> Path:
+        """
+        Return s3 path for given image.
+        """
+        return get_vllm_public_assets(filename=f"{self.name}.{ext}",
+                                      s3_prefix=VLM_IMAGES_DIR)
+
     @property
-    def pil_image(self) -> Image.Image:
+    def pil_image(self, ext="jpg") -> Image.Image:
-        image_path = get_vllm_public_assets(filename=f"{self.name}.jpg",
+
-                                            s3_prefix=VLM_IMAGES_DIR)
+        image_path = self.get_path(ext)
         return Image.open(image_path)
 
     @property
@@ -29,6 +42,9 @@ class ImageAsset:
         """
         Image embeddings, only used for testing purposes with llava 1.5.
         """
-        image_path = get_vllm_public_assets(filename=f"{self.name}.pt",
+        image_path = self.get_path('pt')
-                                            s3_prefix=VLM_IMAGES_DIR)
         return torch.load(image_path, map_location="cpu", weights_only=True)
+
+    def read_bytes(self, ext: str) -> bytes:
+        p = Path(self.get_path(ext))
+        return p.read_bytes()