[Security] Bound decoded and downloaded media per request (#61226)

Signed-off-by: juanpas97 <juan.perezalgaba@gmail.com>
Signed-off-by: Juan Pérez de Algaba <jperezde@redhat.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Juan Pérez de Algaba
2026-10-13 10:17:14 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 877f6df889
commit 7173c64e2e
16 changed files with 565 additions and 50 deletions
+51 -3
View File
@@ -2,15 +2,19 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import AsyncIterator, Iterator
from io import BytesIO
from pathlib import Path
from unittest.mock import patch
import numpy as np
import pytest
from PIL import Image
import vllm.envs as envs
from vllm.connections import HTTPConnection
from vllm.connections import HTTPConnection, MediaRequestDownloadBudget
from vllm.exceptions import VLLMValidationError
from vllm.multimodal.media import AudioMediaIO, MediaConnector
from vllm.multimodal.media import AudioMediaIO, ImageMediaIO, MediaConnector
from vllm.multimodal.media.base import MediaDecodedBudget
from vllm.utils.mem_constants import KiB_bytes, MiB_bytes
pytestmark = [pytest.mark.cpu_test, pytest.mark.skip_global_cleanup]
@@ -87,6 +91,46 @@ def test_audio_load_bytes_rejects_oversized(monkeypatch: pytest.MonkeyPatch):
AudioMediaIO().load_bytes(b"\x00" * (MiB_bytes + 1))
def test_request_audio_budget_limits_multiple_clips_before_decode(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.setattr(envs, "VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES", 24)
budget = MediaDecodedBudget()
clip = np.zeros(4, dtype=np.float32)
def decode(_source, **kwargs):
if kwargs["max_decode_bytes"] < clip.nbytes:
raise ValueError("VLLM_MAX_AUDIO_DECODE_BYTES")
return clip, 16000
with patch("vllm.multimodal.media.audio.load_audio", side_effect=decode) as load:
audio_io = AudioMediaIO(decoded_budget=budget)
audio_io.load_bytes(b"first")
with pytest.raises(ValueError, match="VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES"):
audio_io.load_bytes(b"second")
assert budget.used == clip.nbytes
assert load.call_count == 2
def test_request_decoded_budget_is_shared_by_image_and_audio(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.setattr(envs, "VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES", 24)
budget = MediaDecodedBudget()
image_bytes = BytesIO()
Image.new("RGB", (2, 3)).save(image_bytes, format="PNG")
ImageMediaIO(decoded_budget=budget).load_bytes(image_bytes.getvalue())
with (
patch("vllm.multimodal.media.audio.load_audio") as decode,
pytest.raises(ValueError, match="VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES"),
):
AudioMediaIO(decoded_budget=budget).load_bytes(b"audio")
decode.assert_not_called()
def test_audio_load_file_rejects_oversized(
monkeypatch: pytest.MonkeyPatch, tmp_path: Path
):
@@ -138,12 +182,16 @@ async def test_audio_connector_threads_byte_limit_to_http_reader(
monkeypatch: pytest.MonkeyPatch,
):
monkeypatch.setattr(envs, "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", 1)
budget = MediaRequestDownloadBudget()
class _Connection:
async def async_get_bytes(self, _url: str, **kwargs):
assert kwargs["max_bytes"] == MiB_bytes
assert kwargs["request_budget"] is budget
raise VLLMValidationError("Maximum file size exceeded")
connector = MediaConnector(connection=_Connection())
connector = MediaConnector( # type: ignore[arg-type]
connection=_Connection(), download_budget=budget
)
with pytest.raises(VLLMValidationError, match="Maximum file size exceeded"):
await connector.fetch_audio_async("https://example.test/audio")