[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:
co-authored by
OpenAI Codex
parent
877f6df889
commit
7173c64e2e
@@ -1,5 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import threading
|
||||
from concurrent.futures import ThreadPoolExecutor
|
||||
from io import BytesIO
|
||||
from pathlib import Path
|
||||
from unittest.mock import patch
|
||||
@@ -9,6 +11,7 @@ import pybase64 as base64
|
||||
import pytest
|
||||
import soundfile as sf
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.multimodal.media import AudioMediaIO
|
||||
from vllm.multimodal.media import audio as audio_module
|
||||
from vllm.multimodal.media.audio import (
|
||||
@@ -16,6 +19,7 @@ from vllm.multimodal.media.audio import (
|
||||
load_audio_soundfile,
|
||||
load_audio_torchcodec,
|
||||
)
|
||||
from vllm.multimodal.media.base import MediaDecodedBudget
|
||||
|
||||
from ...conftest import AudioTestAssets
|
||||
|
||||
@@ -118,6 +122,102 @@ def _make_flac_bytes(frames: int, channels: int, samplerate: int) -> bytes:
|
||||
return buf.getvalue()
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_request_budget_rejects_second_decoded_audio_clip(monkeypatch):
|
||||
monkeypatch.setattr(envs, "VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES", 384)
|
||||
clip = _make_flac_bytes(frames=64, channels=1, samplerate=16000)
|
||||
audio_io = AudioMediaIO(decoded_budget=MediaDecodedBudget())
|
||||
|
||||
audio_io.load_bytes(clip)
|
||||
with pytest.raises(ValueError, match="VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES"):
|
||||
audio_io.load_bytes(clip)
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_request_budget_allows_parallel_audio_decodes(monkeypatch):
|
||||
monkeypatch.setattr(envs, "VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES", 512)
|
||||
monkeypatch.setattr(envs, "VLLM_MAX_AUDIO_DECODE_BYTES", 256)
|
||||
budget = MediaDecodedBudget()
|
||||
audio_io = AudioMediaIO(decoded_budget=budget)
|
||||
decoding = threading.Barrier(2)
|
||||
|
||||
def decode(*args, **kwargs):
|
||||
assert kwargs["max_decode_bytes"] == 256
|
||||
decoding.wait(timeout=5)
|
||||
return np.zeros(32, dtype=np.float32), 16000
|
||||
|
||||
with (
|
||||
patch.object(audio_module, "load_audio", side_effect=decode),
|
||||
ThreadPoolExecutor(max_workers=2) as pool,
|
||||
):
|
||||
results = list(pool.map(audio_io.load_bytes, [b"a", b"b"]))
|
||||
|
||||
assert len(results) == 2
|
||||
assert budget.used == 256
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_request_budget_waits_for_inflight_audio_decode(monkeypatch):
|
||||
monkeypatch.setattr(envs, "VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES", 256)
|
||||
monkeypatch.setattr(envs, "VLLM_MAX_AUDIO_DECODE_BYTES", 256)
|
||||
budget = MediaDecodedBudget()
|
||||
audio_io = AudioMediaIO(decoded_budget=budget)
|
||||
decoding = threading.Event()
|
||||
finish_decoding = threading.Event()
|
||||
second_reservation = threading.Event()
|
||||
reserve = budget.reserve
|
||||
decode_calls = 0
|
||||
|
||||
def track_reservation(max_bytes):
|
||||
if decoding.is_set():
|
||||
second_reservation.set()
|
||||
return reserve(max_bytes)
|
||||
|
||||
def decode(*args, **kwargs):
|
||||
nonlocal decode_calls
|
||||
decode_calls += 1
|
||||
decoding.set()
|
||||
assert finish_decoding.wait(timeout=5)
|
||||
return np.zeros(64, dtype=np.float32), 16000
|
||||
|
||||
with (
|
||||
patch.object(budget, "reserve", side_effect=track_reservation),
|
||||
patch.object(audio_module, "load_audio", side_effect=decode),
|
||||
ThreadPoolExecutor(max_workers=2) as pool,
|
||||
):
|
||||
first = pool.submit(audio_io.load_bytes, b"a")
|
||||
assert decoding.wait(timeout=5)
|
||||
second = pool.submit(audio_io.load_bytes, b"b")
|
||||
try:
|
||||
assert second_reservation.wait(timeout=5)
|
||||
finally:
|
||||
finish_decoding.set()
|
||||
first.result(timeout=5)
|
||||
with pytest.raises(ValueError, match="VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES"):
|
||||
second.result(timeout=5)
|
||||
|
||||
assert decode_calls == 1
|
||||
assert budget.used == 256
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_request_budget_releases_failed_audio_reservation(monkeypatch):
|
||||
monkeypatch.setattr(envs, "VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES", 256)
|
||||
monkeypatch.setattr(envs, "VLLM_MAX_AUDIO_DECODE_BYTES", 256)
|
||||
budget = MediaDecodedBudget()
|
||||
audio_io = AudioMediaIO(decoded_budget=budget)
|
||||
audio = (np.zeros(64, dtype=np.float32), 16000)
|
||||
|
||||
with patch.object(
|
||||
audio_module, "load_audio", side_effect=[ValueError("invalid audio"), audio]
|
||||
):
|
||||
with pytest.raises(ValueError, match="invalid audio"):
|
||||
audio_io.load_bytes(b"a")
|
||||
assert audio_io.load_bytes(b"b")[0].nbytes == 256
|
||||
|
||||
assert budget.used == 256
|
||||
|
||||
|
||||
def test_small_file_passes_memory_guard():
|
||||
"""A small valid file should pass both duration and memory guards."""
|
||||
payload = _make_flac_bytes(frames=16000, channels=1, samplerate=16000)
|
||||
|
||||
Reference in New Issue
Block a user