[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
@@ -1,6 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import threading
|
||||
from abc import ABC, abstractmethod
|
||||
from collections.abc import Iterable, Iterator
|
||||
from dataclasses import dataclass, field
|
||||
@@ -9,9 +10,79 @@ from typing import Any, Generic, TypeVar, cast
|
||||
|
||||
import numpy as np
|
||||
|
||||
import vllm.envs as envs
|
||||
|
||||
_T = TypeVar("_T")
|
||||
|
||||
|
||||
class MediaDecodedBudget:
|
||||
"""Thread-safe decoded-media byte accumulator for one request."""
|
||||
|
||||
def __init__(self) -> None:
|
||||
self.used = 0
|
||||
self._reserved = 0
|
||||
self._condition = threading.Condition()
|
||||
|
||||
def reserve(self, max_bytes: int) -> int:
|
||||
"""Reserve decoder capacity, waiting for in-flight decodes if needed."""
|
||||
limit = envs.VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES
|
||||
if limit <= 0:
|
||||
return max_bytes
|
||||
max_bytes = max(0, max_bytes)
|
||||
with self._condition:
|
||||
while True:
|
||||
available = limit - self.used - self._reserved
|
||||
if available >= max_bytes or self._reserved == 0:
|
||||
if available <= 0:
|
||||
raise ValueError(
|
||||
"Request decoded media exceeds "
|
||||
"VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES."
|
||||
)
|
||||
reserved = min(max_bytes, available)
|
||||
self._reserved += reserved
|
||||
return reserved
|
||||
self._condition.wait()
|
||||
|
||||
def settle(self, reserved: int, decoded_bytes: int) -> None:
|
||||
"""Replace a decoder reservation with its actual decoded size."""
|
||||
limit = envs.VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES
|
||||
if limit <= 0:
|
||||
return
|
||||
with self._condition:
|
||||
self._reserved -= reserved
|
||||
try:
|
||||
used = self.used + decoded_bytes
|
||||
if decoded_bytes > reserved or used > limit:
|
||||
raise ValueError(
|
||||
"Request decoded media exceeds "
|
||||
f"VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES={limit} "
|
||||
f"({used} bytes)."
|
||||
)
|
||||
self.used = used
|
||||
finally:
|
||||
self._condition.notify_all()
|
||||
|
||||
def charge(self, decoded_bytes: int) -> None:
|
||||
limit = envs.VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES
|
||||
if limit <= 0 or decoded_bytes <= 0:
|
||||
return
|
||||
with self._condition:
|
||||
while (
|
||||
self.used + self._reserved + decoded_bytes > limit
|
||||
and self._reserved > 0
|
||||
and self.used + decoded_bytes <= limit
|
||||
):
|
||||
self._condition.wait()
|
||||
used = self.used + decoded_bytes
|
||||
if used > limit:
|
||||
raise ValueError(
|
||||
"Request decoded media exceeds "
|
||||
f"VLLM_MAX_MEDIA_REQUEST_DECODE_BYTES={limit} "
|
||||
f"({used} bytes)."
|
||||
)
|
||||
self.used = used
|
||||
|
||||
|
||||
@dataclass
|
||||
class MediaWithBytes(Generic[_T]):
|
||||
"""Wrapper that couples a media object with its original encoded bytes.
|
||||
|
||||
Reference in New Issue
Block a user