| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import asyncio |
| 5 | import atexit |
| 6 | import contextlib |
| 7 | import hashlib |
| 8 | import os |
| 9 | import tempfile |
| 10 | import time |
| 11 | from concurrent.futures import ThreadPoolExecutor |
| 12 | from pathlib import Path |
| 13 | from typing import Any, TypeVar |
| 14 | from urllib.request import url2pathname |
| 15 | |
| 16 | import aiohttp |
| 17 | import numpy as np |
| 18 | import numpy.typing as npt |
| 19 | import requests |
| 20 | import torch |
| 21 | from PIL import Image, UnidentifiedImageError |
| 22 | from urllib3.util import Url, parse_url |
| 23 | |
| 24 | import vllm.envs as envs |
| 25 | from vllm.connections import ( |
| 26 | HTTPConnection, |
| 27 | MediaDownloadSizeExceededError, |
| 28 | global_http_connection, |
| 29 | ) |
| 30 | from vllm.exceptions import VLLMUnprocessableEntityError, VLLMValidationError |
| 31 | from vllm.logger import init_logger |
| 32 | from vllm.multimodal.video import get_video_loader_backend_for_processor |
| 33 | from vllm.utils.registry import ExtensionManager |
| 34 | |
| 35 | from .audio import AudioEmbeddingMediaIO, AudioMediaIO |
| 36 | from .base import MediaIO, MediaWithBytes |
| 37 | from .image import ImageEmbeddingMediaIO, ImageMediaIO |
| 38 | from .video import VideoEmbeddingMediaIO, VideoMediaIO |
| 39 | |
| 40 | logger = init_logger(__name__) |
| 41 | |
| 42 | _M = TypeVar("_M") |
| 43 | |
| 44 | global_thread_pool = ThreadPoolExecutor( |
| 45 | max_workers=envs.VLLM_MEDIA_LOADING_THREAD_COUNT |
| 46 | ) |
| 47 | atexit.register(global_thread_pool.shutdown) |
| 48 | |
| 49 | MEDIA_CONNECTOR_REGISTRY = ExtensionManager() |
| 50 | |
| 51 | MODALITY_IO_MAP: dict[str, type[MediaIO]] = { |
| 52 | "audio": AudioMediaIO, |
| 53 | "image": ImageMediaIO, |
| 54 | "video": VideoMediaIO, |
| 55 | } |
| 56 | |
| 57 | |
| 58 | def _wrap_media_fetch_error( |
| 59 | url: str, exc: Exception |
| 60 | ) -> VLLMUnprocessableEntityError | Exception: |
| 61 | """Convert media fetch exceptions to VLLMUnprocessableEntityError. |
| 62 | |
| 63 | This handles HTTP errors that indicate the media resource is invalid |
| 64 | (4xx responses except 408/429, malformed URLs) and converts them to a |
| 65 | 422 Unprocessable Entity error instead of 500. |
| 66 | |
| 67 | Transient errors (5xx, 408, 429, DNS failures, connection errors, |
| 68 | timeouts) are returned as-is to allow retry logic to handle them |
| 69 | appropriately. |
| 70 | |
| 71 | Returns: |
| 72 | VLLMUnprocessableEntityError for permanent client errors (4xx except |
| 73 | 408/429, invalid URL) |
| 74 | Original exception for transient errors (5xx, 408, 429, network blips) |
| 75 | or other exceptions |
| 76 | |
| 77 | """ |
| 78 | if isinstance(exc, VLLMValidationError): |
| 79 | return exc |
| 80 | |
| 81 | if isinstance(exc, aiohttp.ClientResponseError): |
| 82 | if exc.status in (408, 429): |
| 83 | return exc |
| 84 | if exc.status < 500: |
| 85 | return VLLMUnprocessableEntityError( |
| 86 | f"Failed to fetch media from URL: HTTP {exc.status} error", |
| 87 | parameter="image_url", |
| 88 | value=url, |
| 89 | ) |
| 90 | return exc |
| 91 | |
| 92 | if isinstance(exc, requests.exceptions.HTTPError): |
| 93 | if exc.response is not None: |
| 94 | status_code = exc.response.status_code |
| 95 | if status_code in (408, 429): |
| 96 | return exc |
| 97 | if status_code < 500: |
| 98 | return VLLMUnprocessableEntityError( |
| 99 | f"Failed to fetch media from URL: HTTP {status_code} error", |
| 100 | parameter="image_url", |
| 101 | value=url, |
| 102 | ) |
| 103 | return exc |
| 104 | |
| 105 | if isinstance(exc, requests.exceptions.InvalidURL): |
| 106 | return VLLMUnprocessableEntityError( |
| 107 | "Failed to fetch media from URL: Invalid URL format", |
| 108 | parameter="image_url", |
| 109 | value=url, |
| 110 | ) |
| 111 | |
| 112 | if isinstance(exc, MediaDownloadSizeExceededError): |
| 113 | return VLLMUnprocessableEntityError( |
| 114 | f"Failed to fetch media from URL: {exc}", |
| 115 | parameter="image_url", |
| 116 | value=url, |
| 117 | ) |
| 118 | |
| 119 | if isinstance(exc, ValueError): |
| 120 | return VLLMUnprocessableEntityError( |
| 121 | "Failed to fetch media from URL: Invalid URL", |
| 122 | parameter="image_url", |
| 123 | value=url, |
| 124 | ) |
| 125 | return exc |
| 126 | |
| 127 | |
| 128 | def merge_media_io_kwargs( |
| 129 | defaults: dict[str, dict[str, Any]] | None, |
| 130 | overrides: dict[str, dict[str, Any]] | None, |
| 131 | ) -> dict[str, dict[str, Any]] | None: |
| 132 | """Merge config-level and per-request media_io_kwargs per modality. |
| 133 | |
| 134 | Each modality key is merged using the corresponding MediaIO subclass's |
| 135 | ``merge_kwargs``, which may apply modality-specific logic (e.g. |
| 136 | VideoMediaIO clears cross-dependent fps/num_frames fields). |
| 137 | """ |
| 138 | if not defaults and not overrides: |
| 139 | return None |
| 140 | all_keys = set(defaults or {}) | set(overrides or {}) |
| 141 | merged = {} |
| 142 | for key in all_keys: |
| 143 | io_cls = MODALITY_IO_MAP.get(key, MediaIO) |
| 144 | merged[key] = io_cls.merge_kwargs( |
| 145 | (defaults or {}).get(key), |
| 146 | (overrides or {}).get(key), |
| 147 | ) |
| 148 | return merged or None |
| 149 | |
| 150 | |
| 151 | @MEDIA_CONNECTOR_REGISTRY.register("http") |
| 152 | class MediaConnector: |
| 153 | """Configuration values can be user-provided either by --media-io-kwargs or |
| 154 | by the runtime API field "media_io_kwargs". Ensure proper validation and |
| 155 | error handling. |
| 156 | """ |
| 157 | |
| 158 | def __init__( |
| 159 | self, |
| 160 | media_io_kwargs: dict[str, dict[str, Any]] | None = None, |
| 161 | connection: HTTPConnection = global_http_connection, |
| 162 | *, |
| 163 | allowed_local_media_path: str = "", |
| 164 | allowed_media_domains: list[str] | None = None, |
| 165 | ) -> None: |
| 166 | """Args: |
| 167 | media_io_kwargs: Additional args passed to process media |
| 168 | inputs, keyed by modalities. For example, |
| 169 | to set num_frames for video, set |
| 170 | `--media-io-kwargs '{"video":{"num_frames":40}}'` |
| 171 | connection: HTTP connection client to download media contents. |
| 172 | allowed_local_media_path: A local directory to load media files from. |
| 173 | allowed_media_domains: If set, only media URLs that belong to this |
| 174 | domain can be used for multi-modal inputs. |
| 175 | |
| 176 | """ |
| 177 | super().__init__() |
| 178 | |
| 179 | self.media_io_kwargs: dict[str, dict[str, Any]] = ( |
| 180 | media_io_kwargs if media_io_kwargs else {} |
| 181 | ) |
| 182 | self.connection = connection |
| 183 | |
| 184 | if allowed_local_media_path: |
| 185 | allowed_local_media_path_ = Path(allowed_local_media_path).resolve() |
| 186 | |
| 187 | if not allowed_local_media_path_.exists(): |
| 188 | raise ValueError( |
| 189 | "Invalid `--allowed-local-media-path`: The path " |
| 190 | f"{allowed_local_media_path_} does not exist." |
| 191 | ) |
| 192 | if not allowed_local_media_path_.is_dir(): |
| 193 | raise ValueError( |
| 194 | "Invalid `--allowed-local-media-path`: The path " |
| 195 | f"{allowed_local_media_path_} must be a directory." |
| 196 | ) |
| 197 | else: |
| 198 | allowed_local_media_path_ = None |
| 199 | |
| 200 | self.allowed_local_media_path = allowed_local_media_path_ |
| 201 | if allowed_media_domains is None: |
| 202 | allowed_media_domains = [] |
| 203 | self.allowed_media_domains = allowed_media_domains |
| 204 | |
| 205 | # Media download cache (opt-in via VLLM_MEDIA_CACHE) |
| 206 | self._media_cache_dir: str | None = None |
| 207 | self._media_cache_max_bytes: int = 0 |
| 208 | self._media_cache_ttl_secs: float = 0 |
| 209 | media_cache = envs.VLLM_MEDIA_CACHE |
| 210 | if media_cache: |
| 211 | try: |
| 212 | os.makedirs(media_cache, exist_ok=True) |
| 213 | # Verify the directory is writable before enabling caching |
| 214 | with tempfile.NamedTemporaryFile(dir=media_cache, delete=True): |
| 215 | pass |
| 216 | self._media_cache_dir = media_cache |
| 217 | self._media_cache_max_bytes = ( |
| 218 | envs.VLLM_MEDIA_CACHE_MAX_SIZE_MB * 1024 * 1024 |
| 219 | ) |
| 220 | self._media_cache_ttl_secs = envs.VLLM_MEDIA_CACHE_TTL_HOURS * 3600 |
| 221 | logger.info( |
| 222 | "Media cache enabled at %s (max %d MB, TTL %s hours)", |
| 223 | media_cache, |
| 224 | envs.VLLM_MEDIA_CACHE_MAX_SIZE_MB, |
| 225 | envs.VLLM_MEDIA_CACHE_TTL_HOURS, |
| 226 | ) |
| 227 | except OSError: |
| 228 | logger.warning( |
| 229 | "VLLM_MEDIA_CACHE path %s is not writable, media caching disabled", |
| 230 | media_cache, |
| 231 | ) |
| 232 | |
| 233 | def _get_cached_bytes(self, url: str) -> bytes | None: |
| 234 | """Return cached bytes for a URL, or None if not cached/expired.""" |
| 235 | if not self._media_cache_dir: |
| 236 | return None |
| 237 | cache_path = self._media_cache_path(url) |
| 238 | # Check TTL |
| 239 | try: |
| 240 | age = time.time() - cache_path.stat().st_mtime |
| 241 | except OSError: |
| 242 | return None |
| 243 | if age > self._media_cache_ttl_secs: |
| 244 | cache_path.unlink(missing_ok=True) |
| 245 | return None |
| 246 | # Touch mtime for LRU ordering |
| 247 | try: |
| 248 | cache_path.touch() |
| 249 | return cache_path.read_bytes() |
| 250 | except OSError: |
| 251 | return None |
| 252 | |
| 253 | def _put_cached_bytes(self, url: str, data: bytes) -> None: |
| 254 | """Store downloaded bytes and evict if over budget.""" |
| 255 | if not self._media_cache_dir: |
| 256 | return |
| 257 | cache_path = self._media_cache_path(url) |
| 258 | # Atomic write via temp file + rename |
| 259 | tmp_path = None |
| 260 | try: |
| 261 | with tempfile.NamedTemporaryFile( |
| 262 | mode="wb", dir=self._media_cache_dir, delete=False |
| 263 | ) as tmp_file: |
| 264 | tmp_file.write(data) |
| 265 | tmp_path = tmp_file.name |
| 266 | os.rename(tmp_path, str(cache_path)) |
| 267 | except OSError: |
| 268 | # Another process beat us or disk issue |
| 269 | if tmp_path is not None: |
| 270 | with contextlib.suppress(OSError): |
| 271 | os.remove(tmp_path) |
| 272 | return |
| 273 | self._maybe_evict(exclude=cache_path) |
| 274 | |
| 275 | def _maybe_evict(self, exclude: Path | None = None) -> None: |
| 276 | """Evict expired entries first, then LRU until under size limit.""" |
| 277 | cache_dir = Path(self._media_cache_dir) # type: ignore[arg-type] |
| 278 | entries = [] |
| 279 | expired = [] |
| 280 | total_size = 0 |
| 281 | now = time.time() |
| 282 | for f in cache_dir.iterdir(): |
| 283 | if f.name.startswith("."): |
| 284 | continue |
| 285 | try: |
| 286 | stat = f.stat() |
| 287 | except OSError: |
| 288 | continue |
| 289 | age = now - stat.st_mtime |
| 290 | if age > self._media_cache_ttl_secs: |
| 291 | expired.append(f) |
| 292 | continue |
| 293 | total_size += stat.st_size |
| 294 | # Never evict the file we just wrote |
| 295 | if exclude is not None and f.name == exclude.name: |
| 296 | continue |
| 297 | entries.append((stat.st_mtime, stat.st_size, f)) |
| 298 | |
| 299 | # Evict items according to LRU policy |
| 300 | entries.sort(key=lambda e: e[0], reverse=True) |
| 301 | while total_size > self._media_cache_max_bytes and entries: |
| 302 | mtime, size, f = entries.pop() |
| 303 | expired.append(f) |
| 304 | total_size -= size |
| 305 | |
| 306 | for f in expired: |
| 307 | f.unlink(missing_ok=True) |
| 308 | |
| 309 | def _media_cache_path(self, url: str) -> Path: |
| 310 | url_hash = hashlib.sha256(url.encode()).hexdigest()[:20] |
| 311 | ext = Path(url.split("?", 1)[0]).suffix or "" |
| 312 | return Path(self._media_cache_dir) / f"{url_hash}{ext}" # type: ignore[arg-type] |
| 313 | |
| 314 | def _load_data_url( |
| 315 | self, |
| 316 | url: str, |
| 317 | media_io: MediaIO[_M], |
| 318 | ) -> _M: # type: ignore[type-var] |
| 319 | # Format per RFC 2397: |
| 320 | # data:[<mediatype>][;<param>=<value>]*[;base64],<data> |
| 321 | data_spec, sep, data = url[5:].partition(",") |
| 322 | if not sep: |
| 323 | msg = f"Invalid data URL {url[:32]!r}: missing ',' separator." |
| 324 | raise ValueError(msg) |
| 325 | |
| 326 | media_type, sep, encoding = data_spec.rpartition(";") |
| 327 | if not sep or encoding != "base64": |
| 328 | msg = "Only base64 data URLs are supported for now." |
| 329 | raise NotImplementedError(msg) |
| 330 | |
| 331 | media_type = media_type.partition(";")[0] |
| 332 | return media_io.load_base64(media_type, data) |
| 333 | |
| 334 | def _load_file_url( |
| 335 | self, |
| 336 | url_spec: Url, |
| 337 | media_io: MediaIO[_M], |
| 338 | ) -> _M: # type: ignore[type-var] |
| 339 | allowed_local_media_path = self.allowed_local_media_path |
| 340 | if allowed_local_media_path is None: |
| 341 | raise RuntimeError( |
| 342 | "Cannot load local files without `--allowed-local-media-path`." |
| 343 | ) |
| 344 | |
| 345 | url_spec_path = url_spec.path or "" |
| 346 | url_spec_netloc = url_spec.netloc or "" |
| 347 | filepath = Path(url2pathname(url_spec_netloc + url_spec_path)) |
| 348 | if allowed_local_media_path not in filepath.resolve().parents: |
| 349 | raise ValueError( |
| 350 | f"The file path {filepath} must be a subpath " |
| 351 | f"of `--allowed-local-media-path {allowed_local_media_path}`." |
| 352 | ) |
| 353 | |
| 354 | return media_io.load_file(filepath) |
| 355 | |
| 356 | def _assert_url_in_allowed_media_domains(self, url_spec: Url) -> None: |
| 357 | if ( |
| 358 | self.allowed_media_domains |
| 359 | and url_spec.hostname not in self.allowed_media_domains |
| 360 | ): |
| 361 | raise ValueError( |
| 362 | f"The URL must be from one of the allowed domains: " |
| 363 | f"{self.allowed_media_domains}. Input URL domain: " |
| 364 | f"{url_spec.hostname}" |
| 365 | ) |
| 366 | |
| 367 | def load_from_url( |
| 368 | self, |
| 369 | url: str, |
| 370 | media_io: MediaIO[_M], |
| 371 | *, |
| 372 | fetch_timeout: int | None = None, |
| 373 | ) -> _M: # type: ignore[type-var] |
| 374 | if url[:5].lower() == "data:": |
| 375 | return self._load_data_url(url, media_io) |
| 376 | |
| 377 | url_spec = parse_url(url) |
| 378 | |
| 379 | if url_spec.scheme and url_spec.scheme.startswith("http"): |
| 380 | self._assert_url_in_allowed_media_domains(url_spec) |
| 381 | max_bytes = media_io.get_max_bytes() |
| 382 | |
| 383 | cached = self._get_cached_bytes(url) |
| 384 | if cached is not None: |
| 385 | return media_io.load_bytes(cached) |
| 386 | |
| 387 | connection = self.connection |
| 388 | try: |
| 389 | data = connection.get_bytes( |
| 390 | url_spec.url, |
| 391 | timeout=fetch_timeout, |
| 392 | allow_redirects=envs.VLLM_MEDIA_URL_ALLOW_REDIRECTS, |
| 393 | max_bytes=max_bytes, |
| 394 | ) |
| 395 | except Exception as e: |
| 396 | wrapped = _wrap_media_fetch_error(url, e) |
| 397 | if isinstance(wrapped, VLLMUnprocessableEntityError): |
| 398 | raise wrapped from e |
| 399 | raise |
| 400 | |
| 401 | self._put_cached_bytes(url, data) |
| 402 | return media_io.load_bytes(data) |
| 403 | |
| 404 | if url_spec.scheme == "file": |
| 405 | return self._load_file_url(url_spec, media_io) |
| 406 | |
| 407 | msg = "The URL must be either a HTTP, data or file URL." |
| 408 | raise ValueError(msg) |
| 409 | |
| 410 | async def load_from_url_async( |
| 411 | self, |
| 412 | url: str, |
| 413 | media_io: MediaIO[_M], |
| 414 | *, |
| 415 | fetch_timeout: int | None = None, |
| 416 | ) -> _M: |
| 417 | loop = asyncio.get_running_loop() |
| 418 | |
| 419 | if url[:5].lower() == "data:": |
| 420 | future = loop.run_in_executor( |
| 421 | global_thread_pool, self._load_data_url, url, media_io |
| 422 | ) |
| 423 | return await future |
| 424 | |
| 425 | url_spec = parse_url(url) |
| 426 | |
| 427 | if url_spec.scheme and url_spec.scheme.startswith("http"): |
| 428 | self._assert_url_in_allowed_media_domains(url_spec) |
| 429 | max_bytes = media_io.get_max_bytes() |
| 430 | |
| 431 | cached = await loop.run_in_executor( |
| 432 | global_thread_pool, self._get_cached_bytes, url |
| 433 | ) |
| 434 | if cached is not None: |
| 435 | future = loop.run_in_executor( |
| 436 | global_thread_pool, media_io.load_bytes, cached |
| 437 | ) |
| 438 | return await future |
| 439 | |
| 440 | connection = self.connection |
| 441 | try: |
| 442 | data = await connection.async_get_bytes( |
| 443 | url_spec.url, |
| 444 | timeout=fetch_timeout, |
| 445 | allow_redirects=envs.VLLM_MEDIA_URL_ALLOW_REDIRECTS, |
| 446 | max_bytes=max_bytes, |
| 447 | ) |
| 448 | except Exception as e: |
| 449 | wrapped = _wrap_media_fetch_error(url, e) |
| 450 | if isinstance(wrapped, VLLMUnprocessableEntityError): |
| 451 | raise wrapped from e |
| 452 | raise |
| 453 | |
| 454 | await loop.run_in_executor( |
| 455 | global_thread_pool, self._put_cached_bytes, url, data |
| 456 | ) |
| 457 | future = loop.run_in_executor(global_thread_pool, media_io.load_bytes, data) |
| 458 | return await future |
| 459 | |
| 460 | if url_spec.scheme == "file": |
| 461 | future = loop.run_in_executor( |
| 462 | global_thread_pool, self._load_file_url, url_spec, media_io |
| 463 | ) |
| 464 | return await future |
| 465 | msg = "The URL must be either a HTTP, data or file URL." |
| 466 | raise ValueError(msg) |
| 467 | |
| 468 | def fetch_audio( |
| 469 | self, |
| 470 | audio_url: str, |
| 471 | ) -> tuple[np.ndarray, int | float]: |
| 472 | """Load audio from a URL.""" |
| 473 | audio_io = AudioMediaIO(**self.media_io_kwargs.get("audio", {})) |
| 474 | |
| 475 | return self.load_from_url( |
| 476 | audio_url, |
| 477 | audio_io, |
| 478 | fetch_timeout=envs.VLLM_AUDIO_FETCH_TIMEOUT, |
| 479 | ) |
| 480 | |
| 481 | async def fetch_audio_async( |
| 482 | self, |
| 483 | audio_url: str, |
| 484 | ) -> tuple[np.ndarray, int | float]: |
| 485 | """Asynchronously fetch audio from a URL.""" |
| 486 | audio_io = AudioMediaIO(**self.media_io_kwargs.get("audio", {})) |
| 487 | |
| 488 | return await self.load_from_url_async( |
| 489 | audio_url, |
| 490 | audio_io, |
| 491 | fetch_timeout=envs.VLLM_AUDIO_FETCH_TIMEOUT, |
| 492 | ) |
| 493 | |
| 494 | def fetch_image( |
| 495 | self, |
| 496 | image_url: str, |
| 497 | *, |
| 498 | image_mode: str | None = "RGB", |
| 499 | ) -> Image.Image: |
| 500 | """Load a PIL image from an HTTP or base64 data URL. |
| 501 | |
| 502 | By default, the image is converted into RGB format. Set |
| 503 | `media_io_kwargs={"image": {"image_mode": None}}` to keep the |
| 504 | original image mode (e.g. preserving the alpha channel). |
| 505 | """ |
| 506 | image_io = ImageMediaIO( |
| 507 | **({"image_mode": image_mode} | self.media_io_kwargs.get("image", {})) |
| 508 | ) |
| 509 | |
| 510 | try: |
| 511 | return self.load_from_url( |
| 512 | image_url, |
| 513 | image_io, |
| 514 | fetch_timeout=envs.VLLM_IMAGE_FETCH_TIMEOUT, |
| 515 | ) |
| 516 | except UnidentifiedImageError as e: |
| 517 | # convert to ValueError to be properly caught upstream |
| 518 | raise ValueError(str(e)) from e |
| 519 | |
| 520 | async def fetch_image_async( |
| 521 | self, |
| 522 | image_url: str, |
| 523 | *, |
| 524 | image_mode: str | None = "RGB", |
| 525 | ) -> Image.Image: |
| 526 | """Asynchronously load a PIL image from an HTTP or base64 data URL. |
| 527 | |
| 528 | By default, the image is converted into RGB format. Set |
| 529 | `media_io_kwargs={"image": {"image_mode": None}}` to keep the |
| 530 | original image mode (e.g. preserving the alpha channel). |
| 531 | """ |
| 532 | image_io = ImageMediaIO( |
| 533 | **({"image_mode": image_mode} | self.media_io_kwargs.get("image", {})) |
| 534 | ) |
| 535 | |
| 536 | try: |
| 537 | return await self.load_from_url_async( |
| 538 | image_url, |
| 539 | image_io, |
| 540 | fetch_timeout=envs.VLLM_IMAGE_FETCH_TIMEOUT, |
| 541 | ) |
| 542 | except UnidentifiedImageError as e: |
| 543 | # convert to ValueError to be properly caught upstream |
| 544 | raise ValueError(str(e)) from e |
| 545 | |
| 546 | def fetch_video( |
| 547 | self, |
| 548 | video_url: str, |
| 549 | *, |
| 550 | image_mode: str | None = "RGB", |
| 551 | video_processor: str | None = None, |
| 552 | ) -> MediaWithBytes[tuple[npt.NDArray, dict[str, Any]]]: |
| 553 | """Load video from an HTTP or base64 data URL.""" |
| 554 | image_io = ImageMediaIO( |
| 555 | **({"image_mode": image_mode} | self.media_io_kwargs.get("image", {})) |
| 556 | ) |
| 557 | video_io_kwargs = dict(self.media_io_kwargs.get("video", {})) |
| 558 | if "video_backend" not in video_io_kwargs and ( |
| 559 | video_backend := get_video_loader_backend_for_processor(video_processor) |
| 560 | ): |
| 561 | video_io_kwargs["video_backend"] = video_backend |
| 562 | video_io = VideoMediaIO(image_io, **video_io_kwargs) |
| 563 | |
| 564 | return self.load_from_url( |
| 565 | video_url, |
| 566 | video_io, |
| 567 | fetch_timeout=envs.VLLM_VIDEO_FETCH_TIMEOUT, |
| 568 | ) |
| 569 | |
| 570 | async def fetch_video_async( |
| 571 | self, |
| 572 | video_url: str, |
| 573 | *, |
| 574 | image_mode: str | None = "RGB", |
| 575 | video_processor: str | None = None, |
| 576 | ) -> MediaWithBytes[tuple[npt.NDArray, dict[str, Any]]]: |
| 577 | """Asynchronously load video from an HTTP or base64 data URL. |
| 578 | |
| 579 | By default, the image is converted into RGB format. Set |
| 580 | `media_io_kwargs={"image": {"image_mode": None}}` to keep the |
| 581 | original image mode (e.g. preserving the alpha channel). |
| 582 | """ |
| 583 | image_io = ImageMediaIO( |
| 584 | **({"image_mode": image_mode} | self.media_io_kwargs.get("image", {})) |
| 585 | ) |
| 586 | video_io_kwargs = dict(self.media_io_kwargs.get("video", {})) |
| 587 | if "video_backend" not in video_io_kwargs and ( |
| 588 | video_backend := get_video_loader_backend_for_processor(video_processor) |
| 589 | ): |
| 590 | video_io_kwargs["video_backend"] = video_backend |
| 591 | video_io = VideoMediaIO(image_io, **video_io_kwargs) |
| 592 | |
| 593 | return await self.load_from_url_async( |
| 594 | video_url, |
| 595 | video_io, |
| 596 | fetch_timeout=envs.VLLM_VIDEO_FETCH_TIMEOUT, |
| 597 | ) |
| 598 | |
| 599 | def fetch_image_embedding( |
| 600 | self, |
| 601 | data: str, |
| 602 | ) -> torch.Tensor: |
| 603 | """Load image embedding from a URL.""" |
| 604 | image_embedding_io = ImageEmbeddingMediaIO() |
| 605 | |
| 606 | return image_embedding_io.load_base64("", data) |
| 607 | |
| 608 | async def fetch_image_embedding_async( |
| 609 | self, |
| 610 | data: str, |
| 611 | ) -> torch.Tensor: |
| 612 | """Asynchronously load image embedding from a URL.""" |
| 613 | image_embedding_io = ImageEmbeddingMediaIO() |
| 614 | loop = asyncio.get_running_loop() |
| 615 | |
| 616 | return await loop.run_in_executor( |
| 617 | global_thread_pool, image_embedding_io.load_base64, "", data |
| 618 | ) |
| 619 | |
| 620 | def fetch_audio_embedding( |
| 621 | self, |
| 622 | data: str, |
| 623 | ) -> torch.Tensor: |
| 624 | """Load audio embedding from a URL.""" |
| 625 | audio_embedding_io = AudioEmbeddingMediaIO() |
| 626 | |
| 627 | return audio_embedding_io.load_base64("", data) |
| 628 | |
| 629 | async def fetch_audio_embedding_async( |
| 630 | self, |
| 631 | data: str, |
| 632 | ) -> torch.Tensor: |
| 633 | """Asynchronously load audio embedding from a URL.""" |
| 634 | audio_embedding_io = AudioEmbeddingMediaIO() |
| 635 | loop = asyncio.get_running_loop() |
| 636 | |
| 637 | return await loop.run_in_executor( |
| 638 | global_thread_pool, audio_embedding_io.load_base64, "", data |
| 639 | ) |
| 640 | |
| 641 | def fetch_video_embedding( |
| 642 | self, |
| 643 | data: str, |
| 644 | ) -> torch.Tensor: |
| 645 | """Load video embedding from a URL.""" |
| 646 | video_embedding_io = VideoEmbeddingMediaIO() |
| 647 | |
| 648 | return video_embedding_io.load_base64("", data) |
| 649 | |
| 650 | async def fetch_video_embedding_async( |
| 651 | self, |
| 652 | data: str, |
| 653 | ) -> torch.Tensor: |
| 654 | """Asynchronously load video embedding from a URL.""" |
| 655 | video_embedding_io = VideoEmbeddingMediaIO() |
| 656 | loop = asyncio.get_running_loop() |
| 657 | |
| 658 | return await loop.run_in_executor( |
| 659 | global_thread_pool, video_embedding_io.load_base64, "", data |
| 660 | ) |