Skip to content

vllm_mlx.models.mllm

MLX Multimodal Language Model (MLLM) wrapper.

View the complete module source at #L1-L2944.

API details

Each callable below includes its exact signature, type annotations, inputs, defaults, return contract, documented exceptions, implementation source, and parsed docstring sections when the source provides them.

vllm_mlx.models.mllm

MLX Multimodal Language Model (MLLM) wrapper.

This module provides a wrapper around mlx-vlm for multimodal inference, supporting vision, audio, and video understanding on Apple Silicon.

Features: - OpenAI-compatible API format for images and video - Smart video frame extraction with configurable FPS - Base64 and URL image support - Streaming generation - MLLM KV cache for repeated image/video+prompt combinations

vllm_mlx.models.mllm.logger module-attribute 

logger = logging.getLogger(__name__)

vllm_mlx.models.mllm._temp_manager module-attribute 

_temp_manager = TempFileManager()

vllm_mlx.models.mllm.FRAME_FACTOR module-attribute 

FRAME_FACTOR = 2

vllm_mlx.models.mllm.DEFAULT_FPS module-attribute 

DEFAULT_FPS = 2.0

vllm_mlx.models.mllm.MIN_FRAMES module-attribute 

MIN_FRAMES = 4

vllm_mlx.models.mllm.MAX_FRAMES module-attribute 

MAX_FRAMES = 128

vllm_mlx.models.mllm.IMAGE_FACTOR module-attribute 

IMAGE_FACTOR = 28

vllm_mlx.models.mllm.MAX_IMAGE_SIZE module-attribute 

MAX_IMAGE_SIZE = 20 * 1024 * 1024

vllm_mlx.models.mllm.MAX_VIDEO_SIZE module-attribute 

MAX_VIDEO_SIZE = 500 * 1024 * 1024

vllm_mlx.models.mllm.MAX_AUDIO_SIZE module-attribute 

MAX_AUDIO_SIZE = 100 * 1024 * 1024

vllm_mlx.models.mllm.MAX_BASE64_IMAGE_LENGTH module-attribute 

MAX_BASE64_IMAGE_LENGTH = 30 * 1024 * 1024

vllm_mlx.models.mllm.MAX_BASE64_VIDEO_LENGTH module-attribute 

MAX_BASE64_VIDEO_LENGTH = 700 * 1024 * 1024

vllm_mlx.models.mllm.MAX_BASE64_AUDIO_LENGTH module-attribute 

MAX_BASE64_AUDIO_LENGTH = 140 * 1024 * 1024

vllm_mlx.models.mllm._DRAFT_KWARG_NAMES module-attribute 

_DRAFT_KWARG_NAMES = ('draft_model', 'draft_kind', 'draft_block_size')

vllm_mlx.models.mllm._VIDEO_EXT_MAP module-attribute 

_VIDEO_EXT_MAP: dict[str, str] = {'mp4': '.mp4', 'webm': '.webm', 'avi': '.avi', 'mov': '.mov', 'quicktime': '.mov', 'mkv': '.mkv'}

vllm_mlx.models.mllm._AUDIO_EXT_MAP module-attribute 

_AUDIO_EXT_MAP: dict[str, str] = {'wav': '.wav', 'mpeg': '.mp3', 'mp3': '.mp3', 'flac': '.flac', 'ogg': '.ogg', 'webm': '.webm', 'mp4': '.m4a', 'm4a': '.m4a', 'aac': '.m4a'}

vllm_mlx.models.mllm._base64_image_cache module-attribute 

_base64_image_cache: dict[str, str] = {}

vllm_mlx.models.mllm.MLXVisionLanguageModel module-attribute 

MLXVisionLanguageModel = MLXMultimodalLM

vllm_mlx.models.mllm.VLMOutput module-attribute 

VLMOutput = MLLMOutput

vllm_mlx.models.mllm.is_vlm_model module-attribute 

is_vlm_model = MLXMultimodalLM.is_mllm_model

vllm_mlx.models.mllm.TempFileManager 

TempFileManager()

Thread-safe manager for tracking and cleaning up temporary files.

Source code in vllm_mlx/models/mllm.py 
44
45
46
47
def __init__(self):
    self._files: set[str] = set()
    self._lock = threading.Lock()
    atexit.register(self.cleanup_all)

vllm_mlx.models.mllm.TempFileManager._files instance-attribute 

_files: set[str] = set()

vllm_mlx.models.mllm.TempFileManager._lock instance-attribute 

_lock = threading.Lock()

vllm_mlx.models.mllm.TempFileManager.register 

register(path: str) -> str

Register a temp file for tracking. Returns the path for convenience.

Source code in vllm_mlx/models/mllm.py 
49
50
51
52
53
def register(self, path: str) -> str:
    """Register a temp file for tracking. Returns the path for convenience."""
    with self._lock:
        self._files.add(path)
    return path

vllm_mlx.models.mllm.TempFileManager.cleanup 

cleanup(path: str) -> bool

Clean up a specific temp file. Returns True if successful.

Source code in vllm_mlx/models/mllm.py 
55
56
57
58
59
60
61
62
63
64
65
66
67
def cleanup(self, path: str) -> bool:
    """Clean up a specific temp file. Returns True if successful."""
    with self._lock:
        if path in self._files:
            self._files.discard(path)
    try:
        if os.path.exists(path):
            os.unlink(path)
            logger.debug(f"Cleaned up temp file: {path}")
            return True
    except OSError as e:
        logger.warning(f"Failed to clean up temp file {path}: {e}")
    return False

vllm_mlx.models.mllm.TempFileManager.cleanup_all 

cleanup_all() -> int

Clean up all tracked temp files. Returns count of cleaned files.

Source code in vllm_mlx/models/mllm.py 
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
def cleanup_all(self) -> int:
    """Clean up all tracked temp files. Returns count of cleaned files."""
    with self._lock:
        files_to_clean = list(self._files)
        self._files.clear()

    cleaned = 0
    for path in files_to_clean:
        try:
            if os.path.exists(path):
                os.unlink(path)
                cleaned += 1
        except OSError:
            pass

    if cleaned:
        logger.info(f"Cleaned up {cleaned} temp files")
    return cleaned

vllm_mlx.models.mllm.FileSizeExceededError

Bases: Exception

Raised when a downloaded file exceeds the size limit.

vllm_mlx.models.mllm.UnsafeRemoteURLError 

UnsafeRemoteURLError(message: str, *, public_message: str = 'Remote media URL is not allowed')

Bases: ValueError

Raised when a remote media URL targets an unsafe destination.

Source code in vllm_mlx/models/mllm.py 
128
129
130
131
132
133
134
135
def __init__(
    self,
    message: str,
    *,
    public_message: str = "Remote media URL is not allowed",
) -> None:
    super().__init__(message)
    self.public_message = public_message

vllm_mlx.models.mllm.UnsafeRemoteURLError.public_message instance-attribute 

public_message = public_message

vllm_mlx.models.mllm.MultimodalInput dataclass 

MultimodalInput(prompt: str, images: list[str] = list(), videos: list[str] = list(), audio: list[str] = list())

Input for multimodal generation.

vllm_mlx.models.mllm.MultimodalInput.prompt instance-attribute 

prompt: str

vllm_mlx.models.mllm.MultimodalInput.images class-attribute instance-attribute 

images: list[str] = field(default_factory=list)

vllm_mlx.models.mllm.MultimodalInput.videos class-attribute instance-attribute 

videos: list[str] = field(default_factory=list)

vllm_mlx.models.mllm.MultimodalInput.audio class-attribute instance-attribute 

audio: list[str] = field(default_factory=list)

vllm_mlx.models.mllm.MLLMOutput dataclass 

MLLMOutput(text: str, finish_reason: str | None = None, prompt_tokens: int = 0, completion_tokens: int = 0, mtp_drafts: int = 0, mtp_accepted: int = 0)

Output from multimodal language model.

vllm_mlx.models.mllm.MLLMOutput.text instance-attribute 

text: str

vllm_mlx.models.mllm.MLLMOutput.finish_reason class-attribute instance-attribute 

finish_reason: str | None = None

vllm_mlx.models.mllm.MLLMOutput.prompt_tokens class-attribute instance-attribute 

prompt_tokens: int = 0

vllm_mlx.models.mllm.MLLMOutput.completion_tokens class-attribute instance-attribute 

completion_tokens: int = 0

vllm_mlx.models.mllm.MLLMOutput.mtp_drafts class-attribute instance-attribute 

mtp_drafts: int = 0

vllm_mlx.models.mllm.MLLMOutput.mtp_accepted class-attribute instance-attribute 

mtp_accepted: int = 0

vllm_mlx.models.mllm.MLXMultimodalLM 

MLXMultimodalLM(model_name: str, trust_remote_code: bool = False, enable_cache: bool = True, cache_size: int = 50, max_kv_size: int = 0, draft_model: str | None = None, draft_kind: str | None = None, draft_block_size: int | None = None)

Wrapper around mlx-vlm for multimodal inference.

This class provides a unified interface for multimodal language models using Apple's MLX framework. Supports: - Image understanding (single and multi-image) - Video understanding (smart frame extraction) - Audio understanding (for supported models) - OpenAI-compatible API format

Supported models include: - Qwen2-VL / Qwen2.5-VL / Qwen3-VL - LLaVA - Idefics3 - PaliGemma - And more via mlx-vlm

Example

model = MLXMultimodalLM("mlx-community/Qwen2-VL-2B-Instruct-4bit") model.load() output = model.generate( ... prompt="What's in this image?", ... images=["photo.jpg"] ... ) print(output.text)

Initialize the MLX multimodal language model.

Parameters:

  • model_name (str) –

    HuggingFace model name or local path

  • trust_remote_code (bool, default: False ) –

    Whether to trust remote code

  • enable_cache (bool, default: True ) –

    Enable KV cache for repeated image/video+prompt (default: True)

  • cache_size (int, default: 50 ) –

    Maximum cache entries (default: 50)

  • max_kv_size (int, default: 0 ) –

    Maximum KV cache size per sequence (0 = unbounded)

  • draft_model (str | None, default: None ) –

    Optional MLLM speculative draft/assistant model path.

  • draft_kind (str | None, default: None ) –

    Optional mlx-vlm draft kind, for example "mtp".

  • draft_block_size (int | None, default: None ) –

    Optional speculative block size passed to mlx-vlm.

Source code in vllm_mlx/models/mllm.py 
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
def __init__(
    self,
    model_name: str,
    trust_remote_code: bool = False,
    enable_cache: bool = True,
    cache_size: int = 50,
    max_kv_size: int = 0,
    draft_model: str | None = None,
    draft_kind: str | None = None,
    draft_block_size: int | None = None,
):
    """
    Initialize the MLX multimodal language model.

    Args:
        model_name: HuggingFace model name or local path
        trust_remote_code: Whether to trust remote code
        enable_cache: Enable KV cache for repeated image/video+prompt (default: True)
        cache_size: Maximum cache entries (default: 50)
        max_kv_size: Maximum KV cache size per sequence (0 = unbounded)
        draft_model: Optional MLLM speculative draft/assistant model path.
        draft_kind: Optional mlx-vlm draft kind, for example "mtp".
        draft_block_size: Optional speculative block size passed to mlx-vlm.
    """
    self.model_name = model_name
    self.trust_remote_code = trust_remote_code
    self.enable_cache = enable_cache
    self.max_kv_size = max_kv_size
    self.draft_model_path = draft_model
    self.draft_kind = draft_kind
    self.draft_block_size = draft_block_size

    self.model = None
    self.processor = None
    self.config = None
    self._draft_model = None
    self._loaded = False
    self._video_native = False
    self._video_native_with_audio = False

    # Initialize MLLM prefix cache manager (with vision embedding caching)
    self._cache_manager: MLLMPrefixCacheManager | None = None
    if enable_cache:
        self._cache_manager = MLLMPrefixCacheManager(max_entries=cache_size)

vllm_mlx.models.mllm.MLXMultimodalLM.model_name instance-attribute 

model_name = model_name

vllm_mlx.models.mllm.MLXMultimodalLM.trust_remote_code instance-attribute 

trust_remote_code = trust_remote_code

vllm_mlx.models.mllm.MLXMultimodalLM.enable_cache instance-attribute 

enable_cache = enable_cache

vllm_mlx.models.mllm.MLXMultimodalLM.max_kv_size instance-attribute 

max_kv_size = max_kv_size

vllm_mlx.models.mllm.MLXMultimodalLM.draft_model_path instance-attribute 

draft_model_path = draft_model

vllm_mlx.models.mllm.MLXMultimodalLM.draft_kind instance-attribute 

draft_kind = draft_kind

vllm_mlx.models.mllm.MLXMultimodalLM.draft_block_size instance-attribute 

draft_block_size = draft_block_size

vllm_mlx.models.mllm.MLXMultimodalLM.model instance-attribute 

model = None

vllm_mlx.models.mllm.MLXMultimodalLM.processor instance-attribute 

processor = None

vllm_mlx.models.mllm.MLXMultimodalLM.config instance-attribute 

config = None

vllm_mlx.models.mllm.MLXMultimodalLM._draft_model instance-attribute 

_draft_model = None

vllm_mlx.models.mllm.MLXMultimodalLM._loaded instance-attribute 

_loaded = False

vllm_mlx.models.mllm.MLXMultimodalLM._video_native instance-attribute 

_video_native = False

vllm_mlx.models.mllm.MLXMultimodalLM._video_native_with_audio instance-attribute 

_video_native_with_audio = False

vllm_mlx.models.mllm.MLXMultimodalLM._cache_manager instance-attribute 

_cache_manager: MLLMPrefixCacheManager | None = None

vllm_mlx.models.mllm.MLXMultimodalLM.load 

load() -> None

Load the model and processor.

Source code in vllm_mlx/models/mllm.py 
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
def load(self) -> None:
    """Load the model and processor."""
    if self._loaded:
        return

    try:
        from mlx_vlm import load
        from mlx_vlm.utils import load_config

        logger.info(f"Loading MLLM: {self.model_name}")

        self.model, self.processor = load(self.model_name)
        self.config = load_config(self.model_name)
        if self.draft_model_path:
            self._draft_model = self._load_draft_model()
            _install_draft_metrics_hooks(self._draft_model)

        self._loaded = True
        self._video_native = hasattr(
            self.model.config, "video_token_id"
        ) or hasattr(self.model.config, "video_token_index")
        # Omni models expose a sound_encoder; for these, a video_url
        # without a paired audio_url should auto-extract the video's
        # audio track so the model can fuse A/V in one forward pass.
        # Decoupled from _video_native because some omni models (e.g.
        # Nemotron-H Omni) don't expose video_token_id at config level
        # and run through the frames-as-images fallback path.
        self._video_native_with_audio = _model_has_sound_encoder(self.model)
        logger.info(f"MLLM loaded successfully: {self.model_name}")
        if self._video_native:
            logger.info("Native video pipeline enabled (temporal 3D conv + M-RoPE)")
        if self._video_native_with_audio:
            logger.info(
                "Omni model detected: video_url will auto-extract audio for A/V fusion"
            )

    except ImportError:
        raise ImportError(
            "mlx-vlm is required for multimodal inference. "
            "Install with: pip install mlx-vlm"
        )
    except Exception as e:
        logger.error(f"Failed to load MLLM: {e}")
        raise

vllm_mlx.models.mllm.MLXMultimodalLM._load_draft_model 

_load_draft_model()
Source code in vllm_mlx/models/mllm.py 
1325
1326
1327
1328
1329
1330
1331
1332
def _load_draft_model(self):
    if self.draft_kind == "mtp":
        return load_gemma4_assistant_drafter(self.draft_model_path)

    from mlx_vlm.utils import load

    draft_model, _ = load(self.draft_model_path)
    return draft_model

vllm_mlx.models.mllm.MLXMultimodalLM._draft_generation_kwargs 

_draft_generation_kwargs(call_kwargs: dict | None = None) -> dict

Return mlx-vlm drafter kwargs when the request explicitly opts in.

call_kwargs is the outbound mlx-vlm kwargs dict. This method removes vllm-mlx drafter control keys before the dict is forwarded so caller passthrough values cannot conflict with the configured server drafter.

Source code in vllm_mlx/models/mllm.py 
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
def _draft_generation_kwargs(self, call_kwargs: dict | None = None) -> dict:
    """Return mlx-vlm drafter kwargs when the request explicitly opts in.

    ``call_kwargs`` is the outbound mlx-vlm kwargs dict. This method removes
    vllm-mlx drafter control keys before the dict is forwarded so caller
    passthrough values cannot conflict with the configured server drafter.
    """
    draft_requested = False
    if call_kwargs is not None:
        draft_requested = bool(call_kwargs.pop("mllm_draft", False))
        for key in _DRAFT_KWARG_NAMES:
            call_kwargs.pop(key, None)
    if not draft_requested or self._draft_model is None:
        return {}
    # Tests may install the draft model after load(); the hook is idempotent.
    _install_draft_metrics_hooks(self._draft_model)
    kwargs = {"draft_model": self._draft_model}
    if self.draft_kind:
        kwargs["draft_kind"] = self.draft_kind
    if self.draft_block_size is not None:
        kwargs["draft_block_size"] = self.draft_block_size
    return kwargs

vllm_mlx.models.mllm.MLXMultimodalLM._reset_draft_metrics 

_reset_draft_metrics() -> int
Source code in vllm_mlx/models/mllm.py 
1357
1358
1359
1360
1361
1362
1363
1364
def _reset_draft_metrics(self) -> int:
    if self._draft_model is None:
        return 0
    if hasattr(self._draft_model, "accept_lens"):
        self._draft_model.accept_lens = []
    if hasattr(self._draft_model, "_vllm_mlx_draft_counts"):
        self._draft_model._vllm_mlx_draft_counts = []
    return 0

vllm_mlx.models.mllm.MLXMultimodalLM._draft_metrics_since 

_draft_metrics_since(start_accept_lens: int) -> dict[str, int]
Source code in vllm_mlx/models/mllm.py 
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
def _draft_metrics_since(self, start_accept_lens: int) -> dict[str, int]:
    if self._draft_model is None:
        return {"mtp_drafts": 0, "mtp_accepted": 0}
    accept_lens = list(getattr(self._draft_model, "accept_lens", []))
    if start_accept_lens > len(accept_lens):
        new_accept_lens = accept_lens
    else:
        new_accept_lens = accept_lens[start_accept_lens:]
    draft_counts = list(getattr(self._draft_model, "_vllm_mlx_draft_counts", []))
    if start_accept_lens > len(draft_counts):
        new_draft_counts = draft_counts
    else:
        new_draft_counts = draft_counts[start_accept_lens:]
    block_size = (
        int(self.draft_block_size)
        if self.draft_block_size is not None
        else int(
            getattr(getattr(self._draft_model, "config", None), "block_size", 0)
        )
    )
    drafted_per_round = max(block_size - 1, 0)
    mtp_drafts = (
        sum(max(int(value), 0) for value in new_draft_counts)
        if new_draft_counts
        else drafted_per_round * len(new_accept_lens)
    )
    return {
        "mtp_drafts": mtp_drafts,
        "mtp_accepted": sum(int(value) for value in new_accept_lens),
    }

vllm_mlx.models.mllm.MLXMultimodalLM.get_language_model 

get_language_model()

Extract the underlying language model for mlx_lm TextModel construction.

Source code in vllm_mlx/models/mllm.py 
1397
1398
1399
def get_language_model(self):
    """Extract the underlying language model for mlx_lm TextModel construction."""
    return self.model.language_model

vllm_mlx.models.mllm.MLXMultimodalLM.get_tokenizer 

get_tokenizer()

Get the text tokenizer (not the multimodal processor).

Source code in vllm_mlx/models/mllm.py 
1401
1402
1403
def get_tokenizer(self):
    """Get the text tokenizer (not the multimodal processor)."""
    return self.processor.tokenizer

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_images 

_prepare_images(images: list) -> list[str]

Process remote/base64 image inputs into local temp file paths.

Source code in vllm_mlx/models/mllm.py 
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
def _prepare_images(self, images: list) -> list[str]:
    """Process remote/base64 image inputs into local temp file paths."""
    processed = []
    for img in images:
        try:
            path = process_image_input(img)
            processed.append(path)
        except Exception as e:
            logger.warning(f"Failed to process image: {e}")
    return processed

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_audio 

_prepare_audio(audio_inputs: list) -> list[str]

Process audio inputs and return local file paths.

Source code in vllm_mlx/models/mllm.py 
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
def _prepare_audio(self, audio_inputs: list) -> list[str]:
    """Process audio inputs and return local file paths."""
    processed = []
    for audio_input in audio_inputs:
        try:
            path = process_audio_input(audio_input)
            processed.append(path)
        except Exception as e:
            logger.warning(f"Failed to process audio: {e}")
    return processed

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_video 

_prepare_video(video_input: str | dict, fps: float = DEFAULT_FPS, max_frames: int = MAX_FRAMES, resolved_path: str | None = None) -> list[str]

Process video input and extract frames.

Supports: - URLs (http/https) - will be downloaded - Base64 encoded videos (data:video/mp4;base64,...) - OpenAI format dicts: {"url": "..."} or {"video_url": {"url": "..."}}

Parameters:

  • video_input (str | dict) –

    Video in any supported format

  • fps (float, default: DEFAULT_FPS ) –

    Frames per second to extract

  • max_frames (int, default: MAX_FRAMES ) –

    Maximum frames to extract

  • resolved_path (str | None, default: None ) –

    Optional pre-resolved local path. Callers that already ran process_video_input (e.g. for parallel audio extraction) pass it here to avoid re-downloading / re-decoding.

Returns:

  • list[str] –

    List of paths to extracted frame images

Source code in vllm_mlx/models/mllm.py 
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
def _prepare_video(
    self,
    video_input: str | dict,
    fps: float = DEFAULT_FPS,
    max_frames: int = MAX_FRAMES,
    resolved_path: str | None = None,
) -> list[str]:
    """
    Process video input and extract frames.

    Supports:
    - URLs (http/https) - will be downloaded
    - Base64 encoded videos (data:video/mp4;base64,...)
    - OpenAI format dicts: {"url": "..."} or {"video_url": {"url": "..."}}

    Args:
        video_input: Video in any supported format
        fps: Frames per second to extract
        max_frames: Maximum frames to extract
        resolved_path: Optional pre-resolved local path. Callers that
            already ran process_video_input (e.g. for parallel audio
            extraction) pass it here to avoid re-downloading / re-decoding.

    Returns:
        List of paths to extracted frame images
    """
    # Reuse caller's resolved path when supplied; otherwise resolve here
    # (downloads if URL, decodes if base64).
    video_path = resolved_path or process_video_input(video_input)

    # Extract frames
    frames = extract_video_frames_smart(
        video_path,
        fps=fps,
        max_frames=max_frames,
    )
    return save_frames_to_temp(frames)

vllm_mlx.models.mllm.MLXMultimodalLM._collect_video_inputs 

_collect_video_inputs(messages: list[dict]) -> dict[int, list]

Collect video inputs from messages, keyed by message index.

Handles both 'video' and 'video_url' content types, including Pydantic model conversion.

Source code in vllm_mlx/models/mllm.py 
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
def _collect_video_inputs(self, messages: list[dict]) -> dict[int, list]:
    """Collect video inputs from messages, keyed by message index.

    Handles both 'video' and 'video_url' content types, including
    Pydantic model conversion.
    """
    video_inputs: dict[int, list] = {}
    for msg_idx, msg in enumerate(messages):
        content = msg.get("content", "")
        if not isinstance(content, list):
            continue
        for item in content:
            if hasattr(item, "model_dump"):
                item = item.model_dump(exclude_none=True)
            elif hasattr(item, "dict"):
                item = {k: v for k, v in item.dict().items() if v is not None}

            if not isinstance(item, dict):
                continue
            item_type = item.get("type", "")
            if item_type == "video":
                video_inputs.setdefault(msg_idx, []).append(
                    item.get("video", item.get("url", ""))
                )
            elif item_type == "video_url":
                vid_url = item.get("video_url", {})
                if isinstance(vid_url, str):
                    video_inputs.setdefault(msg_idx, []).append(vid_url)
                elif isinstance(vid_url, dict):
                    url = vid_url.get("url", "")
                    if url:
                        video_inputs.setdefault(msg_idx, []).append(url)
    return video_inputs

vllm_mlx.models.mllm.MLXMultimodalLM._collect_audio_inputs 

_collect_audio_inputs(messages: list[dict]) -> dict[int, list]

Collect audio inputs from messages, keyed by message index.

Source code in vllm_mlx/models/mllm.py 
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
def _collect_audio_inputs(self, messages: list[dict]) -> dict[int, list]:
    """Collect audio inputs from messages, keyed by message index."""
    audio_inputs: dict[int, list] = {}
    for msg_idx, msg in enumerate(messages):
        content = msg.get("content", "")
        if not isinstance(content, list):
            continue
        for item in content:
            if hasattr(item, "model_dump"):
                item = item.model_dump(exclude_none=True)
            elif hasattr(item, "dict"):
                item = {k: v for k, v in item.dict().items() if v is not None}

            if not isinstance(item, dict):
                continue

            item_type = item.get("type", "")
            if item_type == "audio":
                audio_value = item.get("audio", item.get("url", ""))
                if audio_value:
                    audio_inputs.setdefault(msg_idx, []).append(audio_value)
            elif item_type == "audio_url":
                audio_url = item.get("audio_url", {})
                if isinstance(audio_url, str):
                    audio_inputs.setdefault(msg_idx, []).append(audio_url)
                elif isinstance(audio_url, dict):
                    url = audio_url.get("url", "")
                    if url:
                        audio_inputs.setdefault(msg_idx, []).append(url)
    return audio_inputs

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_native_video_inputs 

_prepare_native_video_inputs(messages: list[dict], video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, tools: list | None = None) -> tuple[str, dict]

Preprocess messages into prompt + generation kwargs for native video.

Mirrors the preprocessing in mlx_vlm.video_generate.main() so that upstream improvements are easy to adopt. Returns the formatted prompt text and a dict of kwargs ready to pass to video_generate.generate().

Currently Qwen-family-specific (video_token_id / video_token_index).

Source code in vllm_mlx/models/mllm.py 
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
def _prepare_native_video_inputs(
    self,
    messages: list[dict],
    video_fps: float = DEFAULT_FPS,
    video_max_frames: int = MAX_FRAMES,
    tools: list | None = None,
) -> tuple[str, dict]:
    """Preprocess messages into prompt + generation kwargs for native video.

    Mirrors the preprocessing in mlx_vlm.video_generate.main() so that
    upstream improvements are easy to adopt. Returns the formatted prompt
    text and a dict of kwargs ready to pass to video_generate.generate().

    Currently Qwen-family-specific (video_token_id / video_token_index).
    """
    import mlx.core as mx

    try:
        from mlx_vlm.video_generate import process_vision_info
    except ImportError:
        raise ImportError(
            "mlx_vlm.video_generate is required for native video support. "
            "Upgrade with: pip install --upgrade mlx-vlm"
        )

    # Translate OpenAI API messages into process_vision_info format
    native_messages = self._translate_messages_for_native_video(
        messages, video_fps, video_max_frames
    )

    # Use HF processor's chat template (handles timestamp interleaving)
    template_kwargs: dict = {}
    if tools:
        template_kwargs["tools"] = tools

    text = self.processor.apply_chat_template(
        native_messages,
        tokenize=False,
        add_generation_prompt=True,
        **template_kwargs,
    )

    # Extract vision inputs via mlx-vlm's process_vision_info
    image_inputs, video_inputs, fps_info = process_vision_info(
        native_messages, return_video_kwargs=True
    )

    # Collect audio paths emitted by the translation step
    # (explicit audio_url, or auto-extracted from video_url for omni
    # models).
    audio_inputs: list[str] = []
    for nmsg in native_messages:
        ncontent = nmsg.get("content", [])
        if not isinstance(ncontent, list):
            continue
        for nitem in ncontent:
            if isinstance(nitem, dict) and nitem.get("type") == "audio":
                apath = nitem.get("audio")
                if apath:
                    audio_inputs.append(apath)

    # Process through HF processor to get input_ids, pixel_values, grid_thw
    # and (for omni models) sound_clips / input_features.
    processor_kwargs: dict = {
        "text": [text],
        "images": image_inputs,
        "videos": video_inputs,
        "padding": True,
        "return_tensors": "pt",
    }
    if audio_inputs:
        processor_kwargs["audio"] = audio_inputs
    inputs = self.processor(**processor_kwargs)

    input_ids = mx.array(inputs["input_ids"])
    pixel_values = inputs.get(
        "pixel_values_videos", inputs.get("pixel_values", None)
    )
    if pixel_values is not None:
        pixel_values = mx.array(pixel_values)
    mask = mx.array(inputs["attention_mask"])

    gen_kwargs: dict = {}
    if inputs.get("video_grid_thw", None) is not None:
        gen_kwargs["video_grid_thw"] = mx.array(inputs["video_grid_thw"])
    if inputs.get("image_grid_thw", None) is not None:
        gen_kwargs["image_grid_thw"] = mx.array(inputs["image_grid_thw"])

    # Forward audio embeddings/clips from the processor so the omni
    # model's sound encoder gets fed alongside the visual stream.
    for audio_key in (
        "sound_clips",
        "input_features",
        "feature_attention_mask",
        "audio_feature_lengths",
        "sound_feature_lengths",
        "sound_attention_mask",
    ):
        val = inputs.get(audio_key, None)
        if val is not None:
            gen_kwargs[audio_key] = val
    if audio_inputs:
        logger.info(
            f"Native video: forwarding audio ({len(audio_inputs)} clip(s)) "
            f"to omni model via "
            f"{[k for k in gen_kwargs if k in ('sound_clips', 'input_features')]}"
        )

    gen_kwargs["input_ids"] = input_ids
    gen_kwargs["pixel_values"] = pixel_values
    gen_kwargs["mask"] = mask

    grid_thw_info = gen_kwargs.get("video_grid_thw")
    logger.info(
        f"Native video: {input_ids.size} input tokens, "
        f"video_grid_thw={grid_thw_info.tolist() if grid_thw_info is not None else None}"
    )

    return text, gen_kwargs

vllm_mlx.models.mllm.MLXMultimodalLM._generate_native_video 

_generate_native_video(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, tools: list | None = None, **kwargs) -> MLLMOutput

Generate using native video pipeline (Qwen-family models).

Delegates preprocessing to _prepare_native_video_inputs and generation to mlx_vlm.video_generate.generate(), keeping our code aligned with upstream's video pipeline so improvements are easy to adopt.

Source code in vllm_mlx/models/mllm.py 
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
def _generate_native_video(
    self,
    messages: list[dict],
    max_tokens: int = 256,
    temperature: float = 0.7,
    video_fps: float = DEFAULT_FPS,
    video_max_frames: int = MAX_FRAMES,
    tools: list | None = None,
    **kwargs,
) -> MLLMOutput:
    """Generate using native video pipeline (Qwen-family models).

    Delegates preprocessing to _prepare_native_video_inputs and generation
    to mlx_vlm.video_generate.generate(), keeping our code aligned with
    upstream's video pipeline so improvements are easy to adopt.
    """
    try:
        from mlx_vlm.video_generate import generate
    except ImportError:
        raise ImportError(
            "mlx_vlm.video_generate is required for native video support. "
            "Upgrade with: pip install --upgrade mlx-vlm"
        )

    text, gen_kwargs = self._prepare_native_video_inputs(
        messages, video_fps, video_max_frames, tools
    )
    gen_kwargs["temperature"] = temperature

    result = generate(
        self.model,
        self.processor,
        prompt=text,
        max_tokens=max_tokens,
        verbose=False,
        **gen_kwargs,
    )

    if hasattr(result, "text"):
        return MLLMOutput(
            text=result.text,
            finish_reason="stop",
            prompt_tokens=getattr(result, "prompt_tokens", 0),
            completion_tokens=getattr(result, "generation_tokens", 0),
        )
    return MLLMOutput(text=str(result), finish_reason="stop")

vllm_mlx.models.mllm.MLXMultimodalLM._translate_messages_for_native_video 

_translate_messages_for_native_video(messages: list[dict], video_fps: float, video_max_frames: int) -> list[dict]

Translate OpenAI API format messages to process_vision_info format.

Converts video_url/video types and resolves remote/base64 inputs to local paths. Images are preserved as-is (process_vision_info handles them).

Source code in vllm_mlx/models/mllm.py 
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
def _translate_messages_for_native_video(
    self,
    messages: list[dict],
    video_fps: float,
    video_max_frames: int,
) -> list[dict]:
    """Translate OpenAI API format messages to process_vision_info format.

    Converts video_url/video types and resolves remote/base64 inputs to local paths.
    Images are preserved as-is (process_vision_info handles them).
    """
    translated = []
    for msg in messages:
        role = msg.get("role", "user")
        content = msg.get("content", "")

        if isinstance(content, str):
            translated.append({"role": role, "content": content})
            continue

        if not isinstance(content, list):
            translated.append({"role": role, "content": str(content)})
            continue

        # Pre-pass: does this message have an explicit audio_url/audio
        # block? If so, we skip auto-extracting audio from a video_url to
        # honor the caller's explicit choice.
        has_explicit_audio = False
        for item in content:
            if hasattr(item, "model_dump"):
                probe = item.model_dump(exclude_none=True)
            elif hasattr(item, "dict"):
                probe = {k: v for k, v in item.dict().items() if v is not None}
            else:
                probe = item
            if isinstance(probe, dict) and probe.get("type", "") in (
                "audio",
                "audio_url",
            ):
                has_explicit_audio = True
                break

        new_content = []
        for item in content:
            if hasattr(item, "model_dump"):
                item = item.model_dump(exclude_none=True)
            elif hasattr(item, "dict"):
                item = {k: v for k, v in item.dict().items() if v is not None}

            if not isinstance(item, dict):
                new_content.append({"type": "text", "text": str(item)})
                continue

            item_type = item.get("type", "")

            if item_type == "text":
                new_content.append(item)

            elif item_type == "image_url":
                img_url = item.get("image_url", {})
                url = (
                    img_url.get("url", img_url)
                    if isinstance(img_url, dict)
                    else img_url
                )
                # Resolve to local path for process_vision_info
                local_path = process_image_input(url)
                new_content.append({"type": "image", "image": local_path})

            elif item_type == "image":
                img = item.get("image", item.get("url", ""))
                local_path = process_image_input(img)
                new_content.append({"type": "image", "image": local_path})

            elif item_type in ("video", "video_url"):
                # Extract video path/URL from various formats
                if item_type == "video_url":
                    vid_url = item.get("video_url", {})
                    if isinstance(vid_url, str):
                        video_source = vid_url
                    elif isinstance(vid_url, dict):
                        video_source = vid_url.get("url", "")
                    else:
                        continue
                else:
                    video_source = item.get("video", item.get("url", ""))

                if not video_source:
                    continue

                # Resolve to local path
                video_path = process_video_input(video_source)
                new_content.append(
                    {
                        "type": "video",
                        "video": video_path,
                        "fps": video_fps,
                        "max_frames": video_max_frames,
                    }
                )
                # For omni-capable models, pull the video's audio track
                # alongside frames so the model can fuse A/V in one
                # forward pass. We extract from the already-resolved local
                # path (no raw user URL handed to ffmpeg → avoids URL-
                # protocol SSRF via ffmpeg's network demuxers).
                if not has_explicit_audio and getattr(
                    self, "_video_native_with_audio", False
                ):
                    extracted = extract_audio_from_video(video_path)
                    if extracted is not None:
                        new_content.append({"type": "audio", "audio": extracted})

            elif item_type in ("audio", "audio_url"):
                if item_type == "audio_url":
                    aud_url = item.get("audio_url", {})
                    if isinstance(aud_url, str):
                        audio_source = aud_url
                    elif isinstance(aud_url, dict):
                        audio_source = aud_url.get("url", "")
                    else:
                        continue
                else:
                    audio_source = item.get("audio", item.get("url", ""))

                if not audio_source:
                    continue

                audio_path = process_audio_input(audio_source)
                new_content.append({"type": "audio", "audio": audio_path})

            else:
                new_content.append(item)

        translated.append({"role": role, "content": new_content})

    return translated

vllm_mlx.models.mllm.MLXMultimodalLM.generate 

generate(prompt: str, images: list | None = None, videos: list | None = None, audio: list[str] | None = None, max_tokens: int = 256, temperature: float = 0.7, top_p: float = 0.9, video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, use_cache: bool = True, **kwargs) -> MLLMOutput

Generate text from multimodal input.

Parameters:

  • prompt (str) –

    Text prompt/question

  • images (list | None, default: None ) –

    List of image URLs or base64 strings

  • videos (list | None, default: None ) –

    List of video inputs (URLs, base64, or OpenAI format dicts)

  • audio (list[str] | None, default: None ) –

    List of audio file paths

  • max_tokens (int, default: 256 ) –

    Maximum tokens to generate

  • temperature (float, default: 0.7 ) –

    Sampling temperature

  • top_p (float, default: 0.9 ) –

    Top-p sampling parameter

  • video_fps (float, default: DEFAULT_FPS ) –

    FPS for video frame extraction (default: 2.0)

  • video_max_frames (int, default: MAX_FRAMES ) –

    Max frames to extract from video

  • use_cache (bool, default: True ) –

    Whether to use KV cache (default: True)

  • **kwargs –

    Additional generation parameters

Returns:

Example
With local video

output = model.generate("Describe this video", videos=["video.mp4"])

With video URL

output = model.generate("What happens?", videos=["https://example.com/video.mp4"])

With base64 video

output = model.generate("Describe", videos=["data:video/mp4;base64,AAAA..."])

Source code in vllm_mlx/models/mllm.py 
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
def generate(
    self,
    prompt: str,
    images: list | None = None,
    videos: list | None = None,
    audio: list[str] | None = None,
    max_tokens: int = 256,
    temperature: float = 0.7,
    top_p: float = 0.9,
    video_fps: float = DEFAULT_FPS,
    video_max_frames: int = MAX_FRAMES,
    use_cache: bool = True,
    **kwargs,
) -> MLLMOutput:
    """
    Generate text from multimodal input.

    Args:
        prompt: Text prompt/question
        images: List of image URLs or base64 strings
        videos: List of video inputs (URLs, base64, or OpenAI format dicts)
        audio: List of audio file paths
        max_tokens: Maximum tokens to generate
        temperature: Sampling temperature
        top_p: Top-p sampling parameter
        video_fps: FPS for video frame extraction (default: 2.0)
        video_max_frames: Max frames to extract from video
        use_cache: Whether to use KV cache (default: True)
        **kwargs: Additional generation parameters

    Returns:
        MLLMOutput with generated text

    Example:
        # With local video
        output = model.generate("Describe this video", videos=["video.mp4"])

        # With video URL
        output = model.generate("What happens?", videos=["https://example.com/video.mp4"])

        # With base64 video
        output = model.generate("Describe", videos=["data:video/mp4;base64,AAAA..."])
    """
    if not self._loaded:
        self.load()

    from mlx_vlm import generate
    from mlx_vlm.models import cache as vlm_cache
    from mlx_vlm.prompt_utils import apply_chat_template

    images = images or []
    videos = videos or []
    audio = audio or []

    # Process all images (including frames from videos) and audio inputs
    all_images = []
    all_audio = []
    all_sources = []  # Track original sources for cache key

    # Process image inputs
    if images:
        all_images.extend(self._prepare_images(images))
        all_sources.extend(images)

    # Extract frames from videos
    for video_path in videos:
        frames = self._prepare_video(
            video_path,
            fps=video_fps,
            max_frames=video_max_frames,
        )
        all_images.extend(frames)
        # Include video params in cache key
        video_str = video_path if isinstance(video_path, str) else str(video_path)
        all_sources.append(
            f"video:{video_str}:fps{video_fps}:max{video_max_frames}"
        )
        logger.info(f"Added {len(frames)} frames from video: {video_path}")

    if audio:
        all_audio.extend(self._prepare_audio(audio))

    # Apply chat template if needed
    if (all_images or all_audio) and hasattr(self.processor, "apply_chat_template"):
        try:
            formatted_prompt = apply_chat_template(
                self.processor,
                self.config,
                prompt,
                num_images=len(all_images),
                num_audios=len(all_audio),
            )
        except Exception:
            formatted_prompt = prompt
    else:
        formatted_prompt = prompt

    # Check cache for existing KV state
    prompt_cache = None
    cache_hit = False

    if use_cache and all_audio:
        logger.info("MLLM cache disabled for audio inputs")
        use_cache = False

    if use_cache and self._cache_manager is not None and all_sources:
        prompt_cache, cache_hit = self._cache_manager.fetch_cache(
            all_sources, formatted_prompt
        )
        if cache_hit:
            logger.info(f"MLLM cache hit for {len(all_sources)} source(s)")

    # Create new cache if needed
    if prompt_cache is None and self.model is not None:
        try:
            prompt_cache = vlm_cache.make_prompt_cache(
                self.model.language_model,
                max_kv_size=self.max_kv_size or None,
            )
        except Exception:
            prompt_cache = None

    # Generate with cache
    draft_accept_start = self._reset_draft_metrics()
    result = generate(
        self.model,
        self.processor,
        formatted_prompt,
        all_images if all_images else None,
        audio=all_audio if all_audio else None,
        max_tokens=max_tokens,
        temp=temperature,
        top_p=top_p,
        verbose=False,
        prompt_cache=prompt_cache,
        **self._draft_generation_kwargs(kwargs),
        **kwargs,
    )
    draft_metrics = self._draft_metrics_since(draft_accept_start)

    # Store cache for future reuse (only on miss)
    if use_cache and self._cache_manager and all_sources and not cache_hit:
        if prompt_cache is not None:
            try:
                num_tokens = getattr(result, "prompt_tokens", 0)
                self._cache_manager.store_cache(
                    all_sources, formatted_prompt, prompt_cache, num_tokens
                )
                logger.info(f"MLLM cache stored for {len(all_sources)} source(s)")
            except Exception as e:
                logger.debug(f"Failed to store MLLM cache: {e}")

    # Handle GenerationResult object or plain string
    if hasattr(result, "text"):
        output_text = result.text
        prompt_tokens = getattr(result, "prompt_tokens", 0)
        generation_tokens = getattr(result, "generation_tokens", 0)
    else:
        output_text = str(result)
        prompt_tokens = 0
        generation_tokens = 0

    return MLLMOutput(
        text=output_text,
        finish_reason="stop",
        prompt_tokens=prompt_tokens,
        completion_tokens=generation_tokens,
        **draft_metrics,
    )

vllm_mlx.models.mllm.MLXMultimodalLM.stream_generate 

stream_generate(prompt: str, images: list | None = None, videos: list[str] | None = None, audio: list[str] | None = None, max_tokens: int = 256, temperature: float = 0.7, video_fps: float = DEFAULT_FPS, **kwargs) -> Iterator[str]

Stream text generation for multimodal input.

Parameters:

  • prompt (str) –

    Text prompt

  • images (list | None, default: None ) –

    List of image inputs

  • videos (list[str] | None, default: None ) –

    List of video paths

  • audio (list[str] | None, default: None ) –

    List of audio inputs

  • max_tokens (int, default: 256 ) –

    Maximum tokens to generate

  • temperature (float, default: 0.7 ) –

    Sampling temperature

  • video_fps (float, default: DEFAULT_FPS ) –

    FPS for video frame extraction

  • **kwargs –

    Additional parameters

Yields:

  • str –

    Generated text chunks

Source code in vllm_mlx/models/mllm.py 
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
def stream_generate(
    self,
    prompt: str,
    images: list | None = None,
    videos: list[str] | None = None,
    audio: list[str] | None = None,
    max_tokens: int = 256,
    temperature: float = 0.7,
    video_fps: float = DEFAULT_FPS,
    **kwargs,
) -> Iterator[str]:
    """
    Stream text generation for multimodal input.

    Args:
        prompt: Text prompt
        images: List of image inputs
        videos: List of video paths
        audio: List of audio inputs
        max_tokens: Maximum tokens to generate
        temperature: Sampling temperature
        video_fps: FPS for video frame extraction
        **kwargs: Additional parameters

    Yields:
        Generated text chunks
    """
    if not self._loaded:
        self.load()

    try:
        from mlx_vlm import stream_generate
        from mlx_vlm.prompt_utils import apply_chat_template
    except ImportError:
        # Fallback to non-streaming
        output = self.generate(
            prompt=prompt,
            images=images,
            videos=videos,
            audio=audio,
            max_tokens=max_tokens,
            temperature=temperature,
            video_fps=video_fps,
            **kwargs,
        )
        yield output.text
        return

    images = images or []
    videos = videos or []
    audio = audio or []

    # Process images
    all_images = []
    all_audio = []
    if images:
        all_images.extend(self._prepare_images(images))
    for video_path in videos:
        frames = self._prepare_video(video_path, fps=video_fps)
        all_images.extend(frames)
    if audio:
        all_audio.extend(self._prepare_audio(audio))

    # Apply chat template
    if all_images or all_audio:
        try:
            formatted_prompt = apply_chat_template(
                self.processor,
                self.config,
                prompt,
                num_images=len(all_images),
                num_audios=len(all_audio),
            )
        except Exception:
            formatted_prompt = prompt
    else:
        formatted_prompt = prompt

    for chunk in stream_generate(
        self.model,
        self.processor,
        formatted_prompt,
        all_images if all_images else None,
        audio=all_audio if all_audio else None,
        max_tokens=max_tokens,
        temp=temperature,
        **self._draft_generation_kwargs(kwargs),
        **kwargs,
    ):
        yield chunk

vllm_mlx.models.mllm.MLXMultimodalLM.chat 

chat(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, **kwargs) -> MLLMOutput

Chat with OpenAI-compatible message format.

Supports multimodal content in messages: - {"type": "text", "text": "..."} - {"type": "image_url", "image_url": {"url": "..."}} - {"type": "image_url", "image_url": {"url": "data:image/...;base64,..."}}

Parameters:

  • messages (list[dict]) –

    List of chat messages (OpenAI format)

  • max_tokens (int, default: 256 ) –

    Maximum tokens to generate

  • temperature (float, default: 0.7 ) –

    Sampling temperature

  • **kwargs –

    Additional parameters

Returns:

  • MLLMOutput –

    MLLMOutput with assistant's response

Source code in vllm_mlx/models/mllm.py 
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
def chat(
    self,
    messages: list[dict],
    max_tokens: int = 256,
    temperature: float = 0.7,
    **kwargs,
) -> MLLMOutput:
    """
    Chat with OpenAI-compatible message format.

    Supports multimodal content in messages:
    - {"type": "text", "text": "..."}
    - {"type": "image_url", "image_url": {"url": "..."}}
    - {"type": "image_url", "image_url": {"url": "data:image/...;base64,..."}}

    Args:
        messages: List of chat messages (OpenAI format)
        max_tokens: Maximum tokens to generate
        temperature: Sampling temperature
        **kwargs: Additional parameters

    Returns:
        MLLMOutput with assistant's response
    """
    if not self._loaded:
        self.load()

    from mlx_vlm import generate
    from mlx_vlm.prompt_utils import get_chat_template

    # Extract text, images and audio from messages
    # Build chat_messages for multi-turn support WITH proper image/audio tokens per message
    all_image_urls = []  # Raw URLs/paths to process later
    chat_messages = []  # List of properly formatted messages for chat template

    logger.info(f"MLLM.chat() called with {len(messages)} messages")

    # Pop params early so they don't leak into mlx_vlm.generate()
    video_fps = kwargs.pop("video_fps", DEFAULT_FPS)
    video_max_frames = kwargs.pop("video_max_frames", MAX_FRAMES)
    tools = kwargs.pop("tools", None)
    use_cache = kwargs.pop("use_cache", True)
    enable_thinking = kwargs.pop("enable_thinking", True)
    # Honor chat_template_kwargs on the MLLM path (parity with the text path in
    # llm.py). enable_thinking is commonly nested here via
    # --default-chat-template-kwargs {"enable_thinking": false}; it was previously
    # dropped for MLLM models, so reasoning could not be disabled and leaked into
    # the response.
    chat_template_kwargs = kwargs.pop("chat_template_kwargs", None) or {}
    if "enable_thinking" in chat_template_kwargs:
        enable_thinking = chat_template_kwargs.pop("enable_thinking")

    # Collect video and audio inputs from messages
    _msg_video_inputs = self._collect_video_inputs(messages)
    _msg_audio_inputs = self._collect_audio_inputs(messages)

    # Use native video pipeline for supported models
    if self._video_native and _msg_video_inputs:
        return self._generate_native_video(
            messages=messages,
            max_tokens=max_tokens,
            temperature=temperature,
            video_fps=video_fps,
            video_max_frames=video_max_frames,
            tools=tools,
            **kwargs,
        )

    # Fallback: extract frames and treat as individual images
    _msg_video_frame_counts: dict[int, int] = {}
    _msg_extra_audio: dict[int, list[str]] = {}
    all_video_frames: list[str] = []
    all_audio_inputs: list[str] = []
    for msg_idx, vid_inputs in _msg_video_inputs.items():
        total_frames = 0
        has_explicit_audio = bool(_msg_audio_inputs.get(msg_idx))
        for vid_input in vid_inputs:
            # Resolve the video to a local path ONCE per input. Both
            # audio extraction (when this is an omni model with no
            # explicit audio block) and frame extraction need a local
            # file; resolving twice would re-download remote URLs and
            # re-decode base64. Resolving up front also keeps user-
            # supplied raw URLs out of ffmpeg's URL-protocol demuxers
            # (avoids SSRF via http://, rtsp://, etc.).
            try:
                resolved_video_path = process_video_input(vid_input)
            except Exception as exc:
                logger.warning(f"Could not resolve video: {exc}")
                resolved_video_path = None

            if (
                resolved_video_path
                and self._video_native_with_audio
                and not has_explicit_audio
            ):
                extracted_audio = extract_audio_from_video(resolved_video_path)
                if extracted_audio:
                    _msg_extra_audio.setdefault(msg_idx, []).append(extracted_audio)

            frames = self._prepare_video(
                vid_input,
                fps=video_fps,
                max_frames=video_max_frames,
                resolved_path=resolved_video_path,
            )
            all_video_frames.extend(frames)
            total_frames += len(frames)
            logger.info(f"Added {len(frames)} frames from video: {vid_input}")
        _msg_video_frame_counts[msg_idx] = total_frames

    # Merge auto-extracted audio into the per-message audio map so the
    # chat-template token-counting loop downstream sees the right count.
    for msg_idx, extra in _msg_extra_audio.items():
        _msg_audio_inputs.setdefault(msg_idx, []).extend(extra)

    for aud_inputs in _msg_audio_inputs.values():
        all_audio_inputs.extend(aud_inputs)

    chat_messages = _build_mllm_chat_messages(
        messages,
        all_image_urls=all_image_urls,
        video_frame_counts=_msg_video_frame_counts,
    )

    # Process images
    all_images = []
    if all_image_urls:
        all_images.extend(self._prepare_images(all_image_urls))
    # Append pre-processed video frames
    all_images.extend(all_video_frames)
    all_audio = self._prepare_audio(all_audio_inputs) if all_audio_inputs else []

    # Apply chat template directly - messages are already properly structured
    logger.info(
        f"Applying chat template with {len(chat_messages)} messages, {len(all_images)} images, {len(all_audio)} audios"
    )
    for i, cm in enumerate(chat_messages):
        content_preview = str(cm.get("content", ""))[:80]
        logger.info(
            f"  Chat msg {i}: role={cm['role']}, content={content_preview}..."
        )

    # Build template kwargs for tool definitions (tools already popped above)
    template_extra_kwargs = {}
    if tools:
        template_extra_kwargs["tools"] = tools
    # Forward any remaining chat_template_kwargs (parity with the text path)
    template_extra_kwargs.update(chat_template_kwargs)

    try:
        formatted_prompt = get_chat_template(
            self.processor,
            chat_messages,
            add_generation_prompt=True,
            enable_thinking=enable_thinking,
            **template_extra_kwargs,
        )
    except TypeError:
        # The processor's chat template doesn't accept some forwarded
        # chat_template_kwargs — drop them and retry (mirror llm.py).
        for key in chat_template_kwargs:
            template_extra_kwargs.pop(key, None)
        formatted_prompt = get_chat_template(
            self.processor,
            chat_messages,
            add_generation_prompt=True,
            enable_thinking=enable_thinking,
            **template_extra_kwargs,
        )
    except Exception as e:
        logger.warning(
            f"Failed to apply chat template: {e}, using last user message"
        )
        # Fallback to last user message if template fails
        last_user_msg = ""
        for m in reversed(chat_messages):
            if m["role"] == "user":
                content = m.get("content", "")
                if isinstance(content, list):
                    for item in content:
                        if isinstance(item, dict) and item.get("type") == "text":
                            last_user_msg = item.get("text", "")
                            break
                else:
                    last_user_msg = content
                break
        formatted_prompt = last_user_msg

    # Prefix caching with vision embedding support
    # Following LMCache approach: cache vision embeddings to skip encoder on hit
    import time

    from mlx_vlm.models import cache as vlm_cache

    cache_entry = None
    prefix_match_len = 0
    vision_embeddings = None
    cache_hit = False

    # Tokenize prompt for cache lookup
    tokenizer = (
        self.processor.tokenizer
        if hasattr(self.processor, "tokenizer")
        else self.processor
    )
    token_ids = tokenizer.encode(formatted_prompt)

    # Check prefix cache
    if use_cache and all_audio:
        logger.info("Prefix cache disabled for audio inputs")
        use_cache = False

    if use_cache and self._cache_manager is not None and all_images:
        try:
            cache_entry, prefix_match_len = self._cache_manager.fetch(
                all_images, formatted_prompt, token_ids
            )
            if cache_entry:
                cache_hit = True
                vision_embeddings = cache_entry.vision_embeddings
                if vision_embeddings is not None:
                    logger.info(
                        "[PREFIX CACHE] Vision embeddings cached - would skip encoder!"
                    )
                if prefix_match_len > 0:
                    logger.info(
                        f"[PREFIX CACHE] {prefix_match_len} prefix tokens match"
                    )
        except Exception as e:
            logger.warning(f"Cache fetch failed: {e}")

    # Generate - use KV cache if available from previous identical request
    start_time = time.time()

    # Create or reuse prompt cache for prefix caching speedup
    prompt_cache = None
    skip_prompt_processing = False

    if cache_hit and cache_entry and cache_entry.kv_cache:
        # NOTE: mlx-vlm's generate_step() has its own multimodal KV cache with prefix matching
        # (MULTIMODAL_KV_CACHE_ENABLED in mlx_vlm/utils.py). Let it handle caching.
        # We only use vllm-mlx's cache for text-only requests (no images).
        if all_images:
            # Let mlx-vlm's multimodal cache handle this - don't interfere
            logger.info(
                "[PREFIX CACHE] Images present - delegating to mlx-vlm multimodal cache"
            )
            prompt_cache = None  # Fresh cache, mlx-vlm will handle prefix matching
            skip_prompt_processing = False
        else:
            # Text-only: can use skip_prompt_processing for maximum speedup
            logger.info(
                "[PREFIX CACHE] Text-only cache hit - using skip_prompt_processing speedup"
            )
            cached_prompt_cache = cache_entry.kv_cache
            try:
                import copy

                prompt_cache = []
                for layer_cache in cached_prompt_cache:
                    new_cache = copy.copy(layer_cache)
                    if hasattr(layer_cache, "state"):
                        state = layer_cache.state
                        if state is not None:
                            import mlx.core as mx

                            if len(state) >= 2 and state[0] is not None:
                                new_cache.keys = mx.array(state[0])
                                new_cache.values = mx.array(state[1])
                                if len(state) >= 3:
                                    new_cache.offset = state[2]
                                elif hasattr(layer_cache, "offset"):
                                    new_cache.offset = layer_cache.offset
                    prompt_cache.append(new_cache)
                skip_prompt_processing = True
                logger.info(
                    f"[PREFIX CACHE] Skipping {prefix_match_len} token forward pass"
                )
            except Exception as e:
                logger.warning(f"[PREFIX CACHE] Failed to copy cache: {e}")
                prompt_cache = None
                skip_prompt_processing = False

    if prompt_cache is None and self.model is not None:
        # Create fresh cache
        try:
            prompt_cache = vlm_cache.make_prompt_cache(
                self.model.language_model,
                max_kv_size=self.max_kv_size or None,
            )
        except Exception:
            prompt_cache = None

    draft_accept_start = self._reset_draft_metrics()
    result = generate(
        self.model,
        self.processor,
        formatted_prompt,
        all_images if all_images else None,
        audio=all_audio if all_audio else None,
        max_tokens=max_tokens,
        temp=temperature,
        verbose=False,
        prompt_cache=prompt_cache,
        skip_prompt_processing=skip_prompt_processing,
        **self._draft_generation_kwargs(kwargs),
        **kwargs,
    )
    draft_metrics = self._draft_metrics_since(draft_accept_start)

    # Store KV cache for future reuse (on cache miss)
    # IMPORTANT: We need to store only the prompt portion, not generated tokens
    if (
        use_cache
        and self._cache_manager is not None
        and all_images
        and not cache_hit
        and prompt_cache
    ):
        try:
            import copy

            import mlx.core as mx

            # Get prompt token count (before generation)
            prompt_tokens_count = getattr(result, "prompt_tokens", 0)

            # Deep copy the cache and trim to prompt tokens only
            cache_to_store = []
            for layer_cache in prompt_cache:
                new_cache = copy.copy(layer_cache)
                if hasattr(layer_cache, "state"):
                    state = layer_cache.state
                    if (
                        state is not None
                        and len(state) >= 2
                        and state[0] is not None
                    ):
                        # Copy arrays
                        keys = mx.array(state[0])
                        values = mx.array(state[1])
                        # Trim to prompt tokens only (not generated tokens)
                        if (
                            hasattr(layer_cache, "offset")
                            and layer_cache.offset > prompt_tokens_count
                        ):
                            # For caches with offset tracking, slice to prompt length
                            new_cache.keys = keys[:, :, :prompt_tokens_count, :]
                            new_cache.values = values[:, :, :prompt_tokens_count, :]
                            new_cache.offset = prompt_tokens_count
                        else:
                            new_cache.keys = keys
                            new_cache.values = values
                            if len(state) >= 3:
                                new_cache.offset = state[2]
                            elif hasattr(layer_cache, "offset"):
                                new_cache.offset = min(
                                    layer_cache.offset, prompt_tokens_count
                                )
                cache_to_store.append(new_cache)

            self._cache_manager.store(
                images=all_images,
                prompt=formatted_prompt,
                vision_embeddings=None,
                kv_cache=cache_to_store,
                token_ids=token_ids,
                num_image_tokens=256,
                model_name=self.model_name,
            )
            logger.info(
                f"[PREFIX CACHE] Stored KV cache for {len(all_images)} image(s) ({prompt_tokens_count} prompt tokens)"
            )
        except Exception as e:
            logger.warning(f"Failed to cache: {e}")

    # Handle GenerationResult object or plain string
    if hasattr(result, "text"):
        output_text = result.text
        prompt_tokens = getattr(result, "prompt_tokens", 0)
        generation_tokens = getattr(result, "generation_tokens", 0)
    else:
        output_text = str(result)
        prompt_tokens = 0
        generation_tokens = 0

    return MLLMOutput(
        text=output_text,
        finish_reason="stop",
        prompt_tokens=prompt_tokens,
        completion_tokens=generation_tokens,
        **draft_metrics,
    )

vllm_mlx.models.mllm.MLXMultimodalLM.stream_chat 

stream_chat(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, **kwargs) -> Iterator[MLLMOutput]

Stream chat with OpenAI-compatible message format.

Supports multimodal content in messages: - {"type": "text", "text": "..."} - {"type": "image_url", "image_url": {"url": "..."}} - {"type": "image_url", "image_url": {"url": "data:image/...;base64,..."}}

Parameters:

  • messages (list[dict]) –

    List of chat messages (OpenAI format)

  • max_tokens (int, default: 256 ) –

    Maximum tokens to generate

  • temperature (float, default: 0.7 ) –

    Sampling temperature

  • **kwargs –

    Additional parameters

Yields:

  • MLLMOutput –

    MLLMOutput with incremental text chunks

Source code in vllm_mlx/models/mllm.py 
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
def stream_chat(
    self,
    messages: list[dict],
    max_tokens: int = 256,
    temperature: float = 0.7,
    **kwargs,
) -> Iterator[MLLMOutput]:
    """
    Stream chat with OpenAI-compatible message format.

    Supports multimodal content in messages:
    - {"type": "text", "text": "..."}
    - {"type": "image_url", "image_url": {"url": "..."}}
    - {"type": "image_url", "image_url": {"url": "data:image/...;base64,..."}}

    Args:
        messages: List of chat messages (OpenAI format)
        max_tokens: Maximum tokens to generate
        temperature: Sampling temperature
        **kwargs: Additional parameters

    Yields:
        MLLMOutput with incremental text chunks
    """
    if not self._loaded:
        self.load()

    try:
        from mlx_vlm import stream_generate
        from mlx_vlm.prompt_utils import get_chat_template
    except ImportError:
        # Fallback to non-streaming if stream_generate not available
        output = self.chat(
            messages=messages,
            max_tokens=max_tokens,
            temperature=temperature,
            **kwargs,
        )
        yield output
        return

    # Extract text and images from messages
    # Build chat_messages for multi-turn support WITH proper image tokens per message
    all_image_urls = []  # Raw URLs/paths to process later
    chat_messages = []  # List of properly formatted messages for chat template

    # Pop params early so they don't leak into mlx_vlm.generate()
    video_fps = kwargs.pop("video_fps", DEFAULT_FPS)
    video_max_frames = kwargs.pop("video_max_frames", MAX_FRAMES)
    tools = kwargs.pop("tools", None)
    use_cache = kwargs.pop("use_cache", True)
    enable_thinking = kwargs.pop("enable_thinking", True)
    # Honor chat_template_kwargs on the MLLM path (parity with the text path in
    # llm.py). enable_thinking is commonly nested here via
    # --default-chat-template-kwargs {"enable_thinking": false}; it was previously
    # dropped for MLLM models, so reasoning could not be disabled and leaked into
    # the response.
    chat_template_kwargs = kwargs.pop("chat_template_kwargs", None) or {}
    if "enable_thinking" in chat_template_kwargs:
        enable_thinking = chat_template_kwargs.pop("enable_thinking")

    # Collect video and audio inputs from messages
    _msg_video_inputs = self._collect_video_inputs(messages)
    _msg_audio_inputs = self._collect_audio_inputs(messages)

    # Use native video pipeline for supported models.
    # NOTE: Native video yields a single chunk (not incremental streaming)
    # because mlx_vlm.video_generate has no streaming API. The event loop
    # is NOT blocked at the server level — SimpleEngine wraps this in
    # asyncio.to_thread(). True token-level streaming requires upstream
    # mlx-vlm support for video stream_generate.
    if self._video_native and _msg_video_inputs:
        output = self._generate_native_video(
            messages=messages,
            max_tokens=max_tokens,
            temperature=temperature,
            video_fps=video_fps,
            video_max_frames=video_max_frames,
            tools=tools,
            **kwargs,
        )
        yield output
        return

    # Fallback: frames as images
    _msg_video_frame_counts: dict[int, int] = {}
    _msg_extra_audio: dict[int, list[str]] = {}
    all_video_frames: list[str] = []
    all_audio_inputs: list[str] = []
    for msg_idx, vid_inputs in _msg_video_inputs.items():
        total_frames = 0
        has_explicit_audio = bool(_msg_audio_inputs.get(msg_idx))
        for vid_input in vid_inputs:
            # Resolve once; reused for audio extraction and frame prep.
            # See the matching block in chat() for rationale.
            try:
                resolved_video_path = process_video_input(vid_input)
            except Exception as exc:
                logger.warning(f"Could not resolve video: {exc}")
                resolved_video_path = None

            if (
                resolved_video_path
                and self._video_native_with_audio
                and not has_explicit_audio
            ):
                extracted_audio = extract_audio_from_video(resolved_video_path)
                if extracted_audio:
                    _msg_extra_audio.setdefault(msg_idx, []).append(extracted_audio)

            frames = self._prepare_video(
                vid_input,
                fps=video_fps,
                max_frames=video_max_frames,
                resolved_path=resolved_video_path,
            )
            all_video_frames.extend(frames)
            total_frames += len(frames)
            logger.info(f"Added {len(frames)} frames from video: {vid_input}")
        _msg_video_frame_counts[msg_idx] = total_frames

    for msg_idx, extra in _msg_extra_audio.items():
        _msg_audio_inputs.setdefault(msg_idx, []).extend(extra)

    for aud_inputs in _msg_audio_inputs.values():
        all_audio_inputs.extend(aud_inputs)

    chat_messages = _build_mllm_chat_messages(
        messages,
        all_image_urls=all_image_urls,
        video_frame_counts=_msg_video_frame_counts,
    )

    all_images = []
    if all_image_urls:
        all_images.extend(self._prepare_images(all_image_urls))
    all_images.extend(all_video_frames)
    all_audio = self._prepare_audio(all_audio_inputs) if all_audio_inputs else []

    # Build template kwargs for tool definitions (tools already popped above)
    template_extra_kwargs = {}
    if tools:
        template_extra_kwargs["tools"] = tools
    # Forward any remaining chat_template_kwargs (parity with the text path)
    template_extra_kwargs.update(chat_template_kwargs)

    try:
        formatted_prompt = get_chat_template(
            self.processor,
            chat_messages,
            add_generation_prompt=True,
            enable_thinking=enable_thinking,
            **template_extra_kwargs,
        )
    except TypeError:
        # The processor's chat template doesn't accept some forwarded
        # chat_template_kwargs — drop them and retry (mirror llm.py).
        for key in chat_template_kwargs:
            template_extra_kwargs.pop(key, None)
        formatted_prompt = get_chat_template(
            self.processor,
            chat_messages,
            add_generation_prompt=True,
            enable_thinking=enable_thinking,
            **template_extra_kwargs,
        )
    except Exception as e:
        logger.warning(
            f"Failed to apply chat template: {e}, using last user message"
        )
        # Fallback to last user message if template fails
        last_user_msg = ""
        for m in reversed(chat_messages):
            if m["role"] == "user":
                content = m.get("content", "")
                if isinstance(content, list):
                    for item in content:
                        if isinstance(item, dict) and item.get("type") == "text":
                            last_user_msg = item.get("text", "")
                            break
                else:
                    last_user_msg = content
                break
        formatted_prompt = last_user_msg

    # Check cache for existing KV state (uses images as cache key)
    from mlx_vlm.models import cache as vlm_cache

    prompt_cache = None
    cache_hit = False

    if use_cache and all_audio:
        logger.info("Stream chat cache disabled for audio inputs")
        use_cache = False

    if use_cache and self._cache_manager is not None and all_images:
        prompt_cache, cache_hit = self._cache_manager.fetch_cache(
            all_images, formatted_prompt
        )
        if cache_hit:
            logger.debug(f"Stream chat cache hit for {len(all_images)} image(s)")

    # Create new cache if needed
    if prompt_cache is None and self.model is not None:
        try:
            prompt_cache = vlm_cache.make_prompt_cache(
                self.model.language_model,
                max_kv_size=self.max_kv_size or None,
            )
        except Exception:
            prompt_cache = None

    # Stream generate tokens with cache
    accumulated_text = ""
    token_count = 0
    draft_accept_start = self._reset_draft_metrics()

    for chunk in stream_generate(
        self.model,
        self.processor,
        formatted_prompt,
        all_images if all_images else None,
        audio=all_audio if all_audio else None,
        max_tokens=max_tokens,
        temp=temperature,
        prompt_cache=prompt_cache,
        **self._draft_generation_kwargs(kwargs),
        **kwargs,
    ):
        token_count += 1
        # chunk is a GenerationResult with .text attribute containing the new token
        new_text = chunk.text if hasattr(chunk, "text") else str(chunk)
        accumulated_text += new_text

        yield MLLMOutput(
            text=new_text,  # Just the new token for streaming
            finish_reason=None,
            prompt_tokens=getattr(chunk, "prompt_tokens", 0),
            completion_tokens=token_count,
        )

    # Final yield with finish_reason
    yield MLLMOutput(
        text="",
        finish_reason="stop",
        prompt_tokens=getattr(chunk, "prompt_tokens", 0) if "chunk" in dir() else 0,
        completion_tokens=token_count,
        **self._draft_metrics_since(draft_accept_start),
    )

vllm_mlx.models.mllm.MLXMultimodalLM.describe_image 

describe_image(image: str, prompt: str = 'Describe this image in detail.', max_tokens: int = 512, **kwargs) -> str

Convenience method to describe an image.

Parameters:

  • image (str) –

    Image path, URL, or base64 string

  • prompt (str, default: 'Describe this image in detail.' ) –

    Description prompt

  • max_tokens (int, default: 512 ) –

    Maximum tokens

  • **kwargs –

    Additional parameters

Returns:

  • str –

    Image description text

Source code in vllm_mlx/models/mllm.py 
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
def describe_image(
    self,
    image: str,
    prompt: str = "Describe this image in detail.",
    max_tokens: int = 512,
    **kwargs,
) -> str:
    """
    Convenience method to describe an image.

    Args:
        image: Image path, URL, or base64 string
        prompt: Description prompt
        max_tokens: Maximum tokens
        **kwargs: Additional parameters

    Returns:
        Image description text
    """
    output = self.generate(
        prompt=prompt,
        images=[image],
        max_tokens=max_tokens,
        **kwargs,
    )
    return output.text

vllm_mlx.models.mllm.MLXMultimodalLM.answer_about_image 

answer_about_image(image: str, question: str, max_tokens: int = 256, **kwargs) -> str

Answer a question about an image.

Parameters:

  • image (str) –

    Image path, URL, or base64 string

  • question (str) –

    Question about the image

  • max_tokens (int, default: 256 ) –

    Maximum tokens

  • **kwargs –

    Additional parameters

Returns:

  • str –

    Answer text

Source code in vllm_mlx/models/mllm.py 
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
def answer_about_image(
    self,
    image: str,
    question: str,
    max_tokens: int = 256,
    **kwargs,
) -> str:
    """
    Answer a question about an image.

    Args:
        image: Image path, URL, or base64 string
        question: Question about the image
        max_tokens: Maximum tokens
        **kwargs: Additional parameters

    Returns:
        Answer text
    """
    output = self.generate(
        prompt=question,
        images=[image],
        max_tokens=max_tokens,
        **kwargs,
    )
    return output.text

vllm_mlx.models.mllm.MLXMultimodalLM.describe_video 

describe_video(video: str | dict, prompt: str = 'Describe what happens in this video.', fps: float = 2.0, max_frames: int = 32, max_tokens: int = 512, **kwargs) -> str

Describe a video using frame extraction.

Parameters:

  • video (str | dict) –

    Video file path, URL, base64, or OpenAI format dict

  • prompt (str, default: 'Describe what happens in this video.' ) –

    Description prompt

  • fps (float, default: 2.0 ) –

    Frames per second to extract

  • max_frames (int, default: 32 ) –

    Maximum frames to extract

  • max_tokens (int, default: 512 ) –

    Maximum tokens to generate

Returns:

  • str –

    Video description text

Example
URL

model.describe_video("https://example.com/video.mp4")

OpenAI format

model.describe_video({"url": "https://example.com/video.mp4"})

Source code in vllm_mlx/models/mllm.py 
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
def describe_video(
    self,
    video: str | dict,
    prompt: str = "Describe what happens in this video.",
    fps: float = 2.0,
    max_frames: int = 32,
    max_tokens: int = 512,
    **kwargs,
) -> str:
    """
    Describe a video using frame extraction.

    Args:
        video: Video file path, URL, base64, or OpenAI format dict
        prompt: Description prompt
        fps: Frames per second to extract
        max_frames: Maximum frames to extract
        max_tokens: Maximum tokens to generate

    Returns:
        Video description text

    Example:
        # URL
        model.describe_video("https://example.com/video.mp4")

        # OpenAI format
        model.describe_video({"url": "https://example.com/video.mp4"})
    """
    output = self.generate(
        prompt=prompt,
        videos=[video],
        video_fps=fps,
        video_max_frames=max_frames,
        max_tokens=max_tokens,
        **kwargs,
    )
    return output.text

vllm_mlx.models.mllm.MLXMultimodalLM.get_cache_stats 

get_cache_stats() -> dict

Get MLLM cache statistics.

Returns:

  • dict –

    Dictionary with cache stats (hits, misses, hit_rate, tokens_saved, etc.)

Source code in vllm_mlx/models/mllm.py 
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
def get_cache_stats(self) -> dict:
    """
    Get MLLM cache statistics.

    Returns:
        Dictionary with cache stats (hits, misses, hit_rate, tokens_saved, etc.)
    """
    if self._cache_manager is None:
        return {"enabled": False}

    stats = self._cache_manager.get_stats()
    stats["enabled"] = True
    stats["cache_entries"] = len(self._cache_manager)
    stats["max_entries"] = self._cache_manager.max_size
    return stats

vllm_mlx.models.mllm.MLXMultimodalLM.clear_cache 

clear_cache() -> None

Clear the MLLM KV cache.

Source code in vllm_mlx/models/mllm.py 
2848
2849
2850
2851
2852
def clear_cache(self) -> None:
    """Clear the MLLM KV cache."""
    if self._cache_manager is not None:
        self._cache_manager.clear()
        logger.info("MLLM cache cleared")

vllm_mlx.models.mllm.MLXMultimodalLM.get_model_info 

get_model_info() -> dict

Get information about the loaded model.

Source code in vllm_mlx/models/mllm.py 
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
def get_model_info(self) -> dict:
    """Get information about the loaded model."""
    if not self._loaded:
        return {"loaded": False, "model_name": self.model_name}

    info = {
        "loaded": True,
        "model_name": self.model_name,
        "type": "multimodal-language-model",
        "supports_video": True,
        "supports_streaming": True,
        "cache_enabled": self.enable_cache,
    }

    if self.config:
        info["model_type"] = getattr(self.config, "model_type", "unknown")

    if self._cache_manager is not None:
        info["cache_stats"] = self._cache_manager.get_stats()

    return info

vllm_mlx.models.mllm.MLXMultimodalLM.list_supported_model_families staticmethod 

list_supported_model_families() -> dict[str, str]

List supported model families and their patterns.

Any model on HuggingFace containing these patterns in the name is likely compatible with mlx-vlm.

Source code in vllm_mlx/models/mllm.py 
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
@staticmethod
def list_supported_model_families() -> dict[str, str]:
    """
    List supported model families and their patterns.

    Any model on HuggingFace containing these patterns in the name
    is likely compatible with mlx-vlm.
    """
    return {
        "Qwen-VL": "Qwen VL models (Qwen2-VL, Qwen2.5-VL, Qwen3-VL, etc.)",
        "LLaVA": "LLaVA vision-language models",
        "Idefics": "Idefics vision-language models",
        "PaliGemma": "PaliGemma multimodal models",
        "Pixtral": "Mistral's Pixtral vision models",
        "Molmo": "Allen AI's Molmo models",
        "Phi-3-Vision": "Microsoft's Phi-3 Vision models",
        "CogVLM": "Tsinghua's CogVLM models",
        "InternVL": "InternVL models",
        "MiniCPM-V": "OpenBMB's MiniCPM-V models",
        "Florence": "Microsoft Florence vision models",
        "DeepSeek-VL": "DeepSeek's vision-language models (DeepSeek-VL, DeepSeek-VL2)",
    }

vllm_mlx.models.mllm.MLXMultimodalLM.is_mllm_model staticmethod 

is_mllm_model(model_name: str) -> bool

Check if a model name indicates an MLLM model.

Source code in vllm_mlx/models/mllm.py 
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
@staticmethod
def is_mllm_model(model_name: str) -> bool:
    """Check if a model name indicates an MLLM model."""
    mllm_patterns = [
        "-VL-",
        "-VL/",
        "VL-",
        "llava",
        "LLaVA",
        "idefics",
        "Idefics",
        "paligemma",
        "PaliGemma",
        "gemma-3",
        "gemma3",  # Gemma 3 (multimodal)
        "medgemma",
        "MedGemma",  # MedGemma (medical multimodal)
        "pixtral",
        "Pixtral",
        "molmo",
        "Molmo",
        "phi3-vision",
        "phi-3-vision",
        "cogvlm",
        "CogVLM",
        "internvl",
        "InternVL",
        "minicpm-v",
        "MiniCPM-V",
        "florence",
        "Florence",
        "deepseek-vl",
        "DeepSeek-VL",
    ]
    model_lower = model_name.lower()
    return any(pattern.lower() in model_lower for pattern in mllm_patterns)

vllm_mlx.models.mllm.MLXMultimodalLM.__repr__ 

__repr__() -> str
Source code in vllm_mlx/models/mllm.py 
2936
2937
2938
def __repr__(self) -> str:
    status = "loaded" if self._loaded else "not loaded"
    return f"<MLXMultimodalLM model={self.model_name} status={status}>"

vllm_mlx.models.mllm.cleanup_temp_file 

cleanup_temp_file(path: str) -> bool

Clean up a specific temporary file.

Source code in vllm_mlx/models/mllm.py 
93
94
95
def cleanup_temp_file(path: str) -> bool:
    """Clean up a specific temporary file."""
    return _temp_manager.cleanup(path)

vllm_mlx.models.mllm.cleanup_all_temp_files 

cleanup_all_temp_files() -> int

Clean up all tracked temporary files. Returns count of cleaned files.

Source code in vllm_mlx/models/mllm.py 
 98
 99
100
def cleanup_all_temp_files() -> int:
    """Clean up all tracked temporary files. Returns count of cleaned files."""
    return _temp_manager.cleanup_all()

vllm_mlx.models.mllm._normalize_content_part 

_normalize_content_part(item: object) -> object

Convert Pydantic content parts into plain Python objects.

Source code in vllm_mlx/models/mllm.py 
138
139
140
141
142
143
144
def _normalize_content_part(item: object) -> object:
    """Convert Pydantic content parts into plain Python objects."""
    if hasattr(item, "model_dump"):
        return item.model_dump(exclude_none=True)
    if hasattr(item, "dict"):
        return {k: v for k, v in item.dict().items() if v is not None}
    return item

vllm_mlx.models.mllm._extract_media_url 

_extract_media_url(item: dict, item_type: str) -> str
Source code in vllm_mlx/models/mllm.py 
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
def _extract_media_url(item: dict, item_type: str) -> str:
    if item_type == "image_url":
        media_value = item.get("image_url", {})
    elif item_type == "video_url":
        media_value = item.get("video_url", {})
    elif item_type == "audio_url":
        media_value = item.get("audio_url", {})
    elif item_type in {"image", "video", "audio"}:
        media_value = item.get(item_type, item.get("url", ""))
    else:
        return ""

    if isinstance(media_value, dict):
        media_value = media_value.get("url", "")
    return media_value if isinstance(media_value, str) else ""

vllm_mlx.models.mllm._text_content_part 

_text_content_part(text: str) -> dict[str, str]
Source code in vllm_mlx/models/mllm.py 
164
165
def _text_content_part(text: str) -> dict[str, str]:
    return {"type": "text", "text": text, "content": text}

vllm_mlx.models.mllm._append_text_content_part 

_append_text_content_part(built_parts: list[dict[str, str]], text_parts: list[str], text: str) -> None
Source code in vllm_mlx/models/mllm.py 
168
169
170
171
172
173
174
def _append_text_content_part(
    built_parts: list[dict[str, str]], text_parts: list[str], text: str
) -> None:
    if not text:
        return
    built_parts.append(_text_content_part(text))
    text_parts.append(text)

vllm_mlx.models.mllm._build_string_mllm_message_content 

_build_string_mllm_message_content(content: str, role: str) -> tuple[object, bool]
Source code in vllm_mlx/models/mllm.py 
177
178
179
180
181
182
def _build_string_mllm_message_content(content: str, role: str) -> tuple[object, bool]:
    if not content:
        return "", False
    if role == "assistant":
        return content, True
    return [_text_content_part(content)], True

vllm_mlx.models.mllm._append_ordered_mllm_content_part 

_append_ordered_mllm_content_part(raw_item: object, *, built_parts: list[dict[str, str]], text_parts: list[str], all_image_urls: list[str], video_frame_count: int) -> int
Source code in vllm_mlx/models/mllm.py 
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
def _append_ordered_mllm_content_part(
    raw_item: object,
    *,
    built_parts: list[dict[str, str]],
    text_parts: list[str],
    all_image_urls: list[str],
    video_frame_count: int,
) -> int:
    item = _normalize_content_part(raw_item)
    if isinstance(item, str):
        _append_text_content_part(built_parts, text_parts, item)
        return video_frame_count

    if not isinstance(item, dict):
        return video_frame_count

    item_type = item.get("type", "")
    if item_type in {"text", "input_text"}:
        _append_text_content_part(
            built_parts, text_parts, item.get("text", "") or item.get("content", "")
        )
    elif item_type in {"image_url", "image"}:
        media_url = _extract_media_url(item, item_type)
        if media_url:
            all_image_urls.append(media_url)
        built_parts.append({"type": "image"})
    elif item_type in {"audio_url", "audio"}:
        # Audio inputs are collected once by _collect_audio_inputs before
        # message reconstruction; this helper only preserves placeholder order.
        built_parts.append({"type": "audio"})
    elif item_type in {"video", "video_url"}:
        # Native video models bypass this helper. For fallback frame extraction,
        # preserve the video position by inserting that message's frames here.
        built_parts.extend({"type": "image"} for _ in range(video_frame_count))
        return 0
    return video_frame_count

vllm_mlx.models.mllm._build_ordered_mllm_message_content 

_build_ordered_mllm_message_content(content: object, *, role: str, all_image_urls: list[str], video_frame_count: int = 0) -> tuple[object, bool]

Build template content while preserving OpenAI media/text part order.

Source code in vllm_mlx/models/mllm.py 
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
def _build_ordered_mllm_message_content(
    content: object,
    *,
    role: str,
    all_image_urls: list[str],
    video_frame_count: int = 0,
) -> tuple[object, bool]:
    """Build template content while preserving OpenAI media/text part order."""
    if isinstance(content, str):
        return _build_string_mllm_message_content(content, role)

    if not isinstance(content, list):
        return "", False

    built_parts: list[dict[str, str]] = []
    text_parts: list[str] = []
    remaining_video_frames = video_frame_count

    for raw_item in content:
        remaining_video_frames = _append_ordered_mllm_content_part(
            raw_item,
            built_parts=built_parts,
            text_parts=text_parts,
            all_image_urls=all_image_urls,
            video_frame_count=remaining_video_frames,
        )

    if role == "assistant":
        text = "".join(text_parts)
        return text, bool(text)

    return built_parts, bool(built_parts)

vllm_mlx.models.mllm._normalize_mllm_tool_calls 

_normalize_mllm_tool_calls(tool_calls: list) -> list

Normalize replayed assistant tool calls for chat templates.

Mirrors _normalize_tool_call_arguments_for_template in vllm_mlx/engine/batched.py: JSON argument strings become mappings so templates that iterate argument keys render correctly.

Source code in vllm_mlx/models/mllm.py 
257
258
259
260
261
262
263
264
265
266
267
268
def _normalize_mllm_tool_calls(tool_calls: list) -> list:
    """Normalize replayed assistant tool calls for chat templates.

    Mirrors ``_normalize_tool_call_arguments_for_template`` in
    ``vllm_mlx/engine/batched.py``: JSON argument strings become mappings so
    templates that iterate argument keys render correctly.
    """
    plain_calls = [_normalize_content_part(call) for call in tool_calls]
    normalized = normalize_messages_for_chat_template(
        [{"role": "assistant", "tool_calls": plain_calls}]
    )
    return normalized[0].get("tool_calls", plain_calls)

vllm_mlx.models.mllm._build_mllm_chat_messages 

_build_mllm_chat_messages(messages: list[dict], *, all_image_urls: list[str], video_frame_counts: dict[int, int]) -> list[dict]

Build chat-template messages without reordering multimodal content parts.

Source code in vllm_mlx/models/mllm.py 
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
def _build_mllm_chat_messages(
    messages: list[dict],
    *,
    all_image_urls: list[str],
    video_frame_counts: dict[int, int],
) -> list[dict]:
    """Build chat-template messages without reordering multimodal content parts."""
    chat_messages: list[dict] = []
    for msg_idx, msg in enumerate(messages):
        role = msg.get("role", "user")
        if not isinstance(role, str):
            role = str(role)

        content, has_content = _build_ordered_mllm_message_content(
            msg.get("content", ""),
            role=role,
            all_image_urls=all_image_urls,
            video_frame_count=video_frame_counts.get(msg_idx, 0),
        )
        chat_message = {"role": role, "content": content}

        if role == "assistant":
            tool_calls = msg.get("tool_calls")
            if isinstance(tool_calls, list) and tool_calls:
                # Keep tool-call turns even when text content is empty so
                # templates render the assistant -> tool exchange (issue #608).
                chat_message["tool_calls"] = _normalize_mllm_tool_calls(tool_calls)
                reasoning_content = msg.get("reasoning_content")
                if reasoning_content:
                    chat_message["reasoning_content"] = reasoning_content
                has_content = True
        elif role == "tool":
            tool_call_id = msg.get("tool_call_id")
            if tool_call_id:
                chat_message["tool_call_id"] = tool_call_id
                # Tools may legitimately return empty output; keep the message
                # anyway so the assistant tool_call still has its anchor and
                # template forward-scans pair calls to responses (issue #608).
                if not has_content:
                    chat_message["content"] = ""
                    has_content = True

        if has_content:
            chat_messages.append(chat_message)
    return chat_messages

vllm_mlx.models.mllm.load_gemma4_assistant_drafter 

load_gemma4_assistant_drafter(model_path: str)

Load a Gemma 4 assistant drafter for mlx-vlm speculative decoding.

Source code in vllm_mlx/models/mllm.py 
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
def load_gemma4_assistant_drafter(model_path: str):
    """Load a Gemma 4 assistant drafter for mlx-vlm speculative decoding."""
    try:
        import mlx.core as mx
        from mlx_vlm.speculative.drafters import gemma4_assistant as arch
    except ImportError as exc:
        raise ImportError(
            "Gemma 4 assistant drafter support requires an mlx-vlm build that "
            "provides mlx_vlm.speculative.drafters.gemma4_assistant."
        ) from exc

    try:
        mlx_vlm_version = version("mlx-vlm")
    except PackageNotFoundError:
        mlx_vlm_version = "unknown"
    logger.info(
        "Loading Gemma 4 assistant drafter from %s using mlx-vlm %s",
        model_path,
        mlx_vlm_version,
    )

    path = Path(model_path)
    config_path = path / "config.json"
    weight_paths = sorted(path.glob("*.safetensors"))
    if not config_path.exists():
        raise FileNotFoundError(f"Gemma 4 assistant config not found: {config_path}")
    if not weight_paths:
        raise FileNotFoundError(f"Gemma 4 assistant weights not found: {path}")

    config = arch.ModelConfig.from_dict(
        json.loads(config_path.read_text(encoding="utf-8"))
    )
    model = arch.Model(config)
    weights = {}
    for weight_path in weight_paths:
        weights.update(mx.load(str(weight_path)))
    if hasattr(model, "sanitize"):
        weights = model.sanitize(weights)
    model.load_weights(list(weights.items()))
    mx.eval(model.parameters())
    model.eval()
    return model

vllm_mlx.models.mllm._count_draft_tokens 

_count_draft_tokens(draft_tokens) -> int

Best-effort drafted-token count for an mlx-vlm drafter output.

Source code in vllm_mlx/models/mllm.py 
387
388
389
390
391
392
393
394
395
396
397
398
def _count_draft_tokens(draft_tokens) -> int:
    """Best-effort drafted-token count for an mlx-vlm drafter output."""
    shape = getattr(draft_tokens, "shape", None)
    if shape:
        try:
            return max(int(shape[-1]), 0)
        except (TypeError, ValueError):
            pass
    try:
        return max(len(draft_tokens), 0)
    except TypeError:
        return 0

vllm_mlx.models.mllm._install_draft_metrics_hooks 

_install_draft_metrics_hooks(draft_model) -> None

Record actual drafted token counts from mlx-vlm assistant drafters.

Source code in vllm_mlx/models/mllm.py 
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
def _install_draft_metrics_hooks(draft_model) -> None:
    """Record actual drafted token counts from mlx-vlm assistant drafters."""
    if draft_model is None or getattr(draft_model, "_vllm_mlx_metrics_hooked", False):
        return

    if not hasattr(draft_model, "_vllm_mlx_draft_counts"):
        draft_model._vllm_mlx_draft_counts = []

    draft_block = getattr(draft_model, "draft_block", None)
    if callable(draft_block):

        def draft_block_with_metrics(*args, **kwargs):
            draft_tokens = draft_block(*args, **kwargs)
            draft_model._vllm_mlx_draft_counts.append(_count_draft_tokens(draft_tokens))
            return draft_tokens

        draft_model.draft_block = draft_block_with_metrics

    reset = getattr(draft_model, "reset", None)
    if callable(reset):

        def reset_with_metrics(*args, **kwargs):
            draft_model._vllm_mlx_draft_counts = []
            return reset(*args, **kwargs)

        draft_model.reset = reset_with_metrics

    draft_model._vllm_mlx_metrics_hooked = True

vllm_mlx.models.mllm.is_base64_image 

is_base64_image(s: str) -> bool

Check if string is base64-encoded image data.

Source code in vllm_mlx/models/mllm.py 
431
432
433
434
435
def is_base64_image(s: str) -> bool:
    """Check if string is base64-encoded image data."""
    return s.startswith("data:image/") or (
        len(s) > 100 and not s.startswith(("http://", "https://", "/"))
    )

vllm_mlx.models.mllm.is_url 

is_url(s: str) -> bool

Check if string is a URL.

Source code in vllm_mlx/models/mllm.py 
438
439
440
def is_url(s: str) -> bool:
    """Check if string is a URL."""
    return s.startswith(("http://", "https://"))

vllm_mlx.models.mllm.is_base64_video 

is_base64_video(s: str) -> bool

Check if string is base64-encoded video data.

Source code in vllm_mlx/models/mllm.py 
443
444
445
def is_base64_video(s: str) -> bool:
    """Check if string is base64-encoded video data."""
    return s.startswith("data:video/")

vllm_mlx.models.mllm.is_base64_audio 

is_base64_audio(s: str) -> bool

Check if string is base64-encoded audio data.

Source code in vllm_mlx/models/mllm.py 
448
449
450
def is_base64_audio(s: str) -> bool:
    """Check if string is base64-encoded audio data."""
    return s.startswith("data:audio/")

vllm_mlx.models.mllm.decode_base64_image 

decode_base64_image(base64_string: str, max_length: int = MAX_BASE64_IMAGE_LENGTH) -> bytes

Decode base64 image to bytes.

Parameters:

  • base64_string (str) –

    Base64 encoded image (optionally with data URL prefix)

  • max_length (int, default: MAX_BASE64_IMAGE_LENGTH ) –

    Maximum allowed length of base64 string

Returns:

  • bytes –

    Decoded image bytes

Raises:

Source code in vllm_mlx/models/mllm.py 
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
def decode_base64_image(
    base64_string: str, max_length: int = MAX_BASE64_IMAGE_LENGTH
) -> bytes:
    """
    Decode base64 image to bytes.

    Args:
        base64_string: Base64 encoded image (optionally with data URL prefix)
        max_length: Maximum allowed length of base64 string

    Returns:
        Decoded image bytes

    Raises:
        FileSizeExceededError: If base64 string exceeds max_length
    """
    if len(base64_string) > max_length:
        raise FileSizeExceededError(
            f"Base64 image data exceeds maximum size: {len(base64_string) / 1024 / 1024:.1f} MB > "
            f"{max_length / 1024 / 1024:.1f} MB limit"
        )

    # Handle data URL format: data:image/jpeg;base64,/9j/4AAQ...
    if base64_string.startswith("data:"):
        # Extract the base64 part after the comma
        _, data = base64_string.split(",", 1)
        return base64.b64decode(data)
    return base64.b64decode(base64_string)

vllm_mlx.models.mllm._validate_url_safety 

_validate_url_safety(url: str) -> None

Reject remote URLs that target local or private network resources.

Source code in vllm_mlx/models/mllm.py 
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
def _validate_url_safety(url: str) -> None:
    """Reject remote URLs that target local or private network resources."""
    parsed = urlparse(url)
    if parsed.scheme not in {"http", "https"}:
        raise UnsafeRemoteURLError(
            f"Unsupported remote media URL scheme: {parsed.scheme or '<missing>'}"
        )

    hostname = parsed.hostname
    if not hostname:
        raise UnsafeRemoteURLError("Remote media URL must include a hostname")

    if hostname == "localhost" or hostname.endswith(".localhost"):
        raise UnsafeRemoteURLError(
            f"Remote media URL targets a blocked host: {hostname}"
        )

    try:
        resolved_ips = [ipaddress.ip_address(hostname)]
    except ValueError:
        try:
            addrinfo = socket.getaddrinfo(
                hostname,
                parsed.port or (443 if parsed.scheme == "https" else 80),
                type=socket.SOCK_STREAM,
            )
        except socket.gaierror as exc:
            raise UnsafeRemoteURLError(
                f"Failed to resolve remote media host {hostname}: {exc}"
            ) from exc
        resolved_ips = [ipaddress.ip_address(info[4][0]) for info in addrinfo]

    blocked_ips = [str(ip) for ip in resolved_ips if not ip.is_global]
    if blocked_ips:
        raise UnsafeRemoteURLError(
            f"Remote media URL resolves to blocked address(es): {', '.join(sorted(set(blocked_ips)))}"
        )

vllm_mlx.models.mllm._request_with_safe_redirects 

_request_with_safe_redirects(method: str, url: str, *, timeout: int, headers: dict[str, str], stream: bool = False, max_redirects: int = 5)

Issue a requests call while validating every redirect target.

Source code in vllm_mlx/models/mllm.py 
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
def _request_with_safe_redirects(
    method: str,
    url: str,
    *,
    timeout: int,
    headers: dict[str, str],
    stream: bool = False,
    max_redirects: int = 5,
):
    """Issue a requests call while validating every redirect target."""
    current_url = url
    for _ in range(max_redirects + 1):
        _validate_url_safety(current_url)
        response = requests.request(
            method,
            current_url,
            timeout=timeout,
            headers=headers,
            allow_redirects=False,
            verify=True,
            stream=stream,
        )
        if not response.is_redirect and not response.is_permanent_redirect:
            return response

        location = response.headers.get("location")
        response.close()
        if not location:
            raise UnsafeRemoteURLError(
                f"Remote media URL redirect missing Location header: {current_url}"
            )
        current_url = urljoin(current_url, location)

    raise UnsafeRemoteURLError(
        f"Remote media URL exceeded redirect limit ({max_redirects}): {url}"
    )

vllm_mlx.models.mllm.download_image 

download_image(url: str, timeout: int = 30, max_size: int = MAX_IMAGE_SIZE) -> str

Download image from URL and return local path.

Parameters:

  • url (str) –

    Image URL

  • timeout (int, default: 30 ) –

    Download timeout in seconds

  • max_size (int, default: MAX_IMAGE_SIZE ) –

    Maximum allowed file size in bytes

Returns:

  • str –

    Local file path to downloaded image

Raises:

Source code in vllm_mlx/models/mllm.py 
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
def download_image(url: str, timeout: int = 30, max_size: int = MAX_IMAGE_SIZE) -> str:
    """
    Download image from URL and return local path.

    Args:
        url: Image URL
        timeout: Download timeout in seconds
        max_size: Maximum allowed file size in bytes

    Returns:
        Local file path to downloaded image

    Raises:
        FileSizeExceededError: If image exceeds max_size
    """
    headers = {
        "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36"
    }

    # First, make a HEAD request to check Content-Length
    try:
        head_response = _request_with_safe_redirects(
            "HEAD", url, timeout=timeout, headers=headers
        )
        content_length = head_response.headers.get("content-length")
        if content_length and int(content_length) > max_size:
            raise FileSizeExceededError(
                f"Image at {url} exceeds maximum size: {int(content_length) / 1024 / 1024:.1f} MB > "
                f"{max_size / 1024 / 1024:.1f} MB limit"
            )
    except requests.RequestException:
        # HEAD request failed, proceed with GET and check during download
        pass

    response = _request_with_safe_redirects(
        "GET", url, timeout=timeout, headers=headers, stream=True
    )
    response.raise_for_status()

    # Check Content-Length header from GET response
    content_length = response.headers.get("content-length")
    if content_length and int(content_length) > max_size:
        raise FileSizeExceededError(
            f"Image at {url} exceeds maximum size: {int(content_length) / 1024 / 1024:.1f} MB > "
            f"{max_size / 1024 / 1024:.1f} MB limit"
        )

    # Determine extension from content type or URL
    content_type = response.headers.get("content-type", "")
    if "jpeg" in content_type or "jpg" in content_type:
        ext = ".jpg"
    elif "png" in content_type:
        ext = ".png"
    elif "gif" in content_type:
        ext = ".gif"
    elif "webp" in content_type:
        ext = ".webp"
    else:
        # Try to get from URL
        path = urlparse(response.url).path
        ext = Path(path).suffix or ".jpg"

    # Save to temp file with size checking during download
    temp_file = tempfile.NamedTemporaryFile(suffix=ext, delete=False)
    downloaded_size = 0
    try:
        for chunk in response.iter_content(chunk_size=8192):
            downloaded_size += len(chunk)
            if downloaded_size > max_size:
                temp_file.close()
                os.unlink(temp_file.name)
                raise FileSizeExceededError(
                    f"Image at {url} exceeds maximum size during download: "
                    f"{downloaded_size / 1024 / 1024:.1f} MB > {max_size / 1024 / 1024:.1f} MB limit"
                )
            temp_file.write(chunk)
        temp_file.close()
    except FileSizeExceededError:
        raise
    except Exception:
        temp_file.close()
        if os.path.exists(temp_file.name):
            os.unlink(temp_file.name)
        raise

    return _temp_manager.register(temp_file.name)

vllm_mlx.models.mllm._download_media 

_download_media(url: str, media_type: str, ext_map: dict[str, str], default_ext: str, timeout: int, max_size: int) -> str

Download media from URL, enforce size limits, and return a local temp path.

Source code in vllm_mlx/models/mllm.py 
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
def _download_media(
    url: str,
    media_type: str,
    ext_map: dict[str, str],
    default_ext: str,
    timeout: int,
    max_size: int,
) -> str:
    """Download media from URL, enforce size limits, and return a local temp path."""
    headers = {
        "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36"
    }

    logger.info(f"Downloading {media_type} from: {url}")

    try:
        head_response = _request_with_safe_redirects(
            "HEAD", url, timeout=timeout, headers=headers
        )
        content_length = head_response.headers.get("content-length")
        if content_length and int(content_length) > max_size:
            raise FileSizeExceededError(
                f"{media_type.capitalize()} at {url} exceeds maximum size: "
                f"{int(content_length) / 1024 / 1024:.1f} MB > {max_size / 1024 / 1024:.1f} MB limit"
            )
    except requests.RequestException:
        pass

    response = _request_with_safe_redirects(
        "GET", url, timeout=timeout, headers=headers, stream=True
    )
    response.raise_for_status()

    content_length = response.headers.get("content-length")
    if content_length and int(content_length) > max_size:
        raise FileSizeExceededError(
            f"{media_type.capitalize()} at {url} exceeds maximum size: "
            f"{int(content_length) / 1024 / 1024:.1f} MB > {max_size / 1024 / 1024:.1f} MB limit"
        )

    content_type = response.headers.get("content-type", "").lower()
    ext = default_ext
    for key, mapped_ext in ext_map.items():
        if key in content_type:
            ext = mapped_ext
            break
    else:
        path_ext = Path(urlparse(response.url).path).suffix
        if path_ext:
            ext = path_ext

    temp_file = tempfile.NamedTemporaryFile(suffix=ext, delete=False)
    downloaded_size = 0
    try:
        for chunk in response.iter_content(chunk_size=8192):
            downloaded_size += len(chunk)
            if downloaded_size > max_size:
                temp_file.close()
                os.unlink(temp_file.name)
                raise FileSizeExceededError(
                    f"{media_type.capitalize()} at {url} exceeds maximum size during download: "
                    f"{downloaded_size / 1024 / 1024:.1f} MB > {max_size / 1024 / 1024:.1f} MB limit"
                )
            temp_file.write(chunk)
        temp_file.close()
    except FileSizeExceededError:
        raise
    except Exception:
        temp_file.close()
        if os.path.exists(temp_file.name):
            os.unlink(temp_file.name)
        raise

    file_size = Path(temp_file.name).stat().st_size
    logger.info(
        f"{media_type.capitalize()} downloaded: {temp_file.name} ({file_size / 1024 / 1024:.1f} MB)"
    )

    return _temp_manager.register(temp_file.name)

vllm_mlx.models.mllm.download_video 

download_video(url: str, timeout: int = 120, max_size: int = MAX_VIDEO_SIZE) -> str

Download video from URL and return local path.

Source code in vllm_mlx/models/mllm.py 
751
752
753
def download_video(url: str, timeout: int = 120, max_size: int = MAX_VIDEO_SIZE) -> str:
    """Download video from URL and return local path."""
    return _download_media(url, "video", _VIDEO_EXT_MAP, ".mp4", timeout, max_size)

vllm_mlx.models.mllm.download_audio 

download_audio(url: str, timeout: int = 120, max_size: int = MAX_AUDIO_SIZE) -> str

Download audio from URL and return local path.

Source code in vllm_mlx/models/mllm.py 
756
757
758
def download_audio(url: str, timeout: int = 120, max_size: int = MAX_AUDIO_SIZE) -> str:
    """Download audio from URL and return local path."""
    return _download_media(url, "audio", _AUDIO_EXT_MAP, ".wav", timeout, max_size)

vllm_mlx.models.mllm.decode_base64_video 

decode_base64_video(base64_string: str, max_length: int = MAX_BASE64_VIDEO_LENGTH) -> str

Decode base64 video to temp file and return path.

Supports format: data:video/mp4;base64,AAAA...

Parameters:

  • base64_string (str) –

    Base64-encoded video with data URL prefix

  • max_length (int, default: MAX_BASE64_VIDEO_LENGTH ) –

    Maximum allowed length of base64 string

Returns:

  • str –

    Local file path to decoded video

Raises:

Source code in vllm_mlx/models/mllm.py 
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
def decode_base64_video(
    base64_string: str, max_length: int = MAX_BASE64_VIDEO_LENGTH
) -> str:
    """
    Decode base64 video to temp file and return path.

    Supports format: data:video/mp4;base64,AAAA...

    Args:
        base64_string: Base64-encoded video with data URL prefix
        max_length: Maximum allowed length of base64 string

    Returns:
        Local file path to decoded video

    Raises:
        FileSizeExceededError: If base64 string exceeds max_length
    """
    if len(base64_string) > max_length:
        raise FileSizeExceededError(
            f"Base64 video data exceeds maximum size: {len(base64_string) / 1024 / 1024:.1f} MB > "
            f"{max_length / 1024 / 1024:.1f} MB limit"
        )

    # Extract format and data
    if base64_string.startswith("data:video/"):
        # Format: data:video/mp4;base64,AAAA...
        header, data = base64_string.split(",", 1)
        # Extract extension from header (e.g., "data:video/mp4;base64" -> "mp4")
        format_part = header.split(";")[0]  # "data:video/mp4"
        ext = "." + format_part.split("/")[-1]  # ".mp4"
    else:
        # Assume mp4 if no header
        data = base64_string
        ext = ".mp4"

    # Decode, save, and register for cleanup
    video_bytes = base64.b64decode(data)
    temp_file = tempfile.NamedTemporaryFile(suffix=ext, delete=False)
    temp_file.write(video_bytes)
    temp_file.close()

    logger.info(
        f"Base64 video decoded: {temp_file.name} ({len(video_bytes) / 1024 / 1024:.1f} MB)"
    )

    return _temp_manager.register(temp_file.name)

vllm_mlx.models.mllm.decode_base64_audio 

decode_base64_audio(base64_string: str, max_length: int = MAX_BASE64_AUDIO_LENGTH) -> str

Decode base64 audio to temp file and return path.

Supports format: data:audio/wav;base64,AAAA...

Source code in vllm_mlx/models/mllm.py 
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
def decode_base64_audio(
    base64_string: str, max_length: int = MAX_BASE64_AUDIO_LENGTH
) -> str:
    """
    Decode base64 audio to temp file and return path.

    Supports format: data:audio/wav;base64,AAAA...
    """
    if len(base64_string) > max_length:
        raise FileSizeExceededError(
            f"Base64 audio data exceeds maximum size: {len(base64_string) / 1024 / 1024:.1f} MB > "
            f"{max_length / 1024 / 1024:.1f} MB limit"
        )

    if base64_string.startswith("data:audio/"):
        header, data = base64_string.split(",", 1)
        format_part = header.split(";")[0]
        ext = "." + format_part.split("/")[-1]
    else:
        data = base64_string
        ext = ".wav"

    audio_bytes = base64.b64decode(data)
    temp_file = tempfile.NamedTemporaryFile(suffix=ext, delete=False)
    temp_file.write(audio_bytes)
    temp_file.close()
    return _temp_manager.register(temp_file.name)

vllm_mlx.models.mllm.process_video_input 

process_video_input(video: str | dict) -> str

Process video input in various formats and return local path.

Supports: - URL (http/https) - Base64 encoded string (data:video/mp4;base64,...) - OpenAI format dict: {"url": "..."} or {"url": "data:video/...;base64,..."}

Parameters:

  • video (str | dict) –

    Video input in any supported format

Returns:

  • str –

    Local file path to video

Source code in vllm_mlx/models/mllm.py 
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
def process_video_input(video: str | dict) -> str:
    """
    Process video input in various formats and return local path.

    Supports:
    - URL (http/https)
    - Base64 encoded string (data:video/mp4;base64,...)
    - OpenAI format dict: {"url": "..."} or {"url": "data:video/...;base64,..."}

    Args:
        video: Video input in any supported format

    Returns:
        Local file path to video
    """
    # Handle dict format (OpenAI style)
    if isinstance(video, dict):
        url = video.get("url", video.get("video_url", ""))
        if isinstance(url, dict):
            url = url.get("url", "")
        video = url

    if not video:
        raise ValueError("Empty video input")

    # Check if it's a URL
    if is_url(video):
        return download_video(video)

    # Check if it's base64
    if is_base64_video(video):
        return decode_base64_video(video)

    raise ValueError(
        "Unsupported video input. Only http(s) URLs and data:video base64 payloads are allowed."
    )

vllm_mlx.models.mllm.process_audio_input 

process_audio_input(audio: str | dict) -> str

Process audio input in various formats and return local path.

Supports: - Local file path - URL (http/https) - Base64 encoded string (data:audio/wav;base64,...) - OpenAI format dict: {"url": "..."} or {"audio_url": {"url": "..."}}

Source code in vllm_mlx/models/mllm.py 
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
def process_audio_input(audio: str | dict) -> str:
    """
    Process audio input in various formats and return local path.

    Supports:
    - Local file path
    - URL (http/https)
    - Base64 encoded string (data:audio/wav;base64,...)
    - OpenAI format dict: {"url": "..."} or {"audio_url": {"url": "..."}}
    """
    if isinstance(audio, dict):
        url = audio.get("url", audio.get("audio_url", ""))
        if isinstance(url, dict):
            url = url.get("url", "")
        audio = url

    if not audio:
        raise ValueError("Empty audio input")

    if is_base64_audio(audio):
        return decode_base64_audio(audio)

    if is_url(audio):
        return download_audio(audio)

    if len(audio) < 4096 and Path(audio).exists():
        return audio

    raise ValueError(f"Cannot process audio: {audio[:50]}...")

vllm_mlx.models.mllm._video_has_audio_track 

_video_has_audio_track(video_path: str) -> bool

Return True if ffprobe finds an audio stream in the video.

Source code in vllm_mlx/models/mllm.py 
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
def _video_has_audio_track(video_path: str) -> bool:
    """Return True if ffprobe finds an audio stream in the video."""
    import shutil
    import subprocess

    if not shutil.which("ffprobe"):
        return True  # assume yes; extraction will fail loudly if not
    try:
        r = subprocess.run(
            [
                "ffprobe",
                "-loglevel",
                "error",
                "-select_streams",
                "a",
                "-show_entries",
                "stream=codec_type",
                "-of",
                "csv=p=0",
                video_path,
            ],
            capture_output=True,
            timeout=30,
            text=True,
        )
        return bool(r.stdout.strip())
    except (subprocess.SubprocessError, OSError):
        return True

vllm_mlx.models.mllm._model_has_sound_encoder 

_model_has_sound_encoder(model) -> bool

Whether a loaded model exposes a usable sound encoder.

Uses getattr(..., None) is not None rather than hasattr so model wrappers that declare sound_encoder in __init__ but leave it as None until the first encoder pass are correctly treated as not yet enabled. A bare hasattr check would spuriously enable A/V fusion against a missing encoder and crash the processor downstream.

Source code in vllm_mlx/models/mllm.py 
938
939
940
941
942
943
944
945
946
947
def _model_has_sound_encoder(model) -> bool:
    """Whether a loaded model exposes a usable sound encoder.

    Uses ``getattr(..., None) is not None`` rather than ``hasattr`` so model
    wrappers that declare ``sound_encoder`` in ``__init__`` but leave it as
    ``None`` until the first encoder pass are correctly treated as not yet
    enabled. A bare ``hasattr`` check would spuriously enable A/V fusion
    against a missing encoder and crash the processor downstream.
    """
    return getattr(model, "sound_encoder", None) is not None

vllm_mlx.models.mllm.extract_audio_from_video 

extract_audio_from_video(video_path: str) -> str | None

Extract the audio track from a video file as 16 kHz mono WAV.

Returns the path to the WAV (registered with the temp manager so it's cleaned up automatically), or None if the video has no audio or ffmpeg is unavailable.

Source code in vllm_mlx/models/mllm.py 
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
def extract_audio_from_video(video_path: str) -> str | None:
    """Extract the audio track from a video file as 16 kHz mono WAV.

    Returns the path to the WAV (registered with the temp manager so it's
    cleaned up automatically), or None if the video has no audio or ffmpeg
    is unavailable.
    """
    import os
    import shutil
    import subprocess

    if not shutil.which("ffmpeg"):
        logger.warning(
            "ffmpeg not found; cannot fuse audio from video_url. "
            "Install ffmpeg to enable A/V fusion on omni models."
        )
        return None
    if not _video_has_audio_track(video_path):
        return None

    fd, out_path = tempfile.mkstemp(suffix=".wav", prefix="vllmmlx_va_")
    os.close(fd)
    try:
        r = subprocess.run(
            [
                "ffmpeg",
                "-y",
                "-i",
                video_path,
                "-vn",
                "-ac",
                "1",
                "-ar",
                "16000",
                "-c:a",
                "pcm_s16le",
                out_path,
            ],
            stdout=subprocess.DEVNULL,
            stderr=subprocess.DEVNULL,
            timeout=600,
        )
        if r.returncode != 0 or os.path.getsize(out_path) == 0:
            try:
                os.unlink(out_path)
            except OSError:
                pass
            return None
        return _temp_manager.register(out_path)
    except (subprocess.SubprocessError, OSError) as e:
        logger.warning(f"Audio extraction from video failed: {e}")
        try:
            os.unlink(out_path)
        except OSError:
            pass
        return None

vllm_mlx.models.mllm.save_base64_image 

save_base64_image(base64_string: str) -> str

Save base64 image to temp file and return path. Caches identical images.

Source code in vllm_mlx/models/mllm.py 
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
def save_base64_image(base64_string: str) -> str:
    """Save base64 image to temp file and return path. Caches identical images."""
    import hashlib

    # Hash the FULL base64 string — not just a prefix.
    # Using only the first 1000 chars caused cache collisions between
    # different images with identical JPEG headers (e.g. invoices from
    # the same PDF renderer), returning a previous request's image.
    image_hash = hashlib.sha256(base64_string.encode()).hexdigest()

    # Return cached path if available and file still exists
    if image_hash in _base64_image_cache:
        cached_path = _base64_image_cache[image_hash]
        if Path(cached_path).exists():
            return cached_path

    image_bytes = decode_base64_image(base64_string)

    # Detect format from magic bytes
    if image_bytes[:8] == b"\x89PNG\r\n\x1a\n":
        ext = ".png"
    elif image_bytes[:2] == b"\xff\xd8":
        ext = ".jpg"
    elif image_bytes[:6] in (b"GIF87a", b"GIF89a"):
        ext = ".gif"
    elif image_bytes[:4] == b"RIFF" and image_bytes[8:12] == b"WEBP":
        ext = ".webp"
    else:
        ext = ".jpg"  # Default

    temp_file = tempfile.NamedTemporaryFile(suffix=ext, delete=False)
    temp_file.write(image_bytes)
    temp_file.close()

    path = _temp_manager.register(temp_file.name)
    _base64_image_cache[image_hash] = path
    return path

vllm_mlx.models.mllm.process_image_input 

process_image_input(image: str | dict) -> str

Process image input in various formats and return local path.

Supports: - URL (http/https) - Base64 encoded string - OpenAI format dict: {"url": "..."} or {"url": "data:image/...;base64,..."}

Source code in vllm_mlx/models/mllm.py 
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
def process_image_input(image: str | dict) -> str:
    """
    Process image input in various formats and return local path.

    Supports:
    - URL (http/https)
    - Base64 encoded string
    - OpenAI format dict: {"url": "..."} or {"url": "data:image/...;base64,..."}
    """
    # Handle dict format (OpenAI style)
    if isinstance(image, dict):
        url = image.get("url", image.get("image_url", ""))
        if isinstance(url, dict):
            url = url.get("url", "")
        image = url

    if not image:
        raise ValueError("Empty image input")

    # Check if it's base64 FIRST (before Path.exists() which fails on long strings)
    if is_base64_image(image):
        return save_base64_image(image)

    # Check if it's a URL
    if is_url(image):
        return download_image(image)

    raise ValueError(
        "Unsupported image input. Only http(s) URLs and data:image base64 payloads are allowed."
    )

vllm_mlx.models.mllm.round_by_factor 

round_by_factor(x: int, factor: int) -> int

Round to nearest multiple of factor.

Source code in vllm_mlx/models/mllm.py 
1083
1084
1085
def round_by_factor(x: int, factor: int) -> int:
    """Round to nearest multiple of factor."""
    return round(x / factor) * factor

vllm_mlx.models.mllm.ceil_by_factor 

ceil_by_factor(x: float, factor: int) -> int

Ceiling to next multiple of factor.

Source code in vllm_mlx/models/mllm.py 
1088
1089
1090
def ceil_by_factor(x: float, factor: int) -> int:
    """Ceiling to next multiple of factor."""
    return math.ceil(x / factor) * factor

vllm_mlx.models.mllm.floor_by_factor 

floor_by_factor(x: float, factor: int) -> int

Floor to previous multiple of factor.

Source code in vllm_mlx/models/mllm.py 
1093
1094
1095
def floor_by_factor(x: float, factor: int) -> int:
    """Floor to previous multiple of factor."""
    return math.floor(x / factor) * factor

vllm_mlx.models.mllm.smart_nframes 

smart_nframes(total_frames: int, video_fps: float, target_fps: float = DEFAULT_FPS, min_frames: int = MIN_FRAMES, max_frames: int = MAX_FRAMES) -> int

Calculate optimal number of frames to extract from video.

Uses smart sampling based on video length and target FPS.

Source code in vllm_mlx/models/mllm.py 
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
def smart_nframes(
    total_frames: int,
    video_fps: float,
    target_fps: float = DEFAULT_FPS,
    min_frames: int = MIN_FRAMES,
    max_frames: int = MAX_FRAMES,
) -> int:
    """
    Calculate optimal number of frames to extract from video.

    Uses smart sampling based on video length and target FPS.
    """
    # Calculate duration-based frame count
    duration = total_frames / video_fps if video_fps > 0 else 0
    nframes = duration * target_fps

    # Clamp to min/max
    nframes = max(min_frames, min(nframes, max_frames, total_frames))

    # Round to factor
    nframes = max(FRAME_FACTOR, floor_by_factor(nframes, FRAME_FACTOR))

    return int(nframes)

vllm_mlx.models.mllm.extract_video_frames_smart 

extract_video_frames_smart(video_path: str, fps: float = DEFAULT_FPS, max_frames: int = MAX_FRAMES, resize: tuple[int, int] | None = None) -> list[ndarray]

Extract frames from video with smart sampling.

Parameters:

  • video_path (str) –

    Path to video file

  • fps (float, default: DEFAULT_FPS ) –

    Target frames per second (default: 2.0)

  • max_frames (int, default: MAX_FRAMES ) –

    Maximum frames to extract

  • resize (tuple[int, int] | None, default: None ) –

    Optional (width, height) to resize frames

Returns:

  • list[ndarray] –

    List of frame arrays (RGB format)

Source code in vllm_mlx/models/mllm.py 
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
def extract_video_frames_smart(
    video_path: str,
    fps: float = DEFAULT_FPS,
    max_frames: int = MAX_FRAMES,
    resize: tuple[int, int] | None = None,
) -> list[np.ndarray]:
    """
    Extract frames from video with smart sampling.

    Args:
        video_path: Path to video file
        fps: Target frames per second (default: 2.0)
        max_frames: Maximum frames to extract
        resize: Optional (width, height) to resize frames

    Returns:
        List of frame arrays (RGB format)
    """
    try:
        import cv2
    except ImportError:
        raise ImportError("opencv-python is required for video processing")

    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError(f"Cannot open video: {video_path}")

    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    video_fps = cap.get(cv2.CAP_PROP_FPS) or 30.0

    # Calculate number of frames to extract
    nframes = smart_nframes(
        total_frames=total_frames,
        video_fps=video_fps,
        target_fps=fps,
        max_frames=max_frames,
    )

    # Calculate frame indices (evenly spaced)
    indices = np.linspace(0, total_frames - 1, nframes).round().astype(int)

    logger.info(
        f"Video: {total_frames} total frames @ {video_fps:.1f} fps, "
        f"extracting {nframes} frames"
    )

    frames = []
    for idx in indices:
        cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
        ret, frame = cap.read()
        if not ret:
            continue

        # Convert BGR to RGB
        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)

        # Resize if specified
        if resize:
            frame = cv2.resize(frame, resize)

        frames.append(frame)

    cap.release()

    return frames

vllm_mlx.models.mllm.save_frames_to_temp 

save_frames_to_temp(frames: list[ndarray]) -> list[str]

Save frame arrays to temporary files and return paths.

Source code in vllm_mlx/models/mllm.py 
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
def save_frames_to_temp(frames: list[np.ndarray]) -> list[str]:
    """Save frame arrays to temporary files and return paths."""
    try:
        from PIL import Image
    except ImportError:
        raise ImportError("Pillow is required for frame processing")

    paths = []
    for i, frame in enumerate(frames):
        img = Image.fromarray(frame)
        temp_file = tempfile.NamedTemporaryFile(suffix=".jpg", delete=False)
        img.save(temp_file.name, "JPEG", quality=85)
        paths.append(_temp_manager.register(temp_file.name))

    return paths

Complete contract reference

Expand any definition for its exact inputs, annotations, defaults, return contract, directly raised exceptions, source-grounded behavior, and immutable line link. This section includes private and nested definitions that ordinary API generators omit.

vllm_mlx.models.mllm.TempFileManager · class 
vllm_mlx.models.mllm.TempFileManager()

Thread-safe manager for tracking and cleaning up temporary files.

Parameters

This callable has no explicit inputs.

Returns

  • Constructs: vllm_mlx.models.mllm.TempFileManager

Exceptions and behavior

Class TempFileManager declares 4 direct member(s). No direct raise statement appears in this definition.

View source #L41-L86.

vllm_mlx.models.mllm.TempFileManager.__init__ · method 
vllm_mlx.models.mllm.TempFileManager.__init__() -> not annotated

Method TempFileManager.__init__ updates self._files, self._lock; calls set, threading.Lock, atexit.register.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated

Exceptions and behavior

Method TempFileManager.__init__ updates self._files, self._lock; calls set, threading.Lock, atexit.register. No direct raise statement appears in this definition.

View source #L44-L47.

vllm_mlx.models.mllm.TempFileManager.register · method 
vllm_mlx.models.mllm.TempFileManager.register(path: str) -> str

Register a temp file for tracking.

Parameters

Name Type Required Default Description
path str yes none Required positional or keyword input.

Returns

  • Type: str
  • Direct return expressions: path

Exceptions and behavior

Method TempFileManager.register calls self._files.add; returns path. No direct raise statement appears in this definition.

View source #L49-L53.

vllm_mlx.models.mllm.TempFileManager.cleanup · method 
vllm_mlx.models.mllm.TempFileManager.cleanup(path: str) -> bool

Clean up a specific temp file.

Parameters

Name Type Required Default Description
path str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: True; False

Exceptions and behavior

Method TempFileManager.cleanup calls self._files.discard, os.path.exists, os.unlink, logger.debug; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L55-L67.

vllm_mlx.models.mllm.TempFileManager.cleanup_all · method 
vllm_mlx.models.mllm.TempFileManager.cleanup_all() -> int

Clean up all tracked temp files.

Parameters

This callable has no explicit inputs.

Returns

  • Type: int
  • Direct return expressions: cleaned

Exceptions and behavior

Method TempFileManager.cleanup_all calls list, self._files.clear, os.path.exists, os.unlink; returns cleaned. No direct raise statement appears in this definition.

View source #L69-L86.

vllm_mlx.models.mllm.cleanup_temp_file · function 
vllm_mlx.models.mllm.cleanup_temp_file(path: str) -> bool

Clean up a specific temporary file.

Parameters

Name Type Required Default Description
path str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: _temp_manager.cleanup(path)

Exceptions and behavior

Function cleanup_temp_file calls _temp_manager.cleanup; returns _temp_manager.cleanup(path). No direct raise statement appears in this definition.

View source #L93-L95.

vllm_mlx.models.mllm.cleanup_all_temp_files · function 
vllm_mlx.models.mllm.cleanup_all_temp_files() -> int

Clean up all tracked temporary files.

Parameters

This callable has no explicit inputs.

Returns

  • Type: int
  • Direct return expressions: _temp_manager.cleanup_all()

Exceptions and behavior

Function cleanup_all_temp_files calls _temp_manager.cleanup_all; returns _temp_manager.cleanup_all(). No direct raise statement appears in this definition.

View source #L98-L100.

vllm_mlx.models.mllm.FileSizeExceededError · class 
vllm_mlx.models.mllm.FileSizeExceededError()

Raised when a downloaded file exceeds the size limit.

Parameters

This callable has no explicit inputs.

Returns

  • Constructs: vllm_mlx.models.mllm.FileSizeExceededError

Exceptions and behavior

Class FileSizeExceededError derives from Exception and declares 0 direct member(s). No direct raise statement appears in this definition.

View source #L119-L122.

vllm_mlx.models.mllm.UnsafeRemoteURLError · class 
vllm_mlx.models.mllm.UnsafeRemoteURLError(message: str, *, public_message: str = 'Remote media URL is not allowed')

Raised when a remote media URL targets an unsafe destination.

Parameters

Name Type Required Default Description
message str yes none Required positional or keyword input.
public_message str no 'Remote media URL is not allowed' Optional keyword-only input; defaults to 'Remote media URL is not allowed'.

Returns

  • Constructs: vllm_mlx.models.mllm.UnsafeRemoteURLError

Exceptions and behavior

Class UnsafeRemoteURLError derives from ValueError and declares 1 direct member(s). No direct raise statement appears in this definition.

View source #L125-L135.

vllm_mlx.models.mllm.UnsafeRemoteURLError.__init__ · method 
vllm_mlx.models.mllm.UnsafeRemoteURLError.__init__(message: str, *, public_message: str = 'Remote media URL is not allowed') -> None

Method UnsafeRemoteURLError.__init__ updates self.public_message; calls super().__init__, super.

Parameters

Name Type Required Default Description
message str yes none Required positional or keyword input.
public_message str no 'Remote media URL is not allowed' Optional keyword-only input; defaults to 'Remote media URL is not allowed'.

Returns

  • Type: None

Exceptions and behavior

Method UnsafeRemoteURLError.__init__ updates self.public_message; calls super().__init__, super. No direct raise statement appears in this definition.

View source #L128-L135.

vllm_mlx.models.mllm._normalize_content_part · function 
vllm_mlx.models.mllm._normalize_content_part(item: object) -> object

Convert Pydantic content parts into plain Python objects.

Parameters

Name Type Required Default Description
item object yes none Required positional or keyword input.

Returns

  • Type: object
  • Direct return expressions: item.model_dump(exclude_none=True); {k: v for k, v in item.dict().items() if v is not None}; item

Exceptions and behavior

Function _normalize_content_part calls hasattr, item.model_dump, item.dict().items, item.dict; has 3 explicit return paths. No direct raise statement appears in this definition.

View source #L138-L144.

vllm_mlx.models.mllm._extract_media_url · function 
vllm_mlx.models.mllm._extract_media_url(item: dict, item_type: str) -> str

Function _extract_media_url calls item.get, isinstance, media_value.get; has 2 explicit return paths.

Parameters

Name Type Required Default Description
item dict yes none Required positional or keyword input.
item_type str yes none Required positional or keyword input.

Returns

  • Type: str
  • Direct return expressions: ''; media_value if isinstance(media_value, str) else ''

Exceptions and behavior

Function _extract_media_url calls item.get, isinstance, media_value.get; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L147-L161.

vllm_mlx.models.mllm._text_content_part · function 
vllm_mlx.models.mllm._text_content_part(text: str) -> dict[str, str]

Function _text_content_part returns {'type': 'text', 'text': text, 'content': text}.

Parameters

Name Type Required Default Description
text str yes none Required positional or keyword input.

Returns

  • Type: dict[str, str]
  • Direct return expressions: {'type': 'text', 'text': text, 'content': text}

Exceptions and behavior

Function _text_content_part returns {'type': 'text', 'text': text, 'content': text}. No direct raise statement appears in this definition.

View source #L164-L165.

vllm_mlx.models.mllm._append_text_content_part · function 
vllm_mlx.models.mllm._append_text_content_part(built_parts: list[dict[str, str]], text_parts: list[str], text: str) -> None

Function _append_text_content_part calls built_parts.append, _text_content_part, text_parts.append; returns None.

Parameters

Name Type Required Default Description
built_parts list[dict[str, str]] yes none Required positional or keyword input.
text_parts list[str] yes none Required positional or keyword input.
text str yes none Required positional or keyword input.

Returns

  • Type: None
  • Direct return expressions: None

Exceptions and behavior

Function _append_text_content_part calls built_parts.append, _text_content_part, text_parts.append; returns None. No direct raise statement appears in this definition.

View source #L168-L174.

vllm_mlx.models.mllm._build_string_mllm_message_content · function 
vllm_mlx.models.mllm._build_string_mllm_message_content(content: str, role: str) -> tuple[object, bool]

Function _build_string_mllm_message_content calls _text_content_part; has 3 explicit return paths.

Parameters

Name Type Required Default Description
content str yes none Required positional or keyword input.
role str yes none Required positional or keyword input.

Returns

  • Type: tuple[object, bool]
  • Direct return expressions: ('', False); (content, True); ([_text_content_part(content)], True)

Exceptions and behavior

Function _build_string_mllm_message_content calls _text_content_part; has 3 explicit return paths. No direct raise statement appears in this definition.

View source #L177-L182.

vllm_mlx.models.mllm._append_ordered_mllm_content_part · function 
vllm_mlx.models.mllm._append_ordered_mllm_content_part(raw_item: object, *, built_parts: list[dict[str, str]], text_parts: list[str], all_image_urls: list[str], video_frame_count: int) -> int

Function _append_ordered_mllm_content_part calls _normalize_content_part, isinstance, _append_text_content_part, item.get; has 2 explicit return paths.

Parameters

Name Type Required Default Description
raw_item object yes none Required positional or keyword input.
built_parts list[dict[str, str]] yes none Required keyword-only input.
text_parts list[str] yes none Required keyword-only input.
all_image_urls list[str] yes none Required keyword-only input.
video_frame_count int yes none Required keyword-only input.

Returns

  • Type: int
  • Direct return expressions: video_frame_count; 0

Exceptions and behavior

Function _append_ordered_mllm_content_part calls _normalize_content_part, isinstance, _append_text_content_part, item.get; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L185-L220.

vllm_mlx.models.mllm._build_ordered_mllm_message_content · function 
vllm_mlx.models.mllm._build_ordered_mllm_message_content(content: object, *, role: str, all_image_urls: list[str], video_frame_count: int = 0) -> tuple[object, bool]

Build template content while preserving OpenAI media/text part order.

Parameters

Name Type Required Default Description
content object yes none Required positional or keyword input.
role str yes none Required keyword-only input.
all_image_urls list[str] yes none Required keyword-only input.
video_frame_count int no 0 Optional keyword-only input; defaults to 0.

Returns

  • Type: tuple[object, bool]
  • Direct return expressions: _build_string_mllm_message_content(content, role); ('', False); (text, bool(text)); (built_parts, bool(built_parts))

Exceptions and behavior

Function _build_ordered_mllm_message_content calls isinstance, _build_string_mllm_message_content, _append_ordered_mllm_content_part, ''.join; has 4 explicit return paths. No direct raise statement appears in this definition.

View source #L223-L254.

vllm_mlx.models.mllm._normalize_mllm_tool_calls · function 
vllm_mlx.models.mllm._normalize_mllm_tool_calls(tool_calls: list) -> list

Normalize replayed assistant tool calls for chat templates.

Parameters

Name Type Required Default Description
tool_calls list yes none Required positional or keyword input.

Returns

  • Type: list
  • Direct return expressions: normalized[0].get('tool_calls', plain_calls)

Exceptions and behavior

Function _normalize_mllm_tool_calls calls _normalize_content_part, normalize_messages_for_chat_template, normalized[0].get; returns normalized[0].get('tool_calls', plain_calls). No direct raise statement appears in this definition.

View source #L257-L268.

vllm_mlx.models.mllm._build_mllm_chat_messages · function 
vllm_mlx.models.mllm._build_mllm_chat_messages(messages: list[dict], *, all_image_urls: list[str], video_frame_counts: dict[int, int]) -> list[dict]

Build chat-template messages without reordering multimodal content parts.

Parameters

Name Type Required Default Description
messages list[dict] yes none Required positional or keyword input.
all_image_urls list[str] yes none Required keyword-only input.
video_frame_counts dict[int, int] yes none Required keyword-only input.

Returns

  • Type: list[dict]
  • Direct return expressions: chat_messages

Exceptions and behavior

Function _build_mllm_chat_messages calls enumerate, msg.get, isinstance, str; returns chat_messages. No direct raise statement appears in this definition.

View source #L271-L315.

vllm_mlx.models.mllm.MultimodalInput · class 
vllm_mlx.models.mllm.MultimodalInput(prompt: str, images: list[str] = field(default_factory=list), videos: list[str] = field(default_factory=list), audio: list[str] = field(default_factory=list))

Input for multimodal generation.

Parameters

Name Type Required Default Description
prompt str yes none Required constructor field.
images list[str] no field(default_factory=list) Optional constructor field; defaults to field(default_factory=list).
videos list[str] no field(default_factory=list) Optional constructor field; defaults to field(default_factory=list).
audio list[str] no field(default_factory=list) Optional constructor field; defaults to field(default_factory=list).

Returns

  • Constructs: vllm_mlx.models.mllm.MultimodalInput

Exceptions and behavior

Class MultimodalInput declares 0 direct member(s). No direct raise statement appears in this definition.

View source #L319-L325.

vllm_mlx.models.mllm.MLLMOutput · class 
vllm_mlx.models.mllm.MLLMOutput(text: str, finish_reason: str | None = None, prompt_tokens: int = 0, completion_tokens: int = 0, mtp_drafts: int = 0, mtp_accepted: int = 0)

Output from multimodal language model.

Parameters

Name Type Required Default Description
text str yes none Required constructor field.
finish_reason str \| None no None Optional constructor field; defaults to None.
prompt_tokens int no 0 Optional constructor field; defaults to 0.
completion_tokens int no 0 Optional constructor field; defaults to 0.
mtp_drafts int no 0 Optional constructor field; defaults to 0.
mtp_accepted int no 0 Optional constructor field; defaults to 0.

Returns

  • Constructs: vllm_mlx.models.mllm.MLLMOutput

Exceptions and behavior

Class MLLMOutput declares 0 direct member(s). No direct raise statement appears in this definition.

View source #L329-L337.

vllm_mlx.models.mllm.load_gemma4_assistant_drafter · function 
vllm_mlx.models.mllm.load_gemma4_assistant_drafter(model_path: str) -> not annotated

Load a Gemma 4 assistant drafter for mlx-vlm speculative decoding.

Parameters

Name Type Required Default Description
model_path str yes none Required positional or keyword input.

Returns

  • Type: not annotated
  • Direct return expressions: model

Exceptions and behavior

Function load_gemma4_assistant_drafter calls ImportError, version, logger.info, Path; can raise ImportError, FileNotFoundError; returns model. Directly raised exceptions: ImportError, FileNotFoundError.

View source #L340-L381.

vllm_mlx.models.mllm._count_draft_tokens · function 
vllm_mlx.models.mllm._count_draft_tokens(draft_tokens) -> int

Best-effort drafted-token count for an mlx-vlm drafter output.

Parameters

Name Type Required Default Description
draft_tokens not annotated yes none Required positional or keyword input.

Returns

  • Type: int
  • Direct return expressions: max(int(shape[-1]), 0); max(len(draft_tokens), 0); 0

Exceptions and behavior

Function _count_draft_tokens calls getattr, max, int, len; has 3 explicit return paths. No direct raise statement appears in this definition.

View source #L387-L398.

vllm_mlx.models.mllm._install_draft_metrics_hooks · function 
vllm_mlx.models.mllm._install_draft_metrics_hooks(draft_model) -> None

Record actual drafted token counts from mlx-vlm assistant drafters.

Parameters

Name Type Required Default Description
draft_model not annotated yes none Required positional or keyword input.

Returns

  • Type: None
  • Direct return expressions: None

Exceptions and behavior

Function _install_draft_metrics_hooks calls getattr, hasattr, callable; returns None. No direct raise statement appears in this definition.

View source #L401-L428.

vllm_mlx.models.mllm._install_draft_metrics_hooks.draft_block_with_metrics · nested function 
vllm_mlx.models.mllm._install_draft_metrics_hooks.draft_block_with_metrics(*args, **kwargs) -> not annotated

Nested Function _install_draft_metrics_hooks.draft_block_with_metrics calls draft_block, draft_model._vllm_mlx_draft_counts.append, _count_draft_tokens; returns draft_tokens.

Parameters

Name Type Required Default Description
*args not annotated no none Additional variadic positional inputs accepted by this callable.
**kwargs not annotated no none Additional variadic keyword inputs accepted by this callable.

Returns

  • Type: not annotated
  • Direct return expressions: draft_tokens

Exceptions and behavior

Nested Function _install_draft_metrics_hooks.draft_block_with_metrics calls draft_block, draft_model._vllm_mlx_draft_counts.append, _count_draft_tokens; returns draft_tokens. No direct raise statement appears in this definition.

View source #L412-L415.

vllm_mlx.models.mllm._install_draft_metrics_hooks.reset_with_metrics · nested function 
vllm_mlx.models.mllm._install_draft_metrics_hooks.reset_with_metrics(*args, **kwargs) -> not annotated

Nested Function _install_draft_metrics_hooks.reset_with_metrics calls reset; returns reset(*args, **kwargs).

Parameters

Name Type Required Default Description
*args not annotated no none Additional variadic positional inputs accepted by this callable.
**kwargs not annotated no none Additional variadic keyword inputs accepted by this callable.

Returns

  • Type: not annotated
  • Direct return expressions: reset(*args, **kwargs)

Exceptions and behavior

Nested Function _install_draft_metrics_hooks.reset_with_metrics calls reset; returns reset(*args, **kwargs). No direct raise statement appears in this definition.

View source #L422-L424.

vllm_mlx.models.mllm.is_base64_image · function 
vllm_mlx.models.mllm.is_base64_image(s: str) -> bool

Check if string is base64-encoded image data.

Parameters

Name Type Required Default Description
s str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: s.startswith('data:image/') or (len(s) > 100 and (not s.startswith(('http://', 'https://', '/'))))

Exceptions and behavior

Function is_base64_image calls s.startswith, len; returns s.startswith('data:image/') or (len(s) > 100 and (not s.startswith(('http://', 'https://', '/')))). No direct raise statement appears in this definition.

View source #L431-L435.

vllm_mlx.models.mllm.is_url · function 
vllm_mlx.models.mllm.is_url(s: str) -> bool

Check if string is a URL.

Parameters

Name Type Required Default Description
s str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: s.startswith(('http://', 'https://'))

Exceptions and behavior

Function is_url calls s.startswith; returns s.startswith(('http://', 'https://')). No direct raise statement appears in this definition.

View source #L438-L440.

vllm_mlx.models.mllm.is_base64_video · function 
vllm_mlx.models.mllm.is_base64_video(s: str) -> bool

Check if string is base64-encoded video data.

Parameters

Name Type Required Default Description
s str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: s.startswith('data:video/')

Exceptions and behavior

Function is_base64_video calls s.startswith; returns s.startswith('data:video/'). No direct raise statement appears in this definition.

View source #L443-L445.

vllm_mlx.models.mllm.is_base64_audio · function 
vllm_mlx.models.mllm.is_base64_audio(s: str) -> bool

Check if string is base64-encoded audio data.

Parameters

Name Type Required Default Description
s str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: s.startswith('data:audio/')

Exceptions and behavior

Function is_base64_audio calls s.startswith; returns s.startswith('data:audio/'). No direct raise statement appears in this definition.

View source #L448-L450.

vllm_mlx.models.mllm.decode_base64_image · function 
vllm_mlx.models.mllm.decode_base64_image(base64_string: str, max_length: int = MAX_BASE64_IMAGE_LENGTH) -> bytes

Decode base64 image to bytes.

Parameters

Name Type Required Default Description
base64_string str yes none Base64 encoded image (optionally with data URL prefix)
max_length int no MAX_BASE64_IMAGE_LENGTH Maximum allowed length of base64 string

Returns

  • Type: bytes
  • Direct return expressions: base64.b64decode(data); base64.b64decode(base64_string)

Exceptions and behavior

Function decode_base64_image calls len, FileSizeExceededError, base64_string.startswith, base64_string.split; can raise FileSizeExceededError; has 2 explicit return paths. Directly raised exceptions: FileSizeExceededError.

View source #L453-L480.

vllm_mlx.models.mllm._validate_url_safety · function 
vllm_mlx.models.mllm._validate_url_safety(url: str) -> None

Reject remote URLs that target local or private network resources.

Parameters

Name Type Required Default Description
url str yes none Required positional or keyword input.

Returns

  • Type: None

Exceptions and behavior

Function _validate_url_safety calls urlparse, UnsafeRemoteURLError, hostname.endswith, ipaddress.ip_address; can raise UnsafeRemoteURLError. Directly raised exceptions: UnsafeRemoteURLError.

View source #L483-L519.

vllm_mlx.models.mllm._request_with_safe_redirects · function 
vllm_mlx.models.mllm._request_with_safe_redirects(method: str, url: str, *, timeout: int, headers: dict[str, str], stream: bool = False, max_redirects: int = 5) -> not annotated

Issue a requests call while validating every redirect target.

Parameters

Name Type Required Default Description
method str yes none Required positional or keyword input.
url str yes none Required positional or keyword input.
timeout int yes none Required keyword-only input.
headers dict[str, str] yes none Required keyword-only input.
stream bool no False Optional keyword-only input; defaults to False.
max_redirects int no 5 Optional keyword-only input; defaults to 5.

Returns

  • Type: not annotated
  • Direct return expressions: response

Exceptions and behavior

Function _request_with_safe_redirects calls range, _validate_url_safety, requests.request, response.headers.get; can raise UnsafeRemoteURLError; returns response. Directly raised exceptions: UnsafeRemoteURLError.

View source #L522-L557.

vllm_mlx.models.mllm.download_image · function 
vllm_mlx.models.mllm.download_image(url: str, timeout: int = 30, max_size: int = MAX_IMAGE_SIZE) -> str

Download image from URL and return local path.

Parameters

Name Type Required Default Description
url str yes none Image URL
timeout int no 30 Download timeout in seconds
max_size int no MAX_IMAGE_SIZE Maximum allowed file size in bytes

Returns

  • Type: str
  • Direct return expressions: _temp_manager.register(temp_file.name)

Exceptions and behavior

Function download_image calls _request_with_safe_redirects, head_response.headers.get, int, FileSizeExceededError; can raise FileSizeExceededError; returns _temp_manager.register(temp_file.name). Directly raised exceptions: FileSizeExceededError.

View source #L560-L645.

vllm_mlx.models.mllm._download_media · function 
vllm_mlx.models.mllm._download_media(url: str, media_type: str, ext_map: dict[str, str], default_ext: str, timeout: int, max_size: int) -> str

Download media from URL, enforce size limits, and return a local temp path.

Parameters

Name Type Required Default Description
url str yes none Required positional or keyword input.
media_type str yes none Required positional or keyword input.
ext_map dict[str, str] yes none Required positional or keyword input.
default_ext str yes none Required positional or keyword input.
timeout int yes none Required positional or keyword input.
max_size int yes none Required positional or keyword input.

Returns

  • Type: str
  • Direct return expressions: _temp_manager.register(temp_file.name)

Exceptions and behavior

Function _download_media calls logger.info, _request_with_safe_redirects, head_response.headers.get, int; can raise FileSizeExceededError; returns _temp_manager.register(temp_file.name). Directly raised exceptions: FileSizeExceededError.

View source #L670-L748.

vllm_mlx.models.mllm.download_video · function 
vllm_mlx.models.mllm.download_video(url: str, timeout: int = 120, max_size: int = MAX_VIDEO_SIZE) -> str

Download video from URL and return local path.

Parameters

Name Type Required Default Description
url str yes none Required positional or keyword input.
timeout int no 120 Optional positional or keyword input; defaults to 120.
max_size int no MAX_VIDEO_SIZE Optional positional or keyword input; defaults to MAX_VIDEO_SIZE.

Returns

  • Type: str
  • Direct return expressions: _download_media(url, 'video', _VIDEO_EXT_MAP, '.mp4', timeout, max_size)

Exceptions and behavior

Function download_video calls _download_media; returns _download_media(url, 'video', _VIDEO_EXT_MAP, '.mp4', timeout, max_size). No direct raise statement appears in this definition.

View source #L751-L753.

vllm_mlx.models.mllm.download_audio · function 
vllm_mlx.models.mllm.download_audio(url: str, timeout: int = 120, max_size: int = MAX_AUDIO_SIZE) -> str

Download audio from URL and return local path.

Parameters

Name Type Required Default Description
url str yes none Required positional or keyword input.
timeout int no 120 Optional positional or keyword input; defaults to 120.
max_size int no MAX_AUDIO_SIZE Optional positional or keyword input; defaults to MAX_AUDIO_SIZE.

Returns

  • Type: str
  • Direct return expressions: _download_media(url, 'audio', _AUDIO_EXT_MAP, '.wav', timeout, max_size)

Exceptions and behavior

Function download_audio calls _download_media; returns _download_media(url, 'audio', _AUDIO_EXT_MAP, '.wav', timeout, max_size). No direct raise statement appears in this definition.

View source #L756-L758.

vllm_mlx.models.mllm.decode_base64_video · function 
vllm_mlx.models.mllm.decode_base64_video(base64_string: str, max_length: int = MAX_BASE64_VIDEO_LENGTH) -> str

Decode base64 video to temp file and return path.

Parameters

Name Type Required Default Description
base64_string str yes none Base64-encoded video with data URL prefix
max_length int no MAX_BASE64_VIDEO_LENGTH Maximum allowed length of base64 string

Returns

  • Type: str
  • Direct return expressions: _temp_manager.register(temp_file.name)

Exceptions and behavior

Function decode_base64_video calls len, FileSizeExceededError, base64_string.startswith, base64_string.split; can raise FileSizeExceededError; returns _temp_manager.register(temp_file.name). Directly raised exceptions: FileSizeExceededError.

View source #L761-L807.

vllm_mlx.models.mllm.decode_base64_audio · function 
vllm_mlx.models.mllm.decode_base64_audio(base64_string: str, max_length: int = MAX_BASE64_AUDIO_LENGTH) -> str

Decode base64 audio to temp file and return path.

Parameters

Name Type Required Default Description
base64_string str yes none Required positional or keyword input.
max_length int no MAX_BASE64_AUDIO_LENGTH Optional positional or keyword input; defaults to MAX_BASE64_AUDIO_LENGTH.

Returns

  • Type: str
  • Direct return expressions: _temp_manager.register(temp_file.name)

Exceptions and behavior

Function decode_base64_audio calls len, FileSizeExceededError, base64_string.startswith, base64_string.split; can raise FileSizeExceededError; returns _temp_manager.register(temp_file.name). Directly raised exceptions: FileSizeExceededError.

View source #L810-L836.

vllm_mlx.models.mllm.process_video_input · function 
vllm_mlx.models.mllm.process_video_input(video: str | dict) -> str

Process video input in various formats and return local path.

Parameters

Name Type Required Default Description
video str \| dict yes none Video input in any supported format

Returns

  • Type: str
  • Direct return expressions: download_video(video); decode_base64_video(video)

Exceptions and behavior

Function process_video_input calls isinstance, video.get, url.get, ValueError; can raise ValueError; has 2 explicit return paths. Directly raised exceptions: ValueError.

View source #L839-L874.

vllm_mlx.models.mllm.process_audio_input · function 
vllm_mlx.models.mllm.process_audio_input(audio: str | dict) -> str

Process audio input in various formats and return local path.

Parameters

Name Type Required Default Description
audio str \| dict yes none Required positional or keyword input.

Returns

  • Type: str
  • Direct return expressions: decode_base64_audio(audio); download_audio(audio); audio

Exceptions and behavior

Function process_audio_input calls isinstance, audio.get, url.get, ValueError; can raise ValueError; has 3 explicit return paths. Directly raised exceptions: ValueError.

View source #L877-L905.

vllm_mlx.models.mllm._video_has_audio_track · function 
vllm_mlx.models.mllm._video_has_audio_track(video_path: str) -> bool

Return True if ffprobe finds an audio stream in the video.

Parameters

Name Type Required Default Description
video_path str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: True; bool(r.stdout.strip())

Exceptions and behavior

Function _video_has_audio_track calls shutil.which, subprocess.run, bool, r.stdout.strip; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L908-L935.

vllm_mlx.models.mllm._model_has_sound_encoder · function 
vllm_mlx.models.mllm._model_has_sound_encoder(model) -> bool

Whether a loaded model exposes a usable sound encoder.

Parameters

Name Type Required Default Description
model not annotated yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: getattr(model, 'sound_encoder', None) is not None

Exceptions and behavior

Function _model_has_sound_encoder calls getattr; returns getattr(model, 'sound_encoder', None) is not None. No direct raise statement appears in this definition.

View source #L938-L947.

vllm_mlx.models.mllm.extract_audio_from_video · function 
vllm_mlx.models.mllm.extract_audio_from_video(video_path: str) -> str | None

Extract the audio track from a video file as 16 kHz mono WAV.

Parameters

Name Type Required Default Description
video_path str yes none Required positional or keyword input.

Returns

  • Type: str | None
  • Direct return expressions: None; _temp_manager.register(out_path)

Exceptions and behavior

Function extract_audio_from_video calls shutil.which, logger.warning, _video_has_audio_track, tempfile.mkstemp; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L950-L1005.

vllm_mlx.models.mllm.save_base64_image · function 
vllm_mlx.models.mllm.save_base64_image(base64_string: str) -> str

Save base64 image to temp file and return path.

Parameters

Name Type Required Default Description
base64_string str yes none Required positional or keyword input.

Returns

  • Type: str
  • Direct return expressions: cached_path; path

Exceptions and behavior

Function save_base64_image calls hashlib.sha256(base64_string.encode()).hexdigest, hashlib.sha256, base64_string.encode, Path(cached_path).exists; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L1012-L1048.

vllm_mlx.models.mllm.process_image_input · function 
vllm_mlx.models.mllm.process_image_input(image: str | dict) -> str

Process image input in various formats and return local path.

Parameters

Name Type Required Default Description
image str \| dict yes none Required positional or keyword input.

Returns

  • Type: str
  • Direct return expressions: save_base64_image(image); download_image(image)

Exceptions and behavior

Function process_image_input calls isinstance, image.get, url.get, ValueError; can raise ValueError; has 2 explicit return paths. Directly raised exceptions: ValueError.

View source #L1051-L1080.

vllm_mlx.models.mllm.round_by_factor · function 
vllm_mlx.models.mllm.round_by_factor(x: int, factor: int) -> int

Round to nearest multiple of factor.

Parameters

Name Type Required Default Description
x int yes none Required positional or keyword input.
factor int yes none Required positional or keyword input.

Returns

  • Type: int
  • Direct return expressions: round(x / factor) * factor

Exceptions and behavior

Function round_by_factor calls round; returns round(x / factor) * factor. No direct raise statement appears in this definition.

View source #L1083-L1085.

vllm_mlx.models.mllm.ceil_by_factor · function 
vllm_mlx.models.mllm.ceil_by_factor(x: float, factor: int) -> int

Ceiling to next multiple of factor.

Parameters

Name Type Required Default Description
x float yes none Required positional or keyword input.
factor int yes none Required positional or keyword input.

Returns

  • Type: int
  • Direct return expressions: math.ceil(x / factor) * factor

Exceptions and behavior

Function ceil_by_factor calls math.ceil; returns math.ceil(x / factor) * factor. No direct raise statement appears in this definition.

View source #L1088-L1090.

vllm_mlx.models.mllm.floor_by_factor · function 
vllm_mlx.models.mllm.floor_by_factor(x: float, factor: int) -> int

Floor to previous multiple of factor.

Parameters

Name Type Required Default Description
x float yes none Required positional or keyword input.
factor int yes none Required positional or keyword input.

Returns

  • Type: int
  • Direct return expressions: math.floor(x / factor) * factor

Exceptions and behavior

Function floor_by_factor calls math.floor; returns math.floor(x / factor) * factor. No direct raise statement appears in this definition.

View source #L1093-L1095.

vllm_mlx.models.mllm.smart_nframes · function 
vllm_mlx.models.mllm.smart_nframes(total_frames: int, video_fps: float, target_fps: float = DEFAULT_FPS, min_frames: int = MIN_FRAMES, max_frames: int = MAX_FRAMES) -> int

Calculate optimal number of frames to extract from video.

Parameters

Name Type Required Default Description
total_frames int yes none Required positional or keyword input.
video_fps float yes none Required positional or keyword input.
target_fps float no DEFAULT_FPS Optional positional or keyword input; defaults to DEFAULT_FPS.
min_frames int no MIN_FRAMES Optional positional or keyword input; defaults to MIN_FRAMES.
max_frames int no MAX_FRAMES Optional positional or keyword input; defaults to MAX_FRAMES.

Returns

  • Type: int
  • Direct return expressions: int(nframes)

Exceptions and behavior

Function smart_nframes calls max, min, floor_by_factor, int; returns int(nframes). No direct raise statement appears in this definition.

View source #L1098-L1120.

vllm_mlx.models.mllm.extract_video_frames_smart · function 
vllm_mlx.models.mllm.extract_video_frames_smart(video_path: str, fps: float = DEFAULT_FPS, max_frames: int = MAX_FRAMES, resize: tuple[int, int] | None = None) -> list[np.ndarray]

Extract frames from video with smart sampling.

Parameters

Name Type Required Default Description
video_path str yes none Path to video file
fps float no DEFAULT_FPS Target frames per second (default: 2.0)
max_frames int no MAX_FRAMES Maximum frames to extract
resize tuple[int, int] \| None no None Optional (width, height) to resize frames

Returns

  • Type: list[np.ndarray]
  • Direct return expressions: frames

Exceptions and behavior

Function extract_video_frames_smart calls ImportError, cv2.VideoCapture, cap.isOpened, ValueError; can raise ImportError, ValueError; returns frames. Directly raised exceptions: ImportError, ValueError.

View source #L1123-L1187.

vllm_mlx.models.mllm.save_frames_to_temp · function 
vllm_mlx.models.mllm.save_frames_to_temp(frames: list[np.ndarray]) -> list[str]

Save frame arrays to temporary files and return paths.

Parameters

Name Type Required Default Description
frames list[np.ndarray] yes none Required positional or keyword input.

Returns

  • Type: list[str]
  • Direct return expressions: paths

Exceptions and behavior

Function save_frames_to_temp calls ImportError, enumerate, Image.fromarray, tempfile.NamedTemporaryFile; can raise ImportError; returns paths. Directly raised exceptions: ImportError.

View source #L1190-L1204.

vllm_mlx.models.mllm.MLXMultimodalLM · class 
vllm_mlx.models.mllm.MLXMultimodalLM(model_name: str, trust_remote_code: bool = False, enable_cache: bool = True, cache_size: int = 50, max_kv_size: int = 0, draft_model: str | None = None, draft_kind: str | None = None, draft_block_size: int | None = None)

Wrapper around mlx-vlm for multimodal inference.

Parameters

Name Type Required Default Description
model_name str yes none HuggingFace model name or local path
trust_remote_code bool no False Whether to trust remote code
enable_cache bool no True Enable KV cache for repeated image/video+prompt (default: True)
cache_size int no 50 Maximum cache entries (default: 50)
max_kv_size int no 0 Maximum KV cache size per sequence (0 = unbounded)
draft_model str \| None no None Optional MLLM speculative draft/assistant model path.
draft_kind str \| None no None Optional mlx-vlm draft kind, for example "mtp".
draft_block_size int \| None no None Optional speculative block size passed to mlx-vlm.

Returns

  • Constructs: vllm_mlx.models.mllm.MLXMultimodalLM

Exceptions and behavior

Class MLXMultimodalLM declares 29 direct member(s). No direct raise statement appears in this definition.

View source #L1207-L2938.

vllm_mlx.models.mllm.MLXMultimodalLM.__init__ · method 
vllm_mlx.models.mllm.MLXMultimodalLM.__init__(model_name: str, trust_remote_code: bool = False, enable_cache: bool = True, cache_size: int = 50, max_kv_size: int = 0, draft_model: str | None = None, draft_kind: str | None = None, draft_block_size: int | None = None) -> not annotated

Initialize the MLX multimodal language model.

Parameters

Name Type Required Default Description
model_name str yes none HuggingFace model name or local path
trust_remote_code bool no False Whether to trust remote code
enable_cache bool no True Enable KV cache for repeated image/video+prompt (default: True)
cache_size int no 50 Maximum cache entries (default: 50)
max_kv_size int no 0 Maximum KV cache size per sequence (0 = unbounded)
draft_model str \| None no None Optional MLLM speculative draft/assistant model path.
draft_kind str \| None no None Optional mlx-vlm draft kind, for example "mtp".
draft_block_size int \| None no None Optional speculative block size passed to mlx-vlm.

Returns

  • Type: not annotated

Exceptions and behavior

Method MLXMultimodalLM.__init__ updates self.model_name, self.trust_remote_code, self.enable_cache, self.max_kv_size; calls MLLMPrefixCacheManager. No direct raise statement appears in this definition.

View source #L1235-L1278.

vllm_mlx.models.mllm.MLXMultimodalLM.load · method 
vllm_mlx.models.mllm.MLXMultimodalLM.load() -> None

Load the model and processor.

Parameters

This callable has no explicit inputs.

Returns

  • Type: None
  • Direct return expressions: None

Exceptions and behavior

Method MLXMultimodalLM.load updates self.model, self.processor, self.config, self._draft_model; calls logger.info, load, load_config, self._load_draft_model; can raise ImportError; returns None. Directly raised exceptions: ImportError.

View source #L1280-L1323.

vllm_mlx.models.mllm.MLXMultimodalLM._load_draft_model · method 
vllm_mlx.models.mllm.MLXMultimodalLM._load_draft_model() -> not annotated

Method MLXMultimodalLM._load_draft_model calls load_gemma4_assistant_drafter, load; has 2 explicit return paths.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated
  • Direct return expressions: load_gemma4_assistant_drafter(self.draft_model_path); draft_model

Exceptions and behavior

Method MLXMultimodalLM._load_draft_model calls load_gemma4_assistant_drafter, load; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L1325-L1332.

vllm_mlx.models.mllm.MLXMultimodalLM._draft_generation_kwargs · method 
vllm_mlx.models.mllm.MLXMultimodalLM._draft_generation_kwargs(call_kwargs: dict | None = None) -> dict

Return mlx-vlm drafter kwargs when the request explicitly opts in.

Parameters

Name Type Required Default Description
call_kwargs dict \| None no None Optional positional or keyword input; defaults to None.

Returns

  • Type: dict
  • Direct return expressions: {}; kwargs

Exceptions and behavior

Method MLXMultimodalLM._draft_generation_kwargs calls bool, call_kwargs.pop, _install_draft_metrics_hooks; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L1334-L1355.

vllm_mlx.models.mllm.MLXMultimodalLM._reset_draft_metrics · method 
vllm_mlx.models.mllm.MLXMultimodalLM._reset_draft_metrics() -> int

Method MLXMultimodalLM._reset_draft_metrics updates self._draft_model.accept_lens, self._draft_model._vllm_mlx_draft_counts; calls hasattr; returns 0.

Parameters

This callable has no explicit inputs.

Returns

  • Type: int
  • Direct return expressions: 0

Exceptions and behavior

Method MLXMultimodalLM._reset_draft_metrics updates self._draft_model.accept_lens, self._draft_model._vllm_mlx_draft_counts; calls hasattr; returns 0. No direct raise statement appears in this definition.

View source #L1357-L1364.

vllm_mlx.models.mllm.MLXMultimodalLM._draft_metrics_since · method 
vllm_mlx.models.mllm.MLXMultimodalLM._draft_metrics_since(start_accept_lens: int) -> dict[str, int]

Method MLXMultimodalLM._draft_metrics_since calls list, getattr, len, int; has 2 explicit return paths.

Parameters

Name Type Required Default Description
start_accept_lens int yes none Required positional or keyword input.

Returns

  • Type: dict[str, int]
  • Direct return expressions: {'mtp_drafts': 0, 'mtp_accepted': 0}; {'mtp_drafts': mtp_drafts, 'mtp_accepted': sum((int(value) for value in new_accept_lens))}

Exceptions and behavior

Method MLXMultimodalLM._draft_metrics_since calls list, getattr, len, int; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L1366-L1395.

vllm_mlx.models.mllm.MLXMultimodalLM.get_language_model · method 
vllm_mlx.models.mllm.MLXMultimodalLM.get_language_model() -> not annotated

Extract the underlying language model for mlx_lm TextModel construction.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated
  • Direct return expressions: self.model.language_model

Exceptions and behavior

Method MLXMultimodalLM.get_language_model returns self.model.language_model. No direct raise statement appears in this definition.

View source #L1397-L1399.

vllm_mlx.models.mllm.MLXMultimodalLM.get_tokenizer · method 
vllm_mlx.models.mllm.MLXMultimodalLM.get_tokenizer() -> not annotated

Get the text tokenizer (not the multimodal processor).

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated
  • Direct return expressions: self.processor.tokenizer

Exceptions and behavior

Method MLXMultimodalLM.get_tokenizer returns self.processor.tokenizer. No direct raise statement appears in this definition.

View source #L1401-L1403.

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_images · method 
vllm_mlx.models.mllm.MLXMultimodalLM._prepare_images(images: list) -> list[str]

Process remote/base64 image inputs into local temp file paths.

Parameters

Name Type Required Default Description
images list yes none Required positional or keyword input.

Returns

  • Type: list[str]
  • Direct return expressions: processed

Exceptions and behavior

Method MLXMultimodalLM._prepare_images calls process_image_input, processed.append, logger.warning; returns processed. No direct raise statement appears in this definition.

View source #L1405-L1414.

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_audio · method 
vllm_mlx.models.mllm.MLXMultimodalLM._prepare_audio(audio_inputs: list) -> list[str]

Process audio inputs and return local file paths.

Parameters

Name Type Required Default Description
audio_inputs list yes none Required positional or keyword input.

Returns

  • Type: list[str]
  • Direct return expressions: processed

Exceptions and behavior

Method MLXMultimodalLM._prepare_audio calls process_audio_input, processed.append, logger.warning; returns processed. No direct raise statement appears in this definition.

View source #L1416-L1425.

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_video · method 
vllm_mlx.models.mllm.MLXMultimodalLM._prepare_video(video_input: str | dict, fps: float = DEFAULT_FPS, max_frames: int = MAX_FRAMES, resolved_path: str | None = None) -> list[str]

Process video input and extract frames.

Parameters

Name Type Required Default Description
video_input str \| dict yes none Video in any supported format
fps float no DEFAULT_FPS Frames per second to extract
max_frames int no MAX_FRAMES Maximum frames to extract
resolved_path str \| None no None Optional pre-resolved local path. Callers that already ran process_video_input (e.g. for parallel audio extraction) pass it here to avoid re-downloading / re-decoding.

Returns

  • Type: list[str]
  • Direct return expressions: save_frames_to_temp(frames)

Exceptions and behavior

Method MLXMultimodalLM._prepare_video calls process_video_input, extract_video_frames_smart, save_frames_to_temp; returns save_frames_to_temp(frames). No direct raise statement appears in this definition.

View source #L1427-L1463.

vllm_mlx.models.mllm.MLXMultimodalLM._collect_video_inputs · method 
vllm_mlx.models.mllm.MLXMultimodalLM._collect_video_inputs(messages: list[dict]) -> dict[int, list]

Collect video inputs from messages, keyed by message index.

Parameters

Name Type Required Default Description
messages list[dict] yes none Required positional or keyword input.

Returns

  • Type: dict[int, list]
  • Direct return expressions: video_inputs

Exceptions and behavior

Method MLXMultimodalLM._collect_video_inputs calls enumerate, msg.get, isinstance, hasattr; returns video_inputs. No direct raise statement appears in this definition.

View source #L1465-L1497.

vllm_mlx.models.mllm.MLXMultimodalLM._collect_audio_inputs · method 
vllm_mlx.models.mllm.MLXMultimodalLM._collect_audio_inputs(messages: list[dict]) -> dict[int, list]

Collect audio inputs from messages, keyed by message index.

Parameters

Name Type Required Default Description
messages list[dict] yes none Required positional or keyword input.

Returns

  • Type: dict[int, list]
  • Direct return expressions: audio_inputs

Exceptions and behavior

Method MLXMultimodalLM._collect_audio_inputs calls enumerate, msg.get, isinstance, hasattr; returns audio_inputs. No direct raise statement appears in this definition.

View source #L1499-L1528.

vllm_mlx.models.mllm.MLXMultimodalLM._prepare_native_video_inputs · method 
vllm_mlx.models.mllm.MLXMultimodalLM._prepare_native_video_inputs(messages: list[dict], video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, tools: list | None = None) -> tuple[str, dict]

Preprocess messages into prompt + generation kwargs for native video.

Parameters

Name Type Required Default Description
messages list[dict] yes none Required positional or keyword input.
video_fps float no DEFAULT_FPS Optional positional or keyword input; defaults to DEFAULT_FPS.
video_max_frames int no MAX_FRAMES Optional positional or keyword input; defaults to MAX_FRAMES.
tools list \| None no None Optional positional or keyword input; defaults to None.

Returns

  • Type: tuple[str, dict]
  • Direct return expressions: (text, gen_kwargs)

Exceptions and behavior

Method MLXMultimodalLM._prepare_native_video_inputs calls ImportError, self._translate_messages_for_native_video, self.processor.apply_chat_template, process_vision_info; can raise ImportError; returns (text, gen_kwargs). Directly raised exceptions: ImportError.

View source #L1530-L1648.

vllm_mlx.models.mllm.MLXMultimodalLM._generate_native_video · method 
vllm_mlx.models.mllm.MLXMultimodalLM._generate_native_video(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, tools: list | None = None, **kwargs) -> MLLMOutput

Generate using native video pipeline (Qwen-family models).

Parameters

Name Type Required Default Description
messages list[dict] yes none Required positional or keyword input.
max_tokens int no 256 Optional positional or keyword input; defaults to 256.
temperature float no 0.7 Optional positional or keyword input; defaults to 0.7.
video_fps float no DEFAULT_FPS Optional positional or keyword input; defaults to DEFAULT_FPS.
video_max_frames int no MAX_FRAMES Optional positional or keyword input; defaults to MAX_FRAMES.
tools list \| None no None Optional positional or keyword input; defaults to None.
**kwargs not annotated no none Additional variadic keyword inputs accepted by this callable.

Returns

  • Type: MLLMOutput
  • Direct return expressions: MLLMOutput(text=result.text, finish_reason='stop', prompt_tokens=getattr(result, 'prompt_tokens', 0), completion_tokens…; MLLMOutput(text=str(result), finish_reason='stop')

Exceptions and behavior

Method MLXMultimodalLM._generate_native_video calls ImportError, self._prepare_native_video_inputs, generate, hasattr; can raise ImportError; has 2 explicit return paths. Directly raised exceptions: ImportError.

View source #L1650-L1695.

vllm_mlx.models.mllm.MLXMultimodalLM._translate_messages_for_native_video · method 
vllm_mlx.models.mllm.MLXMultimodalLM._translate_messages_for_native_video(messages: list[dict], video_fps: float, video_max_frames: int) -> list[dict]

Translate OpenAI API format messages to process_vision_info format.

Parameters

Name Type Required Default Description
messages list[dict] yes none Required positional or keyword input.
video_fps float yes none Required positional or keyword input.
video_max_frames int yes none Required positional or keyword input.

Returns

  • Type: list[dict]
  • Direct return expressions: translated

Exceptions and behavior

Method MLXMultimodalLM._translate_messages_for_native_video calls msg.get, isinstance, translated.append, str; returns translated. No direct raise statement appears in this definition.

View source #L1697-L1832.

vllm_mlx.models.mllm.MLXMultimodalLM.generate · method 
vllm_mlx.models.mllm.MLXMultimodalLM.generate(prompt: str, images: list | None = None, videos: list | None = None, audio: list[str] | None = None, max_tokens: int = 256, temperature: float = 0.7, top_p: float = 0.9, video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, use_cache: bool = True, **kwargs) -> MLLMOutput

Generate text from multimodal input.

Parameters

Name Type Required Default Description
prompt str yes none Text prompt/question
images list \| None no None List of image URLs or base64 strings
videos list \| None no None List of video inputs (URLs, base64, or OpenAI format dicts)
audio list[str] \| None no None List of audio file paths
max_tokens int no 256 Maximum tokens to generate
temperature float no 0.7 Sampling temperature
top_p float no 0.9 Top-p sampling parameter
video_fps float no DEFAULT_FPS FPS for video frame extraction (default: 2.0)
video_max_frames int no MAX_FRAMES Max frames to extract from video
use_cache bool no True Whether to use KV cache (default: True)
**kwargs not annotated no none Additional generation parameters

Returns

  • Type: MLLMOutput
  • Direct return expressions: MLLMOutput(text=output_text, finish_reason='stop', prompt_tokens=prompt_tokens, completion_tokens=generation_tokens, **…

Exceptions and behavior

Method MLXMultimodalLM.generate calls self.load, all_images.extend, self._prepare_images, all_sources.extend; returns MLLMOutput(text=output_text, finish_reason='stop', prompt_tokens=prompt_tokens, completion_tokens=generation_tokens, **…. No direct raise statement appears in this definition.

View source #L1834-L2002.

vllm_mlx.models.mllm.MLXMultimodalLM.stream_generate · method 
vllm_mlx.models.mllm.MLXMultimodalLM.stream_generate(prompt: str, images: list | None = None, videos: list[str] | None = None, audio: list[str] | None = None, max_tokens: int = 256, temperature: float = 0.7, video_fps: float = DEFAULT_FPS, **kwargs) -> Iterator[str]

Stream text generation for multimodal input.

Parameters

Name Type Required Default Description
prompt str yes none Text prompt
images list \| None no None List of image inputs
videos list[str] \| None no None List of video paths
audio list[str] \| None no None List of audio inputs
max_tokens int no 256 Maximum tokens to generate
temperature float no 0.7 Sampling temperature
video_fps float no DEFAULT_FPS FPS for video frame extraction
**kwargs not annotated no none Additional parameters

Returns

  • Type: Iterator[str]
  • Direct return expressions: None
  • Yields values incrementally.

Exceptions and behavior

Method MLXMultimodalLM.stream_generate calls self.load, self.generate, all_images.extend, self._prepare_images; yields values incrementally; returns None. No direct raise statement appears in this definition.

View source #L2004-L2093.

vllm_mlx.models.mllm.MLXMultimodalLM.chat · method 
vllm_mlx.models.mllm.MLXMultimodalLM.chat(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, **kwargs) -> MLLMOutput

Chat with OpenAI-compatible message format.

Parameters

Name Type Required Default Description
messages list[dict] yes none List of chat messages (OpenAI format)
max_tokens int no 256 Maximum tokens to generate
temperature float no 0.7 Sampling temperature
**kwargs not annotated no none Additional parameters

Returns

  • Type: MLLMOutput
  • Direct return expressions: self._generate_native_video(messages=messages, max_tokens=max_tokens, temperature=temperature, video_fps=video_fps, vid…; MLLMOutput(text=output_text, finish_reason='stop', prompt_tokens=prompt_tokens, completion_tokens=generation_tokens, **…

Exceptions and behavior

Method MLXMultimodalLM.chat calls self.load, logger.info, len, kwargs.pop; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L2095-L2487.

vllm_mlx.models.mllm.MLXMultimodalLM.stream_chat · method 
vllm_mlx.models.mllm.MLXMultimodalLM.stream_chat(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, **kwargs) -> Iterator[MLLMOutput]

Stream chat with OpenAI-compatible message format.

Parameters

Name Type Required Default Description
messages list[dict] yes none List of chat messages (OpenAI format)
max_tokens int no 256 Maximum tokens to generate
temperature float no 0.7 Sampling temperature
**kwargs not annotated no none Additional parameters

Returns

  • Type: Iterator[MLLMOutput]
  • Direct return expressions: None
  • Yields values incrementally.

Exceptions and behavior

Method MLXMultimodalLM.stream_chat calls self.load, self.chat, kwargs.pop, chat_template_kwargs.pop; yields values incrementally; returns None. No direct raise statement appears in this definition.

View source #L2489-L2737.

vllm_mlx.models.mllm.MLXMultimodalLM.describe_image · method 
vllm_mlx.models.mllm.MLXMultimodalLM.describe_image(image: str, prompt: str = 'Describe this image in detail.', max_tokens: int = 512, **kwargs) -> str

Convenience method to describe an image.

Parameters

Name Type Required Default Description
image str yes none Image path, URL, or base64 string
prompt str no 'Describe this image in detail.' Description prompt
max_tokens int no 512 Maximum tokens
**kwargs not annotated no none Additional parameters

Returns

  • Type: str
  • Direct return expressions: output.text

Exceptions and behavior

Method MLXMultimodalLM.describe_image calls self.generate; returns output.text. No direct raise statement appears in this definition.

View source #L2739-L2764.

vllm_mlx.models.mllm.MLXMultimodalLM.answer_about_image · method 
vllm_mlx.models.mllm.MLXMultimodalLM.answer_about_image(image: str, question: str, max_tokens: int = 256, **kwargs) -> str

Answer a question about an image.

Parameters

Name Type Required Default Description
image str yes none Image path, URL, or base64 string
question str yes none Question about the image
max_tokens int no 256 Maximum tokens
**kwargs not annotated no none Additional parameters

Returns

  • Type: str
  • Direct return expressions: output.text

Exceptions and behavior

Method MLXMultimodalLM.answer_about_image calls self.generate; returns output.text. No direct raise statement appears in this definition.

View source #L2766-L2791.

vllm_mlx.models.mllm.MLXMultimodalLM.describe_video · method 
vllm_mlx.models.mllm.MLXMultimodalLM.describe_video(video: str | dict, prompt: str = 'Describe what happens in this video.', fps: float = 2.0, max_frames: int = 32, max_tokens: int = 512, **kwargs) -> str

Describe a video using frame extraction.

Parameters

Name Type Required Default Description
video str \| dict yes none Video file path, URL, base64, or OpenAI format dict
prompt str no 'Describe what happens in this video.' Description prompt
fps float no 2.0 Frames per second to extract
max_frames int no 32 Maximum frames to extract
max_tokens int no 512 Maximum tokens to generate
**kwargs not annotated no none Additional variadic keyword inputs accepted by this callable.

Returns

  • Type: str
  • Direct return expressions: output.text

Exceptions and behavior

Method MLXMultimodalLM.describe_video calls self.generate; returns output.text. No direct raise statement appears in this definition.

View source #L2793-L2830.

vllm_mlx.models.mllm.MLXMultimodalLM.get_cache_stats · method 
vllm_mlx.models.mllm.MLXMultimodalLM.get_cache_stats() -> dict

Get MLLM cache statistics.

Parameters

This callable has no explicit inputs.

Returns

  • Type: dict
  • Direct return expressions: {'enabled': False}; stats

Exceptions and behavior

Method MLXMultimodalLM.get_cache_stats calls self._cache_manager.get_stats, len; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L2832-L2846.

vllm_mlx.models.mllm.MLXMultimodalLM.clear_cache · method 
vllm_mlx.models.mllm.MLXMultimodalLM.clear_cache() -> None

Clear the MLLM KV cache.

Parameters

This callable has no explicit inputs.

Returns

  • Type: None

Exceptions and behavior

Method MLXMultimodalLM.clear_cache calls self._cache_manager.clear, logger.info. No direct raise statement appears in this definition.

View source #L2848-L2852.

vllm_mlx.models.mllm.MLXMultimodalLM.get_model_info · method 
vllm_mlx.models.mllm.MLXMultimodalLM.get_model_info() -> dict

Get information about the loaded model.

Parameters

This callable has no explicit inputs.

Returns

  • Type: dict
  • Direct return expressions: {'loaded': False, 'model_name': self.model_name}; info

Exceptions and behavior

Method MLXMultimodalLM.get_model_info calls getattr, self._cache_manager.get_stats; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L2854-L2874.

vllm_mlx.models.mllm.MLXMultimodalLM.list_supported_model_families · method 
vllm_mlx.models.mllm.MLXMultimodalLM.list_supported_model_families() -> dict[str, str]

List supported model families and their patterns.

Parameters

This callable has no explicit inputs.

Returns

  • Type: dict[str, str]
  • Direct return expressions: {'Qwen-VL': 'Qwen VL models (Qwen2-VL, Qwen2.5-VL, Qwen3-VL, etc.)', 'LLaVA': 'LLaVA vision-language models', 'Idefics'…

Exceptions and behavior

Method MLXMultimodalLM.list_supported_model_families returns {'Qwen-VL': 'Qwen VL models (Qwen2-VL, Qwen2.5-VL, Qwen3-VL, etc.)', 'LLaVA': 'LLaVA vision-language models', 'Idefics'…. No direct raise statement appears in this definition.

View source #L2877-L2897.

vllm_mlx.models.mllm.MLXMultimodalLM.is_mllm_model · method 
vllm_mlx.models.mllm.MLXMultimodalLM.is_mllm_model(model_name: str) -> bool

Check if a model name indicates an MLLM model.

Parameters

Name Type Required Default Description
model_name str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: any((pattern.lower() in model_lower for pattern in mllm_patterns))

Exceptions and behavior

Method MLXMultimodalLM.is_mllm_model calls model_name.lower, any, pattern.lower; returns any((pattern.lower() in model_lower for pattern in mllm_patterns)). No direct raise statement appears in this definition.

View source #L2900-L2934.

vllm_mlx.models.mllm.MLXMultimodalLM.__repr__ · method 
vllm_mlx.models.mllm.MLXMultimodalLM.__repr__() -> str

Method MLXMultimodalLM.__repr__ returns f'<MLXMultimodalLM model={self.model_name} status={status}>'.

Parameters

This callable has no explicit inputs.

Returns

  • Type: str
  • Direct return expressions: f'<MLXMultimodalLM model={self.model_name} status={status}>'

Exceptions and behavior

Method MLXMultimodalLM.__repr__ returns f'<MLXMultimodalLM model={self.model_name} status={status}>'. No direct raise statement appears in this definition.

View source #L2936-L2938.

Complete symbol map

This map also includes private definitions and nested helpers. The signature column exposes every explicit input even when an internal helper has no dedicated parameter prose.

Symbol Kind Signature and inputs What it does Source
TempFileManager class TempFileManager() Thread-safe manager for tracking and cleaning up temporary files. #L41-L86
TempFileManager.__init__ method TempFileManager.__init__() -> not annotated Method TempFileManager.__init__ updates self._files, self._lock; calls set, threading.Lock, atexit.register. #L44-L47
TempFileManager.register method TempFileManager.register(path: str) -> str Register a temp file for tracking. #L49-L53
TempFileManager.cleanup method TempFileManager.cleanup(path: str) -> bool Clean up a specific temp file. #L55-L67
TempFileManager.cleanup_all method TempFileManager.cleanup_all() -> int Clean up all tracked temp files. #L69-L86
cleanup_temp_file function cleanup_temp_file(path: str) -> bool Clean up a specific temporary file. #L93-L95
cleanup_all_temp_files function cleanup_all_temp_files() -> int Clean up all tracked temporary files. #L98-L100
FileSizeExceededError class FileSizeExceededError() Raised when a downloaded file exceeds the size limit. #L119-L122
UnsafeRemoteURLError class UnsafeRemoteURLError(message: str, *, public_message: str = 'Remote media URL is not allowed') Raised when a remote media URL targets an unsafe destination. #L125-L135
UnsafeRemoteURLError.__init__ method UnsafeRemoteURLError.__init__(message: str, *, public_message: str = 'Remote media URL is not allowed') -> None Method UnsafeRemoteURLError.__init__ updates self.public_message; calls super().__init__, super. #L128-L135
_normalize_content_part function _normalize_content_part(item: object) -> object Convert Pydantic content parts into plain Python objects. #L138-L144
_extract_media_url function _extract_media_url(item: dict, item_type: str) -> str Function _extract_media_url calls item.get, isinstance, media_value.get; has 2 explicit return paths. #L147-L161
_text_content_part function _text_content_part(text: str) -> dict[str, str] Function _text_content_part returns {'type': 'text', 'text': text, 'content': text}. #L164-L165
_append_text_content_part function _append_text_content_part(built_parts: list[dict[str, str]], text_parts: list[str], text: str) -> None Function _append_text_content_part calls built_parts.append, _text_content_part, text_parts.append; returns None. #L168-L174
_build_string_mllm_message_content function _build_string_mllm_message_content(content: str, role: str) -> tuple[object, bool] Function _build_string_mllm_message_content calls _text_content_part; has 3 explicit return paths. #L177-L182
_append_ordered_mllm_content_part function _append_ordered_mllm_content_part(raw_item: object, *, built_parts: list[dict[str, str]], text_parts: list[str], all_image_urls: list[str], video_frame_count: int) -> int Function _append_ordered_mllm_content_part calls _normalize_content_part, isinstance, _append_text_content_part, item.get; has 2 explicit return paths. #L185-L220
_build_ordered_mllm_message_content function _build_ordered_mllm_message_content(content: object, *, role: str, all_image_urls: list[str], video_frame_count: int = 0) -> tuple[object, bool] Build template content while preserving OpenAI media/text part order. #L223-L254
_normalize_mllm_tool_calls function _normalize_mllm_tool_calls(tool_calls: list) -> list Normalize replayed assistant tool calls for chat templates. #L257-L268
_build_mllm_chat_messages function _build_mllm_chat_messages(messages: list[dict], *, all_image_urls: list[str], video_frame_counts: dict[int, int]) -> list[dict] Build chat-template messages without reordering multimodal content parts. #L271-L315
MultimodalInput class MultimodalInput(prompt: str, images: list[str] = field(default_factory=list), videos: list[str] = field(default_factory=list), audio: list[str] = field(default_factory=list)) Input for multimodal generation. #L319-L325
MLLMOutput class MLLMOutput(text: str, finish_reason: str \| None = None, prompt_tokens: int = 0, completion_tokens: int = 0, mtp_drafts: int = 0, mtp_accepted: int = 0) Output from multimodal language model. #L329-L337
load_gemma4_assistant_drafter function load_gemma4_assistant_drafter(model_path: str) -> not annotated Load a Gemma 4 assistant drafter for mlx-vlm speculative decoding. #L340-L381
_count_draft_tokens function _count_draft_tokens(draft_tokens) -> int Best-effort drafted-token count for an mlx-vlm drafter output. #L387-L398
_install_draft_metrics_hooks function _install_draft_metrics_hooks(draft_model) -> None Record actual drafted token counts from mlx-vlm assistant drafters. #L401-L428
_install_draft_metrics_hooks.draft_block_with_metrics nested function _install_draft_metrics_hooks.draft_block_with_metrics(*args, **kwargs) -> not annotated Nested Function _install_draft_metrics_hooks.draft_block_with_metrics calls draft_block, draft_model._vllm_mlx_draft_counts.append, _count_draft_tokens; returns draft_tokens. #L412-L415
_install_draft_metrics_hooks.reset_with_metrics nested function _install_draft_metrics_hooks.reset_with_metrics(*args, **kwargs) -> not annotated Nested Function _install_draft_metrics_hooks.reset_with_metrics calls reset; returns reset(*args, **kwargs). #L422-L424
is_base64_image function is_base64_image(s: str) -> bool Check if string is base64-encoded image data. #L431-L435
is_url function is_url(s: str) -> bool Check if string is a URL. #L438-L440
is_base64_video function is_base64_video(s: str) -> bool Check if string is base64-encoded video data. #L443-L445
is_base64_audio function is_base64_audio(s: str) -> bool Check if string is base64-encoded audio data. #L448-L450
decode_base64_image function decode_base64_image(base64_string: str, max_length: int = MAX_BASE64_IMAGE_LENGTH) -> bytes Decode base64 image to bytes. #L453-L480
_validate_url_safety function _validate_url_safety(url: str) -> None Reject remote URLs that target local or private network resources. #L483-L519
_request_with_safe_redirects function _request_with_safe_redirects(method: str, url: str, *, timeout: int, headers: dict[str, str], stream: bool = False, max_redirects: int = 5) -> not annotated Issue a requests call while validating every redirect target. #L522-L557
download_image function download_image(url: str, timeout: int = 30, max_size: int = MAX_IMAGE_SIZE) -> str Download image from URL and return local path. #L560-L645
_download_media function _download_media(url: str, media_type: str, ext_map: dict[str, str], default_ext: str, timeout: int, max_size: int) -> str Download media from URL, enforce size limits, and return a local temp path. #L670-L748
download_video function download_video(url: str, timeout: int = 120, max_size: int = MAX_VIDEO_SIZE) -> str Download video from URL and return local path. #L751-L753
download_audio function download_audio(url: str, timeout: int = 120, max_size: int = MAX_AUDIO_SIZE) -> str Download audio from URL and return local path. #L756-L758
decode_base64_video function decode_base64_video(base64_string: str, max_length: int = MAX_BASE64_VIDEO_LENGTH) -> str Decode base64 video to temp file and return path. #L761-L807
decode_base64_audio function decode_base64_audio(base64_string: str, max_length: int = MAX_BASE64_AUDIO_LENGTH) -> str Decode base64 audio to temp file and return path. #L810-L836
process_video_input function process_video_input(video: str \| dict) -> str Process video input in various formats and return local path. #L839-L874
process_audio_input function process_audio_input(audio: str \| dict) -> str Process audio input in various formats and return local path. #L877-L905
_video_has_audio_track function _video_has_audio_track(video_path: str) -> bool Return True if ffprobe finds an audio stream in the video. #L908-L935
_model_has_sound_encoder function _model_has_sound_encoder(model) -> bool Whether a loaded model exposes a usable sound encoder. #L938-L947
extract_audio_from_video function extract_audio_from_video(video_path: str) -> str \| None Extract the audio track from a video file as 16 kHz mono WAV. #L950-L1005
save_base64_image function save_base64_image(base64_string: str) -> str Save base64 image to temp file and return path. #L1012-L1048
process_image_input function process_image_input(image: str \| dict) -> str Process image input in various formats and return local path. #L1051-L1080
round_by_factor function round_by_factor(x: int, factor: int) -> int Round to nearest multiple of factor. #L1083-L1085
ceil_by_factor function ceil_by_factor(x: float, factor: int) -> int Ceiling to next multiple of factor. #L1088-L1090
floor_by_factor function floor_by_factor(x: float, factor: int) -> int Floor to previous multiple of factor. #L1093-L1095
smart_nframes function smart_nframes(total_frames: int, video_fps: float, target_fps: float = DEFAULT_FPS, min_frames: int = MIN_FRAMES, max_frames: int = MAX_FRAMES) -> int Calculate optimal number of frames to extract from video. #L1098-L1120
extract_video_frames_smart function extract_video_frames_smart(video_path: str, fps: float = DEFAULT_FPS, max_frames: int = MAX_FRAMES, resize: tuple[int, int] \| None = None) -> list[np.ndarray] Extract frames from video with smart sampling. #L1123-L1187
save_frames_to_temp function save_frames_to_temp(frames: list[np.ndarray]) -> list[str] Save frame arrays to temporary files and return paths. #L1190-L1204
MLXMultimodalLM class MLXMultimodalLM(model_name: str, trust_remote_code: bool = False, enable_cache: bool = True, cache_size: int = 50, max_kv_size: int = 0, draft_model: str \| None = None, draft_kind: str \| None = None, draft_block_size: int \| None = None) Wrapper around mlx-vlm for multimodal inference. #L1207-L2938
MLXMultimodalLM.__init__ method MLXMultimodalLM.__init__(model_name: str, trust_remote_code: bool = False, enable_cache: bool = True, cache_size: int = 50, max_kv_size: int = 0, draft_model: str \| None = None, draft_kind: str \| None = None, draft_block_size: int \| None = None) -> not annotated Initialize the MLX multimodal language model. #L1235-L1278
MLXMultimodalLM.load method MLXMultimodalLM.load() -> None Load the model and processor. #L1280-L1323
MLXMultimodalLM._load_draft_model method MLXMultimodalLM._load_draft_model() -> not annotated Method MLXMultimodalLM._load_draft_model calls load_gemma4_assistant_drafter, load; has 2 explicit return paths. #L1325-L1332
MLXMultimodalLM._draft_generation_kwargs method MLXMultimodalLM._draft_generation_kwargs(call_kwargs: dict \| None = None) -> dict Return mlx-vlm drafter kwargs when the request explicitly opts in. #L1334-L1355
MLXMultimodalLM._reset_draft_metrics method MLXMultimodalLM._reset_draft_metrics() -> int Method MLXMultimodalLM._reset_draft_metrics updates self._draft_model.accept_lens, self._draft_model._vllm_mlx_draft_counts; calls hasattr; returns 0. #L1357-L1364
MLXMultimodalLM._draft_metrics_since method MLXMultimodalLM._draft_metrics_since(start_accept_lens: int) -> dict[str, int] Method MLXMultimodalLM._draft_metrics_since calls list, getattr, len, int; has 2 explicit return paths. #L1366-L1395
MLXMultimodalLM.get_language_model method MLXMultimodalLM.get_language_model() -> not annotated Extract the underlying language model for mlx_lm TextModel construction. #L1397-L1399
MLXMultimodalLM.get_tokenizer method MLXMultimodalLM.get_tokenizer() -> not annotated Get the text tokenizer (not the multimodal processor). #L1401-L1403
MLXMultimodalLM._prepare_images method MLXMultimodalLM._prepare_images(images: list) -> list[str] Process remote/base64 image inputs into local temp file paths. #L1405-L1414
MLXMultimodalLM._prepare_audio method MLXMultimodalLM._prepare_audio(audio_inputs: list) -> list[str] Process audio inputs and return local file paths. #L1416-L1425
MLXMultimodalLM._prepare_video method MLXMultimodalLM._prepare_video(video_input: str \| dict, fps: float = DEFAULT_FPS, max_frames: int = MAX_FRAMES, resolved_path: str \| None = None) -> list[str] Process video input and extract frames. #L1427-L1463
MLXMultimodalLM._collect_video_inputs method MLXMultimodalLM._collect_video_inputs(messages: list[dict]) -> dict[int, list] Collect video inputs from messages, keyed by message index. #L1465-L1497
MLXMultimodalLM._collect_audio_inputs method MLXMultimodalLM._collect_audio_inputs(messages: list[dict]) -> dict[int, list] Collect audio inputs from messages, keyed by message index. #L1499-L1528
MLXMultimodalLM._prepare_native_video_inputs method MLXMultimodalLM._prepare_native_video_inputs(messages: list[dict], video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, tools: list \| None = None) -> tuple[str, dict] Preprocess messages into prompt + generation kwargs for native video. #L1530-L1648
MLXMultimodalLM._generate_native_video method MLXMultimodalLM._generate_native_video(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, tools: list \| None = None, **kwargs) -> MLLMOutput Generate using native video pipeline (Qwen-family models). #L1650-L1695
MLXMultimodalLM._translate_messages_for_native_video method MLXMultimodalLM._translate_messages_for_native_video(messages: list[dict], video_fps: float, video_max_frames: int) -> list[dict] Translate OpenAI API format messages to process_vision_info format. #L1697-L1832
MLXMultimodalLM.generate method MLXMultimodalLM.generate(prompt: str, images: list \| None = None, videos: list \| None = None, audio: list[str] \| None = None, max_tokens: int = 256, temperature: float = 0.7, top_p: float = 0.9, video_fps: float = DEFAULT_FPS, video_max_frames: int = MAX_FRAMES, use_cache: bool = True, **kwargs) -> MLLMOutput Generate text from multimodal input. #L1834-L2002
MLXMultimodalLM.stream_generate method MLXMultimodalLM.stream_generate(prompt: str, images: list \| None = None, videos: list[str] \| None = None, audio: list[str] \| None = None, max_tokens: int = 256, temperature: float = 0.7, video_fps: float = DEFAULT_FPS, **kwargs) -> Iterator[str] Stream text generation for multimodal input. #L2004-L2093
MLXMultimodalLM.chat method MLXMultimodalLM.chat(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, **kwargs) -> MLLMOutput Chat with OpenAI-compatible message format. #L2095-L2487
MLXMultimodalLM.stream_chat method MLXMultimodalLM.stream_chat(messages: list[dict], max_tokens: int = 256, temperature: float = 0.7, **kwargs) -> Iterator[MLLMOutput] Stream chat with OpenAI-compatible message format. #L2489-L2737
MLXMultimodalLM.describe_image method MLXMultimodalLM.describe_image(image: str, prompt: str = 'Describe this image in detail.', max_tokens: int = 512, **kwargs) -> str Convenience method to describe an image. #L2739-L2764
MLXMultimodalLM.answer_about_image method MLXMultimodalLM.answer_about_image(image: str, question: str, max_tokens: int = 256, **kwargs) -> str Answer a question about an image. #L2766-L2791
MLXMultimodalLM.describe_video method MLXMultimodalLM.describe_video(video: str \| dict, prompt: str = 'Describe what happens in this video.', fps: float = 2.0, max_frames: int = 32, max_tokens: int = 512, **kwargs) -> str Describe a video using frame extraction. #L2793-L2830
MLXMultimodalLM.get_cache_stats method MLXMultimodalLM.get_cache_stats() -> dict Get MLLM cache statistics. #L2832-L2846
MLXMultimodalLM.clear_cache method MLXMultimodalLM.clear_cache() -> None Clear the MLLM KV cache. #L2848-L2852
MLXMultimodalLM.get_model_info method MLXMultimodalLM.get_model_info() -> dict Get information about the loaded model. #L2854-L2874
MLXMultimodalLM.list_supported_model_families method MLXMultimodalLM.list_supported_model_families() -> dict[str, str] List supported model families and their patterns. #L2877-L2897
MLXMultimodalLM.is_mllm_model method MLXMultimodalLM.is_mllm_model(model_name: str) -> bool Check if a model name indicates an MLLM model. #L2900-L2934
MLXMultimodalLM.__repr__ method MLXMultimodalLM.__repr__() -> str Method MLXMultimodalLM.__repr__ returns f'<MLXMultimodalLM model={self.model_name} status={status}>'. #L2936-L2938