__init__.py 166 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
# ruff: noqa: F401
5
import ast
6
import copy
7
import enum
8
import hashlib
9
import inspect
10
import json
11
import os
12
import textwrap
13
import warnings
14
from collections.abc import Mapping
15
from contextlib import contextmanager
16
from dataclasses import MISSING, Field, field, fields, is_dataclass, replace
17
from functools import cached_property, lru_cache
18
from importlib.util import find_spec
19
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Literal, Optional,
20
                    Protocol, TypeVar, Union, cast, get_args)
21

22
import regex as re
23
import torch
24
from pydantic import (ConfigDict, SkipValidation, field_validator,
25
26
                      model_validator)
from pydantic.dataclasses import dataclass
27
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
28
from typing_extensions import Self, assert_never, runtime_checkable
29

30
import vllm.envs as envs
31
from vllm import version
32
from vllm.config.cache import (BlockSize, CacheConfig, CacheDType, MambaDType,
33
                               PrefixCachingHashAlgo)
34
from vllm.config.compilation import (CompilationConfig, CompilationLevel,
35
                                     CUDAGraphMode, PassConfig)
36
from vllm.config.kv_events import KVEventsConfig
37
from vllm.config.kv_transfer import KVTransferConfig
38
from vllm.config.load import LoadConfig
39
from vllm.config.lora import LoRAConfig
40
41
from vllm.config.parallel import (DistributedExecutorBackend, EPLBConfig,
                                  ParallelConfig)
42
from vllm.config.scheduler import SchedulerConfig, SchedulerPolicy
43
from vllm.config.utils import ConfigType, config
Woosuk Kwon's avatar
Woosuk Kwon committed
44
from vllm.logger import init_logger
45
from vllm.model_executor.layers.quantization import QuantizationMethods
46
from vllm.multimodal import MULTIMODAL_REGISTRY
47
from vllm.platforms import current_platform
48
49
50
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
51
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
52
53
54
    is_interleaved, maybe_override_with_speculators_target_model,
    try_get_generation_config, try_get_safetensors_metadata,
    try_get_tokenizer_config, uses_mrope)
55
56
57
from vllm.transformers_utils.runai_utils import (ObjectStorageModel,
                                                 is_runai_obj_uri)
from vllm.transformers_utils.utils import maybe_model_redirect
58
59
from vllm.utils import (DEFAULT_MAX_NUM_BATCHED_TOKENS,
                        STR_DUAL_CHUNK_FLASH_ATTN_VAL, LayerBlockType,
60
                        LazyLoader, common_broadcastable_dtype, random_uuid)
61

62
if TYPE_CHECKING:
63
    from _typeshed import DataclassInstance
64
    from transformers.configuration_utils import PretrainedConfig
65

66
67
68
    import vllm.model_executor.layers.quantization as me_quant
    import vllm.model_executor.models as me_models
    from vllm.model_executor.layers.quantization import QuantizationMethods
69
70
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
71
    from vllm.v1.sample.logits_processor import LogitsProcessor
72

73
    HfOverrides = Union[dict, Callable[[type], type]]
74
else:
75
    DataclassInstance = Any
76
    PretrainedConfig = Any
77
    QuantizationConfig = Any
78
    QuantizationMethods = Any
79
    BaseModelLoader = Any
80
    LogitsProcessor = Any
81
82
83
84
85
86
    HfOverrides = Union[dict[str, Any], Callable[[type], type]]

    me_quant = LazyLoader("model_executor", globals(),
                          "vllm.model_executor.layers.quantization")
    me_models = LazyLoader("model_executor", globals(),
                           "vllm.model_executor.models")
87

88
logger = init_logger(__name__)
89
DataclassInstanceT = TypeVar("DataclassInstanceT", bound=DataclassInstance)
90

91
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
92
                     "score", "reward", "transcription", "draft"]
93

94
_ResolvedTask = Literal["generate", "transcription", "encode", "embed",
95
                        "classify", "reward", "draft"]
96

97
RunnerOption = Literal["auto", "generate", "pooling", "draft"]
98

99
RunnerType = Literal["generate", "pooling", "draft"]
100

101
102
103
104
105
ConvertOption = Literal["auto", "none", "embed", "classify", "reward"]

ConvertType = Literal["none", "embed", "classify", "reward"]

_RUNNER_TASKS: dict[RunnerType, list[TaskOption]] = {
106
    "generate": ["generate", "transcription"],
107
108
109
110
111
112
113
    "pooling": ["embedding", "embed", "classify", "score", "reward"],
    "draft": ["draft"],
}

_RUNNER_CONVERTS: dict[RunnerType, list[ConvertType]] = {
    "generate": [],
    "pooling": ["embed", "classify", "reward"],
114
    "draft": [],
115
}
116

117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
# Some model suffixes are based on auto classes from Transformers:
# https://huggingface.co/docs/transformers/en/model_doc/auto
# NOTE: Items higher on this list priority over lower ones
_SUFFIX_TO_DEFAULTS: list[tuple[str, tuple[RunnerType, ConvertType]]] = [
    ("ForCausalLM", ("generate", "none")),
    ("ForConditionalGeneration", ("generate", "none")),
    ("ChatModel", ("generate", "none")),
    ("LMHeadModel", ("generate", "none")),
    ("ForTextEncoding", ("pooling", "embed")),
    ("EmbeddingModel", ("pooling", "embed")),
    ("ForSequenceClassification", ("pooling", "classify")),
    ("ForAudioClassification", ("pooling", "classify")),
    ("ForImageClassification", ("pooling", "classify")),
    ("ForVideoClassification", ("pooling", "classify")),
    ("ClassificationModel", ("pooling", "classify")),
    ("ForRewardModeling", ("pooling", "reward")),
    ("RewardModel", ("pooling", "reward")),
    # Let other `*Model`s take priority
    ("Model", ("pooling", "embed")),
]


def iter_architecture_defaults():
    yield from _SUFFIX_TO_DEFAULTS


def try_match_architecture_defaults(
    architecture: str,
    *,
    runner_type: Optional[RunnerType] = None,
    convert_type: Optional[ConvertType] = None,
) -> Optional[tuple[str, tuple[RunnerType, ConvertType]]]:
    for suffix, (default_runner_type,
                 default_convert_type) in iter_architecture_defaults():
        if ((runner_type is None or runner_type == default_runner_type) and
            (convert_type is None or convert_type == default_convert_type)
                and architecture.endswith(suffix)):
            return suffix, (default_runner_type, default_convert_type)

    return None

158

159
@runtime_checkable
160
161
162
163
164
165
class SupportsHash(Protocol):

    def compute_hash(self) -> str:
        ...


166
167
class SupportsMetricsInfo(Protocol):

168
    def metrics_info(self) -> dict[str, str]:
169
170
171
        ...


172
173
174
175
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"
176
    TERRATORCH = "terratorch"
177
178


179
180
181
182
183
184
185
186
187
188
def get_attr_docs(cls: type[Any]) -> dict[str, str]:
    """
    Get any docstrings placed after attribute assignments in a class body.

    https://davidism.com/mit-license/
    """

    def pairwise(iterable):
        """
        Manually implement https://docs.python.org/3/library/itertools.html#itertools.pairwise
189

190
191
192
193
194
195
196
197
198
        Can be removed when Python 3.9 support is dropped.
        """
        iterator = iter(iterable)
        a = next(iterator, None)

        for b in iterator:
            yield a, b
            a = b

199
200
201
202
203
204
205
206
207
208
209
    try:
        cls_node = ast.parse(textwrap.dedent(inspect.getsource(cls))).body[0]
    except (OSError, KeyError, TypeError):
        # HACK: Python 3.13+ workaround - set missing __firstlineno__
        # Workaround can be removed after we upgrade to pydantic==2.12.0
        with open(inspect.getfile(cls)) as f:
            for i, line in enumerate(f):
                if f"class {cls.__name__}" in line and ":" in line:
                    cls.__firstlineno__ = i + 1
                    break
        cls_node = ast.parse(textwrap.dedent(inspect.getsource(cls))).body[0]
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240

    if not isinstance(cls_node, ast.ClassDef):
        raise TypeError("Given object was not a class.")

    out = {}

    # Consider each pair of nodes.
    for a, b in pairwise(cls_node.body):
        # Must be an assignment then a constant string.
        if (not isinstance(a, (ast.Assign, ast.AnnAssign))
                or not isinstance(b, ast.Expr)
                or not isinstance(b.value, ast.Constant)
                or not isinstance(b.value.value, str)):
            continue

        doc = inspect.cleandoc(b.value.value)

        # An assignment can have multiple targets (a = b = v), but an
        # annotated assignment only has one target.
        targets = a.targets if isinstance(a, ast.Assign) else [a.target]

        for target in targets:
            # Must be assigning to a plain name.
            if not isinstance(target, ast.Name):
                continue

            out[target.id] = doc

    return out


241
def get_field(cls: ConfigType, name: str) -> Field:
242
243
244
245
246
247
248
    """Get the default factory field of a dataclass by name. Used for getting
    default factory fields in `EngineArgs`."""
    if not is_dataclass(cls):
        raise TypeError("The given class is not a dataclass.")
    cls_fields = {f.name: f for f in fields(cls)}
    if name not in cls_fields:
        raise ValueError(f"Field '{name}' not found in {cls.__name__}.")
249
    named_field: Field = cls_fields[name]
250
251
252
253
254
255
256
257
    if (default_factory := named_field.default_factory) is not MISSING:
        return field(default_factory=default_factory)
    if (default := named_field.default) is not MISSING:
        return field(default=default)
    raise ValueError(
        f"{cls.__name__}.{name} must have a default value or default factory.")


258
259
260
261
def is_init_field(cls: ConfigType, name: str) -> bool:
    return next(f for f in fields(cls) if f.name == name).init


262
263
TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
264
MMEncoderTPMode = Literal["weights", "data"]
265
266


267
268
269
270
271
272
273
class LogprobsMode(enum.Enum):
    RAW_LOGITS = "raw_logits"
    RAW_LOGPROBS = "raw_logprobs"
    PROCESSED_LOGITS = "processed_logits"
    PROCESSED_LOGPROBS = "processed_logprobs"


274
@config
275
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
276
class ModelConfig:
277
278
    """Configuration for the model."""

279
    model: str = "Qwen/Qwen3-0.6B"
280
281
282
    """Name or path of the Hugging Face model to use. It is also used as the
    content for `model_name` tag in metrics output when `served_model_name` is
    not specified."""
283
284
285
    runner: RunnerOption = "auto"
    """The type of model runner to use. Each vLLM instance only supports one
    model runner, even if the same model can be used for multiple types."""
286
287
288
289
290
291
292
293
294
295
    convert: ConvertOption = "auto"
    """Convert the model using adapters defined in
    [vllm.model_executor.models.adapters][]. The most common use case is to
    adapt a text generation model to be used for pooling tasks."""
    task: Optional[TaskOption] = None
    """[DEPRECATED] The task to use the model for. If the model supports more
    than one model runner, this is used to select which model runner to run.

    Note that the model may support other tasks using the same model runner.
    """
296
    tokenizer: SkipValidation[str] = None  # type: ignore
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
    """Name or path of the Hugging Face tokenizer to use. If unspecified, model
    name or path will be used."""
    tokenizer_mode: TokenizerMode = "auto"
    """Tokenizer mode:\n
    - "auto" will use the fast tokenizer if available.\n
    - "slow" will always use the slow tokenizer.\n
    - "mistral" will always use the tokenizer from `mistral_common`.\n
    - "custom" will use --tokenizer to select the preregistered tokenizer."""
    trust_remote_code: bool = False
    """Trust remote code (e.g., from HuggingFace) when downloading the model
    and tokenizer."""
    dtype: Union[ModelDType, torch.dtype] = "auto"
    """Data type for model weights and activations:\n
    - "auto" will use FP16 precision for FP32 and FP16 models, and BF16
    precision for BF16 models.\n
    - "half" for FP16. Recommended for AWQ quantization.\n
    - "float16" is the same as "half".\n
    - "bfloat16" for a balance between precision and range.\n
    - "float" is shorthand for FP32 precision.\n
    - "float32" for FP32 precision."""
    seed: Optional[int] = None
318
319
    """Random seed for reproducibility. Initialized to None in V0, but
    initialized to 0 in V1."""
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
    hf_config_path: Optional[str] = None
    """Name or path of the Hugging Face config to use. If unspecified, model
    name or path will be used."""
    allowed_local_media_path: str = ""
    """Allowing API requests to read local images or videos from directories
    specified by the server file system. This is a security risk. Should only
    be enabled in trusted environments."""
    revision: Optional[str] = None
    """The specific model version to use. It can be a branch name, a tag name,
    or a commit id. If unspecified, will use the default version."""
    code_revision: Optional[str] = None
    """The specific revision to use for the model code on the Hugging Face Hub.
    It can be a branch name, a tag name, or a commit id. If unspecified, will
    use the default version."""
    rope_scaling: dict[str, Any] = field(default_factory=dict)
335
    """RoPE scaling configuration. For example,
336
337
338
339
340
341
342
343
    `{"rope_type":"dynamic","factor":2.0}`."""
    rope_theta: Optional[float] = None
    """RoPE theta. Use with `rope_scaling`. In some cases, changing the RoPE
    theta improves the performance of the scaled model."""
    tokenizer_revision: Optional[str] = None
    """The specific revision to use for the tokenizer on the Hugging Face Hub.
    It can be a branch name, a tag name, or a commit id. If unspecified, will
    use the default version."""
344
    max_model_len: SkipValidation[int] = None  # type: ignore
345
346
    """Model context length (prompt and output). If unspecified, will be
    automatically derived from the model config.
347

348
349
350
351
352
353
    When passing via `--max-model-len`, supports k/m/g/K/M/G in human-readable
    format. Examples:\n
    - 1k -> 1000\n
    - 1K -> 1024\n
    - 25.6k -> 25,600"""
    spec_target_max_model_len: Optional[int] = None
omahs's avatar
omahs committed
354
    """Specify the maximum length for spec decoding draft models."""
355
    quantization: SkipValidation[Optional[QuantizationMethods]] = None
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
    """Method used to quantize the weights. If `None`, we first check the
    `quantization_config` attribute in the model config file. If that is
    `None`, we assume the model weights are not quantized and use `dtype` to
    determine the data type of the weights."""
    enforce_eager: bool = False
    """Whether to always use eager-mode PyTorch. If True, we will disable CUDA
    graph and always execute the model in eager mode. If False, we will use
    CUDA graph and eager execution in hybrid for maximal performance and
    flexibility."""
    max_seq_len_to_capture: int = 8192
    """Maximum sequence len covered by CUDA graphs. When a sequence has context
    length larger than this, we fall back to eager mode. Additionally for
    encoder-decoder models, if the sequence length of the encoder input is
    larger than this, we fall back to the eager mode."""
    max_logprobs: int = 20
    """Maximum number of log probabilities to return when `logprobs` is
    specified in `SamplingParams`. The default value comes the default for the
373
374
    OpenAI Chat Completions API. -1 means no cap, i.e. all (output_length *
    vocab_size) logprobs are allowed to be returned and it may cause OOM."""
375
    logprobs_mode: LogprobsMode = LogprobsMode.RAW_LOGPROBS
376
377
378
    """Indicates the content returned in the logprobs and prompt_logprobs.
    Supported mode:
    1) raw_logprobs, 2) processed_logprobs, 3) raw_logits, 4) processed_logits.
379
380
381
    Raw means the values before applying any logit processors, like bad words.
    Processed means the values after applying all processors, including
    temperature and top_k/top_p.
382
    """
383
384
385
386
387
388
389
390
391
392
393
394
395
396
    disable_sliding_window: bool = False
    """Whether to disable sliding window. If True, we will disable the sliding
    window functionality of the model, capping to sliding window size. If the
    model does not support sliding window, this argument is ignored."""
    disable_cascade_attn: bool = False
    """Disable cascade attention for V1. While cascade attention does not
    change the mathematical correctness, disabling it could be useful for
    preventing potential numerical issues. Note that even if this is set to
    False, cascade attention will be only used when the heuristic tells that
    it's beneficial."""
    skip_tokenizer_init: bool = False
    """Skip initialization of tokenizer and detokenizer. Expects valid
    `prompt_token_ids` and `None` for prompt from the input. The generated
    output will contain token ids."""
397
398
399
400
    enable_prompt_embeds: bool = False
    """If `True`, enables passing text embeddings as inputs via the
    `prompt_embeds` key. Note that enabling this will double the time required
    for graph compilation."""
401
402
403
404
405
406
407
408
409
410
411
    served_model_name: Optional[Union[str, list[str]]] = None
    """The model name(s) used in the API. If multiple names are provided, the
    server will respond to any of the provided names. The model name in the
    model field of a response will be the first name in this list. If not
    specified, the model name will be the same as the `--model` argument. Noted
    that this name(s) will also be used in `model_name` tag content of
    prometheus metrics, if multiple names provided, metrics tag will take the
    first one."""
    limit_mm_per_prompt: dict[str, int] = field(default_factory=dict)
    """Maximum number of data items per modality per prompt. Only applicable
    for multimodal models."""
412
    interleave_mm_strings: bool = False
413
    """Enable fully interleaved support for multimodal prompts, while using
414
    --chat-template-content-format=string. Defaults to False."""
415
416
417
418
    skip_mm_profiling: bool = False
    """When enabled, skips multimodal memory profiling and only profiles with
    language backbone model during engine initialization.
    """
419
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
420
421
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
422
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
423
424
    use_async_output_proc: bool = True
    """Whether to use async output processor."""
425
    config_format: Union[str, ConfigFormat] = "auto"
426
427
428
429
430
431
432
433
434
435
436
    """The format of the model config to load:\n
    - "auto" will try to load the config in hf format if available else it
    will try to load in mistral format.\n
    - "hf" will load the config in hf format.\n
    - "mistral" will load the config in mistral format."""
    hf_token: Optional[Union[bool, str]] = None
    """The token to use as HTTP bearer authorization for remote files . If
    `True`, will use the token generated when running `huggingface-cli login`
    (stored in `~/.huggingface`)."""
    hf_overrides: HfOverrides = field(default_factory=dict)
    """If a dictionary, contains arguments to be forwarded to the Hugging Face
437
    config. If a callable, it is called to update the HuggingFace config."""
438
439
440
441
442
    mm_processor_kwargs: Optional[dict[str, Any]] = None
    """Arguments to be forwarded to the model's processor for multi-modal data,
    e.g., image processor. Overrides for the multi-modal processor obtained
    from `AutoProcessor.from_pretrained`. The available overrides depend on the
    model that is being run. For example, for Phi-3-Vision: `{"num_crops": 4}`.
443
    """
444
    mm_processor_cache_gb: float = 4
445
446
447
448
449
450
451
452
    """The size (in GiB) of the multi-modal processor cache, which is used to
    avoid re-processing past multi-modal inputs.

    This cache is duplicated for each API process and engine core process,
    resulting in a total memory usage of
    `mm_processor_cache_gb * (api_server_count + data_parallel_size)`.

    Set to `0` to disable this cache completely (not recommended)."""
453
454
455
456
457
458
459
460
461
462
463
464
465
    mm_encoder_tp_mode: MMEncoderTPMode = "weights"
    """Indicates how to optimize multi-modal encoder inference using
    tensor parallelism (TP).

    - `"weights"`: Within the same vLLM engine, split the weights of
        each layer across TP ranks. (default TP behavior)
    - `"data"`: Within the same vLLM engine, split the batched input data
        across TP ranks to process the data in parallel, while hosting
        the full weights on each TP rank.
        This batch-level DP is not to be confused with API request-level
        DP (which is controlled by `--data-parallel-size`).
        This is only supported on a per-model basis and falls back to
        `"weights"` if the encoder does not support DP."""
466
467
468
469
470
471
    pooler_config: Optional["PoolerConfig"] = field(init=False)
    """Pooler config which controls the behaviour of output pooling in pooling
    models."""
    override_pooler_config: Optional[Union[dict, "PoolerConfig"]] = None
    """Initialize non-default pooling config or override default pooling config
    for the pooling model. e.g. `{"pooling_type": "mean", "normalize": false}`.
472
    """
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
    logits_processor_pattern: Optional[str] = None
    """Optional regex pattern specifying valid logits processor qualified names
    that can be passed with the `logits_processors` extra completion argument.
    Defaults to `None`, which allows no processors."""
    generation_config: str = "auto"
    """The folder path to the generation config. Defaults to `"auto"`, the
    generation config will be loaded from model path. If set to `"vllm"`, no
    generation config is loaded, vLLM defaults will be used. If set to a folder
    path, the generation config will be loaded from the specified folder path.
    If `max_new_tokens` is specified in generation config, then it sets a
    server-wide limit on the number of output tokens for all requests."""
    override_generation_config: dict[str, Any] = field(default_factory=dict)
    """Overrides or sets generation config. e.g. `{"temperature": 0.5}`. If
    used with `--generation-config auto`, the override parameters will be
    merged with the default config from the model. If used with
488
    `--generation-config vllm`, only the override parameters are used."""
489
490
491
492
493
494
495
496
    enable_sleep_mode: bool = False
    """Enable sleep mode for the engine (only cuda platform is supported)."""
    model_impl: Union[str, ModelImpl] = ModelImpl.AUTO.value
    """Which implementation of the model to use:\n
    - "auto" will try to use the vLLM implementation, if it exists, and fall
    back to the Transformers implementation if no vLLM implementation is
    available.\n
    - "vllm" will use the vLLM model implementation.\n
497
498
499
    - "transformers" will use the Transformers model implementation.\n
    - "terratorch" will use the TerraTorch model implementation.
    """
500
501
    override_attention_dtype: Optional[str] = None
    """Override dtype for attention"""
502
503
504
    logits_processors: Optional[list[Union[str, type[LogitsProcessor]]]] = None
    """One or more logits processors' fully-qualified class names or class
    definitions"""
505
506
    io_processor_plugin: Optional[str] = None
    """IOProcessor plugin name to load at model startup"""
507

508
509
510
511
512
513
514
515
516
517
518
519
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
520
        factors: list[Any] = []
521
522
523
524
525
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
526
527
528
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
529
        factors.append(self.trust_remote_code)
530
531
532
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
533
534
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
535
536
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
537
538
        str_factors = str(factors)
        assert_hashable(str_factors)
539
540
        return hashlib.sha256(str(factors).encode()).hexdigest()

541
    def __post_init__(self) -> None:
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
        # Set the default seed to 0 in V1.
        # NOTE(woosuk): In V0, we set the default seed to None because the
        # driver worker shares the same process as the user process, and thus
        # setting a seed affects the user process as well.
        # In V1, we use separate processes for workers (unless
        # VLLM_ENABLE_V1_MULTIPROCESSING=0), so setting a seed here
        # doesn't affect the user process. However, without a consistent seed,
        # different tensor parallel workers would sample different tokens,
        # leading to inconsistent results.
        if envs.VLLM_USE_V1 and self.seed is None:
            self.seed = 0
            if not envs.VLLM_ENABLE_V1_MULTIPROCESSING:
                logger.warning(
                    "The global random seed is set to %d. Since "
                    "VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may "
                    "affect the random state of the Python process that "
                    "launched vLLM.", self.seed)

560
561
562
        # Keep set served_model_name before maybe_model_redirect(self.model)
        self.served_model_name = get_served_model_name(self.model,
                                                       self.served_model_name)
563
564
565
566
567
568
569
570
571
572
573
574
        self.model = maybe_model_redirect(self.model)
        # The tokenizer is consistent with the model by default.
        if self.tokenizer is None:
            self.tokenizer = self.model
        if self.tokenizer_revision is None:
            self.tokenizer_revision = self.revision
        self.tokenizer = maybe_model_redirect(self.tokenizer)

        if isinstance(self.hf_config_path, str):
            self.hf_config_path = maybe_model_redirect(self.hf_config_path)

        if callable(self.hf_overrides):
575
            hf_overrides_kw = {}
576
            hf_overrides_fn = self.hf_overrides
577
        else:
578
            hf_overrides_kw = self.hf_overrides
579
            hf_overrides_fn = None
580

581
582
        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
583
            hf_overrides_kw.update(hf_override)
584
            hf_overrides_str = json.dumps(hf_overrides_kw)
585
586
587
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
588
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
589
590
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
591
            hf_overrides_kw.update(hf_override)
592
            hf_overrides_str = json.dumps(hf_overrides_kw)
593
594
595
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
596
597
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

598
599
600
601
602
603
604
605
606
607
        self.maybe_pull_model_tokenizer_for_runai(self.model, self.tokenizer)

        if self.runner != "draft":
            # If we're not running the draft model, check for speculators config
            # If speculators config, set model / tokenizer to be target model
            self.model, self.tokenizer = maybe_override_with_speculators_target_model(  # noqa: E501
                model=self.model,
                tokenizer=self.tokenizer,
                revision=self.revision,
                trust_remote_code=self.trust_remote_code)
608

609
610
611
612
        if (backend := envs.VLLM_ATTENTION_BACKEND
            ) and backend == "FLASHINFER" and find_spec("flashinfer") is None:
            raise ValueError(
                "VLLM_ATTENTION_BACKEND is set to FLASHINFER, but flashinfer "
613
614
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
615
616
                "for instructions on how to install it.")

617
618
        from vllm.platforms import current_platform

619
620
621
622
623
624
        if (self.override_attention_dtype is not None
                and not current_platform.is_rocm()):
            warnings.warn(
                "override-attention-dtype is set but not using ROCm platform",
                stacklevel=2)

625
626
627
628
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
629

630
        hf_config = get_config(self.hf_config_path or self.model,
631
632
633
634
635
636
                               self.trust_remote_code,
                               self.revision,
                               self.code_revision,
                               self.config_format,
                               hf_overrides_kw=hf_overrides_kw,
                               hf_overrides_fn=hf_overrides_fn)
637

638
        self.hf_config = hf_config
639
        self.hf_text_config = get_hf_text_config(self.hf_config)
640
641
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
642
        self.encoder_config = self._get_encoder_config()
643
        self.hf_image_processor_config = get_hf_image_processor_config(
644
            self.model, hf_token=self.hf_token, revision=self.revision)
645

646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
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
        architectures = self.architectures
        registry = self.registry
        is_generative_model = registry.is_text_generation_model(
            architectures, self)
        is_pooling_model = registry.is_pooling_model(architectures, self)

        def _task_to_convert(task: TaskOption) -> ConvertType:
            if task == "embedding" or task == "embed":
                return "embed"
            if task == "classify":
                return "classify"
            if task == "reward":
                return "reward"
            if task == "score":
                new_task = self._get_default_pooling_task(architectures)
                return "classify" if new_task == "classify" else "embed"

            return "none"

        if self.task is not None:
            runner: RunnerOption = "auto"
            convert: ConvertOption = "auto"
            msg_prefix = ("The 'task' option has been deprecated and will be "
                          "removed in v0.13.0 or v1.0, whichever comes first.")
            msg_hint = "Please remove this option."

            is_generative_task = self.task in _RUNNER_TASKS["generate"]
            is_pooling_task = self.task in _RUNNER_TASKS["pooling"]

            if is_generative_model and is_pooling_model:
                if is_generative_task:
                    runner = "generate"
                    convert = "auto"
                    msg_hint = ("Please replace this option with `--runner "
                                "generate` to continue using this model "
                                "as a generative model.")
                elif is_pooling_task:
                    runner = "pooling"
                    convert = "auto"
                    msg_hint = ("Please replace this option with `--runner "
                                "pooling` to continue using this model "
                                "as a pooling model.")
                else:  # task == "auto"
                    pass
            elif is_generative_model or is_pooling_model:
                if is_generative_task:
                    runner = "generate"
                    convert = "auto"
                    msg_hint = "Please remove this option"
                elif is_pooling_task:
                    runner = "pooling"
                    convert = _task_to_convert(self.task)
                    msg_hint = ("Please replace this option with `--convert "
                                f"{convert}` to continue using this model "
                                "as a pooling model.")
                else:  # task == "auto"
                    pass
703
            else:
704
705
706
707
708
709
710
711
                raise AssertionError("The model should be a generative or "
                                     "pooling model when task is set to "
                                     f"{self.task!r}.")

            self.runner = runner
            self.convert = convert

            msg = f"{msg_prefix} {msg_hint}"
712
713
            warnings.warn(msg, DeprecationWarning, stacklevel=2)

714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
        self.runner_type = self._get_runner_type(architectures, self.runner)
        self.convert_type = self._get_convert_type(architectures,
                                                   self.runner_type,
                                                   self.convert)

        if self.runner_type == "generate" and not is_generative_model:
            generate_converts = _RUNNER_CONVERTS["generate"]
            if self.convert_type not in generate_converts:
                # Currently we don't have any converters for generative models
                raise ValueError(
                    "This model does not support `--runner generate`.")
        if self.runner_type == "pooling" and not is_pooling_model:
            pooling_converts = _RUNNER_CONVERTS["pooling"]
            if self.convert_type not in pooling_converts:
                convert_option = "<" + "|".join(pooling_converts) + ">"
                raise ValueError(
                    "This model does not support `--runner pooling`. "
                    f"You can pass `--convert {convert_option} to adapt "
                    "it into a pooling model.")
733

734
735
736
737
738
739
        self.supported_tasks = self._get_supported_tasks(
            architectures, self.runner_type, self.convert_type)

        # Note: Initialize these attributes early because transformers fallback
        # may fail to load dynamic modules in child processes
        model_info, arch = registry.inspect_model_cls(architectures, self)
740
741
        self._model_info = model_info
        self._architecture = arch
742
        logger.info("Resolved architecture: %s", arch)
743
744
745

        self.pooler_config = self._init_pooler_config()

746
        self.dtype: torch.dtype = _get_and_verify_dtype(
747
748
749
750
751
752
            self.model,
            self.hf_config,
            self.dtype,
            is_pooling_model=self.runner_type == "pooling",
            revision=self.revision,
        )
753

754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
        # Interleaved attention is not supported by some backends in V0
        if (not self.disable_sliding_window
                and is_interleaved(self.hf_text_config)
                and not envs.VLLM_USE_V1
                and (backend := envs.VLLM_ATTENTION_BACKEND)
                in ("XFORMERS", "FLASHINFER")):
            logger.warning_once(
                "%s has interleaved attention, which is currently not "
                "supported by the %s backend. Disabling sliding window and "
                "capping the max length to the sliding window size (%d).",
                self.hf_text_config.model_type,
                backend,
                self.hf_text_config.sliding_window,
            )
            self.disable_sliding_window = True
Woosuk Kwon's avatar
Woosuk Kwon committed
769

770
        self.original_max_model_len = self.max_model_len
771
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
772
        self.multimodal_config = self._init_multimodal_config()
773

774
775
776
777
778
        if self.disable_sliding_window:
            # Set after get_and_verify_max_len to ensure that max_model_len
            # can be correctly capped to sliding window size
            self.hf_text_config.sliding_window = None

779
780
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
781

782
783
784
        # Avoid running try_verify_and_update_config multiple times
        self.config_updated = False

785
        self._verify_quantization()
786
        self._verify_cuda_graph()
787
        self._verify_bnb_config()
788

789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
    @field_validator("quantization", mode="before")
    @classmethod
    def validate_quantization_before(cls, value: Any) -> Any:
        if isinstance(value, str):
            return value.lower()
        return value

    @model_validator(mode="after")
    def validate_model_config_after(self: "ModelConfig") -> "ModelConfig":
        if not isinstance(self.tokenizer, str):
            raise ValueError("tokenizer must be a string after __post_init__.")
        if not isinstance(self.max_model_len, int):
            raise ValueError(
                "max_model_len must be an integer after __post_init__.")
        return self

805
806
807
    def _get_transformers_backend_cls(self) -> str:
        """Determine which Transformers backend class will be used if
        `model_impl` is set to `transformers` or `auto`."""
808
809
        if getattr(self, "runner_type", self.runner) == "pooling":
            return "TransformersModel"
810
811
812
813
        if self.hf_config != self.hf_text_config:
            # If 'hf_text_config' is the same as 'hf_config'. If not, it is
            # probably a composite config, i.e. multimodal
            return "TransformersForMultimodalLM"
814
815
816
817
818
        return "TransformersForCausalLM"

    def using_transformers_backend(self) -> bool:
        """Check if the model is using the Transformers backend class."""
        return self.architecture == self._get_transformers_backend_cls()
819

820
821
    @property
    def registry(self):
822
        return me_models.ModelRegistry
823
824
825

    @property
    def architectures(self) -> list[str]:
826
        return getattr(self.hf_config, "architectures", [])
827

828
829
    @property
    def architecture(self) -> str:
830
        """The architecture vllm actually used."""
831
832
        return self._architecture

833
834
835
836
    def maybe_pull_model_tokenizer_for_runai(self, model: str,
                                             tokenizer: str) -> None:
        """Pull model/tokenizer from Object Storage to temporary
        directory when needed.
837

838
        Args:
839
840
            model: Model name or path
            tokenizer: Tokenizer name or path
841
        """
842
        if not (is_runai_obj_uri(model) or is_runai_obj_uri(tokenizer)):
843
844
            return

845
846
847
848
        if is_runai_obj_uri(model):
            object_storage_model = ObjectStorageModel()
            object_storage_model.pull_files(
                model, allow_pattern=["*.model", "*.py", "*.json"])
849
            self.model_weights = model
850
            self.model = object_storage_model.dir
851
852
853

            # If tokenizer is same as model, download to same directory
            if model == tokenizer:
854
855
856
857
858
859
                object_storage_model.pull_files(model,
                                                ignore_pattern=[
                                                    "*.pt", "*.safetensors",
                                                    "*.bin", "*.tensors"
                                                ])
                self.tokenizer = object_storage_model.dir
860
861
862
                return

        # Only download tokenizer if needed and not already handled
863
864
865
        if is_runai_obj_uri(tokenizer):
            object_storage_tokenizer = ObjectStorageModel()
            object_storage_tokenizer.pull_files(
866
867
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors"])
868
            self.tokenizer = object_storage_tokenizer.dir
869

870
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
871
        if self._model_info.supports_multimodal:
872
873
874
875
876
877
878
            if (self.mm_encoder_tp_mode == "data" and
                    not self._model_info.supports_multimodal_encoder_tp_data):
                logger.warning_once(
                    "This model does not support `--mm-encoder-tp-mode data`. "
                    "Falling back to `--mm-encoder-tp-mode weights`.")
                self.mm_encoder_tp_mode = "weights"

879
            return MultiModalConfig(
880
                limit_per_prompt=self.limit_mm_per_prompt,
881
                media_io_kwargs=self.media_io_kwargs,
882
                mm_processor_kwargs=self.mm_processor_kwargs,
883
                mm_processor_cache_gb=self.mm_processor_cache_gb,
884
                mm_encoder_tp_mode=self.mm_encoder_tp_mode,
885
                interleave_mm_strings=self.interleave_mm_strings,
886
887
                skip_mm_profiling=self.skip_mm_profiling,
            )
888
889

        return None
890

891
892
893
894
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

895
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
896
        if self.runner_type == "pooling":
897
898
899
900
901
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
902
903
904
905
906

            base_config = get_pooling_config(self.model, self.revision)
            if base_config is not None:
                # Only set values that are not overridden by the user
                for k, v in base_config.items():
907
908
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
909

910
911
912
913
            default_pooling_type = self._model_info.default_pooling_type
            if pooler_config.pooling_type is None:
                pooler_config.pooling_type = default_pooling_type

914
            return pooler_config
915

916
917
        return None

918
    def _verify_tokenizer_mode(self) -> None:
919
920
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
921
922
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
923
                f"one of {get_args(TokenizerMode)}.")
924
        self.tokenizer_mode = tokenizer_mode
925

926
927
928
929
930
931
932
933
934
935
    def _get_default_runner_type(
        self,
        architectures: list[str],
    ) -> RunnerType:
        registry = self.registry

        # Some Sentence Transformers models use *ForCausalLM archs
        if get_pooling_config(self.model, self.revision):
            return "pooling"

936
        for arch in architectures:
937
938
939
940
941
942
943
944
945
946
947
948
            if arch in registry.get_supported_archs():
                if registry.is_pooling_model(architectures, self):
                    return "pooling"
                if registry.is_text_generation_model(architectures, self):
                    return "generate"

            match = try_match_architecture_defaults(arch)
            if match:
                _, (runner_type, _) = match
                return runner_type

        return "generate"
949

950
    def _get_runner_type(
951
        self,
952
        architectures: list[str],
953
954
955
956
957
958
959
        runner: RunnerOption,
    ) -> RunnerType:
        if runner != "auto":
            return runner

        runner_type = self._get_default_runner_type(architectures)

960
961
962
963
964
965
        # Don't log the most common case
        if runner_type != "generate":
            logger.info(
                "Resolved `--runner auto` to `--runner %s`. "
                "Pass the value explicitly to silence this message.",
                runner_type)
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

        return runner_type

    def _get_default_convert_type(
        self,
        architectures: list[str],
        runner_type: RunnerType,
    ) -> ConvertType:
        registry = self.registry

        for arch in architectures:
            if arch in registry.get_supported_archs():
                if (runner_type == "generate"
                        and registry.is_text_generation_model(
                            architectures, self)):
                    return "none"
                if (runner_type == "pooling"
                        and registry.is_pooling_model(architectures, self)):
                    return "none"

            match = try_match_architecture_defaults(arch,
                                                    runner_type=runner_type)
            if match:
                _, (_, convert_type) = match
                return convert_type

        # This is to handle Sentence Transformers models that use *ForCausalLM
        # and also multi-modal pooling models which are not defined as
        # Sentence Transformers models
        if runner_type == "pooling":
996
997
            return "embed"

998
999
1000
1001
1002
1003
1004
1005
1006
1007
        return "none"

    def _get_convert_type(
        self,
        architectures: list[str],
        runner_type: RunnerType,
        convert: ConvertOption,
    ) -> ConvertType:
        if convert != "auto":
            return convert
1008

1009
1010
        convert_type = self._get_default_convert_type(architectures,
                                                      runner_type)
1011

1012
1013
1014
1015
1016
1017
        # Don't log the most common case
        if convert_type != "none":
            logger.info(
                "Resolved `--convert auto` to `--convert %s`. "
                "Pass the value explicitly to silence this message.",
                convert_type)
1018
1019

        return convert_type
1020

1021
    def _get_supported_generation_tasks(
1022
        self,
1023
1024
        architectures: list[str],
        convert_type: ConvertType,
1025
1026
1027
    ) -> list[_ResolvedTask]:
        registry = self.registry

1028
        if registry.is_transcription_only_model(architectures, self):
1029
1030
            return ["transcription"]

1031
        # TODO: Use get_supported_generation_tasks once V0 is removed
1032
        supported_tasks = list[_ResolvedTask]()
1033
1034
        if (registry.is_text_generation_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["generate"]):
1035
1036
            supported_tasks.append("generate")

1037
1038
        if registry.is_transcription_model(architectures, self):
            supported_tasks.append("transcription")
1039
1040

        return supported_tasks
1041

1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
    def _get_default_pooling_task(
        self,
        architectures: list[str],
    ) -> Literal["embed", "classify", "reward"]:
        if self.registry.is_cross_encoder_model(architectures, self):
            return "classify"

        for arch in architectures:
            match = try_match_architecture_defaults(arch,
                                                    runner_type="pooling")
            if match:
                _, (_, convert_type) = match
                assert convert_type != "none"
                return convert_type

        return "embed"

1059
1060
    def _get_supported_pooling_tasks(
        self,
1061
1062
        architectures: list[str],
        convert_type: ConvertType,
1063
    ) -> list[_ResolvedTask]:
1064
        registry = self.registry
1065

1066
        # TODO: Use get_supported_pooling_tasks once V0 is removed
1067
        supported_tasks = list[_ResolvedTask]()
1068
1069
        if (registry.is_pooling_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["pooling"]):
1070
            supported_tasks.append("encode")
1071

1072
1073
1074
            extra_task = (self._get_default_pooling_task(architectures)
                          if convert_type == "none" else convert_type)
            supported_tasks.append(extra_task)
1075
1076
1077
1078
1079

        return supported_tasks

    def _get_supported_tasks(
        self,
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
        architectures: list[str],
        runner_type: RunnerType,
        convert_type: ConvertType,
    ) -> list[_ResolvedTask]:
        if runner_type == "generate":
            return self._get_supported_generation_tasks(
                architectures, convert_type)
        if runner_type == "pooling":
            return self._get_supported_pooling_tasks(architectures,
                                                     convert_type)
        if runner_type == "draft":
            return ["draft"]
1092

1093
        assert_never(runner_type)
1094

1095
1096
    def _parse_quant_hf_config(self, hf_config: PretrainedConfig):
        quant_cfg = getattr(hf_config, "quantization_config", None)
1097
        if quant_cfg is None:
1098
            # compressed-tensors uses a "compression_config" key
1099
            quant_cfg = getattr(hf_config, "compression_config", None)
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114

        else:
            # Set quant_method for ModelOpt models.
            producer_name = quant_cfg.get("producer", {}).get("name")
            if producer_name == "modelopt":
                quant_algo = quant_cfg.get("quantization",
                                           {}).get("quant_algo")
                if quant_algo == "FP8":
                    quant_cfg["quant_method"] = "modelopt"
                elif quant_algo == "NVFP4":
                    quant_cfg["quant_method"] = "modelopt_fp4"
                elif quant_algo is not None:
                    raise ValueError(
                        f"Unknown ModelOpt quant algo: {quant_algo}")

1115
1116
        return quant_cfg

1117
    def _verify_quantization(self) -> None:
1118
        supported_quantization = me_quant.QUANTIZATION_METHODS
1119
        optimized_quantization_methods = [
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
            "fp8",
            "modelopt",
            "gptq_marlin_24",
            "gptq_marlin",
            "awq_marlin",
            "fbgemm_fp8",
            "compressed-tensors",
            "experts_int8",
            "quark",
            "modelopt_fp4",
            "bitblas",
            "gptq_bitblas",
            "inc",
            "petit_nvfp4",
1134
        ]
1135
        if self.quantization is not None:
1136
1137
            self.quantization = cast(me_quant.QuantizationMethods,
                                     self.quantization)
1138
1139

        # Parse quantization method from the HF model config, if available.
1140
1141
1142
1143
1144
        quant_cfg = self._parse_quant_hf_config(self.hf_config)
        if quant_cfg is None and (text_config := getattr(
                self.hf_config, "text_config", None)):
            # Check the text config as well for multi-modal models.
            quant_cfg = self._parse_quant_hf_config(text_config)
1145

1146
        if quant_cfg is not None:
1147
            # Use the community standard 'quant_method'
1148
            quant_method = quant_cfg.get("quant_method", "").lower()
1149
1150

            # Normalize library names
1151
1152
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
1153

1154
            quant_cfg["quant_method"] = quant_method
1155

1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
            # Quantization methods which are overrides (i.e. they have a
            # `override_quantization_method` method) must be checked in order
            # of preference (this is particularly important for GPTQ).
            overrides = [
                "bitblas",
                "gptq_marlin_24",
                "gptq_marlin",
                "gptq_bitblas",
                "awq_marlin",
                "ipex",
                "moe_wna16",
1167
1168
                "modelopt",
                "modelopt_fp4",
1169
                "petit_nvfp4",
1170
1171
1172
1173
1174
1175
            ]
            quantization_methods = [
                q for q in supported_quantization if q not in overrides
            ]
            # Any custom overrides will be in quantization_methods so we place
            # them at the start of the list so custom overrides have preference
1176
            # over the built-in ones.
1177
1178
            quantization_methods = quantization_methods + overrides

1179
            # Detect which checkpoint is it
1180
            for name in quantization_methods:
1181
                method = me_quant.get_quantization_config(name)
1182
1183
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
1184
1185
1186
1187
                if quantization_override is not None:
                    # Raise error if the override is not custom (custom would
                    # be in QUANTIZATION_METHODS but not QuantizationMethods)
                    # and hasn't been added to the overrides list.
1188
                    if (name in get_args(me_quant.QuantizationMethods)
1189
1190
1191
1192
1193
1194
                            and name not in overrides):
                        raise ValueError(
                            f"Quantization method {name} is an override but "
                            "is has not been added to the `overrides` list "
                            "above. This is necessary to ensure that the "
                            "overrides are checked in order of preference.")
1195
1196
1197
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
1198

1199
            # Verify quantization configurations.
1200
            if self.quantization is None:
1201
1202
                self.quantization = quant_method
            elif self.quantization != quant_method:
1203
1204
                raise ValueError(
                    "Quantization method specified in the model config "
1205
                    f"({quant_method}) does not match the quantization "
1206
1207
1208
1209
1210
1211
1212
1213
                    f"method specified in the `quantization` argument "
                    f"({self.quantization}).")

        if self.quantization is not None:
            if self.quantization not in supported_quantization:
                raise ValueError(
                    f"Unknown quantization method: {self.quantization}. Must "
                    f"be one of {supported_quantization}.")
1214
            from vllm.platforms import current_platform
1215
            current_platform.verify_quantization(self.quantization)
1216
            if self.quantization not in optimized_quantization_methods:
1217
                logger.warning(
1218
                    "%s quantization is not fully "
1219
                    "optimized yet. The speed can be slower than "
1220
                    "non-quantized models.", self.quantization)
1221

1222
    def _verify_cuda_graph(self) -> None:
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
        # The `max_seq_len_to_capture` was incorrectly
        # based on the encoder's input length (448)
        # but not the decoder's larger input length (1500).
        # This change ensures the CUDA Graph captures the correct,
        # larger sequence length, allowing it to work as intended.
        effective_max_seq_len = self.max_model_len
        if self.is_encoder_decoder:
            effective_max_seq_len = max(
                effective_max_seq_len,
                getattr(self.hf_config, "max_source_positions", 0))
1233
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
1234
                                          effective_max_seq_len)
1235
        # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
1236
        ROCM_UNSUPPORTED_MODELS = ['mllama']
1237
1238
1239
1240
1241
1242
        unsupported_rocm = (self.hf_config.model_type
                            in ROCM_UNSUPPORTED_MODELS
                            or self.is_encoder_decoder)

        if (unsupported_rocm and not self.enforce_eager
                and current_platform.is_rocm()):
1243
1244
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1245
                "to eager mode.", self.hf_config.model_type)
1246
            self.enforce_eager = True
1247

1248
1249
    def _verify_bnb_config(self) -> None:
        """
1250
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
1251
        yet support CUDA graph.
1252
        # TODO Remove this when bitsandbytes supports.
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
        """
        is_bitsandbytes = self.quantization == "bitsandbytes"
        has_quantization_config = (getattr(self.hf_config,
                                           "quantization_config", None)
                                   is not None)
        is_8bit = (self.hf_config.quantization_config.get(
            "load_in_8bit", False) if has_quantization_config else False)
        if all([
                is_bitsandbytes,
                has_quantization_config,
                is_8bit,
                not self.enforce_eager,
        ]):
            logger.warning(
1267
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1268
                "fallback to the eager mode.")
1269

1270
1271
            self.enforce_eager = True

1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
    def _verify_with_expert_parallelism(self) -> None:
        num_expert_names = [
            "moe_num_experts",  # Dbrx
            "num_experts",  # Jamba
            "n_routed_experts",  # DeepSeek
            "num_local_experts",  # Mixtral
        ]
        num_experts = 0
        for name in num_expert_names:
            num_experts = getattr(self.hf_text_config, name, 0)
            if num_experts > 0:
                break
        if num_experts < 1:
            raise ValueError(
                "Number of experts in the model must be greater than 0 "
                "when expert parallelism is enabled.")

1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
    def verify_dual_chunk_attention_config(
        self,
        load_config: "LoadConfig",
    ) -> None:
        if hasattr(self.hf_config, "dual_chunk_attention_config"):
            # Try loading the sparse attention config
            from vllm.model_executor.model_loader.weight_utils import (
                get_sparse_attention_config)
            sparse_attn_config = get_sparse_attention_config(self, load_config)
            if sparse_attn_config:
                self.hf_config.dual_chunk_attention_config[
                    "sparse_attention_config"] = sparse_attn_config
                if "sparse_attention_enabled" not in \
                        self.hf_config.dual_chunk_attention_config:
                    self.hf_config.dual_chunk_attention_config[
                        "sparse_attention_enabled"] = True

1306
1307
1308
1309
            if envs.VLLM_ATTENTION_BACKEND != STR_DUAL_CHUNK_FLASH_ATTN_VAL:
                raise ValueError("please set VLLM_ATTENTION_BACKEND to "
                                 f"{STR_DUAL_CHUNK_FLASH_ATTN_VAL}")

1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
    def verify_async_output_proc(self, parallel_config, speculative_config,
                                 device_config) -> None:
        if not self.use_async_output_proc:
            # Nothing to check
            return

        if parallel_config.pipeline_parallel_size > 1:
            self.use_async_output_proc = False
            return

1320
        # Reminder: Please update docs/features/compatibility_matrix.md
1321
        # If the feature combo become valid
1322
        from vllm.platforms import current_platform
1323
        if not current_platform.is_async_output_supported(self.enforce_eager):
1324
1325
1326
1327
1328
1329
1330
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

1331
        # Async postprocessor is not necessary for pooling models
1332
        # since there is no token generation
1333
        if self.runner_type == "pooling":
1334
1335
            self.use_async_output_proc = False

1336
        # Reminder: Please update docs/features/compatibility_matrix.md
1337
        # If the feature combo become valid
1338
1339
1340
        if speculative_config:
            self.use_async_output_proc = False

1341
1342
1343
1344
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
1345
1346
1347
1348
1349
1350

        if parallel_config.distributed_executor_backend == "external_launcher":
            assert self.seed is not None, (
                "Seed must be set when using external launcher backend to "
                "make sure sampling results are the same across workers.")

1351
1352
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
1353
1354
1355
1356
1357
1358
1359
        tensor_parallel_size = parallel_config.tensor_parallel_size
        if total_num_attention_heads % tensor_parallel_size != 0:
            raise ValueError(
                f"Total number of attention heads ({total_num_attention_heads})"
                " must be divisible by tensor parallel size "
                f"({tensor_parallel_size}).")

1360
        if parallel_config.enable_expert_parallel:
1361
1362
            self._verify_with_expert_parallelism()

1363
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1364
        if pipeline_parallel_size > 1:
1365
1366
            if not self.registry.is_pp_supported_model(self.architectures,
                                                       self):
1367
1368
1369
1370
1371
1372
                raise NotImplementedError(
                    "Pipeline parallelism is not supported for this model. "
                    "Supported models implement the `SupportsPP` interface.")

            if self.use_async_output_proc:
                self.use_async_output_proc = False
1373

1374
1375
    def get_sliding_window(self) -> Optional[int]:
        """Get the sliding window size from the HF text config if present."""
1376
        return getattr(self.hf_text_config, "sliding_window", None)
1377
1378

    def get_vocab_size(self) -> int:
1379
        return getattr(self.hf_text_config, "vocab_size", 0)
1380

1381
    def get_hidden_size(self) -> int:
1382
        return getattr(self.hf_text_config, "hidden_size", 0)
1383

1384
1385
    @property
    def is_deepseek_mla(self) -> bool:
1386
1387
1388
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
bigmoyan's avatar
bigmoyan committed
1389
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp', 'kimi_k2'):
1390
1391
1392
1393
1394
1395
1396
1397
            return self.hf_text_config.kv_lora_rank is not None
        elif self.hf_text_config.model_type == 'eagle':
            # if the model is an EAGLE module, check for the
            # underlying architecture
            return self.hf_text_config.model.model_type in \
                    ('deepseek_v2', 'deepseek_v3') \
                and self.hf_text_config.kv_lora_rank is not None
        return False
1398

1399
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1400
        # TODO remove hard code
1401
        if self.is_deepseek_mla:
1402
1403
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1404
            if self.use_mla:
1405
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1406
1407
1408
1409
1410
            else:
                qk_nope_head_dim = getattr(self.hf_text_config,
                                           "qk_nope_head_dim", 0)
                if qk_rope_head_dim and qk_nope_head_dim:
                    return qk_rope_head_dim + qk_nope_head_dim
1411

1412
1413
1414
1415
1416
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1417
1418
1419
        if self.is_attention_free:
            return 0

1420
1421
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1422
            return self.hf_text_config.head_dim
1423

1424
1425
1426
1427
1428
        # NOTE: Some models (such as PLaMo2.1) use `hidden_size_per_head`
        if getattr(self.hf_text_config, "hidden_size_per_head",
                   None) is not None:
            return self.hf_text_config.hidden_size_per_head

1429
        # FIXME(woosuk): This may not be true for all models.
1430
1431
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1432

1433
1434
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1435
        # For GPTBigCode & Falcon:
1436
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1437
1438
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1439
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1440
        new_decoder_arch_falcon = (
1441
            self.hf_config.model_type in falcon_model_types
1442
            and getattr(self.hf_config, "new_decoder_architecture", False))
1443
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1444
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1445
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1446
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1447
            return 1
1448

1449
        # For DBRX and MPT
1450
1451
1452
1453
1454
        if self.hf_config.model_type == "mpt":
            if "kv_n_heads" in self.hf_config.attn_config:
                return self.hf_config.attn_config["kv_n_heads"]
            return self.hf_config.num_attention_heads
        if self.hf_config.model_type == "dbrx":
1455
1456
1457
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1458
1459
1460
1461
1462
1463
1464
1465
        if self.hf_config.model_type == "nemotron-nas":
            for block in self.hf_config.block_configs:
                if not block.attention.no_op:
                    return self.hf_config.num_attention_heads \
                        // block.attention.n_heads_in_group

            raise RuntimeError("Couldn't determine number of kv heads")

1466
1467
1468
        if self.is_attention_free:
            return 0

1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
        attributes = [
            # For Falcon:
            "n_head_kv",
            "num_kv_heads",
            # For LLaMA-2:
            "num_key_value_heads",
            # For ChatGLM:
            "multi_query_group_num",
        ]
        for attr in attributes:
1479
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1480
1481
1482
1483
1484
            if num_kv_heads is not None:
                return num_kv_heads

        # For non-grouped-query attention models, the number of KV heads is
        # equal to the number of attention heads.
1485
        return self.hf_text_config.num_attention_heads
1486
1487
1488

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1489
1490
1491
1492
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1493
1494
1495
1496
1497
1498
1499
        total_num_kv_heads = self.get_total_num_kv_heads()
        # If tensor parallelism is used, we divide the number of KV heads by
        # the tensor parallel size. We will replicate the KV heads in the
        # case where the number of KV heads is smaller than the tensor
        # parallel size so each GPU has at least one KV head.
        return max(1,
                   total_num_kv_heads // parallel_config.tensor_parallel_size)
1500

1501
1502
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1503
1504
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1505

1506
    def get_layers_start_end_indices(
1507
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1508
        from vllm.distributed.utils import get_pp_indices
1509
        if (self.hf_text_config.model_type == "deepseek_mtp"
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1510
                or self.hf_config.model_type == "mimo_mtp"
1511
                or self.hf_config.model_type == "glm4_moe_mtp"
1512
1513
                or self.hf_config.model_type == "ernie_mtp"
                or self.hf_config.model_type == "qwen3_next_mtp"):
1514
1515
1516
1517
1518
            total_num_hidden_layers = getattr(self.hf_text_config,
                                              "num_nextn_predict_layers", 0)
        else:
            total_num_hidden_layers = getattr(self.hf_text_config,
                                              "num_hidden_layers", 0)
1519
1520
1521
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1522
1523
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1524
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1525

1526
1527
1528
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
        start, end = self.get_layers_start_end_indices(parallel_config)
        return end - start
Mor Zusman's avatar
Mor Zusman committed
1529

1530
1531
1532
1533
1534
1535
1536
1537
    def get_num_layers_by_block_type(
        self,
        parallel_config: "ParallelConfig",
        block_type: LayerBlockType = LayerBlockType.attention,
    ) -> int:
        # This function relies on 'layers_block_type' in hf_config,
        # for w/o this attribute, we will need to have workarounds like so
        attn_block_type = block_type == LayerBlockType.attention
1538
1539
1540
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
        start, end = self.get_layers_start_end_indices(parallel_config)

        if is_transformer:
            # Handle the basic case first
            return end - start if attn_block_type else 0
        elif self.is_attention_free:
            # Attention free
            # Note that this code assumes there
            # is only one type of attention-free block type.
            return 0 if attn_block_type else end - start
1551
1552
1553
1554
        elif self.has_noops:
            block_configs = self.hf_config.block_configs
            return sum(not bc.attention.no_op
                       for bc in block_configs[start:end])
1555
        else:
1556
            # Hybrid model Jamba
1557
            layers_block_type_value = getattr(self.hf_text_config,
1558
                                              "layers_block_type", None)
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
            if layers_block_type_value is not None:
                if hasattr(self.hf_text_config,
                           "model_type") and (self.hf_text_config.model_type
                                              == "zamba2"):
                    if attn_block_type:
                        return sum(t == "hybrid"
                                   for t in layers_block_type_value[start:end])
                    else:
                        return self.get_num_layers(parallel_config)
                return sum(t == block_type.value
                           for t in layers_block_type_value[start:end])

            # Hybrid model Minimax
            attn_type_list = getattr(self.hf_config, "attn_type_list", None)
            if attn_type_list:
                return sum(t == 1 for t in attn_type_list[start:end])

1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
            # Hybrid model Qwen3Next
            layer_types_value = getattr(self.hf_config, "layer_types", None)
            if layer_types_value is not None:
                if getattr(block_type, "value", block_type) == "attention":
                    return sum(t == "full_attention"
                               for t in layer_types_value[start:end])
                elif getattr(block_type, "value",
                             block_type) == "linear_attention":
                    return sum(t == "linear_attention"
                               for t in layer_types_value[start:end])
                else:
                    return sum(t == getattr(block_type, "value", block_type)
                               for t in layer_types_value[start:end])

            if (layers_block_type_value is None and attn_type_list is None
                    and layer_types_value is None):
1592
1593
                raise ValueError(
                    "The model is an hybrid without a"
1594
1595
                    "layers_block_type or an attn_type_list, or a layer_types "
                    "in the hf_config, cannot determine the num of "
1596
1597
                    f"{block_type.value} layers")

1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
    def get_mamba_chunk_size(self) -> Optional[int]:
        """
        Returns the mamba chunk size if it exists
        """
        # used by e.g. Bamba, FalconH1, Granite, PLaMo2
        chunk_size = getattr(self.hf_text_config, "mamba_chunk_size", None)
        if chunk_size is None:
            # used by e.g. Mamba2, NemotronH, Zamba
            chunk_size = getattr(self.hf_text_config, "chunk_size", None)
        return chunk_size

1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
    def get_multimodal_config(self) -> "MultiModalConfig":
        """
        Get the multimodal configuration of the model.

        Raises:
            ValueError: If the model is not multimodal.
        """
        if self.multimodal_config is None:
            raise ValueError("The model is not multimodal.")

        return self.multimodal_config

1621
    def try_get_generation_config(self) -> dict[str, Any]:
1622
1623
1624
        """
        This method attempts to retrieve the non-default values of the
        generation config for this model.
1625

1626
1627
1628
1629
1630
1631
1632
1633
        The generation config can contain information about special tokens, as
        well as sampling parameters. Which is why this method exists separately
        to `get_diff_sampling_param`.

        Returns:
            A dictionary containing the non-default generation config.
        """
        if self.generation_config in {"auto", "vllm"}:
1634
            config = try_get_generation_config(
1635
                self.hf_config_path or self.model,
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
                trust_remote_code=self.trust_remote_code,
                revision=self.revision,
            )
        else:
            config = try_get_generation_config(
                self.generation_config,
                trust_remote_code=self.trust_remote_code,
            )

        if config is None:
            return {}

        return config.to_diff_dict()

1650
    def get_diff_sampling_param(self) -> dict[str, Any]:
1651
        """
1652
1653
1654
1655
1656
1657
1658
1659
1660
        This method returns a dictionary containing the non-default sampling
        parameters with `override_generation_config` applied.

        The default sampling parameters are:

        - vLLM's neutral defaults if `self.generation_config="vllm"`
        - the model's defaults if `self.generation_config="auto"`
        - as defined in `generation_config.json` if
            `self.generation_config="path/to/generation_config/dir"`
1661
1662

        Returns:
1663
            A dictionary containing the non-default sampling parameters.
1664
        """
1665
        if self.generation_config == "vllm":
1666
1667
1668
1669
1670
1671
1672
            config = {}
        else:
            config = self.try_get_generation_config()

        # Overriding with given generation config
        config.update(self.override_generation_config)

1673
1674
1675
1676
1677
1678
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1679
            "max_new_tokens",
1680
1681
1682
1683
1684
1685
        ]
        if any(p in config for p in available_params):
            diff_sampling_param = {
                p: config.get(p)
                for p in available_params if config.get(p) is not None
            }
1686
1687
1688
1689
1690
            # Huggingface definition of max_new_tokens is equivalent
            # to vLLM's max_tokens
            if "max_new_tokens" in diff_sampling_param:
                diff_sampling_param["max_tokens"] = diff_sampling_param.pop(
                    "max_new_tokens")
1691
1692
        else:
            diff_sampling_param = {}
1693
1694
1695
1696
1697
1698
1699

        if diff_sampling_param:
            logger.warning_once(
                "Default sampling parameters have been overridden by the "
                "model's Hugging Face generation config recommended from the "
                "model creator. If this is not intended, please relaunch "
                "vLLM instance with `--generation-config vllm`.")
1700
1701
        return diff_sampling_param

1702
    @property
1703
    def is_encoder_decoder(self) -> bool:
1704
        """Extract the HF encoder/decoder model flag."""
1705
        """
1706
        For Mllama, VLLM overrides HF's is_encoder_decoder flag and sets it to
1707
1708
        True to enable cross-attention
        """
1709
1710
1711
1712
1713
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1714

1715
1716
1717
1718
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1719
1720
1721
1722
    @property
    def is_multimodal_raw_input_only_model(self) -> bool:
        return self._model_info.supports_multimodal_raw_input_only

1723
1724
    @property
    def is_cross_encoder(self) -> bool:
1725
1726
        return (self._model_info.supports_cross_encoding
                or self.convert_type == "classify")
1727

1728
    @property
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
    def is_pp_supported(self) -> bool:
        return self._model_info.supports_pp

    @property
    def is_attention_free(self) -> bool:
        return self._model_info.is_attention_free

    @property
    def is_hybrid(self) -> bool:
        return self._model_info.is_hybrid

    @property
    def has_noops(self) -> bool:
        return self._model_info.has_noops

    @property
    def has_inner_state(self):
        return self._model_info.has_inner_state
1747

1748
1749
    @property
    def is_v1_compatible(self) -> bool:
1750
1751
1752
1753
1754
        return not self._model_info.supports_v0_only

    @property
    def use_mla(self) -> bool:
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1755

1756
1757
    @property
    def is_matryoshka(self) -> bool:
1758
        return (bool(getattr(self.hf_config, "matryoshka_dimensions", None))
1759
1760
                or getattr(self.hf_config, "is_matryoshka", False))

1761
1762
1763
1764
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)

1765
1766
1767
1768
1769
1770
    @property
    def use_pad_token(self) -> bool:
        # cross_encoder models defaults to using pad_token.
        # `llm as reranker` models defaults to not using pad_token.
        return getattr(self.hf_config, "use_pad_token", True)

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
    @property
    def head_dtype(self) -> torch.dtype:
        """
        "head" refers to the last Linear layer(s) of an LLM,
        such as the lm_head in a generation model,
        or the score or classifier in a classification model.

        The default head_dtype based on runner_type.\n
        - The pooling model defaults to using fp32 head,
        you can use --hf-overrides '{"head_dtype": "model"}' to disable it.\n
        - The generate model defaults to not using fp32 head,
        you can use --hf-overrides '{"head_dtype": "float32"}' to enable it.
        """
        head_dtype = _get_head_dtype(config=self.hf_config,
                                     dtype=self.dtype,
                                     runner_type=self.runner_type)

        if head_dtype not in current_platform.supported_dtypes:
            logger.warning_once(
                "The current platform does not support [%s] head dtype, "
                "fallback to model dtype [%s].", head_dtype, self.dtype)
            return self.dtype

        logger.debug_once("head dtype: %s", head_dtype)
        return head_dtype

1797
    def get_and_verify_max_len(self, max_model_len: int):
1798
1799
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
1800
        tokenizer_config = None
1801
1802
        if (self.runner_type == "pooling" and getattr(
                self.hf_config, "position_embedding_type", "") == "absolute"):
1803
1804
1805
1806
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
                revision=self.tokenizer_revision)
1807
1808
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
1809
            tokenizer_config=tokenizer_config,
1810
1811
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
1812
            sliding_window=self.get_sliding_window(),
1813
1814
            spec_target_max_model_len=self.spec_target_max_model_len,
            encoder_config=self.encoder_config)
1815
        logger.info("Using max model len %s", max_model_len)
1816
1817
        return max_model_len

1818

1819
Device = Literal["auto", "cuda", "cpu", "tpu", "xpu"]
1820
1821
1822


@config
1823
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
1824
class DeviceConfig:
1825
1826
    """Configuration for the device to use for vLLM execution."""

1827
    device: SkipValidation[Optional[Union[Device, torch.device]]] = "auto"
1828
    """Device type for vLLM execution.
1829
1830
1831
    This parameter is deprecated and will be
    removed in a future release.
    It will now be set automatically based
1832
    on the current platform."""
1833
1834
1835
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
1836

1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # the device/platform information will be summarized
        # by torch/vllm automatically.
1852
        factors: list[Any] = []
1853
1854
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1855
1856
        return hash_str

1857
1858
    def __post_init__(self):
        if self.device == "auto":
1859
            # Automated device type detection
1860
            from vllm.platforms import current_platform
1861
            self.device_type = current_platform.device_type
1862
            if not self.device_type:
1863
1864
1865
1866
                raise RuntimeError(
                    "Failed to infer device type, please set "
                    "the environment variable `VLLM_LOGGING_LEVEL=DEBUG` "
                    "to turn on verbose logging to help debug the issue.")
1867
1868
        else:
            # Device type is assigned explicitly
1869
1870
1871
1872
            if isinstance(self.device, str):
                self.device_type = self.device
            elif isinstance(self.device, torch.device):
                self.device_type = self.device.type
1873
1874

        # Some device types require processing inputs on CPU
1875
        if self.device_type in ["tpu"]:
1876
            self.device = None
1877
1878
1879
1880
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1881

1882
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
1883
                            "mlp_speculator", "draft_model", "deepseek_mtp",
1884
                            "ernie_mtp", "qwen3_next_mtp"]
1885
1886
1887


@config
1888
@dataclass
1889
class SpeculativeConfig:
1890
    """Configuration for speculative decoding."""
1891

1892
    # General speculative decoding control
1893
    num_speculative_tokens: SkipValidation[int] = None  # type: ignore
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
    """The number of speculative tokens, if provided. It will default to the
    number in the draft model config if present, otherwise, it is required."""
    model: Optional[str] = None
    """The name of the draft model, eagle head, or additional weights, if
    provided."""
    method: Optional[SpeculativeMethod] = None
    """The name of the speculative method to use. If users provide and set the
    `model` param, the speculative method type will be detected automatically
    if possible, if `model` param is not provided, the method name must be
    provided.

    If using `ngram` method, the related configuration `prompt_lookup_max` and
    `prompt_lookup_min` should be considered."""
1907
    draft_tensor_parallel_size: Optional[int] = None
1908
1909
    """The degree of the tensor parallelism for the draft model. Can only be 1
    or the same as the target model's tensor parallel size."""
1910
    disable_logprobs: bool = True
1911
1912
1913
    """If set to True, token log probabilities are not returned during
    speculative decoding. If set to False, token log probabilities are returned
    according to the log probability settings in SamplingParams."""
1914

1915
    # Draft model configuration
1916
    quantization: Optional[me_quant.QuantizationMethods] = None
1917
1918
1919
    """Quantization method that was used to quantize the draft model weights.
    If `None`, we assume the model weights are not quantized. Note that it only
    takes effect when using the draft model-based speculative method."""
1920
    max_model_len: Optional[int] = None
1921
1922
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
1923
    revision: Optional[str] = None
1924
1925
1926
    """The specific model version to use for the draft model. It can be a
    branch name, a tag name, or a commit id. If unspecified, will use the
    default version."""
1927
    code_revision: Optional[str] = None
1928
1929
1930
    """The specific revision to use for the draft model code on Hugging Face
    Hub. It can be a branch name, a tag name, or a commit id. If unspecified,
    will use the default version."""
1931

1932
    # Advanced control
1933
    disable_by_batch_size: Optional[int] = None
1934
1935
1936
1937
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
1938
    prompt_lookup_max: Optional[int] = None
1939
1940
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
1941
    prompt_lookup_min: Optional[int] = None
1942
1943
1944
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

1945
    speculative_token_tree: Optional[str] = None
1946
    """Specifies the tree structure for speculative token generation.
1947
    """
1948
    # required configuration params passed from engine
1949
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
1950
    """The configuration of the target model."""
1951
1952
    target_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
1953
    """The parallel configuration for the target model."""
1954
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
1955
1956
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
1957
    disable_log_stats: SkipValidation[bool] = None  # type: ignore
1958
1959
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
1960
1961

    # params generated in the post-init stage
1962
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
1963
    """The configuration of the draft model initialized internal."""
1964
1965
    draft_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
1966
    """The parallel configuration for the draft model initialized internal."""
1967

1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
1980
        factors: list[Any] = []
1981
1982
1983
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
1984
1985
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1986
1987
        return hash_str

1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
    @staticmethod
    def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
        if hf_config.model_type == "deepseek_v3":
            hf_config.model_type = "deepseek_mtp"
        if hf_config.model_type == "deepseek_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
            hf_config.update({
                "n_predict": n_predict,
                "architectures": ["DeepSeekMTPModel"]
            })
1998
1999
2000
2001
2002
2003
2004
2005
2006

        if hf_config.architectures[0] == "MiMoForCausalLM":
            hf_config.model_type = "mimo_mtp"
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
            hf_config.update({
                "num_hidden_layers": 0,
                "n_predict": n_predict,
                "architectures": ["MiMoMTPModel"]
            })
Yuxuan Zhang's avatar
Yuxuan Zhang committed
2007
2008
2009
2010
2011
2012
2013
2014
2015

        if hf_config.architectures[0] == "Glm4MoeForCausalLM":
            hf_config.model_type = "glm4_moe_mtp"
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
            hf_config.update({
                "num_hidden_layers": 0,
                "n_predict": n_predict,
                "architectures": ["Glm4MoeMTPModel"]
            })
2016

2017
2018
2019
2020
2021
2022
2023
2024
        if hf_config.model_type == "ernie4_5_moe":
            hf_config.model_type = "ernie_mtp"
        if hf_config.model_type == "ernie_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
            hf_config.update({
                "n_predict": n_predict,
                "architectures": ["ErnieMTPModel"]
            })
2025
2026
2027
2028
2029
2030
2031
2032
2033

        if hf_config.model_type == "qwen3_next":
            hf_config.model_type = "qwen3_next_mtp"
        if hf_config.model_type == "qwen3_next_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
            hf_config.update({
                "n_predict": n_predict,
                "architectures": ["Qwen3NextMTP"]
            })
2034

2035
2036
        return hf_config

2037
    def __post_init__(self):
2038

2039
2040
2041
2042
2043
2044
2045
        # Note: "method" is a new parameter that helps to extend the
        # configuration of non-model-based proposers, and the "model" parameter
        # will be used to set the draft model, eagle head, or additional weight
        # when needed. If users do not specify "method", the speculative method
        # will be detected automatically if possible. If the speculative method
        # can not be detected, it will be considered as the "draft_model" by
        # default.
2046
2047
2048
2049

        if self.model is None and self.num_speculative_tokens is not None:
            # TODO(Shangming): Refactor mtp configuration logic when supporting
            # mtp acceleration for more models besides deepseek_v3
2050
            if self.target_model_config and \
2051
2052
                (self.target_model_config.hf_text_config.model_type \
                        == "deepseek_v3" or
2053
                    self.target_model_config.hf_text_config.model_type in
2054
                        ("mimo","ernie4_5_moe", "qwen3_next")):
2055
2056
                # use the draft model from the same model:
                self.model = self.target_model_config.model
2057
2058
2059
2060
                # Align the quantization of draft model for cases such as
                # --quantization fp8 with a bf16 checkpoint.
                if not self.quantization:
                    self.quantization = self.target_model_config.quantization
2061
2062
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2063
            else:
2064
2065
2066
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2067
2068
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2069
2070
2071
2072
2073
2074
2075
        if self.method is None and (self.model is not None
                                    and self.model in ("ngram", "[ngram]")):
            self.method = "ngram"

        if self.method in ("ngram", "[ngram]"):
            # Unified to "ngram" internally
            self.method = "ngram"
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
            # Set default values if not provided
            if (self.prompt_lookup_min is None
                    and self.prompt_lookup_max is None):
                # TODO(woosuk): Tune these values. They are arbitrarily chosen.
                self.prompt_lookup_min = 5
                self.prompt_lookup_max = 5
            elif self.prompt_lookup_min is None:
                assert self.prompt_lookup_max is not None
                self.prompt_lookup_min = self.prompt_lookup_max
            elif self.prompt_lookup_max is None:
                assert self.prompt_lookup_min is not None
                self.prompt_lookup_max = self.prompt_lookup_min

            # Validate values
2090
            if self.prompt_lookup_min < 1:
2091
2092
2093
2094
2095
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must be > 0")
            if self.prompt_lookup_max < 1:
                raise ValueError(
                    f"prompt_lookup_max={self.prompt_lookup_max} must be > 0")
2096
            if self.prompt_lookup_min > self.prompt_lookup_max:
2097
2098
2099
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2100

2101
2102
2103
            # TODO: current we still need extract vocab_size from target model
            # config, in future, we may try refactor it out, and set
            # draft related config as None here.
2104
2105
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2106
        else:
2107
2108
2109
2110
2111
2112
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
2113
                    runner="draft",
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
                    tokenizer=self.target_model_config.tokenizer,
                    tokenizer_mode=self.target_model_config.tokenizer_mode,
                    trust_remote_code=self.target_model_config.
                    trust_remote_code,
                    allowed_local_media_path=self.target_model_config.
                    allowed_local_media_path,
                    dtype=self.target_model_config.dtype,
                    seed=self.target_model_config.seed,
                    revision=self.revision,
                    code_revision=self.code_revision,
                    tokenizer_revision=self.target_model_config.
                    tokenizer_revision,
                    spec_target_max_model_len=self.target_model_config.
                    max_model_len,
                    quantization=self.quantization,
                    enforce_eager=self.target_model_config.enforce_eager,
                    max_seq_len_to_capture=self.target_model_config.
                    max_seq_len_to_capture,
                    max_logprobs=self.target_model_config.max_logprobs,
                    hf_overrides=SpeculativeConfig.hf_config_override,
                )
2135

2136
                # Automatically detect the method
2137
                if self.method in ('eagle', 'eagle3'):
2138
                    pass
2139
2140
2141
2142
2143
                # examples:
                # yuhuili/EAGLE-LLaMA3-Instruct-8B
                # yuhuili/EAGLE3-LLaMA3.1-Instruct-8B
                # AngelSlim/Qwen3-8B_eagle3
                elif "eagle-" in self.draft_model_config.model.lower():
2144
                    self.method = "eagle"
2145
2146
                elif "eagle3" in self.draft_model_config.model.lower():
                    self.method = "eagle3"
2147
2148
2149
2150
2151
                elif self.draft_model_config.hf_config.model_type == "medusa":
                    self.method = "medusa"
                elif (self.draft_model_config.hf_config.model_type ==
                      "mlp_speculator"):
                    self.method = "mlp_speculator"
2152
                elif (self.draft_model_config.hf_config.model_type
Yuxuan Zhang's avatar
Yuxuan Zhang committed
2153
                      in ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp")):
Jiayi Yao's avatar
Jiayi Yao committed
2154
2155
2156
2157
2158
2159
2160
                    self.method = "deepseek_mtp"
                    if self.num_speculative_tokens > 1:
                        logger.warning(
                                "All Deepseek MTP models only have " \
                                "one layer. Might need some code changes " \
                                "to support multiple layers."
                            )
2161
2162
2163
2164
2165
2166
2167
2168
2169
                elif (self.draft_model_config.hf_config.model_type ==
                      "ernie_mtp"):
                    self.method = "ernie_mtp"
                    if self.num_speculative_tokens > 1:
                        logger.warning(
                                "All Ernie MTP models only have " \
                                "one layer. Might need some code changes " \
                                "to support multiple layers."
                            )
2170
2171
2172
2173
2174
2175
2176
2177
2178
                elif (self.draft_model_config.hf_config.model_type ==
                      "qwen3_next_mtp"):
                    self.method = "qwen3_next_mtp"
                    if self.num_speculative_tokens > 1:
                        logger.warning(
                                "All Qwen3Next MTP models only have " \
                                "one layer. Might need some code changes " \
                                "to support multiple layers."
                            )
2179
                else:
2180
                    self.method = "draft_model"
2181
2182
2183
2184
2185
                    raise NotImplementedError(
                        "Speculative decoding with draft model is not "
                        "supported yet. Please consider using other "
                        "speculative decoding methods such as ngram, medusa, "
                        "eagle, or deepseek_mtp.")
2186
2187

                # Replace hf_config for EAGLE draft_model
2188
                if self.method in ("eagle", "eagle3"):
2189
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2190
                        raise ValueError(
2191
2192
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2193

2194
2195
                    from vllm.transformers_utils.configs import (
                        SpeculatorsConfig)
2196
2197
                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
2198

2199
                    if isinstance(self.draft_model_config.hf_config,
2200
                                  (EAGLEConfig, SpeculatorsConfig)):
2201
2202
2203
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2204
                            self.draft_model_config.hf_config,
2205
2206
                            method=self.method,
                            model_type="eagle")
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
                        self.draft_model_config.hf_config = eagle_config

                if (self.num_speculative_tokens is not None
                        and hasattr(self.draft_model_config.hf_config,
                                    "num_lookahead_tokens")):
                    self.draft_model_config.hf_config.num_lookahead_tokens = \
                    self.num_speculative_tokens

                n_predict = getattr(self.draft_model_config.hf_config,
                                    "n_predict", None)
                if n_predict is not None:
                    if self.num_speculative_tokens is None:
                        # Default to max value defined in draft model config.
                        self.num_speculative_tokens = n_predict
                    elif self.num_speculative_tokens > n_predict and \
                            self.num_speculative_tokens % n_predict != 0:
                        # Ensure divisibility for MTP module reuse.
                        raise ValueError(
                            f"num_speculative_tokens:{self.num_speculative_tokens}"
                            f" must be divisible by {n_predict=}")

2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
                if self.speculative_token_tree is None:
                    # Generate chain of tokens.
                    self.speculative_token_tree = str([
                        (i + 1) * (0, )
                        for i in range(self.num_speculative_tokens)
                    ])
                else:
                    # Sort the token tree breadth-first.
                    tree_choices = ast.literal_eval(
                        self.speculative_token_tree)
                    self.speculative_token_tree = str(
                        sorted(tree_choices, key=lambda t: (len(t), t)))

2241
2242
2243
2244
2245
2246
                self.draft_tensor_parallel_size = \
                    SpeculativeConfig._verify_and_get_draft_tp(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size,
                        self.draft_model_config.hf_config
                )
2247

2248
2249
2250
2251
2252
2253
                self.draft_model_config.max_model_len = (
                    SpeculativeConfig._maybe_override_draft_max_model_len(
                        self.max_model_len,
                        self.draft_model_config.max_model_len,
                        self.target_model_config.max_model_len,
                    ))
2254

2255
2256
2257
2258
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
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
    @staticmethod
    def _maybe_override_draft_max_model_len(
        speculative_max_model_len: Optional[int],
        draft_max_model_len: int,
        target_max_model_len: int,
    ) -> int:
        """Determine the max sequence len for the draft model. This is usually
        the draft_max_model_len, but may be the target_max_model_len if it is
        less than the draft_max_model_len, or may be speculative_max_model_len
        if it is specified.

        This is necessary so that sequences do not exceed the capacity of the
        draft model or the target model.

        speculative_max_model_len is mainly used for testing that sequences can
        skip speculation.
        """

        if speculative_max_model_len is not None:

            if speculative_max_model_len > draft_max_model_len:
                raise ValueError(f"{speculative_max_model_len=} cannot be "
                                 f"larger than {draft_max_model_len=}")

            if speculative_max_model_len > target_max_model_len:
                raise ValueError(f"{speculative_max_model_len=} cannot be "
                                 f"larger than {target_max_model_len=}")

            return speculative_max_model_len

        return min(
            draft_max_model_len,
            target_max_model_len,
        )

2295
    @staticmethod
2296
    def _verify_and_get_draft_tp(
2297
2298
2299
2300
2301
2302
            target_parallel_config: ParallelConfig,
            speculative_draft_tensor_parallel_size: Optional[int],
            draft_hf_config: PretrainedConfig) -> int:
        """
        Verifies and adjusts the tensor parallel size for a draft model
        specified using speculative_draft_tensor_parallel_size.
2303
        """
2304
2305
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2306
        if speculative_draft_tensor_parallel_size is None:
2307
2308
2309
2310
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2311
2312
2313
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2314
2315
2316
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2317
2318
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2319
            raise ValueError(
2320
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2321
                f"other value than 1 or target model tensor_parallel_size")
2322
        return speculative_draft_tensor_parallel_size
2323

2324
2325
2326
2327
2328
2329
2330
2331
2332
    @staticmethod
    def create_draft_parallel_config(
        target_parallel_config: ParallelConfig,
        speculative_draft_tensor_parallel_size: int,
    ) -> ParallelConfig:
        """Create a parallel config for use by the draft worker.

        This is mostly a copy of the target parallel config, except the tp_size.
        """
2333
2334
2335
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2336
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2337
2338
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
            max_parallel_loading_workers=target_parallel_config.
            max_parallel_loading_workers,
            disable_custom_all_reduce=target_parallel_config.
            disable_custom_all_reduce,
            ray_workers_use_nsight=target_parallel_config.
            ray_workers_use_nsight,
            placement_group=target_parallel_config.placement_group,
        )

        return draft_parallel_config

2350
2351
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2352
2353
2354
2355
2356
2357
        if self.num_speculative_tokens is None:
            raise ValueError(
                "num_speculative_tokens must be provided with "
                "speculative model unless the draft model config contains an "
                "n_predict parameter.")

2358
2359
2360
2361
2362
2363
2364
        if self.num_speculative_tokens <= 0:
            raise ValueError("Expected num_speculative_tokens to be greater "
                             f"than zero ({self.num_speculative_tokens}).")

        if self.draft_model_config:
            self.draft_model_config.verify_with_parallel_config(
                self.draft_parallel_config)
2365
2366
2367
2368
2369
2370

        if (self.disable_by_batch_size is not None
                and self.disable_by_batch_size < 2):
            raise ValueError("Expect the batch size threshold of disabling "
                             "speculative decoding is > 1, but got "
                             f"{self.disable_by_batch_size=}")
2371

2372
        eagle3_target_supported = ["llama", "qwen"]
2373
2374
2375
2376
        if self.method == "eagle3" and self.target_model_config and not any(
                supported_model in
                self.target_model_config.hf_text_config.model_type
                for supported_model in eagle3_target_supported):
2377
            raise ValueError(
2378
                f"Eagle3 is only supported for {eagle3_target_supported} models. "  # noqa: E501
2379
2380
                f"Got {self.target_model_config.hf_text_config.model_type=}")

2381
2382
        return self

2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
    @property
    def num_lookahead_slots(self) -> int:
        """The number of additional slots the scheduler should allocate per
        step, in addition to the slots allocated for each known token.

        This is equal to the number of speculative tokens, as each speculative
        token must be scored.
        """
        return self.num_speculative_tokens

2393
    def use_eagle(self) -> bool:
2394
2395
        return self.method in ("eagle", "eagle3", "deepseek_mtp", "ernie_mtp",
                               "qwen3_next_mtp")
2396

2397
    def __repr__(self) -> str:
2398
2399
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
2400
        num_spec_tokens = self.num_speculative_tokens
2401
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
2402
2403


2404
@config
2405
@dataclass
2406
class MultiModalConfig:
2407
2408
    """Controls the behavior of multimodal models."""

2409
2410
    limit_per_prompt: dict[str, int] = \
        cast(dict[str, int], get_field(ModelConfig, "limit_mm_per_prompt"))
2411
    """
2412
    The maximum number of input items allowed per prompt for each modality.
2413
    Defaults to 1 (V0) or 999 (V1) for each modality.
2414
2415

    For example, to allow up to 16 images and 2 videos per prompt:
2416
    `{"image": 16, "video": 2}`
2417
2418
    """

2419
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
2420
2421
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
2422
2423
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """

2424
2425
2426
    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
2427
    `transformers.AutoProcessor.from_pretrained`.
2428
2429
2430
2431

    The available overrides depend on the model that is being run.

    For example, for Phi-3-Vision:
2432
    `{"num_crops": 4}`.
2433
2434
    """

2435
    mm_processor_cache_gb: float = 4
2436
    """
2437
2438
2439
2440
2441
2442
2443
    The size (in GiB) of the multi-modal processor cache, which is used to

    This cache is duplicated for each API process and engine core process,
    resulting in a total memory usage of
    `mm_processor_cache_gb * (api_server_count + data_parallel_size)`.

    Set to `0` to disable this cache completely (not recommended).
2444
2445
    """

2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
    mm_encoder_tp_mode: MMEncoderTPMode = "weights"
    """
    Indicates how to optimize multi-modal encoder inference using
    tensor parallelism (TP).

    - `"weights"`: Within the same vLLM engine, split the weights of
        each layer across TP ranks. (default TP behavior)
    - `"data"`: Within the same vLLM engine, split the batched input data
        across TP ranks to process the data in parallel, while hosting
        the full weights on each TP rank.
        This batch-level DP is not to be confused with API request-level
        DP (which is controlled by `--data-parallel-size`).
        This is only supported on a per-model basis and falls back to
        `"weights"` if the encoder does not support DP.
    """

2462
2463
2464
    interleave_mm_strings: bool = False
    """
    Enable fully interleaved support for multimodal prompts.
2465
2466
2467
2468
    """

    skip_mm_profiling: bool = False
    """
2469
    When enabled, skips multimodal memory profiling and only profiles with
2470
2471
2472
2473
2474
    language backbone model during engine initialization.

    This reduces engine startup time but shifts the responsibility to users for
    estimating the peak memory usage of the activation of multimodal encoder and
    embedding cache.
2475
2476
    """

2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
2491
        factors: list[Any] = []
2492
2493
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2494
2495
        return hash_str

2496
2497
2498
2499
2500
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
2501
2502
2503
2504
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
2505

2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
    def merge_mm_processor_kwargs(
        self,
        inference_kwargs: Mapping[str, object],
    ) -> dict[str, object]:
        """
        Get the keyword arguments to pass to the multi-modal processor
        according to the extra arguments passed during inference.
        """
        kwargs = self.mm_processor_kwargs or {}
        return kwargs | dict(inference_kwargs)
2516

2517

2518
@config
2519
2520
@dataclass
class PoolerConfig:
2521
    """Controls the behavior of output pooling in pooling models."""
2522
2523

    pooling_type: Optional[str] = None
2524
    """
2525
    The pooling method of the pooling model. This should be a key in
2526
    [`vllm.model_executor.layers.pooler.PoolingType`][].
2527
2528
    """

2529
    ## for embeddings models
2530
2531
    normalize: Optional[bool] = None
    """
2532
    Whether to normalize the embeddings outputs. Defaults to True.
2533
2534
2535
    """
    dimensions: Optional[int] = None
    """
2536
    Reduce the dimensions of embeddings if model
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
    support matryoshka representation. Defaults to None.
    """
    enable_chunked_processing: Optional[bool] = None
    """
    Whether to enable chunked processing for long inputs that exceed the model's
    maximum position embeddings. When enabled, long inputs will be split into
    chunks, processed separately, and then aggregated using weighted averaging.
    This allows embedding models to handle arbitrarily long text without CUDA
    errors. Defaults to False.
    """
    max_embed_len: Optional[int] = None
    """
    Maximum input length allowed for embedding generation. When set, allows
    inputs longer than max_embed_len to be accepted for embedding models.
    When an input exceeds max_embed_len, it will be handled according to 
    the original max_model_len validation logic. 
    Defaults to None (i.e. set to max_model_len).
2554
2555
    """

2556
2557
    ## for classification models
    activation: Optional[bool] = None
2558
    """
2559
    Whether to apply activation function to the classification outputs.
2560
2561
2562
2563
2564
    Defaults to True.
    """
    logit_bias: Optional[float] = None
    """
    If provided, apply classification logit biases. Defaults to None.
2565
2566
    """

2567
2568
2569
    ## for reward models
    softmax: Optional[bool] = None
    """
2570
    Whether to apply softmax to the reward outputs.
2571
    Defaults to True.
2572
    """
2573
2574
    step_tag_id: Optional[int] = None
    """
2575
    If set, only the score corresponding to the ``step_tag_id`` in the
2576
2577
2578
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """
2579
    returned_token_ids: Optional[list[int]] = None
2580
    """
2581
2582
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2583
2584
2585
    ``math-shepherd-mistral-7b-prm`` model.
    """

2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
2600
        factors: list[Any] = []
2601
2602
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2603
2604
        return hash_str

2605

2606
2607
2608
2609
2610
2611
2612
2613
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2614
2615
2616
2617
# model_type -> reason
_FLOAT16_NOT_SUPPORTED_MODELS = {
    "gemma2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "gemma3": "Numerical instability. Please use bfloat16 or float32 instead.",
2618
2619
    "gemma3_text":
    "Numerical instability. Please use bfloat16 or float32 instead.",
2620
2621
2622
    "plamo2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "glm4": "Numerical instability. Please use bfloat16 or float32 instead.",
}
2623

2624

2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
def _is_valid_dtype(model_type: str, dtype: torch.dtype):
    if model_type in _FLOAT16_NOT_SUPPORTED_MODELS and dtype == torch.float16:  # noqa: E501, SIM103
        return False

    return True


def _check_valid_dtype(model_type: str, dtype: torch.dtype):
    if model_type in _FLOAT16_NOT_SUPPORTED_MODELS and dtype == torch.float16:
        reason = _FLOAT16_NOT_SUPPORTED_MODELS[model_type]
        raise ValueError(f"The model type {model_type!r} "
                         f"does not support float16. Reason: {reason}")

    return True


def _find_dtype(
    model_id: str,
2643
    config: PretrainedConfig,
2644
2645
2646
    *,
    revision: Optional[str],
):
2647
2648
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
2649
    config_dtype = getattr(config, "torch_dtype", None)
2650

2651
    # Fallbacks for multi-modal models if the root config
2652
    # does not define torch_dtype
2653
2654
    if config_dtype is None:
        config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
2655
2656
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)
2657
2658
    if config_dtype is None and hasattr(config, "encoder_config"):
        config_dtype = getattr(config.encoder_config, "torch_dtype", None)
2659

2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
    # Try to read the dtype of the weights if they are in safetensors format
    if config_dtype is None:
        repo_mt = try_get_safetensors_metadata(model_id, revision=revision)

        if repo_mt and (files_mt := repo_mt.files_metadata):
            param_dtypes: set[torch.dtype] = {
                _SAFETENSORS_TO_TORCH_DTYPE[dtype_str]
                for file_mt in files_mt.values()
                for dtype_str in file_mt.parameter_count
                if dtype_str in _SAFETENSORS_TO_TORCH_DTYPE
            }

            if param_dtypes:
                return common_broadcastable_dtype(param_dtypes)

2675
2676
2677
    if config_dtype is None:
        config_dtype = torch.float32

2678
    return config_dtype
2679

Shinichi Hemmi's avatar
Shinichi Hemmi committed
2680

2681
2682
2683
2684
2685
2686
2687
def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform
2688

2689
2690
2691
2692
    supported_dtypes = [
        dtype for dtype in current_platform.supported_dtypes
        if _is_valid_dtype(model_type, dtype)
    ]
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
2738
2739
2740
2741
2742
2743
2744
2745
2746
    if is_pooling_model and torch.float16 in supported_dtypes:
        preferred_dtype = torch.float16
    else:
        preferred_dtype = supported_dtypes[0]

    # Downcast for float32 models
    if config_dtype == torch.float32:
        config_dtype = preferred_dtype

    if config_dtype in supported_dtypes:
        return config_dtype

    # Ensure device compatibility
    device_name = current_platform.get_device_name()
    device_capability = current_platform.get_device_capability()

    if device_capability is None:
        device_str = f"{device_name!r}"
    else:
        version_str = device_capability.as_version_str()
        device_str = f"{device_name!r} (with compute capability {version_str})"

    logger.warning(
        "Your device %s doesn't support %s. "
        "Falling back to %s for compatibility.",
        device_str,
        config_dtype,
        preferred_dtype,
    )

    return preferred_dtype


def _get_and_verify_dtype(
    model_id: str,
    config: PretrainedConfig,
    dtype: Union[str, torch.dtype],
    *,
    is_pooling_model: bool,
    revision: Optional[str] = None,
) -> torch.dtype:
    config_dtype = _find_dtype(model_id, config, revision=revision)
    model_type = config.model_type

    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            # Set default dtype from model config
            torch_dtype = _resolve_auto_dtype(
                model_type,
                config_dtype,
                is_pooling_model=is_pooling_model,
            )
2747
        else:
2748
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
2749
                raise ValueError(f"Unknown dtype: {dtype!r}")
2750
2751
2752
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
2753
    else:
2754
        raise ValueError(f"Unknown dtype: {dtype}")
2755

2756
2757
    _check_valid_dtype(model_type, torch_dtype)

2758
2759
2760
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
2761
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
2762
2763
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
2764
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
2765
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
2766
            # Casting between float16 and bfloat16 is allowed with a warning.
2767
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
2768
2769

    return torch_dtype
2770
2771


2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
def _get_head_dtype(config: PretrainedConfig, dtype: torch.dtype,
                    runner_type: str) -> torch.dtype:
    head_dtype: Optional[Union[str,
                               torch.dtype]] = getattr(config, "head_dtype",
                                                       None)

    if head_dtype == "model":
        return dtype
    elif isinstance(head_dtype, str):
        head_dtype = head_dtype.lower()
        if head_dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
            raise ValueError(f"Unknown dtype: {head_dtype!r}")
        return _STR_DTYPE_TO_TORCH_DTYPE[head_dtype]
    elif isinstance(head_dtype, torch.dtype):
        return head_dtype
    elif head_dtype is None:
        if torch.float32 not in current_platform.supported_dtypes:
            return dtype
        if runner_type == "pooling":
            return torch.float32
        return dtype
    else:
        raise ValueError(f"Unknown dtype: {head_dtype}")


2797
2798
def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
2799
    tokenizer_config: Optional[dict],
2800
    max_model_len: Optional[int],
2801
    disable_sliding_window: bool,
2802
    sliding_window: Optional[int],
2803
    spec_target_max_model_len: Optional[int] = None,
2804
    encoder_config: Optional[Any] = None,
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
) -> int:
    """Get and verify the model's maximum length."""
    derived_max_model_len = float("inf")
    possible_keys = [
        # OPT
        "max_position_embeddings",
        # GPT-2
        "n_positions",
        # MPT
        "max_seq_len",
2815
2816
        # ChatGLM2
        "seq_length",
2817
2818
        # Command-R
        "model_max_length",
2819
2820
        # Whisper
        "max_target_positions",
2821
2822
2823
2824
2825
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
2826
    # Choose the smallest "max_length" from the possible keys
2827
    max_len_key = None
2828
    for key in possible_keys:
2829
2830
2831
2832
2833
        max_len = getattr(hf_config, key, None)
        if max_len is not None:
            max_len_key = key if max_len < derived_max_model_len \
                else max_len_key
            derived_max_model_len = min(derived_max_model_len, max_len)
Jennifer Zhao's avatar
Jennifer Zhao committed
2834
2835
2836
2837
    # For Command-R / Cohere, Cohere2 / Aya Vision models
    if tmp_max_len := getattr(hf_config, "model_max_length", None):
        max_len_key = "model_max_length"
        derived_max_model_len = tmp_max_len
2838
2839
2840

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
2841
2842
2843
2844
    if (disable_sliding_window and sliding_window is not None
            and sliding_window < derived_max_model_len):
        max_len_key = "sliding_window"
        derived_max_model_len = sliding_window
2845

2846
2847
2848
2849
2850
2851
2852
    # Consider model_max_length in tokenizer_config
    if tokenizer_config:
        tokenizer_model_max_length = tokenizer_config.get(
            "model_max_length", derived_max_model_len)
        derived_max_model_len = min(derived_max_model_len,
                                    tokenizer_model_max_length)

2853
2854
    # If none of the keys were found in the config, use a default and
    # log a warning.
2855
    if derived_max_model_len == float("inf"):
2856
2857
2858
2859
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

2860
2861
2862
2863
2864
        if spec_target_max_model_len is not None:
            # If this is a speculative draft model, we use the max model len
            # from the target model.
            return spec_target_max_model_len

2865
2866
2867
2868
        default_max_len = 2048
        logger.warning(
            "The model's config.json does not contain any of the following "
            "keys to determine the original maximum length of the model: "
2869
            "%s. Assuming the model's maximum length is %d.", possible_keys,
2870
            default_max_len)
2871
        derived_max_model_len = default_max_len
2872

2873
    rope_scaling = getattr(hf_config, "rope_scaling", None)
2874
2875
2876
    # NOTE(woosuk): Gemma3's max_model_len (128K) is already scaled by RoPE
    # scaling, so we skip applying the scaling factor again.
    if rope_scaling is not None and "gemma3" not in hf_config.model_type:
2877
2878
2879
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889

        if rope_type not in ("su", "longrope", "llama3"):
            if disable_sliding_window:
                # TODO(robertgshaw): Find a model that supports rope_scaling
                # with sliding window to see if this case should be allowed.
                raise NotImplementedError(
                    "Disabling sliding window is not supported for models "
                    "with rope_scaling. Please raise an issue so we can "
                    "investigate.")

2890
2891
2892
2893
            # NOTE: rope_type == "default" does not define factor
            # https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/modeling_rope_utils.py
            scaling_factor = rope_scaling.get("factor", 1.0)

2894
2895
2896
2897
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
2898

2899
2900
2901
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2902
2903
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2904
    if max_model_len is None:
2905
        max_model_len = int(derived_max_model_len)
2906
2907
2908
2909
2910
2911
2912
2913
        if current_platform.is_tpu():
            logger.warning(
                "--max-model-len is not specified, "
                "it's currently using model's default length %s, "
                "which might be too large."
                "Please input with --max-model-len based on your "
                "request input length and output length, to avoid "
                "unnecessary degradation.", max_model_len)
2914
    elif max_model_len > derived_max_model_len:
2915
2916
2917
2918
2919
        # Some models might have a separate key for specifying model_max_length
        # that will be bigger than derived_max_model_len. We compare user input
        # with model_max_length and allow this override when it's smaller.
        model_max_length = getattr(hf_config, "model_max_length", None)
        if model_max_length is not None and max_model_len <= model_max_length:
2920
2921
2922
2923
2924
2925
2926
            if disable_sliding_window:
                # TODO(robertgshaw): Find a model that has model_max_length
                # with sliding window to see if this case should be allowed.
                raise NotImplementedError(
                    "Disabling sliding window is not supported for models "
                    "model_max_length in the config. Please raise an issue "
                    "so we can investigate.")
2927
        else:
2928
            msg = (
2929
                f"User-specified max_model_len ({max_model_len}) is greater "
2930
2931
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2932
2933
2934
2935
2936
2937
2938
2939
                f"{model_max_length} in model's config.json).")
            warning = (
                "VLLM_ALLOW_LONG_MAX_MODEL_LEN must be used with extreme "
                "caution. If the model uses relative position encoding (RoPE), "
                "positions exceeding derived_max_model_len lead to nan. If the "
                "model uses absolute position encoding, positions exceeding "
                "derived_max_model_len will cause a CUDA array out-of-bounds "
                "error.")
2940
            if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
2941
                logger.warning_once("%s %s", msg, warning)
2942
2943
2944
            else:
                raise ValueError(
                    f"{msg} To allow overriding this maximum, set "
2945
                    f"the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1. {warning}")
2946
    return int(max_model_len)
2947
2948


2949
def get_served_model_name(model: str,
2950
                          served_model_name: Optional[Union[str, list[str]]]):
2951
    """
2952
2953
2954
2955
    If the input is a non-empty list, the first model_name in
    `served_model_name` is taken.
    If the input is a non-empty string, it is used directly.
    For cases where the input is either an empty string or an
2956
2957
2958
2959
2960
2961
2962
2963
2964
    empty list, the fallback is to use `self.model`.
    """
    if not served_model_name:
        return model
    if isinstance(served_model_name, list):
        return served_model_name[0]
    return served_model_name


2965
2966
GuidedDecodingBackend = Literal["auto", "xgrammar", "guidance", "outlines",
                                "lm-format-enforcer"]
2967
2968
2969


@config
2970
2971
@dataclass
class DecodingConfig:
2972
    """Dataclass which contains the decoding strategy of the engine."""
2973

2974
    backend: GuidedDecodingBackend = "auto"
2975
2976
2977
2978
    """Which engine will be used for guided decoding (JSON schema / regex etc)
    by default. With "auto", we will make opinionated choices based on request
    contents and what the backend libraries currently support, so the behavior
    is subject to change in each release."""
2979

2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
    disable_fallback: bool = False
    """If `True`, vLLM will not fallback to a different backend on error."""

    disable_any_whitespace: bool = False
    """If `True`, the model will not generate any whitespace during guided
    decoding. This is only supported for xgrammar and guidance backends."""

    disable_additional_properties: bool = False
    """If `True`, the `guidance` backend will not use `additionalProperties`
    in the JSON schema. This is only supported for the `guidance` backend and
    is used to better align its behaviour with `outlines` and `xgrammar`."""

2992
    reasoning_backend: str = ""
2993
    """Select the reasoning parser depending on the model that you're using.
2994
    This is used to parse the reasoning content into OpenAI API format."""
2995

2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
3010
        factors: list[Any] = []
3011
3012
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3013
3014
        return hash_str

3015
    def __post_init__(self):
3016
3017
3018
3019
3020
3021
3022
3023
        if (self.disable_any_whitespace
                and self.backend not in ("xgrammar", "guidance")):
            raise ValueError("disable_any_whitespace is only supported for "
                             "xgrammar and guidance backends.")
        if (self.disable_additional_properties and self.backend != "guidance"):
            raise ValueError("disable_additional_properties is only supported "
                             "for the guidance backend.")

3024

3025
3026
3027
3028
DetailedTraceModules = Literal["model", "worker", "all"]


@config
3029
3030
@dataclass
class ObservabilityConfig:
3031
    """Configuration for observability - metrics and tracing."""
3032

3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
    show_hidden_metrics_for_version: Optional[str] = None
    """Enable deprecated Prometheus metrics that have been hidden since the
    specified version. For example, if a previously deprecated metric has been
    hidden since the v0.7.0 release, you use
    `--show-hidden-metrics-for-version=0.7` as a temporary escape hatch while
    you migrate to new metrics. The metric is likely to be removed completely
    in an upcoming release."""

    @cached_property
    def show_hidden_metrics(self) -> bool:
        """Check if the hidden metrics should be shown."""
        if self.show_hidden_metrics_for_version is None:
            return False
        return version._prev_minor_version_was(
            self.show_hidden_metrics_for_version)
3048

3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
    otlp_traces_endpoint: Optional[str] = None
    """Target URL to which OpenTelemetry traces will be sent."""

    collect_detailed_traces: Optional[list[DetailedTraceModules]] = None
    """It makes sense to set this only if `--otlp-traces-endpoint` is set. If
    set, it will collect detailed traces for the specified modules. This
    involves use of possibly costly and or blocking operations and hence might
    have a performance impact.

    Note that collecting detailed timing information for each request can be
    expensive."""

    @cached_property
    def collect_model_forward_time(self) -> bool:
        """Whether to collect model forward time for the request."""
        return (self.collect_detailed_traces is not None
                and ("model" in self.collect_detailed_traces
                     or "all" in self.collect_detailed_traces))

    @cached_property
    def collect_model_execute_time(self) -> bool:
        """Whether to collect model execute time for the request."""
        return (self.collect_detailed_traces is not None
                and ("worker" in self.collect_detailed_traces
                     or "all" in self.collect_detailed_traces))
3074

3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
3089
        factors: list[Any] = []
3090
3091
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3092
3093
        return hash_str

3094
    def __post_init__(self):
3095
3096
3097
3098
3099
        if (self.collect_detailed_traces is not None
                and len(self.collect_detailed_traces) == 1
                and "," in self.collect_detailed_traces[0]):
            self._parse_collect_detailed_traces()

3100
        from vllm.tracing import is_otel_available, otel_import_error_traceback
3101
3102
3103
3104
3105
        if not is_otel_available() and self.otlp_traces_endpoint is not None:
            raise ValueError(
                "OpenTelemetry is not available. Unable to configure "
                "'otlp_traces_endpoint'. Ensure OpenTelemetry packages are "
                f"installed. Original error:\n{otel_import_error_traceback}")
3106

3107
3108
3109
3110
3111
3112
    def _parse_collect_detailed_traces(self):
        assert isinstance(self.collect_detailed_traces, list)
        self.collect_detailed_traces = cast(
            list[DetailedTraceModules],
            self.collect_detailed_traces[0].split(","))

3113

3114
@config
3115
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3116
3117
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
3118
3119
3120
    simplifies passing around the distinct configurations in the codebase.
    """

3121
3122
3123
    # TODO: use default_factory once default constructing ModelConfig doesn't
    # try to download a model
    model_config: ModelConfig = None  # type: ignore
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
    """Model configuration."""
    cache_config: CacheConfig = field(default_factory=CacheConfig)
    """Cache configuration."""
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig)
    """Parallel configuration."""
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig)
    """Scheduler configuration."""
    device_config: DeviceConfig = field(default_factory=DeviceConfig)
    """Device configuration."""
    load_config: LoadConfig = field(default_factory=LoadConfig)
    """Load configuration."""
3135
    lora_config: Optional[LoRAConfig] = None
3136
3137
3138
    """LoRA configuration."""
    speculative_config: Optional[SpeculativeConfig] = None
    """Speculative decoding configuration."""
3139
    decoding_config: DecodingConfig = field(default_factory=DecodingConfig)
3140
    """Decoding configuration."""
3141
    observability_config: Optional[ObservabilityConfig] = None
3142
    """Observability configuration."""
3143
    quant_config: Optional[QuantizationConfig] = None
3144
3145
3146
    """Quantization configuration."""
    compilation_config: CompilationConfig = field(
        default_factory=CompilationConfig)
3147
    """`torch.compile` and cudagraph capture configuration for the model.
3148

3149
3150
    As a shorthand, `-O<n>` can be used to directly specify the compilation
    level `n`: `-O3` is equivalent to `-O.level=3` (same as `-O='{"level":3}'`).
3151
    Currently, -O <n> and -O=<n> are supported as well but this will likely be
3152
    removed in favor of clearer -O<n> syntax in the future.
3153
3154
3155

    NOTE: level 0 is the default level without any optimization. level 1 and 2
    are for internal testing only. level 3 is the recommended level for
3156
    production, also default in V1.
3157
3158
3159
3160
3161
3162

    You can specify the full compilation config like so:
    `{"level": 3, "cudagraph_capture_sizes": [1, 2, 4, 8]}`
    """
    kv_transfer_config: Optional[KVTransferConfig] = None
    """The configurations for distributed KV cache transfer."""
3163
    kv_events_config: Optional[KVEventsConfig] = None
3164
    """The configurations for event publishing."""
3165
    # some opaque config, only used to provide additional information
3166
3167
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
3168
3169
3170
3171
    additional_config: Union[dict, SupportsHash] = field(default_factory=dict)
    """Additional config for specified platform. Different platforms may
    support different configs. Make sure the configs are valid for the platform
    you are using. Contents must be hashable."""
3172
    instance_id: str = ""
3173
    """The ID of the vLLM instance."""
3174

3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
3187
        factors: list[Any] = []
3188
3189

        # summarize vllm config
3190
        vllm_factors: list[Any] = []
3191
3192
        from vllm import __version__
        vllm_factors.append(__version__)
3193
        vllm_factors.append(envs.VLLM_USE_V1)
3194
3195
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3196
3197
        else:
            vllm_factors.append("None")
3198
3199
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3200
3201
        else:
            vllm_factors.append("None")
3202
3203
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3204
3205
        else:
            vllm_factors.append("None")
3206
3207
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3208
3209
        else:
            vllm_factors.append("None")
3210
3211
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3212
3213
        else:
            vllm_factors.append("None")
3214
3215
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3216
3217
        else:
            vllm_factors.append("None")
3218
3219
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3220
3221
3222
3223
3224
            # LoRA creates static buffers based on max_num_batched_tokens.
            # The tensor sizes and strides get captured in the torch.compile
            # graph explicitly.
            vllm_factors.append(
                str(self.scheduler_config.max_num_batched_tokens))
3225
3226
        else:
            vllm_factors.append("None")
3227
3228
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3229
3230
        else:
            vllm_factors.append("None")
3231
3232
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3233
3234
        else:
            vllm_factors.append("None")
3235
3236
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3237
3238
        else:
            vllm_factors.append("None")
3239
3240
3241
3242
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3243
3244
        else:
            vllm_factors.append("None")
3245
3246
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3247
3248
3249
        else:
            vllm_factors.append("None")
        if self.additional_config:
3250
3251
3252
3253
3254
3255
3256
3257
            if isinstance(additional_config := self.additional_config, dict):
                additional_config_hash = hashlib.md5(
                    json.dumps(additional_config, sort_keys=True).encode(),
                    usedforsecurity=False,
                ).hexdigest()
            else:
                additional_config_hash = additional_config.compute_hash()
            vllm_factors.append(additional_config_hash)
3258
3259
        else:
            vllm_factors.append("None")
3260
3261
        factors.append(vllm_factors)

3262
3263
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
3264
3265
        return hash_str

3266
3267
3268
3269
3270
3271
    def pad_for_cudagraph(self, batch_size: int) -> int:
        # if batch_size > self.compilation_config.max_capture_size,
        # it should raise an IndexError.
        # the caller should make sure the batch_size is within the range,
        # i.e., batch_size <= self.compilation_config.max_capture_size
        return self.compilation_config.bs_to_padded_graph_size[batch_size]
3272

3273
3274
3275
3276
3277
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3278
        from vllm.platforms import current_platform
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
        if model_config.quantization is not None:
            from vllm.model_executor.model_loader.weight_utils import (
                get_quant_config)
            quant_config = get_quant_config(model_config, load_config)
            capability_tuple = current_platform.get_device_capability()

            if capability_tuple is not None:
                capability = capability_tuple.to_int()
                if capability < quant_config.get_min_capability():
                    raise ValueError(
                        f"The quantization method {model_config.quantization} "
                        "is not supported for the current GPU. Minimum "
                        f"capability: {quant_config.get_min_capability()}. "
                        f"Current capability: {capability}.")
            supported_dtypes = quant_config.get_supported_act_dtypes()
            if model_config.dtype not in supported_dtypes:
                raise ValueError(
                    f"{model_config.dtype} is not supported for quantization "
                    f"method {model_config.quantization}. Supported dtypes: "
                    f"{supported_dtypes}")
            return quant_config
        return None
3301

3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
    @staticmethod
    def get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        import copy

        # For some reason, the _ version of this modifies the model_config
        # object, so using deepcopy to avoid this problem.
        return VllmConfig._get_quantization_config(copy.deepcopy(model_config),
                                                   load_config)

3313
3314
3315
3316
3317
3318
3319
3320
3321
    def with_hf_config(
        self,
        hf_config: PretrainedConfig,
        architectures: Optional[list[str]] = None,
    ) -> "VllmConfig":
        if architectures is not None:
            hf_config = copy.deepcopy(hf_config)
            hf_config.architectures = architectures

3322
3323
3324
3325
3326
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

3327
3328
3329
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
3330
3331
3332

        self.try_verify_and_update_config()

3333
3334
3335
3336
3337
        if self.model_config is not None:
            self.model_config.verify_async_output_proc(self.parallel_config,
                                                       self.speculative_config,
                                                       self.device_config)
            self.model_config.verify_with_parallel_config(self.parallel_config)
3338
3339
            self.model_config.verify_dual_chunk_attention_config(
                self.load_config)
3340

3341
        self.cache_config.verify_with_parallel_config(self.parallel_config)
3342

3343
        if self.lora_config is not None:
3344
            self.lora_config.verify_with_cache_config(self.cache_config)
3345
            self.lora_config.verify_with_model_config(self.model_config)
3346

3347
        if self.quant_config is None and self.model_config is not None:
3348
3349
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
3350

3351
        from vllm.platforms import current_platform
3352
        if self.model_config is not None and \
3353
3354
3355
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
3356
            logger.warning_once(
3357
3358
3359
3360
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
        # If the user does not explicitly set a compilation level, then
        # we use the default level. The default level depends on other
        # settings (see the below code).
        if self.compilation_config.level is None:
            if envs.VLLM_USE_V1:
                if (self.model_config is not None
                        and not self.model_config.enforce_eager):
                    self.compilation_config.level = CompilationLevel.PIECEWISE
                else:
                    self.compilation_config.level = \
                            CompilationLevel.NO_COMPILATION
3372

3373
3374
3375
3376
3377
            else:
                # NB: Passing both --enforce-eager and a compilation level
                # in V0 means the compilation level wins out.
                self.compilation_config.level = CompilationLevel.NO_COMPILATION

3378
3379
3380
3381
3382
        # async tp is built on top of sequence parallelism
        # and requires it to be enabled.
        if self.compilation_config.pass_config.enable_async_tp:
            self.compilation_config.pass_config.enable_sequence_parallelism = \
                True
3383
3384
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
3385

3386
        if current_platform.is_cuda_alike() or current_platform.is_xpu():
3387
3388
3389
3390
3391
3392
3393
3394
3395
            # if cudagraph_mode is not explicitly set by users, set default
            # value
            if self.compilation_config.cudagraph_mode is None:
                if envs.VLLM_USE_V1 and self.compilation_config.level \
                    == CompilationLevel.PIECEWISE:
                    self.compilation_config.cudagraph_mode = \
                        CUDAGraphMode.PIECEWISE
                else:
                    self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
3396

3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
            # disable cudagraph when enforce eager execution
            if self.model_config is not None and \
                    self.model_config.enforce_eager:
                logger.info("Cudagraph is disabled under eager mode")
                self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
            elif envs.VLLM_USE_V1:
                self.compilation_config.cudagraph_num_of_warmups = 1

            self._set_cudagraph_sizes()
        else:
            self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
3408

3409
        if self.cache_config.cpu_offload_gb > 0 and \
3410
3411
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
3412
            logger.warning(
3413
                "CPU offload is not supported with `torch.compile` in v0 yet."
3414
3415
3416
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
        if self.cache_config.kv_sharing_fast_prefill:
            if not envs.VLLM_USE_V1:
                raise NotImplementedError(
                    "Fast prefill optimization for KV sharing is not supported "
                    "in V0 currently.")

            if self.speculative_config is not None and \
                self.speculative_config.use_eagle():
                raise NotImplementedError(
                    "Fast prefill optimization for KV sharing is not "
                    "compatible with EAGLE as EAGLE requires correct logits "
                    "for all tokens while fast prefill gives incorrect logits "
                    "for prompt tokens.")

            logger.warning_once(
                "--kv-sharing-fast-prefill requires changes on model side for "
                "correctness and to realize prefill savings. ")

3435
3436
3437
3438
3439
3440
        if ((not envs.VLLM_USE_V1) and self.lora_config is not None
                and self.compilation_config.level
                != CompilationLevel.NO_COMPILATION):
            logger.warning(
                "LoRA for V0 is not supported with `torch.compile` yet. "
                "Disabling `torch.compile`.")
3441
3442
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3443
3444
        disable_chunked_prefill_reasons: list[str] = []

3445
3446
3447
3448
3449
3450
3451
        if self.model_config:
            if self.model_config.pooler_config:
                pooling_type = self.model_config.pooler_config.pooling_type
                if pooling_type is None or pooling_type.lower() != "last":
                    disable_chunked_prefill_reasons.append(
                        "Only \"last\" pooling supports chunked "
                        "prefill and prefix caching; disabling both.")
3452
3453
3454
3455
                if not getattr(self.model_config.hf_config, "is_causal", True):
                    disable_chunked_prefill_reasons.append(
                        "Only models using causal attention supports chunked "
                        "prefill and prefix caching; disabling both.")
3456
3457
3458
3459
3460
3461
3462
3463
            elif self.model_config.is_encoder_decoder:
                self.scheduler_config.max_num_encoder_input_tokens = \
                    MULTIMODAL_REGISTRY.get_encdec_max_encoder_len(self.model_config)
                logger.debug(
                    "Encoder-decoder model detected: setting "
                    "`max_num_encoder_input_tokens` to encoder length (%s)",
                    self.scheduler_config.max_num_encoder_input_tokens)
                self.scheduler_config.disable_chunked_mm_input = True
3464
                disable_chunked_prefill_reasons.append(
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
                    "Encoder-decoder models do not support chunked prefill nor"
                    " prefix caching; disabling both.")
                if (self.model_config.architecture
                        == "WhisperForConditionalGeneration"
                        and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD")
                        != "spawn"):
                    logger.warning(
                        "Whisper is known to have issues with "
                        "forked workers. If startup is hanging, "
                        "try setting 'VLLM_WORKER_MULTIPROC_METHOD' "
                        "to 'spawn'.")
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485

        if disable_chunked_prefill_reasons:
            for reason in disable_chunked_prefill_reasons:
                logger.info(reason)
            self.scheduler_config.chunked_prefill_enabled = False
            self.scheduler_config.long_prefill_token_threshold = 0

            if self.cache_config is not None:
                self.cache_config.enable_prefix_caching = False

3486
        if (self.kv_events_config is not None
3487
3488
3489
3490
3491
                and self.kv_events_config.enable_kv_cache_events
                and not self.cache_config.enable_prefix_caching):
            logger.warning(
                "KV cache events are on, but prefix caching is not enabled."
                "Use --enable-prefix-caching to enable.")
3492
3493
        if (self.kv_events_config is not None
                and self.kv_events_config.publisher != "null"
3494
3495
3496
3497
3498
                and not self.kv_events_config.enable_kv_cache_events):
            logger.warning("KV cache events are disabled,"
                           "but the scheduler is configured to publish them."
                           "Modify KVEventsConfig.enable_kv_cache_events"
                           "to True to enable.")
3499
3500
        current_platform.check_and_update_config(self)

3501
        # final check of cudagraph mode after platform-specific update
3502
        if envs.VLLM_USE_V1 and current_platform.is_cuda_alike():
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
            if self.compilation_config.cudagraph_mode == CUDAGraphMode.FULL \
                and self.model_config is not None and \
                not self.model_config.disable_cascade_attn:
                logger.info("CUDAGraphMode.FULL is not supported with "
                            "cascade attention currently. Disabling cascade"
                            "attention.")
                self.model_config.disable_cascade_attn = True

            if self.compilation_config.cudagraph_mode\
                .requires_piecewise_compilation():
                assert self.compilation_config.level == \
                    CompilationLevel.PIECEWISE, \
                    "Compilation level should be CompilationLevel.PIECEWISE "\
                    "when cudagraph_mode piecewise cudagraphs is used, "\
                    f"cudagraph_mode={self.compilation_config.cudagraph_mode}"

3519
3520
3521
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

3522
3523
3524
3525
3526
        # Do this after all the updates to compilation_config.level
        if envs.VLLM_USE_V1 and \
            self.compilation_config.level == CompilationLevel.PIECEWISE:
            self.compilation_config.set_splitting_ops_for_v1()

3527
3528
3529
3530
3531
        if (envs.VLLM_USE_V1
                and not self.scheduler_config.disable_hybrid_kv_cache_manager):
            # logger should only print warning message for hybrid models. As we
            # can't know whether the model is hybrid or not now, so we don't log
            # warning message here and will log it later.
3532
            if not current_platform.support_hybrid_kv_cache():
3533
                # Hybrid KV cache manager is not supported on non-GPU platforms.
3534
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3535
3536
            if self.kv_transfer_config is not None:
                # Hybrid KV cache manager is not compatible with KV transfer.
3537
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3538
3539
            if self.kv_events_config is not None:
                # Hybrid KV cache manager is not compatible with KV events.
3540
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3541
            if self.model_config is not None and \
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
                self.model_config.attention_chunk_size is not None:
                if self.speculative_config is not None and \
                    self.speculative_config.use_eagle():
                    # Hybrid KV cache manager is not yet supported with chunked
                    # local attention + eagle.
                    self.scheduler_config.disable_hybrid_kv_cache_manager = True
                elif \
                    not envs.VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE:
                    logger.warning(
                        "There is a latency regression when using chunked local"
                        " attention with the hybrid KV cache manager. Disabling"
                        " it, by default. To enable it, set the environment "
                        "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE=1."
                    )
                    # Hybrid KV cache manager is not yet supported with chunked
                    # local attention.
                    self.scheduler_config.disable_hybrid_kv_cache_manager = True
3559

3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
    def update_sizes_for_sequence_parallelism(self,
                                              possible_sizes: list) -> list:
        # remove the sizes that not multiple of tp_size when
        # enable sequence parallelism
        removed_sizes = [
            size for size in possible_sizes
            if size % self.parallel_config.tensor_parallel_size != 0
        ]
        if removed_sizes:
            logger.warning(
                "Batch sizes %s are removed because they are not "
                "multiple of tp_size %d when "
                "sequence parallelism is enabled", removed_sizes,
                self.parallel_config.tensor_parallel_size)

        return [
            size for size in possible_sizes
            if size % self.parallel_config.tensor_parallel_size == 0
        ]

3580
3581
    def _set_cudagraph_sizes(self):
        """
3582
3583
        vLLM defines the default candidate list of batch sizes for CUDA graph
        capture as:
3584

3585
3586
3587
3588
        ```python
        max_graph_size = min(max_num_seqs * 2, 512)
        # 1, 2, 4, then multiples of 8 up to max_graph_size
        cuda_graph_sizes = [1, 2, 4, 8, 16, 24, 32, 40, ..., max_graph_size]
3589

3590
3591
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
3592

3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
        These sizes are used to capture and reuse CUDA graphs for
        performance-critical paths (e.g., decoding). Capturing enables
        significantly faster kernel dispatch by avoiding Python overhead. The
        list is then filtered based on `max_num_batched_tokens` (e.g., 8192 on
        most GPUs), which controls the total allowed number of tokens in a
        batch. Since each sequence may have a variable number of tokens, the
        maximum usable batch size will depend on actual sequence lengths.

        Example:
            With `max_num_batched_tokens = 8192`, and typical sequences
            averaging ~32 tokens, most practical batch sizes fall below 256.
            However, the system will still allow capture sizes up to 512 if
            shape and memory permit.

        Note:
            If users explicitly specify cudagraph capture sizes in the
            compilation config, those will override this default logic.
            At runtime:

            - If batch size <= one of the `cudagraph_capture_sizes`, the closest
            padded CUDA graph will be used.
            - If batch size > largest `cudagraph_capture_sizes`, cudagraph will
            not be used.
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
        """

        # calculate the default `batch_size_capture_list`
        if not envs.VLLM_USE_V1:
            batch_size_capture_list = []
            if self.scheduler_config is not None and \
                self.model_config is not None and \
                    not self.model_config.enforce_eager:

                possible_sizes = [1, 2, 4] + [8 * i for i in range(1, 1025)]
3626
3627
3628
3629
3630
                if self.parallel_config.tensor_parallel_size > 1 and \
                    self.compilation_config.pass_config.enable_sequence_parallelism:
                    possible_sizes = self.update_sizes_for_sequence_parallelism(
                        possible_sizes)

3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
                # find the minimum size that is larger than max_num_seqs,
                # which then becomes the max_batchsize_to_capture
                larger_sizes = [
                    x for x in possible_sizes
                    if x >= self.scheduler_config.max_num_seqs
                ]
                if larger_sizes:
                    max_batchsize_to_capture = larger_sizes[0]
                else:
                    max_batchsize_to_capture = possible_sizes[-1]

                # filter out the sizes that are
                # larger than max_batchsize_to_capture
                batch_size_capture_list = [
                    size for size in possible_sizes
                    if size <= max_batchsize_to_capture
                ]
        else:
            batch_size_capture_list = []
            if self.model_config is not None and \
                not self.model_config.enforce_eager:
3652
3653
3654
3655
3656
3657
3658
3659
                cuda_graph_sizes = self.scheduler_config.cuda_graph_sizes
                if len(cuda_graph_sizes) == 1:
                    batch_size_capture_list = [1, 2, 4] + [
                        i for i in range(8, cuda_graph_sizes[0] + 1, 8)
                    ]
                elif len(cuda_graph_sizes) > 1:
                    batch_size_capture_list = sorted(cuda_graph_sizes)
                else:
Cyrus Leung's avatar
Cyrus Leung committed
3660
                    raise TypeError(f"Invalid value for {cuda_graph_sizes=}.")
3661
3662
3663
3664
                if self.parallel_config.tensor_parallel_size > 1 and \
                    self.compilation_config.pass_config.enable_sequence_parallelism:
                    batch_size_capture_list = \
                        self.update_sizes_for_sequence_parallelism(batch_size_capture_list)
3665
3666
3667
3668
3669
                max_num_tokens = self.scheduler_config.max_num_batched_tokens
                batch_size_capture_list = [
                    size for size in batch_size_capture_list
                    if size <= max_num_tokens
                ]
3670
3671
3672
3673

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

3674
    def recalculate_max_model_len(self, max_model_len: int):
3675
        # Can only be called in try_verify_and_update_config
3676
        model_config = self.model_config
3677
        max_model_len = model_config.get_and_verify_max_len(max_model_len)
3678
3679
        self.model_config.max_model_len = max_model_len
        self.scheduler_config.max_model_len = max_model_len
3680
3681

    def try_verify_and_update_config(self):
3682
3683
3684
        if self.model_config is None:
            return

3685
3686
3687
3688
3689
        # Avoid running try_verify_and_update_config multiple times
        if getattr(self.model_config, "config_updated", False):
            return
        self.model_config.config_updated = True

3690
        architecture = self.model_config.architecture
3691
3692
3693
        if architecture is None:
            return

3694
3695
        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP, HybridAttentionMambaModelConfig)
3696
3697
3698
        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_config(self)
3699

3700
3701
3702
        if self.model_config.is_hybrid:
            HybridAttentionMambaModelConfig.verify_and_update_config(self)

3703
        if self.model_config.convert_type == "classify":
3704
3705
3706
3707
3708
            # Maybe convert ForCausalLM into ForSequenceClassification model.
            from vllm.model_executor.models.adapters import (
                SequenceClassificationConfig)
            SequenceClassificationConfig.verify_and_update_config(self)

3709
    def __str__(self):
3710
        return (
3711
3712
3713
3714
3715
            f"model={self.model_config.model!r}, "
            f"speculative_config={self.speculative_config!r}, "
            f"tokenizer={self.model_config.tokenizer!r}, "
            f"skip_tokenizer_init={self.model_config.skip_tokenizer_init}, "
            f"tokenizer_mode={self.model_config.tokenizer_mode}, "
3716
            f"revision={self.model_config.revision}, "
3717
            f"tokenizer_revision={self.model_config.tokenizer_revision}, "
3718
3719
            f"trust_remote_code={self.model_config.trust_remote_code}, "
            f"dtype={self.model_config.dtype}, "
3720
3721
            f"max_seq_len={self.model_config.max_model_len}, "
            f"download_dir={self.load_config.download_dir!r}, "
3722
            f"load_format={self.load_config.load_format}, "
3723
3724
            f"tensor_parallel_size={self.parallel_config.tensor_parallel_size}, "  # noqa
            f"pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, "  # noqa
3725
            f"data_parallel_size={self.parallel_config.data_parallel_size}, "  # noqa
3726
3727
3728
3729
            f"disable_custom_all_reduce={self.parallel_config.disable_custom_all_reduce}, "  # noqa
            f"quantization={self.model_config.quantization}, "
            f"enforce_eager={self.model_config.enforce_eager}, "
            f"kv_cache_dtype={self.cache_config.cache_dtype}, "
3730
            f"device_config={self.device_config.device}, "
3731
3732
3733
3734
3735
3736
3737
            f"decoding_config={self.decoding_config!r}, "
            f"observability_config={self.observability_config!r}, "
            f"seed={self.model_config.seed}, "
            f"served_model_name={self.model_config.served_model_name}, "
            f"enable_prefix_caching={self.cache_config.enable_prefix_caching}, "
            f"chunked_prefill_enabled={self.scheduler_config.chunked_prefill_enabled}, "  # noqa
            f"use_async_output_proc={self.model_config.use_async_output_proc}, "
3738
3739
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
3740
3741
3742


_current_vllm_config: Optional[VllmConfig] = None
3743
_current_prefix: Optional[str] = None
3744
3745
3746


@contextmanager
3747
3748
3749
def set_current_vllm_config(vllm_config: VllmConfig,
                            check_compile=False,
                            prefix: Optional[str] = None):
3750
    """
3751
    Temporarily set the current vLLM config.
3752
    Used during model initialization.
3753
    We save the current vLLM config in a global variable,
3754
    so that all modules can access it, e.g. custom ops
3755
    can access the vLLM config to determine how to dispatch.
3756
    """
3757
    global _current_vllm_config, _current_prefix
3758
    old_vllm_config = _current_vllm_config
3759
    old_prefix = _current_prefix
3760
3761
3762
3763
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
3764
        _current_prefix = prefix
3765
        yield
3766
3767
3768
    except Exception:
        raise
    else:
3769
3770
3771
3772
        logger.debug("enabled custom ops: %s",
                     vllm_config.compilation_config.enabled_custom_ops)
        logger.debug("disabled custom ops: %s",
                     vllm_config.compilation_config.disabled_custom_ops)
3773
3774
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
3775
3776
3777
3778
3779
3780
3781
3782
3783
            and compilation_counter.num_models_seen == num_models_seen:
            # If the model supports compilation,
            # compilation_counter.num_models_seen should be increased
            # by at least 1.
            # If it is not increased, it means the model does not support
            # compilation (does not have @support_torch_compile decorator).
            logger.warning(
                "`torch.compile` is turned on, but the model %s"
                " does not support it. Please open an issue on GitHub"
3784
                " if you want it to be supported.",
3785
                vllm_config.model_config.model)
3786
    finally:
3787
        _current_vllm_config = old_vllm_config
3788
        _current_prefix = old_prefix
3789
3790
3791
3792
3793
3794
3795
3796
        # Clear the compilation config cache when context changes
        get_cached_compilation_config.cache_clear()


@lru_cache(maxsize=1)
def get_cached_compilation_config():
    """Cache config to avoid repeated calls to get_current_vllm_config()"""
    return get_current_vllm_config().compilation_config
3797
3798
3799
3800
3801
3802
3803


def get_current_vllm_config() -> VllmConfig:
    if _current_vllm_config is None:
        # in ci, usually when we test custom ops/modules directly,
        # we don't set the vllm config. In that case, we set a default
        # config.
3804
        logger.warning("Current vLLM config is not set.")
3805
3806
3807
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
3808
3809


3810
3811
3812
3813
3814
3815
3816
3817
3818
def get_current_model_prefix() -> str:
    """
    Get the prefix of the model that's currently being initialized.
    """
    assert _current_prefix is not None, \
        "Current model prefix is not set. "
    return _current_prefix


3819
3820
3821
3822
3823
def contains_object_print(text):
    """
    Check if the text looks like a printed Python object, e.g.
    contains any substring matching the pattern: "at 0xFFFFFFF>"
    We match against 0x followed by 2-16 hex chars (there's
3824
    a max of 16 on a 64-bit system).
3825
3826
3827
3828
3829

    Args:
        text (str): The text to check

    Returns:
3830
        result (bool): `True` if a match is found, `False` otherwise.
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
    """
    pattern = r'at 0x[a-fA-F0-9]{2,16}>'
    match = re.search(pattern, text)
    return match is not None


def assert_hashable(text):
    if not contains_object_print(text):
        return True
    raise AssertionError(
        f"vLLM tried to hash some configs that may have Python objects ids "
        f"in them. This is a bug, please file an issue. "
        f"Text being hashed: {text}")
3844
3845
3846
3847
3848


T = TypeVar("T")


3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
def get_layers_from_vllm_config(
        vllm_config: VllmConfig,
        layer_type: type[T],
        layer_names: Optional[list[str]] = None) -> dict[str, T]:
    """
    Get layers from the vLLM config.

    Args:
        vllm_config: The vLLM config.
        layer_type: The type of the layer to get.
        layer_names: The names of the layers to get. If None, return all layers.
    """

    if layer_names is None:
        layer_names = list(
            vllm_config.compilation_config.static_forward_context.keys())

    forward_context = vllm_config.compilation_config.static_forward_context

3868
    return {
3869
3870
3871
        layer_name: forward_context[layer_name]
        for layer_name in layer_names
        if isinstance(forward_context[layer_name], layer_type)
3872
    }
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902


@config
@dataclass
class SpeechToTextConfig:
    """Configuration for speech-to-text models."""

    sample_rate: float = 16_000
    """Sample rate (Hz) to resample input audio to. Most speech models expect
    16kHz audio input. The input audio will be automatically resampled to this
    rate before processing."""

    max_audio_clip_s: int = 30
    """Maximum duration in seconds for a single audio clip without chunking.
    Audio longer than this will be split into smaller chunks if
    `allow_audio_chunking` evaluates to True, otherwise it will be rejected."""

    overlap_chunk_second: int = 1
    """Overlap duration in seconds between consecutive audio chunks when
    splitting long audio. This helps maintain context across chunk boundaries
    and improves transcription quality at split points."""

    min_energy_split_window_size: Optional[int] = 1600
    """Window size in samples for finding low-energy (quiet) regions to split
    audio chunks. The algorithm looks for the quietest moment within this
    window to minimize cutting through speech. Default 1600 samples ≈ 100ms
    at 16kHz. If None, no chunking will be done."""

    @property
    def allow_audio_chunking(self) -> bool:
3903
        return self.min_energy_split_window_size is not None
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921


def update_config(config: DataclassInstanceT,
                  overrides: dict[str, Any]) -> DataclassInstanceT:
    processed_overrides = {}
    for field_name, value in overrides.items():
        assert hasattr(
            config, field_name), f"{type(config)} has no field `{field_name}`"
        current_value = getattr(config, field_name)
        if is_dataclass(current_value) and not is_dataclass(value):
            assert isinstance(value, dict), (
                f"Overrides to {type(config)}.{field_name} must be a dict"
                f"  or {type(current_value)}, but got {type(value)}")
            value = update_config(
                current_value,  # type: ignore[type-var]
                value)
        processed_overrides[field_name] = value
    return replace(config, **processed_overrides)