__init__.py 169 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 textwrap
Robert Shaw's avatar
Robert Shaw committed
12
import uuid
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.parallel import DistributedExecutorBackend, ParallelConfig
37
from vllm.config.scheduler import SchedulerConfig, SchedulerPolicy
38
from vllm.config.utils import ConfigType, config
Woosuk Kwon's avatar
Woosuk Kwon committed
39
from vllm.logger import init_logger
40
from vllm.model_executor.layers.quantization import QuantizationMethods
41
from vllm.platforms import current_platform
42
43
44
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
45
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
46
    is_interleaved, maybe_override_with_speculators_target_model,
47
48
    try_get_generation_config, try_get_safetensors_metadata,
    try_get_tokenizer_config, uses_mrope)
49
from vllm.transformers_utils.s3_utils import S3Model
50
from vllm.transformers_utils.utils import is_s3, maybe_model_redirect
51
from vllm.utils import (DEFAULT_MAX_NUM_BATCHED_TOKENS, LayerBlockType,
52
                        LazyLoader, common_broadcastable_dtype, random_uuid)
53

54
if TYPE_CHECKING:
55
    from _typeshed import DataclassInstance
56
    from transformers.configuration_utils import PretrainedConfig
57

58
59
60
    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
61
62
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
63
    from vllm.model_executor.model_loader import LoadFormats
64
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
65
    from vllm.v1.sample.logits_processor import LogitsProcessor
66

67
    HfOverrides = Union[dict, Callable[[type], type]]
68
else:
69
    DataclassInstance = Any
70
    PretrainedConfig = Any
71
    QuantizationConfig = Any
72
    QuantizationMethods = Any
73
    BaseModelLoader = Any
74
    LoadFormats = Any
75
    TensorizerConfig = Any
76
    LogitsProcessor = Any
77
78
79
80
81
82
    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")
83

84
logger = init_logger(__name__)
85
DataclassInstanceT = TypeVar("DataclassInstanceT", bound=DataclassInstance)
86

87
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
88
                     "score", "reward", "transcription", "draft"]
89

90
_ResolvedTask = Literal["generate", "transcription", "encode", "embed",
91
                        "classify", "reward", "draft"]
92

93
RunnerOption = Literal["auto", "generate", "pooling", "draft"]
94

95
RunnerType = Literal["generate", "pooling", "draft"]
96

97
98
99
100
101
ConvertOption = Literal["auto", "none", "embed", "classify", "reward"]

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

_RUNNER_TASKS: dict[RunnerType, list[TaskOption]] = {
102
    "generate": ["generate", "transcription"],
103
104
105
106
107
108
109
    "pooling": ["embedding", "embed", "classify", "score", "reward"],
    "draft": ["draft"],
}

_RUNNER_CONVERTS: dict[RunnerType, list[ConvertType]] = {
    "generate": [],
    "pooling": ["embed", "classify", "reward"],
110
    "draft": [],
111
}
112

113
114
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
# 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

154

155
@runtime_checkable
156
157
158
159
160
161
class SupportsHash(Protocol):

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


162
163
class SupportsMetricsInfo(Protocol):

164
    def metrics_info(self) -> dict[str, str]:
165
166
167
        ...


168
169
170
171
172
173
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"


174
175
176
177
178
179
180
181
182
183
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
184

185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
        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

    cls_node = ast.parse(textwrap.dedent(inspect.getsource(cls))).body[0]

    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


226
def get_field(cls: ConfigType, name: str) -> Field:
227
228
229
230
231
232
233
    """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__}.")
234
    named_field: Field = cls_fields[name]
235
236
237
238
239
240
241
242
    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.")


243
244
245
246
def is_init_field(cls: ConfigType, name: str) -> bool:
    return next(f for f in fields(cls) if f.name == name).init


247
248
TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
249
250
LogprobsMode = Literal["raw_logprobs", "raw_logits", "processed_logprobs",
                       "processed_logits"]
251
252
253


@config
254
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
255
class ModelConfig:
256
257
    """Configuration for the model."""

258
    model: str = "Qwen/Qwen3-0.6B"
259
260
261
    """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."""
262
263
264
    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."""
265
266
267
268
269
270
271
272
273
274
    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.
    """
275
    tokenizer: SkipValidation[str] = None  # type: ignore
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
    """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
297
298
    """Random seed for reproducibility. Initialized to None in V0, but
    initialized to 0 in V1."""
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
    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)
314
    """RoPE scaling configuration. For example,
315
316
317
318
319
320
321
322
    `{"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."""
323
    max_model_len: SkipValidation[int] = None  # type: ignore
324
325
    """Model context length (prompt and output). If unspecified, will be
    automatically derived from the model config.
326

327
328
329
330
331
332
    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
333
    """Specify the maximum length for spec decoding draft models."""
334
    quantization: SkipValidation[Optional[QuantizationMethods]] = None
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
    """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
352
353
    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."""
354
355
356
357
358
359
360
    logprobs_mode: LogprobsMode = "raw_logprobs"
    """Indicates the content returned in the logprobs and prompt_logprobs.
    Supported mode:
    1) raw_logprobs, 2) processed_logprobs, 3) raw_logits, 4) processed_logits.
    Raw means the values before applying logit processors, like bad words.
    Processed means the values after applying such processors.
    """
361
362
363
364
365
366
367
368
369
370
371
372
373
374
    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."""
375
376
377
378
    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."""
379
380
381
382
383
384
385
386
387
388
389
    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."""
390
    interleave_mm_strings: bool = False
391
    """Enable fully interleaved support for multimodal prompts, while using
392
    --chat-template-content-format=string. Defaults to False."""
393
394
395
396
    skip_mm_profiling: bool = False
    """When enabled, skips multimodal memory profiling and only profiles with
    language backbone model during engine initialization.
    """
397
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
398
399
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
400
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
401
402
403
404
405
406
407
408
409
410
411
412
413
414
    use_async_output_proc: bool = True
    """Whether to use async output processor."""
    config_format: Union[str, ConfigFormat] = ConfigFormat.AUTO.value
    """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
415
    config. If a callable, it is called to update the HuggingFace config."""
416
417
418
419
420
    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}`.
421
    """
422
423
424
425
426
427
428
429
430
    mm_processor_cache_gb: int = 4
    """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)."""
431
432
433
434
    override_neuron_config: dict[str, Any] = field(default_factory=dict)
    """Initialize non-default neuron config or override default neuron config
    that are specific to Neuron devices, this argument will be used to
    configure the neuron config that can not be gathered from the vllm
435
    arguments. e.g. `{"cast_logits_dtype": "bfloat16"}`."""
436
437
438
439
440
441
    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}`.
442
    """
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
    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
458
    `--generation-config vllm`, only the override parameters are used."""
459
460
461
462
463
464
465
466
467
    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
    - "transformers" will use the Transformers model implementation."""
468
469
    override_attention_dtype: Optional[str] = None
    """Override dtype for attention"""
470
471
472
    logits_processors: Optional[list[Union[str, type[LogitsProcessor]]]] = None
    """One or more logits processors' fully-qualified class names or class
    definitions"""
zhuwenwen's avatar
zhuwenwen committed
473
474
475
    enable_chunked_prefill: Optional[bool] = None
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""
476

477
478
479
480
481
482
483
484
485
486
487
488
    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.
        """
489
        factors: list[Any] = []
490
491
492
493
494
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
495
496
497
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
498
        factors.append(self.trust_remote_code)
499
500
501
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
502
503
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
504
505
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
zhuwenwen's avatar
zhuwenwen committed
506
        factors.append(self.enable_chunked_prefill)
507
508
        str_factors = str(factors)
        assert_hashable(str_factors)
509
510
        return hashlib.sha256(str(factors).encode()).hexdigest()

511
    def __post_init__(self) -> None:
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
        # 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)

530
531
532
533
534
535
536
537
538
        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)

539
540
541
        # 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)
542
543
544
545
546
547
548
549
550
551
552
553
        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):
554
            hf_overrides_kw = {}
555
            hf_overrides_fn = self.hf_overrides
556
        else:
557
            hf_overrides_kw = self.hf_overrides
558
            hf_overrides_fn = None
559

560
561
        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
562
            hf_overrides_kw.update(hf_override)
563
            hf_overrides_str = json.dumps(hf_overrides_kw)
564
565
566
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
567
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
568
569
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
570
            hf_overrides_kw.update(hf_override)
571
            hf_overrides_str = json.dumps(hf_overrides_kw)
572
573
574
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
575
576
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

577
        self.maybe_pull_model_tokenizer_for_s3(self.model, self.tokenizer)
578

579
580
581
582
        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 "
583
584
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
585
586
                "for instructions on how to install it.")

587
588
        from vllm.platforms import current_platform

589
590
591
592
593
594
        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)

595
596
597
598
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
599

600
601
602
        if isinstance(self.config_format, str):
            self.config_format = ConfigFormat(self.config_format)

603
        hf_config = get_config(self.hf_config_path or self.model,
604
605
606
607
608
609
                               self.trust_remote_code,
                               self.revision,
                               self.code_revision,
                               self.config_format,
                               hf_overrides_kw=hf_overrides_kw,
                               hf_overrides_fn=hf_overrides_fn)
610

611
        self.hf_config = hf_config
612
        self.hf_text_config = get_hf_text_config(self.hf_config)
613
614
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
615
        self.encoder_config = self._get_encoder_config()
616
        self.hf_image_processor_config = get_hf_image_processor_config(
617
            self.model, hf_token=self.hf_token, revision=self.revision)
618

619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
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
        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
676
            else:
677
678
679
680
681
682
683
684
                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}"
685
686
            warnings.warn(msg, DeprecationWarning, stacklevel=2)

687
688
689
690
        self.runner_type = self._get_runner_type(architectures, self.runner)
        self.convert_type = self._get_convert_type(architectures,
                                                   self.runner_type,
                                                   self.convert)
691

692
693
694
695
696
697
698
699
700
701
702
703
704
705
        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.")
706

707
708
709
710
711
712
        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)
713
714
        self._model_info = model_info
        self._architecture = arch
715
        logger.info("Resolved architecture: %s", arch)
716
717
718
719
720
721
722
723
724
725

        self.pooler_config = self._init_pooler_config()

        self.dtype = _get_and_verify_dtype(
            self.model,
            self.hf_config,
            self.dtype,
            is_pooling_model=self.runner_type == "pooling",
            revision=self.revision,
        )
726

727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
        # 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
742

743
        self.original_max_model_len = self.max_model_len
744
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
745
        self.multimodal_config = self._init_multimodal_config()
746

747
748
749
750
751
        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

752
753
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
754

755
756
757
        if (not current_platform.is_neuron() and self.override_neuron_config):
            raise ValueError(
                "`override_neuron_config` is only supported on Neuron.")
758

759
760
761
        # Avoid running try_verify_and_update_config multiple times
        self.config_updated = False

762
        self._verify_quantization()
763
        self._verify_cuda_graph()
764
        self._verify_bnb_config()
765

766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
    @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

782
783
784
    def _get_transformers_backend_cls(self) -> str:
        """Determine which Transformers backend class will be used if
        `model_impl` is set to `transformers` or `auto`."""
785
786
        if getattr(self, "runner_type", self.runner) == "pooling":
            return "TransformersModel"
787
788
789
790
        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"
791
792
793
794
795
        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()
796

797
798
    @property
    def registry(self):
799
        return me_models.ModelRegistry
800
801
802

    @property
    def architectures(self) -> list[str]:
803
        return getattr(self.hf_config, "architectures", [])
804

805
806
    @property
    def architecture(self) -> str:
807
        """The architecture vllm actually used."""
808
809
        return self._architecture

810
811
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
812
        """Pull model/tokenizer from S3 to temporary directory when needed.
813

814
        Args:
815
816
            model: Model name or path
            tokenizer: Tokenizer name or path
817
        """
818
819
        if not (is_s3(model) or is_s3(tokenizer)):
            return
820

821
822
823
824
825
826
        if is_s3(model):
            s3_model = S3Model()
            s3_model.pull_files(model,
                                allow_pattern=["*.model", "*.py", "*.json"])
            self.model_weights = model
            self.model = s3_model.dir
827

828
829
            # If tokenizer is same as model, download to same directory
            if model == tokenizer:
830
831
832
833
834
                s3_model.pull_files(model,
                                    ignore_pattern=[
                                        "*.pt", "*.safetensors", "*.bin",
                                        "*.tensors"
                                    ])
835
836
837
838
839
840
841
                self.tokenizer = s3_model.dir
                return

        # Only download tokenizer if needed and not already handled
        if is_s3(tokenizer):
            s3_tokenizer = S3Model()
            s3_tokenizer.pull_files(
842
843
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors"])
844
            self.tokenizer = s3_tokenizer.dir
845

846
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
847
        if self._model_info.supports_multimodal:
848
            return MultiModalConfig(
849
                limit_per_prompt=self.limit_mm_per_prompt,
850
                media_io_kwargs=self.media_io_kwargs,
851
                mm_processor_kwargs=self.mm_processor_kwargs,
852
                mm_processor_cache_gb=self.mm_processor_cache_gb,
853
854
                interleave_mm_strings=self.interleave_mm_strings,
                skip_mm_profiling=self.skip_mm_profiling)
855
856

        return None
857

858
    def set_mm_processor_cache_gb(self, value: int) -> None:
859
860
        mm_config = self.get_multimodal_config()

861
862
        self.mm_processor_cache_gb = value
        mm_config.mm_processor_cache_gb = value
863

864
865
866
867
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

868
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
869
        if self.runner_type == "pooling":
870
871
872
873
874
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
875
876
877
878
879

            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():
880
881
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
882

883
884
885
            default_pooling_type = self._model_info.default_pooling_type
            if pooler_config.pooling_type is None:
                pooler_config.pooling_type = default_pooling_type
886

887
            return pooler_config
888

889
890
        return None

891
    def _verify_tokenizer_mode(self) -> None:
892
893
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
894
895
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
896
                f"one of {get_args(TokenizerMode)}.")
897
        self.tokenizer_mode = tokenizer_mode
898

899
900
901
902
903
904
905
906
907
908
    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"

909
        for arch in architectures:
910
911
912
913
914
915
916
917
918
919
920
921
            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"
922

923
    def _get_runner_type(
924
        self,
925
        architectures: list[str],
926
927
928
929
930
931
        runner: RunnerOption,
    ) -> RunnerType:
        if runner != "auto":
            return runner

        runner_type = self._get_default_runner_type(architectures)
932

933
934
935
936
937
938
        # 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)
939
940
941
942

        return runner_type

    def _get_default_convert_type(
943
        self,
944
        architectures: list[str],
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
        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":
969
970
            return "embed"

971
        return "none"
972

973
974
975
976
977
978
979
980
    def _get_convert_type(
        self,
        architectures: list[str],
        runner_type: RunnerType,
        convert: ConvertOption,
    ) -> ConvertType:
        if convert != "auto":
            return convert
981

982
983
        convert_type = self._get_default_convert_type(architectures,
                                                      runner_type)
984

985
986
987
988
989
990
        # 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)
991
992

        return convert_type
993

994
    def _get_supported_generation_tasks(
995
        self,
996
997
        architectures: list[str],
        convert_type: ConvertType,
998
999
1000
    ) -> list[_ResolvedTask]:
        registry = self.registry

1001
        if registry.is_transcription_only_model(architectures, self):
1002
1003
            return ["transcription"]

1004
        # TODO: Use get_supported_generation_tasks once V0 is removed
1005
        supported_tasks = list[_ResolvedTask]()
1006
1007
        if (registry.is_text_generation_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["generate"]):
1008
            supported_tasks.append("generate")
1009

1010
1011
        if registry.is_transcription_model(architectures, self):
            supported_tasks.append("transcription")
1012
1013

        return supported_tasks
1014

1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
    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"

1032
1033
    def _get_supported_pooling_tasks(
        self,
1034
1035
        architectures: list[str],
        convert_type: ConvertType,
1036
    ) -> list[_ResolvedTask]:
1037
        registry = self.registry
1038

1039
        # TODO: Use get_supported_pooling_tasks once V0 is removed
1040
        supported_tasks = list[_ResolvedTask]()
1041
1042
        if (registry.is_pooling_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["pooling"]):
1043
            supported_tasks.append("encode")
1044

1045
1046
1047
            extra_task = (self._get_default_pooling_task(architectures)
                          if convert_type == "none" else convert_type)
            supported_tasks.append(extra_task)
1048

1049
        return supported_tasks
1050

1051
1052
    def _get_supported_tasks(
        self,
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
        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"]
1065

1066
        assert_never(runner_type)
1067

1068
1069
1070
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
1071
            # compressed-tensors uses a "compression_config" key
1072
            quant_cfg = getattr(self.hf_config, "compression_config", None)
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087

        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}")

1088
1089
        return quant_cfg

1090
    def _verify_quantization(self) -> None:
1091
        supported_quantization = me_quant.QUANTIZATION_METHODS
1092
        optimized_quantization_methods = [
1093
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
1094
            "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
1095
            "quark", "modelopt_fp4", "bitblas", "gptq_bitblas", "inc"
1096
        ]
1097
        if self.quantization is not None:
1098
1099
            self.quantization = cast(me_quant.QuantizationMethods,
                                     self.quantization)
1100
1101

        # Parse quantization method from the HF model config, if available.
1102
1103
        quant_cfg = self._parse_quant_hf_config()

1104
        if quant_cfg is not None:
1105
            # Use the community standard 'quant_method'
1106
            quant_method = quant_cfg.get("quant_method", "").lower()
1107
1108

            # Normalize library names
1109
1110
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
1111

1112
            quant_cfg["quant_method"] = quant_method
1113

1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
            # 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 = [
                "marlin",
                "bitblas",
                "gptq_marlin_24",
                "gptq_marlin",
                "gptq_bitblas",
                "awq_marlin",
                "ipex",
                "moe_wna16",
1126
1127
                "modelopt",
                "modelopt_fp4",
1128
1129
1130
1131
1132
1133
1134
1135
            ]
            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
            # over the built in ones.
            quantization_methods = quantization_methods + overrides
1136
1137

            # Detect which checkpoint is it
1138
            for name in quantization_methods:
1139
                method = me_quant.get_quantization_config(name)
1140
1141
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
1142
1143
1144
1145
                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.
1146
                    if (name in get_args(me_quant.QuantizationMethods)
1147
1148
1149
1150
1151
1152
                            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.")
1153
1154
1155
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
1156

1157
            # Verify quantization configurations.
1158
            if self.quantization is None:
1159
1160
                self.quantization = quant_method
            elif self.quantization != quant_method:
1161
1162
                raise ValueError(
                    "Quantization method specified in the model config "
1163
                    f"({quant_method}) does not match the quantization "
1164
1165
1166
1167
1168
1169
1170
1171
                    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}.")
1172
            from vllm.platforms import current_platform
1173
            current_platform.verify_quantization(self.quantization)
1174
            if self.quantization not in optimized_quantization_methods:
1175
                logger.warning(
1176
                    "%s quantization is not fully "
1177
                    "optimized yet. The speed can be slower than "
1178
                    "non-quantized models.", self.quantization)
1179

1180
    def _verify_cuda_graph(self) -> None:
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
        # 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))
1191
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
1192
                                          effective_max_seq_len)
1193
        # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
1194
        ROCM_UNSUPPORTED_MODELS = ['mllama']
1195
1196
1197
1198
1199
1200
        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()):
1201
1202
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1203
                "to eager mode.", self.hf_config.model_type)
1204
            self.enforce_eager = True
1205

1206
1207
    def _verify_bnb_config(self) -> None:
        """
1208
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
1209
        yet support CUDA graph.
1210
        # TODO Remove this when bitsandbytes supports.
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
        """
        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(
1225
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1226
                "fallback to the eager mode.")
1227

1228
1229
            self.enforce_eager = True

1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
    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.")

1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
    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

1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
    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

1274
        # Reminder: Please update docs/features/compatibility_matrix.md
1275
        # If the feature combo become valid
1276
        from vllm.platforms import current_platform
1277
        if not current_platform.is_async_output_supported(self.enforce_eager):
1278
1279
1280
1281
1282
1283
1284
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

1285
        # Async postprocessor is not necessary for pooling models
1286
        # since there is no token generation
1287
        if self.runner_type == "pooling":
1288
1289
            self.use_async_output_proc = False

1290
        # Reminder: Please update docs/features/compatibility_matrix.md
1291
        # If the feature combo become valid
1292
1293
1294
        if speculative_config:
            self.use_async_output_proc = False

1295
1296
1297
1298
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
1299
1300
1301
1302
1303
1304

        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.")

1305
1306
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
1307
1308
1309
1310
1311
1312
1313
        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}).")

1314
        if parallel_config.enable_expert_parallel:
1315
1316
            self._verify_with_expert_parallelism()

1317
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1318
        if pipeline_parallel_size > 1:
1319
1320
            if not self.registry.is_pp_supported_model(self.architectures,
                                                       self):
1321
1322
1323
1324
1325
1326
                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
1327

1328
1329
    def get_sliding_window(self) -> Optional[int]:
        """Get the sliding window size from the HF text config if present."""
1330
        return getattr(self.hf_text_config, "sliding_window", None)
1331
1332

    def get_vocab_size(self) -> int:
1333
        return getattr(self.hf_text_config, "vocab_size", 0)
1334

1335
    def get_hidden_size(self) -> int:
1336
        return getattr(self.hf_text_config, "hidden_size", 0)
1337

1338
1339
    @property
    def is_deepseek_mla(self) -> bool:
1340
1341
1342
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
bigmoyan's avatar
bigmoyan committed
1343
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp', 'kimi_k2'):
1344
1345
1346
1347
1348
1349
1350
1351
            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
1352

1353
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1354
        # TODO remove hard code
1355
        if self.is_deepseek_mla:
1356
1357
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1358
            if self.use_mla:
1359
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1360
1361
1362
1363
1364
            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
1365

1366
1367
1368
1369
1370
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1371
1372
1373
        if self.is_attention_free:
            return 0

1374
1375
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1376
            return self.hf_text_config.head_dim
1377

1378
        # FIXME(woosuk): This may not be true for all models.
1379
1380
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1381

1382
1383
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1384
        # For GPTBigCode & Falcon:
1385
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1386
1387
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1388
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1389
        new_decoder_arch_falcon = (
1390
            self.hf_config.model_type in falcon_model_types
1391
            and getattr(self.hf_config, "new_decoder_architecture", False))
1392
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1393
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1394
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1395
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1396
            return 1
1397

1398
        # For DBRX and MPT
1399
1400
1401
1402
1403
        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":
1404
1405
1406
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1407
1408
1409
1410
1411
1412
1413
1414
        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")

1415
1416
1417
        if self.is_attention_free:
            return 0

1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
        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:
1428
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1429
1430
1431
1432
1433
            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.
1434
        return self.hf_text_config.num_attention_heads
1435
1436
1437

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1438
1439
1440
1441
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1442
1443
1444
1445
1446
1447
1448
        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)
1449

1450
1451
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1452
1453
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1454

1455
    def get_layers_start_end_indices(
1456
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1457
        from vllm.distributed.utils import get_pp_indices
1458
        if (self.hf_text_config.model_type == "deepseek_mtp"
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1459
1460
                or self.hf_config.model_type == "mimo_mtp"
                or self.hf_config.model_type == "glm4_moe_mtp"):
1461
1462
1463
1464
1465
            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)
1466
1467
1468
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1469
1470
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1471
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1472

1473
1474
1475
    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
1476

1477
1478
1479
1480
1481
1482
1483
1484
    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
1485
1486
1487
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
        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
1498
1499
1500
1501
        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])
1502
        else:
1503
            # Hybrid model Jamba
1504
1505
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
            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])

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

            return sum(t == 1 for t in attn_type_list[start:end])
Mor Zusman's avatar
Mor Zusman committed
1531

1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
    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

1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
    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

1555
    def try_get_generation_config(self) -> dict[str, Any]:
1556
1557
1558
        """
        This method attempts to retrieve the non-default values of the
        generation config for this model.
1559

1560
1561
1562
1563
1564
1565
1566
1567
        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"}:
1568
            config = try_get_generation_config(
1569
                self.hf_config_path or self.model,
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
                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()

1584
    def get_diff_sampling_param(self) -> dict[str, Any]:
1585
        """
1586
1587
1588
1589
1590
1591
1592
1593
1594
        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"`
1595
1596

        Returns:
1597
            A dictionary containing the non-default sampling parameters.
1598
        """
1599
        if self.generation_config == "vllm":
1600
1601
1602
1603
1604
1605
1606
            config = {}
        else:
            config = self.try_get_generation_config()

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

1607
1608
1609
1610
1611
1612
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1613
            "max_new_tokens",
1614
1615
1616
1617
1618
1619
        ]
        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
            }
1620
1621
1622
1623
1624
            # 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")
1625
1626
        else:
            diff_sampling_param = {}
1627
1628
1629
1630
1631
1632
1633

        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`.")
1634
1635
        return diff_sampling_param

1636
    @property
1637
    def is_encoder_decoder(self) -> bool:
1638
        """Extract the HF encoder/decoder model flag."""
1639
        """
1640
        For Mllama, VLLM overrides HF's is_encoder_decoder flag and sets it to
1641
        True to enable cross-attention
1642
        Neuron needs all multimodal data to be in the decoder and does not
1643
1644
1645
1646
1647
1648
        need to explicitly enable cross-attention
        """
        if (current_platform.is_neuron()
                and self.hf_config.model_type == "mllama"):
            return False

1649
1650
1651
1652
1653
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1654

1655
1656
1657
1658
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1659
1660
1661
1662
1663
1664
1665
    @property
    def processor_return_mm_hashes(self) -> bool:
        """Whether the multi-modal processor should output hashes."""
        mm_config = self.multimodal_config
        if mm_config is None:
            return False

1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
        return mm_config.mm_processor_cache_gb > 0

    @property
    def enable_mm_processor_cache(self) -> bool:
        """Whether the multi-modal processor cache should be enabled."""
        mm_config = self.multimodal_config
        if mm_config is None:
            return False

        return mm_config.mm_processor_cache_gb > 0
1676
1677
1678
1679
1680
1681
1682
1683

    def get_mm_input_cache_gb(self) -> int:
        mm_config = self.multimodal_config
        if mm_config is None:
            return 0

        return envs.VLLM_MM_INPUT_CACHE_GIB

1684
1685
    @property
    def is_cross_encoder(self) -> bool:
1686
1687
        return (self._model_info.supports_cross_encoding
                or self.convert_type == "classify")
1688

1689
    @property
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
    def is_pp_supported(self) -> bool:
        return self._model_info.supports_pp

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

    @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
1712

1713
1714
    @property
    def is_v1_compatible(self) -> bool:
1715
1716
1717
1718
1719
        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
1720

1721
1722
    @property
    def is_matryoshka(self) -> bool:
1723
        return (bool(getattr(self.hf_config, "matryoshka_dimensions", None))
1724
1725
                or getattr(self.hf_config, "is_matryoshka", False))

1726
1727
1728
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)
1729

1730
1731
1732
1733
1734
1735
    @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)

1736
    def get_and_verify_max_len(self, max_model_len: int):
1737
1738
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
1739
        tokenizer_config = None
1740
1741
        if (self.runner_type == "pooling" and getattr(
                self.hf_config, "position_embedding_type", "") == "absolute"):
1742
1743
1744
1745
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
                revision=self.tokenizer_revision)
1746
1747
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
1748
            tokenizer_config=tokenizer_config,
1749
1750
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
1751
            sliding_window=self.get_sliding_window(),
1752
1753
            spec_target_max_model_len=self.spec_target_max_model_len,
            encoder_config=self.encoder_config)
1754
        logger.info("Using max model len %s", max_model_len)
1755
1756
        return max_model_len

1757

1758
@config
1759
1760
@dataclass
class LoadConfig:
1761
1762
    """Configuration for loading the model weights."""

1763
    load_format: Union[str, LoadFormats] = "auto"
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
    """The format of the model weights to load:\n
    - "auto" will try to load the weights in the safetensors format and fall
    back to the pytorch bin format if safetensors format is not available.\n
    - "pt" will load the weights in the pytorch bin format.\n
    - "safetensors" will load the weights in the safetensors format.\n
    - "npcache" will load the weights in pytorch format and store a numpy cache
    to speed up the loading.\n
    - "dummy" will initialize the weights with random values, which is mainly
    for profiling.\n
    - "tensorizer" will use CoreWeave's tensorizer library for fast weight
    loading. See the Tensorize vLLM Model script in the Examples section for
    more information.\n
    - "runai_streamer" will load the Safetensors weights using Run:ai Model
    Streamer.\n
    - "bitsandbytes" will load the weights using bitsandbytes quantization.\n
    - "sharded_state" will load weights from pre-sharded checkpoint files,
    supporting efficient loading of tensor-parallel models.\n
    - "gguf" will load weights from GGUF format files (details specified in
    https://github.com/ggml-org/ggml/blob/master/docs/gguf.md).\n
    - "mistral" will load weights from consolidated safetensors files used by
1784
1785
    Mistral models.
    - Other custom values can be supported via plugins."""
1786
    download_dir: Optional[str] = None
1787
1788
    """Directory to download and load the weights, default to the default
    cache directory of Hugging Face."""
1789
1790
    model_loader_extra_config: Union[dict, TensorizerConfig] = field(
        default_factory=dict)
1791
    """Extra config for model loader. This will be passed to the model loader
1792
    corresponding to the chosen load_format."""
1793
1794
1795
    device: Optional[str] = None
    """Device to which model weights will be loaded, default to
    device_config.device"""
1796
    ignore_patterns: Optional[Union[list[str], str]] = None
1797
1798
    """The list of patterns to ignore when loading the model. Default to
    "original/**/*" to avoid repeated loading of llama's checkpoints."""
1799
    use_tqdm_on_load: bool = True
1800
1801
    """Whether to enable tqdm for showing progress bar when loading model
    weights."""
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
    pt_load_map_location: Union[str, dict[str, str]] = "cpu"
    """
    pt_load_map_location: the map location for loading pytorch checkpoint, to
    support loading checkpoints can only be loaded on certain devices like
    "cuda", this is equivalent to {"": "cuda"}. Another supported format is
    mapping from different devices like from GPU 1 to GPU 0:
    {"cuda:1": "cuda:0"}. Note that when passed from command line, the strings
    in dictionary needs to be double quoted for json parsing. For more details,
    see original doc for `map_location` in https://pytorch.org/docs/stable/generated/torch.load.html
    """
1812

1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
    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.
1827
        factors: list[Any] = []
1828
1829
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1830
1831
        return hash_str

1832
    def __post_init__(self):
1833
        self.load_format = self.load_format.lower()
1834
1835
1836
1837
1838
1839
1840
        if self.ignore_patterns is not None and len(self.ignore_patterns) > 0:
            logger.info(
                "Ignoring the following patterns when downloading weights: %s",
                self.ignore_patterns)
        else:
            self.ignore_patterns = ["original/**/*"]

1841
Device = Literal["auto", "cuda", "neuron", "cpu", "tpu", "xpu"]
1842
1843
1844


@config
1845
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
1846
class DeviceConfig:
1847
1848
    """Configuration for the device to use for vLLM execution."""

1849
    device: SkipValidation[Optional[Union[Device, torch.device]]] = "auto"
1850
    """Device type for vLLM execution.
1851
1852
1853
    This parameter is deprecated and will be
    removed in a future release.
    It will now be set automatically based
1854
    on the current platform."""
1855
1856
1857
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
1858

1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
    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.
1874
        factors: list[Any] = []
1875
1876
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1877
1878
        return hash_str

1879
1880
    def __post_init__(self):
        if self.device == "auto":
1881
            # Automated device type detection
1882
            from vllm.platforms import current_platform
1883
            self.device_type = current_platform.device_type
1884
            if not self.device_type:
1885
1886
1887
1888
                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.")
1889
1890
        else:
            # Device type is assigned explicitly
1891
1892
1893
1894
            if isinstance(self.device, str):
                self.device_type = self.device
            elif isinstance(self.device, torch.device):
                self.device_type = self.device.type
1895
1896

        # Some device types require processing inputs on CPU
1897
        if self.device_type in ["neuron"]:
1898
            self.device = torch.device("cpu")
1899
1900
        elif self.device_type in ["tpu"]:
            self.device = None
1901
1902
1903
1904
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1905

1906
1907
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
                            "mlp_speculator", "draft_model", "deepseek_mtp"]
1908
1909
1910


@config
1911
@dataclass
1912
class SpeculativeConfig:
1913
    """Configuration for speculative decoding."""
1914

1915
    # General speculative decoding control
1916
    num_speculative_tokens: SkipValidation[int] = None  # type: ignore
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
    """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."""
1930
    draft_tensor_parallel_size: Optional[int] = None
1931
1932
    """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."""
1933
    disable_logprobs: bool = True
1934
1935
1936
    """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."""
1937

1938
    # Draft model configuration
1939
    quantization: Optional[me_quant.QuantizationMethods] = None
1940
1941
1942
    """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."""
1943
    max_model_len: Optional[int] = None
1944
1945
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
1946
    revision: Optional[str] = None
1947
1948
1949
    """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."""
1950
    code_revision: Optional[str] = None
1951
1952
1953
    """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."""
1954

1955
    # Advanced control
1956
    disable_by_batch_size: Optional[int] = None
1957
1958
1959
1960
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
1961
    prompt_lookup_max: Optional[int] = None
1962
1963
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
1964
    prompt_lookup_min: Optional[int] = None
1965
1966
1967
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

1968
    speculative_token_tree: Optional[str] = None
1969
    """Specifies the tree structure for speculative token generation.
1970
    """
1971
    # required configuration params passed from engine
1972
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
1973
    """The configuration of the target model."""
1974
1975
    target_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
1976
    """The parallel configuration for the target model."""
1977
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
1978
1979
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
1980
    disable_log_stats: SkipValidation[bool] = None  # type: ignore
1981
1982
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
1983
1984

    # params generated in the post-init stage
1985
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
1986
    """The configuration of the draft model initialized internal."""
1987
1988
    draft_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
1989
    """The parallel configuration for the draft model initialized internal."""
1990

1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
    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.
        """
2003
        factors: list[Any] = []
2004
2005
2006
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2007
2008
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2009
2010
        return hash_str

2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
    @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"]
            })
2021
2022
2023
2024
2025
2026
2027
2028
2029

        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
2030
2031
2032
2033
2034
2035
2036
2037
2038

        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"]
            })
2039

2040
2041
        return hf_config

2042
    def __post_init__(self):
2043

2044
2045
2046
2047
2048
2049
2050
        # 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.
2051
2052
2053
2054

        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
2055
            if self.target_model_config and \
2056
2057
2058
2059
                (self.target_model_config.hf_text_config.model_type \
                        == "deepseek_v3" or
                    self.target_model_config.hf_text_config.model_type \
                        == "mimo"):
2060
2061
2062
2063
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2064
            else:
2065
2066
2067
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2068
2069
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2070
2071
2072
2073
2074
2075
2076
        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"
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
            # 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
2091
            if self.prompt_lookup_min < 1:
2092
2093
2094
2095
2096
                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")
2097
            if self.prompt_lookup_min > self.prompt_lookup_max:
2098
2099
2100
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2101

2102
2103
2104
            # 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.
2105
2106
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2107
        else:
2108
2109
2110
2111
2112
2113
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
2114
                    runner="draft",
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
                    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,
                )
2136

2137
                # Automatically detect the method
2138
                if self.method in ('eagle', 'eagle3'):
2139
                    pass
2140
2141
                elif "eagle-" in self.draft_model_config.model.lower() or \
                        "eagle3-" in self.draft_model_config.model.lower():
2142
2143
2144
2145
2146
2147
                    self.method = "eagle"
                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"
2148
                elif (self.draft_model_config.hf_config.model_type
Yuxuan Zhang's avatar
Yuxuan Zhang committed
2149
                      in ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp")):
Jiayi Yao's avatar
Jiayi Yao committed
2150
2151
2152
2153
2154
2155
2156
                    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."
                            )
2157
                else:
2158
                    self.method = "draft_model"
2159
2160
2161
2162
2163
                    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.")
2164
2165

                # Replace hf_config for EAGLE draft_model
2166
                if self.method in ("eagle", "eagle3"):
2167
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2168
                        raise ValueError(
2169
2170
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2171

2172
2173
                    from vllm.transformers_utils.configs import (
                        SpeculatorsConfig)
2174
2175
                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
2176

2177
                    if isinstance(self.draft_model_config.hf_config,
2178
                                  (EAGLEConfig, SpeculatorsConfig)):
2179
2180
2181
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2182
                            self.draft_model_config.hf_config,
2183
2184
                            method=self.method,
                            model_type="eagle")
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
                        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=}")

2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
                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)))

2219
2220
2221
2222
2223
2224
                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
                )
2225

2226
2227
2228
2229
2230
2231
                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,
                    ))
2232

2233
2234
2235
2236
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2237

2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
    @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,
        )

2273
    @staticmethod
2274
    def _verify_and_get_draft_tp(
2275
2276
2277
2278
2279
2280
            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.
2281
        """
2282
2283
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2284
        if speculative_draft_tensor_parallel_size is None:
2285
2286
2287
2288
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2289
2290
2291
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2292
2293
2294
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2295
2296
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2297
            raise ValueError(
2298
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2299
                f"other value than 1 or target model tensor_parallel_size")
2300
        return speculative_draft_tensor_parallel_size
2301

2302
2303
2304
2305
2306
2307
2308
2309
2310
    @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.
        """
2311
2312
2313
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2314
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2315
2316
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
            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

2328
2329
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2330
2331
2332
2333
2334
2335
        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.")

2336
2337
2338
2339
2340
2341
2342
        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)
2343
2344
2345
2346
2347
2348

        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=}")
2349

2350
        eagle3_target_supported = ["llama", "qwen"]
2351
2352
2353
2354
        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):
2355
            raise ValueError(
2356
                f"Eagle3 is only supported for {eagle3_target_supported} models. "  # noqa: E501
2357
2358
                f"Got {self.target_model_config.hf_text_config.model_type=}")

2359
2360
        return self

2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
    @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

2371
    def use_eagle(self) -> bool:
Jiayi Yao's avatar
Jiayi Yao committed
2372
        return self.method in ("eagle", "eagle3", "deepseek_mtp")
2373

2374
    def __repr__(self) -> str:
2375
2376
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
2377
        num_spec_tokens = self.num_speculative_tokens
2378
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
2379
2380


2381
2382
2383
2384
LoRADType = Literal["auto", "float16", "bfloat16"]


@config
2385
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
2386
class LoRAConfig:
2387
2388
2389
2390
2391
2392
    """Configuration for LoRA."""

    max_lora_rank: int = 16
    """Max LoRA rank."""
    max_loras: int = 1
    """Max number of LoRAs in a single batch."""
2393
    fully_sharded_loras: bool = False
2394
2395
2396
2397
    """By default, only half of the LoRA computation is sharded with tensor
    parallelism. Enabling this will use the fully sharded layers. At high
    sequence length, max rank or tensor parallel size, this is likely faster.
    """
2398
    max_cpu_loras: Optional[int] = None
2399
2400
2401
2402
    """Maximum number of LoRAs to store in CPU memory. Must be >= than
    `max_loras`."""
    lora_dtype: Union[torch.dtype, LoRADType] = "auto"
    """Data type for LoRA. If auto, will default to base model dtype."""
2403
    lora_extra_vocab_size: int = 256
2404
2405
    """Maximum size of extra vocabulary that can be present in a LoRA adapter
    (added to the base model vocabulary)."""
2406
2407
    lora_vocab_padding_size: ClassVar[int] = current_platform\
        .get_lora_vocab_padding_size()
2408

2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
    default_mm_loras: Optional[dict[str, str]] = None
    """Dictionary mapping specific modalities to LoRA model paths; this field
    is only applicable to multimodal models and should be leveraged when a
    model always expects a LoRA to be active when a given modality is present.
    Note that currently, if a request provides multiple additional
    modalities, each of which have their own LoRA, we do NOT apply
    default_mm_loras because we currently only support one lora adapter
    per prompt. When run in offline mode, the lora IDs for n modalities
    will be automatically assigned to 1-n with the names of the modalities
    in alphabetic order."""
2419
    bias_enabled: bool = False
2420
    """Enable bias for LoRA adapters."""
2421

2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
    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.
        """
2434
        factors: list[Any] = []
2435
2436
2437
2438
2439
        factors.append(self.max_lora_rank)
        factors.append(self.max_loras)
        factors.append(self.fully_sharded_loras)
        factors.append(self.lora_dtype)
        factors.append(self.lora_extra_vocab_size)
2440
        factors.append(self.lora_vocab_padding_size)
2441
        factors.append(self.bias_enabled)
2442
2443
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2444
2445
        return hash_str

2446
    def __post_init__(self):
2447
        # Setting the maximum rank to 512 should be able to satisfy the vast
2448
        # majority of applications.
2449
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
2450
        possible_lora_extra_vocab_size = (256, 512)
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
        if self.max_lora_rank not in possible_max_ranks:
            raise ValueError(
                f"max_lora_rank ({self.max_lora_rank}) must be one of "
                f"{possible_max_ranks}.")
        if self.lora_extra_vocab_size not in possible_lora_extra_vocab_size:
            raise ValueError(
                f"lora_extra_vocab_size ({self.lora_extra_vocab_size}) "
                f"must be one of {possible_lora_extra_vocab_size}.")
        if self.max_loras < 1:
            raise ValueError(f"max_loras ({self.max_loras}) must be >= 1.")
        if self.max_cpu_loras is None:
            self.max_cpu_loras = self.max_loras
        elif self.max_cpu_loras < self.max_loras:
            raise ValueError(
                f"max_cpu_loras ({self.max_cpu_loras}) must be >= "
zspo's avatar
zspo committed
2466
                f"max_loras ({self.max_loras})")
2467

2468
    def verify_with_cache_config(self, cache_config: CacheConfig):
2469
2470
2471
        if cache_config.cpu_offload_gb > 0 and not envs.VLLM_USE_V1:
            raise ValueError(
                "V0 LoRA does not support CPU offload, please use V1.")
2472

2473
2474
2475
2476
2477
2478
    def verify_with_model_config(self, model_config: ModelConfig):
        if self.lora_dtype in (None, "auto"):
            self.lora_dtype = model_config.dtype
        elif isinstance(self.lora_dtype, str):
            self.lora_dtype = getattr(torch, self.lora_dtype)

2479

2480
@config
2481
@dataclass
2482
class MultiModalConfig:
2483
2484
    """Controls the behavior of multimodal models."""

2485
2486
    limit_per_prompt: dict[str, int] = \
        cast(dict[str, int], get_field(ModelConfig, "limit_mm_per_prompt"))
2487
    """
2488
    The maximum number of input items allowed per prompt for each modality.
2489
    Defaults to 1 (V0) or 999 (V1) for each modality.
2490
2491

    For example, to allow up to 16 images and 2 videos per prompt:
2492
    `{"image": 16, "video": 2}`
2493
2494
    """

2495
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
2496
2497
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
2498
2499
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """

2500
2501
2502
    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
2503
    `transformers.AutoProcessor.from_pretrained`.
2504
2505
2506
2507

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

    For example, for Phi-3-Vision:
2508
    `{"num_crops": 4}`.
2509
2510
    """

2511
    mm_processor_cache_gb: int = 4
2512
    """
2513
2514
2515
2516
2517
2518
2519
    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).
2520
2521
    """

2522
2523
2524
2525
2526
    interleave_mm_strings: bool = False
    """
    Enable fully interleaved support for multimodal prompts.
    """

2527
2528
2529
2530
2531
2532
2533
2534
    skip_mm_profiling: bool = False
    """
    When enabled, skips multimodal memory profiling and only profiles with 
    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.
2535
2536
    """

2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
    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.
2551
        factors: list[Any] = []
2552
2553
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2554
2555
        return hash_str

2556
2557
2558
2559
2560
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
2561
2562
2563
2564
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
2565

2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
    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)
2576

2577

2578
@config
2579
2580
@dataclass
class PoolerConfig:
2581
    """Controls the behavior of output pooling in pooling models."""
2582
2583

    pooling_type: Optional[str] = None
2584
    """
2585
    The pooling method of the pooling model. This should be a key in
2586
    [`vllm.model_executor.layers.pooler.PoolingType`][].
2587
2588
    """

2589
    ## for embeddings models
2590
2591
    normalize: Optional[bool] = None
    """
2592
2593
2594
2595
2596
2597
    Whether to normalize the embeddings outputs. 
    """
    dimensions: Optional[int] = None
    """
    Reduce the dimensions of embeddings if model 
    support matryoshka representation.
2598
2599
    """

2600
2601
    ## for classification models
    activation: Optional[bool] = None
2602
    """
2603
    Whether to apply activation function to the classification outputs. 
2604
2605
    """

2606
2607
2608
2609
2610
    ## for reward models
    softmax: Optional[bool] = None
    """
    Whether to apply softmax to the reward outputs. 
    """
2611
2612
    step_tag_id: Optional[int] = None
    """
2613
    If set, only the score corresponding to the ``step_tag_id`` in the
2614
2615
2616
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """
2617
    returned_token_ids: Optional[list[int]] = None
2618
    """
2619
2620
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2621
2622
2623
    ``math-shepherd-mistral-7b-prm`` model.
    """

2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
    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.
    This parameter enables accepting long inputs without requiring 
    VLLM_ALLOW_LONG_MAX_MODEL_LEN environment variable. 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).
    """

2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
    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.
2657
        factors: list[Any] = []
2658
2659
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2660
2661
        return hash_str

2662

2663
2664
2665
2666
2667
2668
2669
2670
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2671
2672
2673
2674
2675
2676
2677
# 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.",
    "plamo2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "glm4": "Numerical instability. Please use bfloat16 or float32 instead.",
}
2678

2679

2680
2681
2682
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
2683

2684
    return True
2685

2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697

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,
2698
    config: PretrainedConfig,
2699
2700
2701
    *,
    revision: Optional[str],
):
2702
2703
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
2704
    config_dtype = getattr(config, "torch_dtype", None)
2705

2706
    # Fallbacks for multi-modal models if the root config
2707
    # does not define torch_dtype
2708
2709
    if config_dtype is None:
        config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
2710
2711
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)
2712
2713
    if config_dtype is None and hasattr(config, "encoder_config"):
        config_dtype = getattr(config.encoder_config, "torch_dtype", None)
2714

2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
    # 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)
2729

2730
2731
2732
    if config_dtype is None:
        config_dtype = torch.float32

2733
    return config_dtype
2734

Shinichi Hemmi's avatar
Shinichi Hemmi committed
2735

2736
2737
2738
2739
2740
2741
2742
def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform
2743

2744
2745
2746
2747
    supported_dtypes = [
        dtype for dtype in current_platform.supported_dtypes
        if _is_valid_dtype(model_type, dtype)
    ]
2748

2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
    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,
            )
2802
        else:
2803
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
2804
                raise ValueError(f"Unknown dtype: {dtype!r}")
2805
2806
2807
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
2808
    else:
2809
        raise ValueError(f"Unknown dtype: {dtype}")
2810

2811
2812
    _check_valid_dtype(model_type, torch_dtype)

2813
2814
2815
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
2816
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
2817
2818
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
2819
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
2820
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
2821
            # Casting between float16 and bfloat16 is allowed with a warning.
2822
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
2823
2824

    return torch_dtype
2825
2826
2827
2828


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
2829
    tokenizer_config: Optional[dict],
2830
    max_model_len: Optional[int],
2831
    disable_sliding_window: bool,
2832
    sliding_window: Optional[int],
2833
    spec_target_max_model_len: Optional[int] = None,
2834
    encoder_config: Optional[Any] = None,
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
) -> 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",
2845
2846
        # ChatGLM2
        "seq_length",
2847
2848
        # Command-R
        "model_max_length",
2849
2850
        # Whisper
        "max_target_positions",
2851
2852
2853
2854
2855
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
2856
    # Choose the smallest "max_length" from the possible keys
2857
    max_len_key = None
2858
    for key in possible_keys:
2859
2860
2861
2862
2863
        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
2864
2865
2866
2867
    # 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
2868
2869
2870

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
2871
2872
2873
2874
    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
2875

2876
2877
2878
2879
2880
2881
2882
    # 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)

2883
2884
    # If none of the keys were found in the config, use a default and
    # log a warning.
2885
    if derived_max_model_len == float("inf"):
2886
2887
2888
2889
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

2890
2891
2892
2893
2894
        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

2895
2896
2897
2898
        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: "
2899
            "%s. Assuming the model's maximum length is %d.", possible_keys,
2900
            default_max_len)
2901
        derived_max_model_len = default_max_len
2902

2903
    rope_scaling = getattr(hf_config, "rope_scaling", None)
2904
2905
2906
    # 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:
2907
2908
2909
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919

        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.")

2920
2921
2922
2923
            # 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)

2924
2925
2926
2927
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
2928

2929
2930
2931
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2932
2933
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2934
    if max_model_len is None:
2935
        max_model_len = int(derived_max_model_len)
2936
2937
2938
2939
2940
2941
2942
2943
        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)
2944
    elif max_model_len > derived_max_model_len:
2945
2946
2947
2948
2949
        # 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:
2950
2951
2952
2953
2954
2955
2956
            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.")
2957
        else:
2958
            msg = (
2959
                f"User-specified max_model_len ({max_model_len}) is greater "
2960
2961
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2962
                f"{model_max_length} in model's config.json). This may lead "
2963
2964
2965
2966
2967
2968
2969
2970
2971
                "to incorrect model outputs or CUDA errors.")
            if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
                logger.warning(
                    "%s Make sure the value is correct and within the "
                    "model context size.", msg)
            else:
                raise ValueError(
                    f"{msg} To allow overriding this maximum, set "
                    "the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1")
2972
    return int(max_model_len)
2973
2974


2975
def get_served_model_name(model: str,
2976
                          served_model_name: Optional[Union[str, list[str]]]):
2977
    """
2978
2979
2980
2981
    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
2982
2983
2984
2985
2986
2987
2988
2989
2990
    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


2991
GuidedDecodingBackend = Literal["auto", "xgrammar", "guidance", "outlines"]
2992
2993
2994


@config
2995
2996
@dataclass
class DecodingConfig:
2997
    """Dataclass which contains the decoding strategy of the engine."""
2998

2999
    backend: GuidedDecodingBackend = "auto"
3000
3001
3002
3003
    """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."""
3004

3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
    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`."""

3017
    reasoning_backend: str = ""
3018
    """Select the reasoning parser depending on the model that you're using.
3019
    This is used to parse the reasoning content into OpenAI API format."""
3020

3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
    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.
3035
        factors: list[Any] = []
3036
3037
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3038
3039
        return hash_str

3040
    def __post_init__(self):
3041
3042
3043
3044
3045
3046
3047
3048
        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.")

3049

3050
DetailedTraceModules = Literal["model", "worker", "all"]
3051
3052


3053
@config
3054
3055
@dataclass
class ObservabilityConfig:
3056
    """Configuration for observability - metrics and tracing."""
3057

3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
    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)
3073

3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
    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))
3099

3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
    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.
3114
        factors: list[Any] = []
3115
3116
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3117
3118
        return hash_str

3119
    def __post_init__(self):
3120
3121
3122
3123
3124
        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()

3125
        from vllm.tracing import is_otel_available, otel_import_error_traceback
3126
3127
3128
3129
3130
        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}")
3131

3132
3133
3134
3135
3136
3137
    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(","))

3138

3139
3140
3141
3142
KVProducer = Literal["kv_producer", "kv_both"]
KVConsumer = Literal["kv_consumer", "kv_both"]
KVRole = Literal[KVProducer, KVConsumer]

3143

3144
3145
3146
@config
@dataclass
class KVTransferConfig:
3147
3148
3149
    """Configuration for distributed KV cache transfer."""

    kv_connector: Optional[str] = None
3150
3151
    """The KV connector for vLLM to transmit KV caches between vLLM instances.
    """
3152

3153
    engine_id: Optional[str] = None
Robert Shaw's avatar
Robert Shaw committed
3154
    """The engine id for KV transfers."""
3155
3156

    kv_buffer_device: Optional[str] = "cuda"
3157
3158
    """The device used by kv connector to buffer the KV cache.
    Currently only support 'cuda'."""
3159
3160

    kv_buffer_size: float = 1e9
3161
3162
    """The buffer size for TorchDistributedConnector. Measured in number of
    bytes. Recommended value: 1e9 (about 1GB)."""
3163

3164
3165
    kv_role: Optional[KVRole] = None
    """Whether this vLLM instance produces, consumes KV cache, or both. Choices
Robert Shaw's avatar
Robert Shaw committed
3166
    are 'kv_producer', 'kv_consumer', and 'kv_both'."""
3167
3168

    kv_rank: Optional[int] = None
3169
3170
3171
    """The rank of this vLLM instance in the KV cache transfer. Typical value:
    0 for prefill instance, 1 for decode instance.
    Currently only 1P1D is supported."""
3172
3173

    kv_parallel_size: int = 1
3174
3175
    """The number of parallel instances for KV cache transfer. For
    PyNcclConnector, this should be 2."""
3176
3177

    kv_ip: str = "127.0.0.1"
3178
    """The KV connector ip, used to build distributed connection."""
3179
3180

    kv_port: int = 14579
3181
    """The KV connector port, used to build distributed connection."""
3182

3183
3184
    kv_connector_extra_config: dict[str, Any] = field(default_factory=dict)
    """any extra config that the connector may need."""
3185

3186
3187
3188
    kv_connector_module_path: Optional[str] = None
    """The Python module path to dynamically load the KV connector from.
    Only supported in V1."""
3189

3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
    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.
3204
        factors: list[Any] = []
3205
3206
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3207
3208
        return hash_str

3209
    def __post_init__(self) -> None:
3210
3211
3212
        if self.engine_id is None:
            self.engine_id = str(uuid.uuid4())

3213
3214
3215
        if self.kv_role is not None and self.kv_role not in get_args(KVRole):
            raise ValueError(f"Unsupported kv_role: {self.kv_role}. "
                             f"Supported roles are {get_args(KVRole)}")
3216
3217
3218

        if self.kv_connector is not None and self.kv_role is None:
            raise ValueError("Please specify kv_disagg_role when kv_connector "
3219
                             f"is set, supported roles are {get_args(KVRole)}")
3220
3221
3222
3223

    @property
    def is_kv_transfer_instance(self) -> bool:
        return self.kv_connector is not None and \
3224
            self.kv_role in get_args(KVRole)
3225
3226
3227
3228

    @property
    def is_kv_producer(self) -> bool:
        return self.kv_connector is not None and \
3229
            self.kv_role in get_args(KVProducer)
3230
3231
3232
3233

    @property
    def is_kv_consumer(self) -> bool:
        return self.kv_connector is not None and \
3234
            self.kv_role in get_args(KVConsumer)
3235

3236
3237
3238
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

3239

3240
3241
3242
@config
@dataclass
class KVEventsConfig:
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
    """Configuration for KV event publishing."""

    enable_kv_cache_events: bool = False
    """If True, enable KV cache events for tracking block storage and removal.
    Events can be published externally by zmq using the event publisher config.
    """

    publisher: str = "null"
    """The publisher to use for publishing kv events. Can be "null", "zmq".
    """

    endpoint: str = "tcp://*:5557"
    """The zmq endpoint to use for publishing kv events.
    """

    replay_endpoint: Optional[str] = None
    """The zmq endpoint to use for replaying kv events.
    """

    buffer_steps: int = 10_000
    """The number of steps to cache for replay endpoint. Will only save
    events from the last N steps for the replay endpoint.
    """

    hwm: int = 100_000
    """The zmq high water mark for the event publisher. After queueing N events,
    events will start dropping if the consumer is not keeping up.
    """

    max_queue_size: int = 100_000
    """The maximum number of events to queue while waiting for publishing.
    """

    topic: str = ""
    """The topic to use for the event publisher. Consumers can subscribe to
    this topic to receive events.
    """


3282
@config
3283
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3284
3285
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
3286
3287
3288
    simplifies passing around the distinct configurations in the codebase.
    """

3289
3290
3291
    # TODO: use default_factory once default constructing ModelConfig doesn't
    # try to download a model
    model_config: ModelConfig = None  # type: ignore
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
    """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."""
3303
    lora_config: Optional[LoRAConfig] = None
3304
3305
3306
    """LoRA configuration."""
    speculative_config: Optional[SpeculativeConfig] = None
    """Speculative decoding configuration."""
3307
    decoding_config: DecodingConfig = field(default_factory=DecodingConfig)
3308
    """Decoding configuration."""
3309
    observability_config: Optional[ObservabilityConfig] = None
3310
    """Observability configuration."""
3311
    quant_config: Optional[QuantizationConfig] = None
3312
3313
3314
    """Quantization configuration."""
    compilation_config: CompilationConfig = field(
        default_factory=CompilationConfig)
3315
    """`torch.compile` and cudagraph capture configuration for the model.
3316

3317
3318
    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}'`).
3319
    Currently, -O <n> and -O=<n> are supported as well but this will likely be
3320
    removed in favor of clearer -O<n> syntax in the future.
3321
3322
3323

    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
3324
    production, also default in V1.
3325
3326
3327
3328
3329
3330

    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."""
3331
    kv_events_config: Optional[KVEventsConfig] = None
3332
    """The configurations for event publishing."""
3333
    # some opaque config, only used to provide additional information
3334
3335
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
3336
3337
3338
3339
    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."""
3340
    instance_id: str = ""
3341
    """The ID of the vLLM instance."""
3342

3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
    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.
        """
3355
        factors: list[Any] = []
3356
3357

        # summarize vllm config
3358
        vllm_factors: list[Any] = []
3359
3360
        from vllm import __version__
        vllm_factors.append(__version__)
3361
        vllm_factors.append(envs.VLLM_USE_V1)
3362
3363
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3364
3365
        else:
            vllm_factors.append("None")
3366
3367
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3368
3369
        else:
            vllm_factors.append("None")
3370
3371
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3372
3373
        else:
            vllm_factors.append("None")
3374
3375
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3376
3377
        else:
            vllm_factors.append("None")
3378
3379
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3380
3381
        else:
            vllm_factors.append("None")
3382
3383
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3384
3385
        else:
            vllm_factors.append("None")
3386
3387
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3388
3389
3390
3391
3392
            # 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))
3393
3394
        else:
            vllm_factors.append("None")
3395
3396
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3397
3398
        else:
            vllm_factors.append("None")
3399
3400
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3401
3402
        else:
            vllm_factors.append("None")
3403
3404
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3405
3406
        else:
            vllm_factors.append("None")
3407
3408
3409
3410
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3411
3412
        else:
            vllm_factors.append("None")
3413
3414
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3415
3416
3417
        else:
            vllm_factors.append("None")
        if self.additional_config:
3418
3419
3420
3421
3422
3423
3424
3425
            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)
3426
3427
        else:
            vllm_factors.append("None")
3428
3429
        factors.append(vllm_factors)

3430
3431
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
3432
3433
        return hash_str

3434
3435
3436
3437
3438
3439
    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]
3440

3441
3442
3443
3444
3445
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3446
        from vllm.platforms import current_platform
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
        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
3469

3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
    @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)

3481
3482
3483
3484
3485
3486
3487
3488
3489
    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

3490
3491
3492
3493
3494
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

3495
3496
3497
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
3498
3499
3500

        self.try_verify_and_update_config()

3501
3502
3503
3504
3505
        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)
3506
3507
            self.model_config.verify_dual_chunk_attention_config(
                self.load_config)
3508

3509
        self.cache_config.verify_with_parallel_config(self.parallel_config)
3510

3511
        if self.lora_config is not None:
3512
            self.lora_config.verify_with_cache_config(self.cache_config)
3513
            self.lora_config.verify_with_model_config(self.model_config)
3514

3515
        if self.quant_config is None and self.model_config is not None:
3516
3517
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
3518

3519
        from vllm.platforms import current_platform
3520
        if self.model_config is not None and \
3521
3522
3523
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
3524
            logger.warning_once(
3525
3526
3527
3528
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
        # 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
3540

3541
3542
3543
3544
3545
            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

3546
3547
3548
3549
3550
        # 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
3551
3552
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
3553

3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
        if current_platform.is_cuda_alike():
            # 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
3564

3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
            # 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
3576

3577
        if self.cache_config.cpu_offload_gb > 0 and \
3578
3579
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
3580
            logger.warning(
3581
                "CPU offload is not supported with `torch.compile` in v0 yet."
3582
3583
3584
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3585
3586
3587
3588
3589
3590
        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`.")
3591
3592
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3593
3594
3595
3596
3597
3598
3599
3600
        disable_chunked_prefill_reasons: list[str] = []

        if self.model_config and 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.")
3601
3602
3603
3604
            elif 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.")
3605
3606
3607
3608
3609
3610
3611
3612
3613

        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
3614

3615
        if (self.kv_events_config is not None
3616
3617
3618
3619
3620
                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.")
3621
3622
        if (self.kv_events_config is not None
                and self.kv_events_config.publisher != "null"
3623
3624
3625
3626
3627
                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.")
3628
3629
        current_platform.check_and_update_config(self)

3630
        # final check of cudagraph mode after platform-specific update
3631
        if envs.VLLM_USE_V1 and current_platform.is_cuda_alike():
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
            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}"

3648
3649
3650
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

3651
3652
3653
3654
3655
        # 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()

3656
3657
3658
3659
3660
3661
3662
        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.
            if not (current_platform.is_cuda() or current_platform.is_rocm()):
                # Hybrid KV cache manager is not supported on non-GPU platforms.
3663
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3664
3665
            if self.kv_transfer_config is not None:
                # Hybrid KV cache manager is not compatible with KV transfer.
3666
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3667
3668
            if self.kv_events_config is not None:
                # Hybrid KV cache manager is not compatible with KV events.
3669
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3670
            if self.model_config is not None and \
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
                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
3688

3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
    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
        ]

3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
    def _set_cudagraph_sizes(self):
        """
        cudagraph batchsize padding logic:

        `[1, 2, 4] + [8 * i for i in range(1, 1025)]` is a list of all possible
        batch sizes that cudagraph will capture.

        Depending on the engine's configuration of `max_num_seqs`, the
        candidate batch sizes to capture cudagraph will shrink to the subset
        which just cover the range of `[1, max_num_seqs]`. In the common case,
        `max_num_seqs` is 256, and the cudagraph batch sizes will be
        `[1, 2, 4, 8, 16, 24, 32, 40, ..., 256]`.

        However, if users specify the cudagraph capture sizes through
        compilation config, we will use the specified sizes instead.

3725
3726
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
3727
3728

        During runtime, if batchsize is larger than
3729
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3730
3731
        no cudagraph will be used.
        If the batch size is no larger than
3732
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
        we can quickly find the padded graph size for a given batch size by
        looking up `vllm_config.compilation_config.bs_to_padded_graph_size`.
        """

        # 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)]
3745
3746
3747
3748
3749
                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)

3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
                # 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:
3771
3772
3773
3774
3775
3776
3777
3778
                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
3779
                    raise TypeError(f"Invalid value for {cuda_graph_sizes=}.")
3780
3781
3782
3783
                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)
3784
3785
3786
3787
3788
                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
                ]
3789
3790
3791
3792

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

3793
    def recalculate_max_model_len(self, max_model_len: int):
3794
        # Can only be called in try_verify_and_update_config
3795
        model_config = self.model_config
3796
        max_model_len = model_config.get_and_verify_max_len(max_model_len)
3797
3798
        self.model_config.max_model_len = max_model_len
        self.scheduler_config.max_model_len = max_model_len
3799
3800

    def try_verify_and_update_config(self):
3801
3802
3803
        if self.model_config is None:
            return

3804
3805
3806
3807
3808
        # Avoid running try_verify_and_update_config multiple times
        if getattr(self.model_config, "config_updated", False):
            return
        self.model_config.config_updated = True

3809
        architecture = self.model_config.architecture
3810
3811
3812
        if architecture is None:
            return

3813
3814
        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP, HybridAttentionMambaModelConfig)
3815
3816
3817
        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_config(self)
3818

3819
3820
3821
        if self.model_config.is_hybrid:
            HybridAttentionMambaModelConfig.verify_and_update_config(self)

3822
        if self.model_config.convert_type == "classify":
3823
3824
3825
3826
            # Maybe convert ForCausalLM into ForSequenceClassification model.
            from vllm.model_executor.models.adapters import (
                SequenceClassificationConfig)
            SequenceClassificationConfig.verify_and_update_config(self)
3827

3828
    def __str__(self):
3829
        return (
3830
3831
3832
3833
3834
            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}, "
3835
            f"revision={self.model_config.revision}, "
3836
3837
            f"override_neuron_config={self.model_config.override_neuron_config}, "  # noqa
            f"tokenizer_revision={self.model_config.tokenizer_revision}, "
3838
3839
            f"trust_remote_code={self.model_config.trust_remote_code}, "
            f"dtype={self.model_config.dtype}, "
3840
3841
            f"max_seq_len={self.model_config.max_model_len}, "
            f"download_dir={self.load_config.download_dir!r}, "
3842
            f"load_format={self.load_config.load_format}, "
3843
3844
            f"tensor_parallel_size={self.parallel_config.tensor_parallel_size}, "  # noqa
            f"pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, "  # noqa
3845
3846
3847
3848
            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}, "
3849
            f"device_config={self.device_config.device}, "
3850
3851
3852
3853
3854
3855
3856
            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}, "
3857
3858
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
3859
3860
3861


_current_vllm_config: Optional[VllmConfig] = None
3862
_current_prefix: Optional[str] = None
3863
3864
3865


@contextmanager
3866
3867
3868
def set_current_vllm_config(vllm_config: VllmConfig,
                            check_compile=False,
                            prefix: Optional[str] = None):
3869
    """
3870
    Temporarily set the current vLLM config.
3871
    Used during model initialization.
3872
    We save the current vLLM config in a global variable,
3873
    so that all modules can access it, e.g. custom ops
3874
    can access the vLLM config to determine how to dispatch.
3875
    """
3876
    global _current_vllm_config, _current_prefix
3877
    old_vllm_config = _current_vllm_config
3878
    old_prefix = _current_prefix
3879
3880
3881
3882
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
3883
        _current_prefix = prefix
3884
        yield
3885
3886
3887
    except Exception:
        raise
    else:
3888
3889
3890
3891
        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)
3892
3893
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
3894
3895
3896
3897
3898
3899
3900
3901
3902
            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"
3903
                " if you want it to be supported.",
3904
                vllm_config.model_config.model)
3905
    finally:
3906
        _current_vllm_config = old_vllm_config
3907
        _current_prefix = old_prefix
3908
3909
3910
3911
3912
3913
3914
3915
        # 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
3916
3917
3918
3919
3920
3921
3922


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.
3923
        logger.warning("Current vLLM config is not set.")
3924
3925
3926
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
3927
3928


3929
3930
3931
3932
3933
3934
3935
3936
3937
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


3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
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
    a max of 16 on a 64 bit system).

    Args:
        text (str): The text to check

    Returns:
3949
        result (bool): `True` if a match is found, `False` otherwise.
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
    """
    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}")
3963
3964
3965
3966
3967


T = TypeVar("T")


3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
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

3987
    return {
3988
3989
3990
        layer_name: forward_context[layer_name]
        for layer_name in layer_names
        if isinstance(forward_context[layer_name], layer_type)
3991
    }
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021


@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:
4022
        return self.min_energy_split_window_size is not None
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040


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)