__init__.py 168 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

zhuwenwen's avatar
zhuwenwen committed
4
import os
5

6
# ruff: noqa: F401
7
import ast
8
import copy
9
import enum
10
import hashlib
11
import inspect
12
import json
13
import os
14
import textwrap
15
import warnings
16
from collections.abc import Mapping
17
from contextlib import contextmanager
18
from dataclasses import MISSING, Field, field, fields, is_dataclass, replace
19
from functools import cached_property, lru_cache
20
from importlib.util import find_spec
21

zhuwenwen's avatar
zhuwenwen committed
22
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Literal, Optional, List,
23
                    Protocol, TypeVar, Union, cast, get_args)
24

25
import regex as re
26
import torch
27
from pydantic import (ConfigDict, SkipValidation, field_validator,
28
29
                      model_validator)
from pydantic.dataclasses import dataclass
30
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
31
from typing_extensions import Self, assert_never, runtime_checkable
32

33
import vllm.envs as envs
34
from vllm import version
35
from vllm.config.cache import (BlockSize, CacheConfig, CacheDType, MambaDType,
36
                               PrefixCachingHashAlgo)
37
from vllm.config.compilation import (CompilationConfig, CompilationLevel,
38
                                     CUDAGraphMode, PassConfig)
39
from vllm.config.kv_events import KVEventsConfig
40
from vllm.config.kv_transfer import KVTransferConfig
41
from vllm.config.load import LoadConfig
42
from vllm.config.lora import LoRAConfig
43
44
from vllm.config.parallel import (DistributedExecutorBackend, EPLBConfig,
                                  ParallelConfig)
45
from vllm.config.scheduler import SchedulerConfig, SchedulerPolicy
46
from vllm.config.utils import ConfigType, config
Woosuk Kwon's avatar
Woosuk Kwon committed
47
from vllm.logger import init_logger
48
from vllm.model_executor.layers.quantization import QuantizationMethods
49
from vllm.multimodal import MULTIMODAL_REGISTRY
50
from vllm.platforms import current_platform
51
52
53
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
54
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
55
    is_interleaved, maybe_override_with_speculators_target_model,
56
57
    try_get_generation_config, try_get_safetensors_metadata,
    try_get_tokenizer_config, uses_mrope)
58
59
60
from vllm.transformers_utils.runai_utils import (ObjectStorageModel,
                                                 is_runai_obj_uri)
from vllm.transformers_utils.utils import maybe_model_redirect
61
62
from vllm.utils import (DEFAULT_MAX_NUM_BATCHED_TOKENS,
                        STR_DUAL_CHUNK_FLASH_ATTN_VAL, LayerBlockType,
63
                        LazyLoader, common_broadcastable_dtype, random_uuid)
zhuwenwen's avatar
zhuwenwen committed
64
from vllm.utils import SUPPORT_TC
65

66
if TYPE_CHECKING:
67
    from _typeshed import DataclassInstance
68
    from transformers.configuration_utils import PretrainedConfig
69

70
71
72
    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
73
74
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
75
    from vllm.v1.sample.logits_processor import LogitsProcessor
76

77
    HfOverrides = Union[dict, Callable[[type], type]]
78
else:
79
    DataclassInstance = Any
80
    PretrainedConfig = Any
81
    QuantizationConfig = Any
82
    QuantizationMethods = Any
83
    BaseModelLoader = Any
84
    LogitsProcessor = Any
85
86
87
88
89
90
    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")
91

92
93
logger = init_logger(__name__)

zhuwenwen's avatar
zhuwenwen committed
94
models_path_prefix = os.getenv('VLLM_OPTEST_MODELS_PATH') or os.getenv("OPTEST_MODELS_PATH")
95

96
DataclassInstanceT = TypeVar("DataclassInstanceT", bound=DataclassInstance)
97

98
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
99
                     "score", "reward", "transcription", "draft"]
100

101
_ResolvedTask = Literal["generate", "transcription", "encode", "embed",
102
                        "classify", "reward", "draft"]
103

104
RunnerOption = Literal["auto", "generate", "pooling", "draft"]
105

106
RunnerType = Literal["generate", "pooling", "draft"]
107

108
109
110
111
112
ConvertOption = Literal["auto", "none", "embed", "classify", "reward"]

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

_RUNNER_TASKS: dict[RunnerType, list[TaskOption]] = {
113
    "generate": ["generate", "transcription"],
114
115
116
117
118
119
120
    "pooling": ["embedding", "embed", "classify", "score", "reward"],
    "draft": ["draft"],
}

_RUNNER_CONVERTS: dict[RunnerType, list[ConvertType]] = {
    "generate": [],
    "pooling": ["embed", "classify", "reward"],
121
    "draft": [],
122
}
123

124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
# 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

165

166
@runtime_checkable
167
168
169
170
171
172
class SupportsHash(Protocol):

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


173
174
class SupportsMetricsInfo(Protocol):

175
    def metrics_info(self) -> dict[str, str]:
176
177
178
        ...


179
180
181
182
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"
183
    TERRATORCH = "terratorch"
184
185


186
187
188
189
190
191
192
193
194
195
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
196

197
198
199
200
201
202
203
204
205
        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

206
207
208
209
210
211
212
213
214
215
216
    try:
        cls_node = ast.parse(textwrap.dedent(inspect.getsource(cls))).body[0]
    except (OSError, KeyError, TypeError):
        # HACK: Python 3.13+ workaround - set missing __firstlineno__
        # Workaround can be removed after we upgrade to pydantic==2.12.0
        with open(inspect.getfile(cls)) as f:
            for i, line in enumerate(f):
                if f"class {cls.__name__}" in line and ":" in line:
                    cls.__firstlineno__ = i + 1
                    break
        cls_node = ast.parse(textwrap.dedent(inspect.getsource(cls))).body[0]
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247

    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


248
def get_field(cls: ConfigType, name: str) -> Field:
249
250
251
252
253
254
255
    """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__}.")
256
    named_field: Field = cls_fields[name]
257
258
259
260
261
262
263
264
    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.")


265
266
def is_init_field(cls: ConfigType, name: str) -> bool:
    return next(f for f in fields(cls) if f.name == name).init
267

268

zhuwenwen's avatar
zhuwenwen committed
269
TokenizerMode = Literal["auto", "cpm", "slow", "mistral", "custom"]
270
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
271
MMEncoderTPMode = Literal["weights", "data"]
272
273


274
275
276
277
278
class LogprobsMode(enum.Enum):
    RAW_LOGITS = "raw_logits"
    RAW_LOGPROBS = "raw_logprobs"
    PROCESSED_LOGITS = "processed_logits"
    PROCESSED_LOGPROBS = "processed_logprobs"
279
280
281


@config
282
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
283
class ModelConfig:
284
    """Configuration for the model."""
285
286
    
    model: str = os.path.join(models_path_prefix, "Qwen/Qwen3-0.6B") if models_path_prefix is not None else "Qwen/Qwen3-0.6B"
287
288
289
    """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."""
290
291
292
    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."""
293
294
295
296
297
298
299
300
301
302
    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.
    """
303
    tokenizer: SkipValidation[str] = None  # type: ignore
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
    """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
325
326
    """Random seed for reproducibility. Initialized to None in V0, but
    initialized to 0 in V1."""
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
    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)
342
    """RoPE scaling configuration. For example,
343
344
345
346
347
348
349
350
    `{"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."""
351
    max_model_len: SkipValidation[int] = None  # type: ignore
352
353
    """Model context length (prompt and output). If unspecified, will be
    automatically derived from the model config.
354

355
356
357
358
359
360
    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
361
    """Specify the maximum length for spec decoding draft models."""
362
    quantization: SkipValidation[Optional[QuantizationMethods]] = None
363
364
365
366
367
368
369
370
371
    """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."""
372
    max_seq_len_to_capture: Optional[int] = None # 8192
373
374
375
376
377
378
379
    """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
380
381
    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."""
382
    logprobs_mode: LogprobsMode = LogprobsMode.RAW_LOGPROBS
383
384
385
    """Indicates the content returned in the logprobs and prompt_logprobs.
    Supported mode:
    1) raw_logprobs, 2) processed_logprobs, 3) raw_logits, 4) processed_logits.
386
387
388
    Raw means the values before applying any logit processors, like bad words.
    Processed means the values after applying all processors, including
    temperature and top_k/top_p.
389
    """
390
391
392
393
394
395
396
397
398
399
400
401
402
403
    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."""
404
405
406
407
    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."""
408
409
410
411
412
413
414
415
416
417
418
    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."""
419
    interleave_mm_strings: bool = False
420
    """Enable fully interleaved support for multimodal prompts, while using
421
    --chat-template-content-format=string. Defaults to False."""
422
423
424
425
    skip_mm_profiling: bool = False
    """When enabled, skips multimodal memory profiling and only profiles with
    language backbone model during engine initialization.
    """
426
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
427
428
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
429
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
430
431
    use_async_output_proc: bool = True
    """Whether to use async output processor."""
432
    config_format: Union[str, ConfigFormat] = "auto"
433
434
435
436
437
438
439
440
441
442
443
    """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
444
    config. If a callable, it is called to update the HuggingFace config."""
445
446
447
448
449
    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}`.
450
    """
451
    mm_processor_cache_gb: float = 4
452
453
454
455
456
457
458
459
    """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)."""
460
461
462
463
464
465
466
467
468
469
470
471
472
    mm_encoder_tp_mode: MMEncoderTPMode = "weights"
    """Indicates how to optimize multi-modal encoder inference using
    tensor parallelism (TP).

    - `"weights"`: Within the same vLLM engine, split the weights of
        each layer across TP ranks. (default TP behavior)
    - `"data"`: Within the same vLLM engine, split the batched input data
        across TP ranks to process the data in parallel, while hosting
        the full weights on each TP rank.
        This batch-level DP is not to be confused with API request-level
        DP (which is controlled by `--data-parallel-size`).
        This is only supported on a per-model basis and falls back to
        `"weights"` if the encoder does not support DP."""
473
474
475
476
477
478
    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}`.
479
    """
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
    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
495
    `--generation-config vllm`, only the override parameters are used."""
496
497
498
499
500
501
502
503
    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
504
505
506
    - "transformers" will use the Transformers model implementation.\n
    - "terratorch" will use the TerraTorch model implementation.
    """
507
508
    override_attention_dtype: Optional[str] = None
    """Override dtype for attention"""
509
510
511
    logits_processors: Optional[list[Union[str, type[LogitsProcessor]]]] = None
    """One or more logits processors' fully-qualified class names or class
    definitions"""
512
513
    io_processor_plugin: Optional[str] = None
    """IOProcessor plugin name to load at model startup"""
514

515
516
517
518
    enable_chunked_prefill: Optional[bool] = None
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""

519
520
521
522
523
524
525
526
527
528
529
530
    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.
        """
531
        factors: list[Any] = []
532
533
534
535
536
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
537
538
539
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
540
        factors.append(self.trust_remote_code)
541
542
543
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
544
545
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
546
547
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
548
        factors.append(self.enable_chunked_prefill)
549
550
        str_factors = str(factors)
        assert_hashable(str_factors)
551
552
        return hashlib.sha256(str(factors).encode()).hexdigest()

553
    def __post_init__(self) -> None:
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
        # 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)

572
573
574
        # 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)
575
576
577
578
579
580
581
582
583
584
585
586
        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):
587
            hf_overrides_kw = {}
588
            hf_overrides_fn = self.hf_overrides
589
        else:
590
            hf_overrides_kw = self.hf_overrides
591
            hf_overrides_fn = None
592

593
594
        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
595
            hf_overrides_kw.update(hf_override)
596
            hf_overrides_str = json.dumps(hf_overrides_kw)
597
598
599
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
600
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
601
602
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
603
            hf_overrides_kw.update(hf_override)
604
            hf_overrides_str = json.dumps(hf_overrides_kw)
605
606
607
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
608
609
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

610
611
612
613
614
615
616
617
618
619
        self.maybe_pull_model_tokenizer_for_runai(self.model, self.tokenizer)

        if self.runner != "draft":
            # If we're not running the draft model, check for speculators config
            # If speculators config, set model / tokenizer to be target model
            self.model, self.tokenizer = maybe_override_with_speculators_target_model(  # noqa: E501
                model=self.model,
                tokenizer=self.tokenizer,
                revision=self.revision,
                trust_remote_code=self.trust_remote_code)
620

621
622
623
624
        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 "
625
626
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
627
628
                "for instructions on how to install it.")

629
630
        from vllm.platforms import current_platform

631
632
633
634
635
636
        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)

637
638
639
640
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
641

642
        hf_config = get_config(self.hf_config_path or self.model,
643
644
645
646
647
648
                               self.trust_remote_code,
                               self.revision,
                               self.code_revision,
                               self.config_format,
                               hf_overrides_kw=hf_overrides_kw,
                               hf_overrides_fn=hf_overrides_fn)
649

650
        self.hf_config = hf_config
651
        self.hf_text_config = get_hf_text_config(self.hf_config)
652
653
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
654
        self.encoder_config = self._get_encoder_config()
655
        self.hf_image_processor_config = get_hf_image_processor_config(
656
            self.model, hf_token=self.hf_token, revision=self.revision)
657

658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
        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
715
            else:
716
717
718
719
720
721
722
723
                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}"
724
725
            warnings.warn(msg, DeprecationWarning, stacklevel=2)

726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
        self.runner_type = self._get_runner_type(architectures, self.runner)
        self.convert_type = self._get_convert_type(architectures,
                                                   self.runner_type,
                                                   self.convert)

        if self.runner_type == "generate" and not is_generative_model:
            generate_converts = _RUNNER_CONVERTS["generate"]
            if self.convert_type not in generate_converts:
                # Currently we don't have any converters for generative models
                raise ValueError(
                    "This model does not support `--runner generate`.")
        if self.runner_type == "pooling" and not is_pooling_model:
            pooling_converts = _RUNNER_CONVERTS["pooling"]
            if self.convert_type not in pooling_converts:
                convert_option = "<" + "|".join(pooling_converts) + ">"
                raise ValueError(
                    "This model does not support `--runner pooling`. "
                    f"You can pass `--convert {convert_option} to adapt "
                    "it into a pooling model.")
745

746
747
748
749
750
751
        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)
752
753
        self._model_info = model_info
        self._architecture = arch
754
        logger.info("Resolved architecture: %s", arch)
755
756
757

        self.pooler_config = self._init_pooler_config()

758
        self.dtype: torch.dtype = _get_and_verify_dtype(
759
760
761
762
763
764
            self.model,
            self.hf_config,
            self.dtype,
            is_pooling_model=self.runner_type == "pooling",
            revision=self.revision,
        )
Woosuk Kwon's avatar
Woosuk Kwon committed
765

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

782
        self.original_max_model_len = self.max_model_len
783
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
784
        self.multimodal_config = self._init_multimodal_config()
785

786
787
788
789
790
        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

791
792
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
793

794
795
796
        # Avoid running try_verify_and_update_config multiple times
        self.config_updated = False

797
        self._verify_quantization()
798
        self._verify_cuda_graph()
799
        self._verify_bnb_config()
800

801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
    @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

817
818
819
    def _get_transformers_backend_cls(self) -> str:
        """Determine which Transformers backend class will be used if
        `model_impl` is set to `transformers` or `auto`."""
820
821
        if getattr(self, "runner_type", self.runner) == "pooling":
            return "TransformersModel"
822
823
824
825
        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"
826
827
828
829
830
        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()
831

832
833
    @property
    def registry(self):
834
        return me_models.ModelRegistry
835
836
837

    @property
    def architectures(self) -> list[str]:
838
        return getattr(self.hf_config, "architectures", [])
839

840
841
    @property
    def architecture(self) -> str:
842
        """The architecture vllm actually used."""
843
844
        return self._architecture

845
846
847
848
    def maybe_pull_model_tokenizer_for_runai(self, model: str,
                                             tokenizer: str) -> None:
        """Pull model/tokenizer from Object Storage to temporary
        directory when needed.
849
850

        Args:
851
852
            model: Model name or path
            tokenizer: Tokenizer name or path
853
        """
854
        if not (is_runai_obj_uri(model) or is_runai_obj_uri(tokenizer)):
855
            return
856

857
858
859
860
        if is_runai_obj_uri(model):
            object_storage_model = ObjectStorageModel()
            object_storage_model.pull_files(
                model, allow_pattern=["*.model", "*.py", "*.json"])
861
            self.model_weights = model
862
            self.model = object_storage_model.dir
863
864
865

            # If tokenizer is same as model, download to same directory
            if model == tokenizer:
866
867
868
869
870
871
                object_storage_model.pull_files(model,
                                                ignore_pattern=[
                                                    "*.pt", "*.safetensors",
                                                    "*.bin", "*.tensors"
                                                ])
                self.tokenizer = object_storage_model.dir
872
                return
873

874
        # Only download tokenizer if needed and not already handled
875
876
877
        if is_runai_obj_uri(tokenizer):
            object_storage_tokenizer = ObjectStorageModel()
            object_storage_tokenizer.pull_files(
878
879
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors"])
880
            self.tokenizer = object_storage_tokenizer.dir
881

882
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
883
        if self._model_info.supports_multimodal:
884
885
886
887
888
889
890
            if (self.mm_encoder_tp_mode == "data" and
                    not self._model_info.supports_multimodal_encoder_tp_data):
                logger.warning_once(
                    "This model does not support `--mm-encoder-tp-mode data`. "
                    "Falling back to `--mm-encoder-tp-mode weights`.")
                self.mm_encoder_tp_mode = "weights"

891
            return MultiModalConfig(
892
                limit_per_prompt=self.limit_mm_per_prompt,
893
                media_io_kwargs=self.media_io_kwargs,
894
                mm_processor_kwargs=self.mm_processor_kwargs,
895
                mm_processor_cache_gb=self.mm_processor_cache_gb,
896
                mm_encoder_tp_mode=self.mm_encoder_tp_mode,
897
                interleave_mm_strings=self.interleave_mm_strings,
898
899
                skip_mm_profiling=self.skip_mm_profiling,
            )
900
901

        return None
902

903
904
905
906
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

907
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
908
        if self.runner_type == "pooling":
909
910
911
912
913
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
914
915
916
917
918

            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():
919
920
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
921

922
923
924
            default_pooling_type = self._model_info.default_pooling_type
            if pooler_config.pooling_type is None:
                pooler_config.pooling_type = default_pooling_type
925

926
            return pooler_config
927

928
929
        return None

930
    def _verify_tokenizer_mode(self) -> None:
931
932
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
933
934
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
935
                f"one of {get_args(TokenizerMode)}.")
936
        self.tokenizer_mode = tokenizer_mode
937

938
939
940
941
942
943
944
945
946
947
    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"

948
        for arch in architectures:
949
950
951
952
953
954
955
956
957
958
959
960
            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"
961

962
    def _get_runner_type(
963
        self,
964
        architectures: list[str],
965
966
967
968
        runner: RunnerOption,
    ) -> RunnerType:
        if runner != "auto":
            return runner
969

970
971
        runner_type = self._get_default_runner_type(architectures)

972
973
974
975
976
977
        # 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)
978
979
980
981

        return runner_type

    def _get_default_convert_type(
982
        self,
983
        architectures: list[str],
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
        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":
1008
1009
            return "embed"

1010
        return "none"
1011

1012
1013
1014
1015
1016
1017
1018
1019
    def _get_convert_type(
        self,
        architectures: list[str],
        runner_type: RunnerType,
        convert: ConvertOption,
    ) -> ConvertType:
        if convert != "auto":
            return convert
1020

1021
1022
        convert_type = self._get_default_convert_type(architectures,
                                                      runner_type)
1023

1024
1025
1026
1027
1028
1029
        # 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)
1030
1031

        return convert_type
1032

1033
    def _get_supported_generation_tasks(
1034
        self,
1035
1036
        architectures: list[str],
        convert_type: ConvertType,
1037
1038
    ) -> list[_ResolvedTask]:
        registry = self.registry
1039

1040
        if registry.is_transcription_only_model(architectures, self):
1041
1042
            return ["transcription"]

1043
        # TODO: Use get_supported_generation_tasks once V0 is removed
1044
        supported_tasks = list[_ResolvedTask]()
1045
1046
        if (registry.is_text_generation_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["generate"]):
1047
1048
            supported_tasks.append("generate")

1049
1050
        if registry.is_transcription_model(architectures, self):
            supported_tasks.append("transcription")
1051
1052

        return supported_tasks
1053

1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
    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"

1071
1072
    def _get_supported_pooling_tasks(
        self,
1073
1074
        architectures: list[str],
        convert_type: ConvertType,
1075
    ) -> list[_ResolvedTask]:
1076
        registry = self.registry
1077

1078
        # TODO: Use get_supported_pooling_tasks once V0 is removed
1079
        supported_tasks = list[_ResolvedTask]()
1080
1081
        if (registry.is_pooling_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["pooling"]):
1082
            supported_tasks.append("encode")
1083

1084
1085
1086
            extra_task = (self._get_default_pooling_task(architectures)
                          if convert_type == "none" else convert_type)
            supported_tasks.append(extra_task)
1087

1088
        return supported_tasks
1089

1090
1091
    def _get_supported_tasks(
        self,
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
        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"]
1104

1105
        assert_never(runner_type)
1106

1107
1108
    def _parse_quant_hf_config(self, hf_config: PretrainedConfig):
        quant_cfg = getattr(hf_config, "quantization_config", None)
1109
        if quant_cfg is None:
1110
            # compressed-tensors uses a "compression_config" key
1111
            quant_cfg = getattr(hf_config, "compression_config", None)
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126

        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}")

1127
1128
        return quant_cfg

1129
    def _verify_quantization(self) -> None:
1130
        supported_quantization = me_quant.QUANTIZATION_METHODS
1131
        optimized_quantization_methods = [
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
            "fp8",
            "modelopt",
            "gptq_marlin_24",
            "gptq_marlin",
            "awq_marlin",
            "fbgemm_fp8",
            "compressed-tensors",
            "experts_int8",
            "quark",
            "modelopt_fp4",
            "bitblas",
            "gptq_bitblas",
            "inc",
            "petit_nvfp4",
1146
1147
            "slimquant_w4a8", 
            "slimquant_w4a8_marlin"
1148
        ]
1149
        if self.quantization is not None:
1150
1151
            self.quantization = cast(me_quant.QuantizationMethods,
                                     self.quantization)
1152
1153

        # Parse quantization method from the HF model config, if available.
1154
1155
1156
1157
1158
        quant_cfg = self._parse_quant_hf_config(self.hf_config)
        if quant_cfg is None and (text_config := getattr(
                self.hf_config, "text_config", None)):
            # Check the text config as well for multi-modal models.
            quant_cfg = self._parse_quant_hf_config(text_config)
1159

1160
        if quant_cfg is not None:
1161
            # Use the community standard 'quant_method'
1162
            quant_method = quant_cfg.get("quant_method", "").lower()
1163
1164

            # Normalize library names
1165
1166
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
1167

1168
            quant_cfg["quant_method"] = quant_method
1169

1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
            # Quantization methods which are overrides (i.e. they have a
            # `override_quantization_method` method) must be checked in order
            # of preference (this is particularly important for GPTQ).
            overrides = [
                "bitblas",
                "gptq_marlin_24",
                "gptq_marlin",
                "gptq_bitblas",
                "awq_marlin",
                "ipex",
                "moe_wna16",
1181
                "slimquant_w4a8_marlin"
1182
1183
                "modelopt",
                "modelopt_fp4",
1184
                "petit_nvfp4",
1185
1186
1187
1188
1189
1190
            ]
            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
1191
            # over the built-in ones.
1192
            quantization_methods = quantization_methods + overrides
1193
1194

            # Detect which checkpoint is it
1195
            for name in quantization_methods:
1196
                method = me_quant.get_quantization_config(name)
1197
1198
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
1199
1200
1201
1202
                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.
1203
                    if (name in get_args(me_quant.QuantizationMethods)
1204
1205
1206
1207
1208
1209
                            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.")
1210
1211
1212
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
1213

1214
            # Verify quantization configurations.
1215
            if self.quantization is None:
1216
1217
                self.quantization = quant_method
            elif self.quantization != quant_method:
1218
1219
                raise ValueError(
                    "Quantization method specified in the model config "
1220
                    f"({quant_method}) does not match the quantization "
1221
1222
1223
1224
1225
1226
1227
1228
                    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}.")
1229
            from vllm.platforms import current_platform
1230
            current_platform.verify_quantization(self.quantization)
1231
            if self.quantization not in optimized_quantization_methods:
1232
                logger.warning(
1233
                    "%s quantization is not fully "
1234
                    "optimized yet. The speed can be slower than "
1235
                    "non-quantized models.", self.quantization)
1236

1237
    def _verify_cuda_graph(self) -> None:
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
        # 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))
zhuwenwen's avatar
zhuwenwen committed
1248
1249
1250
        if self.max_seq_len_to_capture is None:
            self.max_seq_len_to_capture = self.max_model_len
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
1251
                                          effective_max_seq_len)
1252
        # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
1253
        ROCM_UNSUPPORTED_MODELS = ['mllama']
1254
1255
1256
1257
1258
1259
        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()):
1260
1261
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1262
                "to eager mode.", self.hf_config.model_type)
1263
            self.enforce_eager = True
1264

1265
1266
    def _verify_bnb_config(self) -> None:
        """
1267
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
1268
        yet support CUDA graph.
1269
        # TODO Remove this when bitsandbytes supports.
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
        """
        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(
1284
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1285
                "fallback to the eager mode.")
1286

1287
1288
            self.enforce_eager = True

1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
    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.")

1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
    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

1323
1324
1325
1326
            if envs.VLLM_ATTENTION_BACKEND != STR_DUAL_CHUNK_FLASH_ATTN_VAL:
                raise ValueError("please set VLLM_ATTENTION_BACKEND to "
                                 f"{STR_DUAL_CHUNK_FLASH_ATTN_VAL}")

1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
    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

1337
        # Reminder: Please update docs/features/compatibility_matrix.md
1338
        # If the feature combo become valid
1339
        from vllm.platforms import current_platform
1340
        if not current_platform.is_async_output_supported(self.enforce_eager):
1341
1342
1343
1344
1345
1346
1347
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

1348
        # Async postprocessor is not necessary for pooling models
1349
        # since there is no token generation
1350
        if self.runner_type == "pooling":
1351
1352
            self.use_async_output_proc = False

1353
        # Reminder: Please update docs/features/compatibility_matrix.md
1354
        # If the feature combo become valid
1355
1356
1357
        if speculative_config:
            self.use_async_output_proc = False

1358
1359
1360
1361
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
1362
1363
1364
1365
1366
1367

        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.")

1368
1369
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
1370
1371
1372
1373
1374
1375
1376
        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}).")

1377
        if parallel_config.enable_expert_parallel:
1378
1379
            self._verify_with_expert_parallelism()

1380
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1381
        if pipeline_parallel_size > 1:
1382
1383
            if not self.registry.is_pp_supported_model(self.architectures,
                                                       self):
1384
1385
1386
                raise NotImplementedError(
                    "Pipeline parallelism is not supported for this model. "
                    "Supported models implement the `SupportsPP` interface.")
1387

1388
1389
            if self.use_async_output_proc:
                self.use_async_output_proc = False
1390

1391
1392
    def get_sliding_window(self) -> Optional[int]:
        """Get the sliding window size from the HF text config if present."""
1393
        return getattr(self.hf_text_config, "sliding_window", None)
1394
1395

    def get_vocab_size(self) -> int:
1396
        return getattr(self.hf_text_config, "vocab_size", 0)
1397

1398
    def get_hidden_size(self) -> int:
1399
        return getattr(self.hf_text_config, "hidden_size", 0)
1400

1401
1402
    @property
    def is_deepseek_mla(self) -> bool:
1403
1404
1405
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
bigmoyan's avatar
bigmoyan committed
1406
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp', 'kimi_k2'):
1407
1408
1409
1410
1411
1412
1413
1414
            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
1415

1416
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1417
        # TODO remove hard code
1418
        if self.is_deepseek_mla:
1419
1420
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1421
            if self.use_mla:
1422
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1423
1424
1425
1426
1427
            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
1428

1429
1430
1431
1432
1433
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1434
1435
1436
        if self.is_attention_free:
            return 0

1437
1438
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1439
            return self.hf_text_config.head_dim
1440

1441
1442
1443
1444
1445
        # NOTE: Some models (such as PLaMo2.1) use `hidden_size_per_head`
        if getattr(self.hf_text_config, "hidden_size_per_head",
                   None) is not None:
            return self.hf_text_config.hidden_size_per_head

1446
        # FIXME(woosuk): This may not be true for all models.
1447
1448
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1449

1450
1451
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1452
        # For GPTBigCode & Falcon:
1453
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1454
1455
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1456
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1457
        new_decoder_arch_falcon = (
1458
            self.hf_config.model_type in falcon_model_types
1459
            and getattr(self.hf_config, "new_decoder_architecture", False))
1460
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1461
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1462
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1463
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1464
            return 1
1465

1466
        # For DBRX and MPT
1467
1468
1469
1470
1471
        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":
1472
1473
1474
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1475
1476
1477
1478
1479
1480
1481
1482
        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")

1483
1484
1485
        if self.is_attention_free:
            return 0

1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
        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:
1496
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1497
1498
1499
1500
1501
            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.
1502
        return self.hf_text_config.num_attention_heads
1503
1504
1505

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1506
1507
1508
1509
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1510
1511
1512
1513
1514
1515
1516
        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)
1517

1518
1519
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1520
1521
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1522

1523
    def get_layers_start_end_indices(
1524
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1525
        from vllm.distributed.utils import get_pp_indices
1526
        if (self.hf_text_config.model_type == "deepseek_mtp"
zhuwenwen's avatar
zhuwenwen committed
1527
                or self.hf_config.model_type == "mimo_mtp"
1528
                or self.hf_config.model_type == "glm4_moe_mtp"
1529
1530
                or self.hf_config.model_type == "ernie_mtp"
                or self.hf_config.model_type == "qwen3_next_mtp"):
1531
1532
1533
1534
1535
            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)
1536
1537
1538
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1539
1540
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1541
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1542

1543
1544
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
        start, end = self.get_layers_start_end_indices(parallel_config)
1545
        return end - start
1546

1547
1548
1549
1550
1551
1552
1553
1554
    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
1555
1556
1557
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
        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
1568
1569
1570
1571
        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])
1572
        else:
1573
            # Hybrid model Jamba
1574
            layers_block_type_value = getattr(self.hf_text_config,
1575
                                              "layers_block_type", None)
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
            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])

1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
            # Hybrid model Qwen3Next
            layer_types_value = getattr(self.hf_config, "layer_types", None)
            if layer_types_value is not None:
                if getattr(block_type, "value", block_type) == "attention":
                    return sum(t == "full_attention"
                               for t in layer_types_value[start:end])
                elif getattr(block_type, "value",
                             block_type) == "linear_attention":
                    return sum(t == "linear_attention"
                               for t in layer_types_value[start:end])
                else:
                    return sum(t == getattr(block_type, "value", block_type)
                               for t in layer_types_value[start:end])

            if (layers_block_type_value is None and attn_type_list is None
                    and layer_types_value is None):
1609
1610
                raise ValueError(
                    "The model is an hybrid without a"
1611
1612
                    "layers_block_type or an attn_type_list, or a layer_types "
                    "in the hf_config, cannot determine the num of "
1613
                    f"{block_type.value} layers")
1614

1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
    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

1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
    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

1638
    def try_get_generation_config(self) -> dict[str, Any]:
1639
1640
1641
        """
        This method attempts to retrieve the non-default values of the
        generation config for this model.
1642

1643
1644
1645
1646
1647
1648
1649
1650
        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"}:
1651
            config = try_get_generation_config(
1652
                self.hf_config_path or self.model,
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
                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()

1667
    def get_diff_sampling_param(self) -> dict[str, Any]:
1668
        """
1669
1670
1671
1672
1673
1674
1675
1676
1677
        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"`
1678
1679

        Returns:
1680
            A dictionary containing the non-default sampling parameters.
1681
        """
1682
        if self.generation_config == "vllm":
1683
1684
1685
1686
1687
1688
1689
            config = {}
        else:
            config = self.try_get_generation_config()

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

1690
1691
1692
1693
1694
1695
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1696
            "max_new_tokens",
1697
1698
1699
1700
1701
1702
        ]
        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
            }
1703
1704
1705
1706
1707
            # 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")
1708
1709
        else:
            diff_sampling_param = {}
1710
1711
1712
1713
1714
1715
1716

        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`.")
1717
1718
        return diff_sampling_param

1719
    @property
1720
    def is_encoder_decoder(self) -> bool:
1721
        """Extract the HF encoder/decoder model flag."""
1722
        """
1723
        For Mllama, VLLM overrides HF's is_encoder_decoder flag and sets it to
1724
1725
        True to enable cross-attention
        """
1726
        return is_encoder_decoder(self.hf_config)
1727
1728

    @property
1729
1730
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1731

1732
1733
1734
1735
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1736
    @property
1737
1738
    def is_multimodal_raw_input_only_model(self) -> bool:
        return self._model_info.supports_multimodal_raw_input_only
1739

1740
1741
    @property
    def is_cross_encoder(self) -> bool:
1742
1743
        return (self._model_info.supports_cross_encoding
                or self.convert_type == "classify")
1744

1745
    @property
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
    def is_pp_supported(self) -> bool:
        return self._model_info.supports_pp

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

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

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

    @property
    def has_inner_state(self):
        return self._model_info.has_inner_state
1764

1765
1766
    @property
    def is_v1_compatible(self) -> bool:
1767
1768
1769
1770
        return not self._model_info.supports_v0_only

    @property
    def use_mla(self) -> bool:
1771
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE and SUPPORT_TC
1772

1773
1774
    @property
    def is_matryoshka(self) -> bool:
1775
        return (bool(getattr(self.hf_config, "matryoshka_dimensions", None))
1776
1777
                or getattr(self.hf_config, "is_matryoshka", False))

1778
1779
1780
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)
1781

1782
1783
1784
1785
1786
1787
    @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)

1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
    @property
    def head_dtype(self) -> torch.dtype:
        """
        "head" refers to the last Linear layer(s) of an LLM,
        such as the lm_head in a generation model,
        or the score or classifier in a classification model.

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

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

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

1814
    def get_and_verify_max_len(self, max_model_len: int):
1815
1816
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
1817
        tokenizer_config = None
1818
1819
        if (self.runner_type == "pooling" and getattr(
                self.hf_config, "position_embedding_type", "") == "absolute"):
1820
1821
1822
1823
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
                revision=self.tokenizer_revision)
1824
1825
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
1826
            tokenizer_config=tokenizer_config,
1827
1828
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
1829
            sliding_window=self.get_sliding_window(),
1830
1831
            spec_target_max_model_len=self.spec_target_max_model_len,
            encoder_config=self.encoder_config)
1832
        logger.info("Using max model len %s", max_model_len)
1833
1834
        return max_model_len

1835

1836
Device = Literal["auto", "cuda", "cpu", "tpu", "xpu"]
1837
1838
1839


@config
1840
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
1841
class DeviceConfig:
1842
1843
    """Configuration for the device to use for vLLM execution."""

1844
    device: SkipValidation[Optional[Union[Device, torch.device]]] = "auto"
1845
    """Device type for vLLM execution.
1846
1847
1848
    This parameter is deprecated and will be
    removed in a future release.
    It will now be set automatically based
1849
    on the current platform."""
1850
1851
1852
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
1853

1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
    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.
1869
        factors: list[Any] = []
1870
1871
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1872
        return hash_str
1873

1874
1875
    def __post_init__(self):
        if self.device == "auto":
1876
            # Automated device type detection
1877
            from vllm.platforms import current_platform
1878
            self.device_type = current_platform.device_type
1879
            if not self.device_type:
1880
1881
1882
1883
                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.")
1884
1885
        else:
            # Device type is assigned explicitly
1886
1887
1888
1889
            if isinstance(self.device, str):
                self.device_type = self.device
            elif isinstance(self.device, torch.device):
                self.device_type = self.device.type
1890
1891

        # Some device types require processing inputs on CPU
1892
        if self.device_type in ["tpu"]:
1893
            self.device = None
1894
1895
1896
1897
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1898

1899
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
1900
                            "mlp_speculator", "draft_model", "deepseek_mtp",
1901
                            "ernie_mtp", "qwen3_next_mtp"]
1902
1903
1904


@config
1905
@dataclass
1906
class SpeculativeConfig:
1907
    """Configuration for speculative decoding."""
1908

1909
    # General speculative decoding control
1910
    num_speculative_tokens: SkipValidation[int] = None  # type: ignore
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
    """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."""
1924
    draft_tensor_parallel_size: Optional[int] = None
1925
1926
    """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."""
1927
    disable_logprobs: bool = True
1928
1929
1930
    """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."""
1931

1932
    # Draft model configuration
1933
    quantization: Optional[me_quant.QuantizationMethods] = None
1934
1935
1936
    """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."""
1937
    max_model_len: Optional[int] = None
1938
1939
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
1940
    revision: Optional[str] = None
1941
1942
1943
    """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."""
1944
    code_revision: Optional[str] = None
1945
1946
1947
    """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."""
1948

1949
    # Advanced control
1950
    disable_by_batch_size: Optional[int] = None
1951
1952
1953
1954
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
1955
    prompt_lookup_max: Optional[int] = None
1956
1957
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
1958
    prompt_lookup_min: Optional[int] = None
1959
1960
1961
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

1962
    speculative_token_tree: Optional[str] = None
1963
    """Specifies the tree structure for speculative token generation.
1964
    """
1965
    # required configuration params passed from engine
1966
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
1967
    """The configuration of the target model."""
1968
1969
    target_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
1970
    """The parallel configuration for the target model."""
1971
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
1972
1973
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
1974
    disable_log_stats: SkipValidation[bool] = None  # type: ignore
1975
1976
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
1977
1978

    # params generated in the post-init stage
1979
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
1980
    """The configuration of the draft model initialized internal."""
1981
1982
    draft_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
1983
    """The parallel configuration for the draft model initialized internal."""
1984

1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
    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.
        """
1997
        factors: list[Any] = []
1998
1999
2000
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2001
2002
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2003
2004
        return hash_str

2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
    @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"]
            })
2015
2016
2017
2018
2019
2020
2021
2022
2023

        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
2024

zhuwenwen's avatar
zhuwenwen committed
2025
2026
2027
2028
2029
2030
2031
2032
        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"]
            })
2033

2034
2035
2036
2037
2038
2039
2040
2041
        if hf_config.model_type == "ernie4_5_moe":
            hf_config.model_type = "ernie_mtp"
        if hf_config.model_type == "ernie_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
            hf_config.update({
                "n_predict": n_predict,
                "architectures": ["ErnieMTPModel"]
            })
2042
2043
2044
2045
2046
2047
2048
2049
2050

        if hf_config.model_type == "qwen3_next":
            hf_config.model_type = "qwen3_next_mtp"
        if hf_config.model_type == "qwen3_next_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
            hf_config.update({
                "n_predict": n_predict,
                "architectures": ["Qwen3NextMTP"]
            })
2051

2052
2053
        return hf_config

2054
    def __post_init__(self):
2055

2056
2057
2058
2059
2060
2061
2062
        # 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.
2063
2064
2065
2066

        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
2067
            if self.target_model_config and \
2068
2069
                (self.target_model_config.hf_text_config.model_type \
                        == "deepseek_v3" or
2070
                    self.target_model_config.hf_text_config.model_type in
2071
                        ("mimo","ernie4_5_moe", "qwen3_next")):
2072
2073
                # use the draft model from the same model:
                self.model = self.target_model_config.model
2074
2075
2076
2077
                # Align the quantization of draft model for cases such as
                # --quantization fp8 with a bf16 checkpoint.
                if not self.quantization:
                    self.quantization = self.target_model_config.quantization
2078
2079
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2080
            else:
2081
2082
2083
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2084
2085
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2086
2087
2088
2089
2090
2091
2092
        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"
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
            # 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
2107
            if self.prompt_lookup_min < 1:
2108
2109
2110
2111
2112
                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")
2113
            if self.prompt_lookup_min > self.prompt_lookup_max:
2114
2115
2116
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2117

2118
2119
2120
            # 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.
2121
2122
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2123
        else:
2124
2125
2126
2127
2128
2129
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
2130
                    runner="draft",
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
                    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,
2146
                    enforce_eager=True if envs.VLLM_SPEC_DECODE_EAGER else self.target_model_config.enforce_eager,
2147
2148
2149
2150
2151
                    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,
                )
2152

2153
                # Automatically detect the method
2154
                if self.method in ('eagle', 'eagle3'):
2155
                    pass
2156
2157
2158
2159
2160
                # examples:
                # yuhuili/EAGLE-LLaMA3-Instruct-8B
                # yuhuili/EAGLE3-LLaMA3.1-Instruct-8B
                # AngelSlim/Qwen3-8B_eagle3
                elif "eagle-" in self.draft_model_config.model.lower():
2161
                    self.method = "eagle"
2162
2163
                elif "eagle3" in self.draft_model_config.model.lower():
                    self.method = "eagle3"
2164
2165
2166
2167
2168
                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"
2169
                elif (self.draft_model_config.hf_config.model_type
Yuxuan Zhang's avatar
Yuxuan Zhang committed
2170
                      in ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp")):
Jiayi Yao's avatar
Jiayi Yao committed
2171
2172
2173
2174
2175
2176
2177
                    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."
                            )
2178
2179
2180
2181
2182
2183
2184
2185
2186
                elif (self.draft_model_config.hf_config.model_type ==
                      "ernie_mtp"):
                    self.method = "ernie_mtp"
                    if self.num_speculative_tokens > 1:
                        logger.warning(
                                "All Ernie MTP models only have " \
                                "one layer. Might need some code changes " \
                                "to support multiple layers."
                            )
2187
2188
2189
2190
2191
2192
2193
2194
2195
                elif (self.draft_model_config.hf_config.model_type ==
                      "qwen3_next_mtp"):
                    self.method = "qwen3_next_mtp"
                    if self.num_speculative_tokens > 1:
                        logger.warning(
                                "All Qwen3Next MTP models only have " \
                                "one layer. Might need some code changes " \
                                "to support multiple layers."
                            )
2196
                else:
2197
                    self.method = "draft_model"
2198
2199
2200
2201
2202
                    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.")
2203
2204

                # Replace hf_config for EAGLE draft_model
2205
                if self.method in ("eagle", "eagle3"):
2206
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2207
                        raise ValueError(
2208
2209
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2210

2211
2212
                    from vllm.transformers_utils.configs import (
                        SpeculatorsConfig)
2213
2214
                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
2215

2216
                    if isinstance(self.draft_model_config.hf_config,
2217
                                  (EAGLEConfig, SpeculatorsConfig)):
2218
2219
2220
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2221
                            self.draft_model_config.hf_config,
2222
2223
                            method=self.method,
                            model_type="eagle")
2224
2225
2226
2227
2228
2229
2230
                        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
zhuwenwen's avatar
zhuwenwen committed
2231
2232
2233
2234
                    
                # if (self.num_speculative_heads is not None
                #     and hasattr(self.draft_model_config.hf_config, "num_lookahead_heads")):
                #     self.draft_model_config.hf_config.num_lookahead_heads = self.num_speculative_heads
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248

                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=}")

2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
                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)))

2262
2263
2264
2265
2266
2267
                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
                )
2268

2269
2270
2271
2272
2273
2274
                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,
                    ))
2275

2276
2277
2278
2279
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2280

2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
    @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,
        )

2316
    @staticmethod
2317
    def _verify_and_get_draft_tp(
2318
2319
2320
2321
2322
2323
            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.
2324
        """
2325
2326
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2327
        if speculative_draft_tensor_parallel_size is None:
2328
2329
2330
2331
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2332
2333
2334
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2335
2336
2337
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2338
2339
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2340
            raise ValueError(
2341
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2342
                f"other value than 1 or target model tensor_parallel_size")
2343
        return speculative_draft_tensor_parallel_size
2344

2345
2346
2347
2348
2349
2350
2351
2352
2353
    @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.
        """
2354
2355
2356
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2357
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2358
2359
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
            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

2371
2372
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2373
2374
2375
2376
2377
2378
        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.")

2379
2380
2381
2382
2383
2384
2385
        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)
2386
2387
2388
2389
2390
2391

        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=}")
2392

2393
        eagle3_target_supported = ["llama", "qwen"]
2394
2395
2396
2397
        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):
2398
            raise ValueError(
2399
                f"Eagle3 is only supported for {eagle3_target_supported} models. "  # noqa: E501
2400
2401
                f"Got {self.target_model_config.hf_text_config.model_type=}")

2402
2403
        return self

2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
    @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

2414
    def use_eagle(self) -> bool:
2415
2416
        return self.method in ("eagle", "eagle3", "deepseek_mtp", "ernie_mtp",
                               "qwen3_next_mtp")
2417

2418
    def __repr__(self) -> str:
2419
2420
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
2421
        num_spec_tokens = self.num_speculative_tokens
2422
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
2423
2424


2425
@config
2426
@dataclass
2427
class MultiModalConfig:
2428
2429
    """Controls the behavior of multimodal models."""

2430
2431
    limit_per_prompt: dict[str, int] = \
        cast(dict[str, int], get_field(ModelConfig, "limit_mm_per_prompt"))
2432
    """
2433
    The maximum number of input items allowed per prompt for each modality.
2434
    Defaults to 1 (V0) or 999 (V1) for each modality.
2435
2436

    For example, to allow up to 16 images and 2 videos per prompt:
2437
    `{"image": 16, "video": 2}`
2438
2439
    """

2440
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
2441
2442
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
2443
2444
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """

2445
2446
2447
    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
2448
    `transformers.AutoProcessor.from_pretrained`.
2449
2450
2451
2452

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

    For example, for Phi-3-Vision:
2453
    `{"num_crops": 4}`.
2454
2455
    """

2456
    mm_processor_cache_gb: float = 4
2457
    """
2458
2459
2460
2461
2462
2463
2464
    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).
2465
2466
    """

2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
    mm_encoder_tp_mode: MMEncoderTPMode = "weights"
    """
    Indicates how to optimize multi-modal encoder inference using
    tensor parallelism (TP).

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

2483
2484
2485
2486
2487
    interleave_mm_strings: bool = False
    """
    Enable fully interleaved support for multimodal prompts.
    """

2488
2489
    skip_mm_profiling: bool = False
    """
2490
    When enabled, skips multimodal memory profiling and only profiles with
2491
2492
2493
2494
2495
    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.
2496
2497
    """

2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
    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.
2512
        factors: list[Any] = []
2513
2514
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2515
2516
        return hash_str

2517
2518
2519
2520
2521
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
2522
2523
2524
2525
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
2526

2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
    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)
2537

2538

2539
@config
2540
2541
@dataclass
class PoolerConfig:
2542
    """Controls the behavior of output pooling in pooling models."""
2543
2544

    pooling_type: Optional[str] = None
2545
    """
2546
    The pooling method of the pooling model. This should be a key in
2547
    [`vllm.model_executor.layers.pooler.PoolingType`][].
2548
2549
    """

2550
    ## for embeddings models
2551
2552
    normalize: Optional[bool] = None
    """
2553
    Whether to normalize the embeddings outputs. Defaults to True.
2554
2555
2556
    """
    dimensions: Optional[int] = None
    """
2557
    Reduce the dimensions of embeddings if model
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
    support matryoshka representation. Defaults to None.
    """
    enable_chunked_processing: Optional[bool] = None
    """
    Whether to enable chunked processing for long inputs that exceed the model's
    maximum position embeddings. When enabled, long inputs will be split into
    chunks, processed separately, and then aggregated using weighted averaging.
    This allows embedding models to handle arbitrarily long text without CUDA
    errors. Defaults to False.
    """
    max_embed_len: Optional[int] = None
    """
    Maximum input length allowed for embedding generation. When set, allows
    inputs longer than max_embed_len to be accepted for embedding models.
    When an input exceeds max_embed_len, it will be handled according to 
    the original max_model_len validation logic. 
    Defaults to None (i.e. set to max_model_len).
2575
2576
    """

2577
2578
    ## for classification models
    activation: Optional[bool] = None
2579
    """
2580
    Whether to apply activation function to the classification outputs.
2581
2582
2583
2584
2585
    Defaults to True.
    """
    logit_bias: Optional[float] = None
    """
    If provided, apply classification logit biases. Defaults to None.
2586
2587
    """

2588
2589
2590
    ## for reward models
    softmax: Optional[bool] = None
    """
2591
    Whether to apply softmax to the reward outputs.
2592
    Defaults to True.
2593
    """
2594
2595
    step_tag_id: Optional[int] = None
    """
2596
    If set, only the score corresponding to the ``step_tag_id`` in the
2597
2598
2599
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """
2600
    returned_token_ids: Optional[list[int]] = None
2601
    """
2602
2603
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2604
2605
2606
    ``math-shepherd-mistral-7b-prm`` model.
    """

2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
    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.
2621
        factors: list[Any] = []
2622
2623
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2624
2625
        return hash_str

2626

2627
2628
2629
2630
2631
2632
2633
2634
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2635
2636
2637
2638
# 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.",
2639
2640
    "gemma3_text":
    "Numerical instability. Please use bfloat16 or float32 instead.",
2641
2642
2643
    "plamo2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "glm4": "Numerical instability. Please use bfloat16 or float32 instead.",
}
2644

2645

2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
def _is_valid_dtype(model_type: str, dtype: torch.dtype):
    if model_type in _FLOAT16_NOT_SUPPORTED_MODELS and dtype == torch.float16:  # noqa: E501, SIM103
        return False

    return True


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

    return True


def _find_dtype(
    model_id: str,
2664
    config: PretrainedConfig,
2665
2666
2667
    *,
    revision: Optional[str],
):
2668
2669
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
2670
    config_dtype = getattr(config, "torch_dtype", None)
2671

2672
    # Fallbacks for multi-modal models if the root config
2673
    # does not define torch_dtype
2674
2675
    if config_dtype is None:
        config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
2676
2677
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)
2678
2679
    if config_dtype is None and hasattr(config, "encoder_config"):
        config_dtype = getattr(config.encoder_config, "torch_dtype", None)
2680

2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
    # 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)
2695

2696
2697
2698
    if config_dtype is None:
        config_dtype = torch.float32

2699
    return config_dtype
2700

Shinichi Hemmi's avatar
Shinichi Hemmi committed
2701

2702
2703
2704
2705
2706
2707
2708
def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform
2709

2710
2711
2712
2713
    supported_dtypes = [
        dtype for dtype in current_platform.supported_dtypes
        if _is_valid_dtype(model_type, dtype)
    ]
2714

2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
    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

2759
2760
2761
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
2762
2763
2764
2765
2766
2767
            # Set default dtype from model config
            torch_dtype = _resolve_auto_dtype(
                model_type,
                config_dtype,
                is_pooling_model=is_pooling_model,
            )
2768
        else:
2769
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
2770
                raise ValueError(f"Unknown dtype: {dtype!r}")
2771
2772
2773
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
2774
    else:
2775
        raise ValueError(f"Unknown dtype: {dtype}")
2776

2777
2778
    _check_valid_dtype(model_type, torch_dtype)

2779
2780
2781
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
2782
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
2783
2784
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
2785
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
2786
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
2787
            # Casting between float16 and bfloat16 is allowed with a warning.
2788
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
2789
2790

    return torch_dtype
2791
2792


2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
def _get_head_dtype(config: PretrainedConfig, dtype: torch.dtype,
                    runner_type: str) -> torch.dtype:
    head_dtype: Optional[Union[str,
                               torch.dtype]] = getattr(config, "head_dtype",
                                                       None)

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


2818
2819
def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
2820
    tokenizer_config: Optional[dict],
2821
    max_model_len: Optional[int],
2822
    disable_sliding_window: bool,
2823
    sliding_window: Optional[int],
2824
    spec_target_max_model_len: Optional[int] = None,
2825
    encoder_config: Optional[Any] = None,
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
) -> 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",
2836
2837
        # ChatGLM2
        "seq_length",
2838
2839
        # Command-R
        "model_max_length",
2840
2841
        # Whisper
        "max_target_positions",
2842
2843
2844
2845
2846
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
2847
    # Choose the smallest "max_length" from the possible keys
2848
    max_len_key = None
2849
    for key in possible_keys:
2850
2851
2852
2853
2854
        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
2855
2856
2857
2858
    # 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
2859
2860
2861

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
2862
2863
2864
2865
    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
2866

2867
2868
2869
2870
2871
2872
2873
    # 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)

2874
2875
    # If none of the keys were found in the config, use a default and
    # log a warning.
2876
    if derived_max_model_len == float("inf"):
2877
2878
2879
2880
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

2881
2882
2883
2884
2885
        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

2886
2887
2888
2889
        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: "
2890
            "%s. Assuming the model's maximum length is %d.", possible_keys,
2891
            default_max_len)
2892
        derived_max_model_len = default_max_len
2893

2894
    rope_scaling = getattr(hf_config, "rope_scaling", None)
2895
2896
2897
    # 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:
2898
2899
2900
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910

        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.")

2911
2912
2913
2914
            # 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)

2915
2916
2917
2918
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
2919

2920
2921
2922
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2923
2924
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2925
    if max_model_len is None:
2926
        max_model_len = int(derived_max_model_len)
2927
2928
2929
2930
2931
2932
2933
2934
        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)
2935
    elif max_model_len > derived_max_model_len:
2936
2937
2938
2939
2940
        # 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:
2941
2942
2943
2944
2945
2946
2947
            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.")
2948
        else:
2949
            msg = (
2950
                f"User-specified max_model_len ({max_model_len}) is greater "
2951
2952
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2953
2954
2955
2956
2957
2958
2959
2960
                f"{model_max_length} in model's config.json).")
            warning = (
                "VLLM_ALLOW_LONG_MAX_MODEL_LEN must be used with extreme "
                "caution. If the model uses relative position encoding (RoPE), "
                "positions exceeding derived_max_model_len lead to nan. If the "
                "model uses absolute position encoding, positions exceeding "
                "derived_max_model_len will cause a CUDA array out-of-bounds "
                "error.")
2961
            if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
2962
                logger.warning_once("%s %s", msg, warning)
2963
2964
2965
            else:
                raise ValueError(
                    f"{msg} To allow overriding this maximum, set "
2966
                    f"the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1. {warning}")
2967
    return int(max_model_len)
2968
2969


2970
def get_served_model_name(model: str,
2971
                          served_model_name: Optional[Union[str, list[str]]]):
2972
    """
2973
2974
2975
2976
    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
2977
2978
2979
2980
2981
2982
2983
2984
2985
    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


2986
2987
GuidedDecodingBackend = Literal["auto", "xgrammar", "guidance", "outlines",
                                "lm-format-enforcer"]
2988
2989
2990


@config
2991
2992
@dataclass
class DecodingConfig:
2993
    """Dataclass which contains the decoding strategy of the engine."""
2994

2995
    backend: GuidedDecodingBackend = "auto"
2996
2997
2998
2999
    """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."""
3000

3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
    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`."""

3013
    reasoning_backend: str = ""
3014
    """Select the reasoning parser depending on the model that you're using.
3015
    This is used to parse the reasoning content into OpenAI API format."""
3016

3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
    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.
3031
        factors: list[Any] = []
3032
3033
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3034
        return hash_str
3035
3036

    def __post_init__(self):
3037
3038
3039
3040
3041
3042
3043
3044
        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.")

3045

3046
DetailedTraceModules = Literal["model", "worker", "all"]
3047
3048


3049
@config
3050
3051
@dataclass
class ObservabilityConfig:
3052
3053
    """Configuration for observability - metrics and tracing."""

3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
    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)
3069

3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
    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))
3095

3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
    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.
3110
        factors: list[Any] = []
3111
3112
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3113
3114
        return hash_str

3115
    def __post_init__(self):
3116
3117
3118
3119
3120
        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()

3121
        from vllm.tracing import is_otel_available, otel_import_error_traceback
3122
3123
3124
3125
3126
        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}")
3127

3128
3129
3130
3131
3132
3133
    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(","))

3134

3135
@config
3136
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3137
3138
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
3139
3140
3141
    simplifies passing around the distinct configurations in the codebase.
    """

3142
3143
3144
    # TODO: use default_factory once default constructing ModelConfig doesn't
    # try to download a model
    model_config: ModelConfig = None  # type: ignore
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
    """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."""
3156
    lora_config: Optional[LoRAConfig] = None
3157
3158
3159
    """LoRA configuration."""
    speculative_config: Optional[SpeculativeConfig] = None
    """Speculative decoding configuration."""
3160
    decoding_config: DecodingConfig = field(default_factory=DecodingConfig)
3161
    """Decoding configuration."""
3162
    observability_config: Optional[ObservabilityConfig] = None
3163
    """Observability configuration."""
3164
    quant_config: Optional[QuantizationConfig] = None
3165
3166
3167
    """Quantization configuration."""
    compilation_config: CompilationConfig = field(
        default_factory=CompilationConfig)
3168
    """`torch.compile` and cudagraph capture configuration for the model.
3169

3170
3171
    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}'`).
3172
    Currently, -O <n> and -O=<n> are supported as well but this will likely be
3173
    removed in favor of clearer -O<n> syntax in the future.
3174
3175
3176

    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
3177
    production, also default in V1.
3178
3179
3180
3181
3182
3183

    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."""
3184
    kv_events_config: Optional[KVEventsConfig] = None
3185
    """The configurations for event publishing."""
3186
    # some opaque config, only used to provide additional information
3187
3188
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
3189
3190
3191
3192
    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."""
3193
    instance_id: str = ""
3194
    """The ID of the vLLM instance."""
3195

3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
    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.
        """
3208
        factors: list[Any] = []
3209
3210

        # summarize vllm config
3211
        vllm_factors: list[Any] = []
3212
3213
        from vllm import __version__
        vllm_factors.append(__version__)
3214
        vllm_factors.append(envs.VLLM_USE_V1)
3215
3216
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3217
3218
        else:
            vllm_factors.append("None")
3219
3220
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3221
3222
        else:
            vllm_factors.append("None")
3223
3224
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3225
3226
        else:
            vllm_factors.append("None")
3227
3228
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3229
3230
        else:
            vllm_factors.append("None")
3231
3232
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3233
3234
        else:
            vllm_factors.append("None")
3235
3236
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3237
3238
        else:
            vllm_factors.append("None")
3239
3240
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3241
3242
3243
3244
3245
            # 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))
3246
3247
        else:
            vllm_factors.append("None")
3248
3249
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3250
3251
        else:
            vllm_factors.append("None")
3252
3253
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3254
3255
        else:
            vllm_factors.append("None")
3256
3257
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3258
3259
        else:
            vllm_factors.append("None")
3260
3261
3262
3263
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3264
3265
        else:
            vllm_factors.append("None")
3266
3267
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3268
3269
3270
        else:
            vllm_factors.append("None")
        if self.additional_config:
3271
3272
3273
3274
3275
3276
3277
3278
            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)
3279
3280
        else:
            vllm_factors.append("None")
3281
3282
        factors.append(vllm_factors)

3283
3284
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
3285
3286
        return hash_str

3287
3288
3289
3290
3291
3292
    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]
3293

3294
3295
3296
3297
3298
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3299
        from vllm.platforms import current_platform
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
        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
3322

3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
    @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)

3334
3335
3336
3337
3338
3339
3340
3341
3342
    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

3343
3344
3345
3346
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)
3347
3348
3349
3350

    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
3351
3352
3353

        self.try_verify_and_update_config()

3354
3355
3356
3357
3358
        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)
3359
3360
            self.model_config.verify_dual_chunk_attention_config(
                self.load_config)
3361

3362
        self.cache_config.verify_with_parallel_config(self.parallel_config)
3363

3364
        if self.lora_config is not None:
3365
            self.lora_config.verify_with_cache_config(self.cache_config)
3366
3367
            self.lora_config.verify_with_model_config(self.model_config)

3368
        if self.quant_config is None and self.model_config is not None:
3369
3370
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
3371

3372
        from vllm.platforms import current_platform
3373
        if self.model_config is not None and \
3374
3375
3376
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
3377
            logger.warning_once(
3378
3379
3380
3381
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
        # 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
3393

3394
3395
3396
3397
3398
            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

3399
3400
3401
3402
3403
        # 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
3404
3405
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
3406

3407
        if current_platform.is_cuda_alike() or current_platform.is_xpu():
3408
3409
3410
3411
3412
3413
3414
3415
3416
            # 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
3417

3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
            # 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
3429

3430
        if self.cache_config.cpu_offload_gb > 0 and \
3431
3432
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
3433
            logger.warning(
3434
                "CPU offload is not supported with `torch.compile` in v0 yet."
3435
3436
3437
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
        if self.cache_config.kv_sharing_fast_prefill:
            if not envs.VLLM_USE_V1:
                raise NotImplementedError(
                    "Fast prefill optimization for KV sharing is not supported "
                    "in V0 currently.")

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

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

3456
3457
3458
3459
3460
3461
        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`.")
3462
3463
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3464
3465
        disable_chunked_prefill_reasons: list[str] = []

3466
3467
3468
3469
3470
3471
3472
        if self.model_config:
            if self.model_config.pooler_config:
                pooling_type = self.model_config.pooler_config.pooling_type
                if pooling_type is None or pooling_type.lower() != "last":
                    disable_chunked_prefill_reasons.append(
                        "Only \"last\" pooling supports chunked "
                        "prefill and prefix caching; disabling both.")
3473
3474
3475
3476
                if not getattr(self.model_config.hf_config, "is_causal", True):
                    disable_chunked_prefill_reasons.append(
                        "Only models using causal attention supports chunked "
                        "prefill and prefix caching; disabling both.")
3477
3478
3479
3480
3481
3482
3483
3484
            elif self.model_config.is_encoder_decoder:
                self.scheduler_config.max_num_encoder_input_tokens = \
                    MULTIMODAL_REGISTRY.get_encdec_max_encoder_len(self.model_config)
                logger.debug(
                    "Encoder-decoder model detected: setting "
                    "`max_num_encoder_input_tokens` to encoder length (%s)",
                    self.scheduler_config.max_num_encoder_input_tokens)
                self.scheduler_config.disable_chunked_mm_input = True
3485
                disable_chunked_prefill_reasons.append(
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
                    "Encoder-decoder models do not support chunked prefill nor"
                    " prefix caching; disabling both.")
                if (self.model_config.architecture
                        == "WhisperForConditionalGeneration"
                        and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD")
                        != "spawn"):
                    logger.warning(
                        "Whisper is known to have issues with "
                        "forked workers. If startup is hanging, "
                        "try setting 'VLLM_WORKER_MULTIPROC_METHOD' "
                        "to 'spawn'.")
3497
3498
3499
3500
3501
3502
3503
3504
3505

        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
3506

3507
        if (self.kv_events_config is not None
3508
3509
3510
3511
3512
                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.")
3513
3514
        if (self.kv_events_config is not None
                and self.kv_events_config.publisher != "null"
3515
3516
3517
3518
3519
                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.")
3520
3521
        current_platform.check_and_update_config(self)

3522
        # final check of cudagraph mode after platform-specific update
3523
        if envs.VLLM_USE_V1 and current_platform.is_cuda_alike():
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
            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}"

3540
3541
3542
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

3543
3544
3545
3546
3547
        # 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()

3548
3549
3550
3551
3552
        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.
3553
            if not current_platform.support_hybrid_kv_cache():
3554
                # Hybrid KV cache manager is not supported on non-GPU platforms.
3555
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3556
3557
            if self.kv_transfer_config is not None:
                # Hybrid KV cache manager is not compatible with KV transfer.
3558
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3559
3560
            if self.kv_events_config is not None:
                # Hybrid KV cache manager is not compatible with KV events.
3561
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
3562
            if self.model_config is not None and \
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
                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
3580

3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
    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
        ]

3601
3602
    def _set_cudagraph_sizes(self):
        """
3603
3604
        vLLM defines the default candidate list of batch sizes for CUDA graph
        capture as:
3605

3606
3607
3608
3609
        ```python
        max_graph_size = min(max_num_seqs * 2, 512)
        # 1, 2, 4, then multiples of 8 up to max_graph_size
        cuda_graph_sizes = [1, 2, 4, 8, 16, 24, 32, 40, ..., max_graph_size]
3610

3611
3612
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
3613

3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
        These sizes are used to capture and reuse CUDA graphs for
        performance-critical paths (e.g., decoding). Capturing enables
        significantly faster kernel dispatch by avoiding Python overhead. The
        list is then filtered based on `max_num_batched_tokens` (e.g., 8192 on
        most GPUs), which controls the total allowed number of tokens in a
        batch. Since each sequence may have a variable number of tokens, the
        maximum usable batch size will depend on actual sequence lengths.

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

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

            - If batch size <= one of the `cudagraph_capture_sizes`, the closest
            padded CUDA graph will be used.
            - If batch size > largest `cudagraph_capture_sizes`, cudagraph will
            not be used.
3637
        """
3638
3639
3640
3641
3642
3643
3644
3645
3646

        # 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)]
3647
3648
3649
3650
3651
                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)

3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
                # 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:
zhuwenwen's avatar
zhuwenwen committed
3673
                if self.model_config.use_mla and self.compilation_config.full_cuda_graph and self.scheduler_config.max_num_seqs<=512:
3674
                    cuda_graph_sizes = [self.scheduler_config.max_num_seqs]
3675
3676
                else:
                    cuda_graph_sizes = self.scheduler_config.cuda_graph_sizes 
3677
3678
3679
3680
3681
3682
3683
                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
3684
                    raise TypeError(f"Invalid value for {cuda_graph_sizes=}.")
3685
3686
3687
3688
                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)
3689
3690
3691
3692
3693
                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
                ]
3694
3695
3696
3697
3698
                
                # add for spec decode
                if self.speculative_config is not None and self.speculative_config.num_lookahead_slots > 0:
                    batch_size_capture_list = list(map(lambda x: x * (1 + self.speculative_config.num_lookahead_slots),
                                                        batch_size_capture_list))
3699
3700
3701
3702

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

3703
    def recalculate_max_model_len(self, max_model_len: int):
3704
        # Can only be called in try_verify_and_update_config
3705
        model_config = self.model_config
3706
        max_model_len = model_config.get_and_verify_max_len(max_model_len)
3707
3708
        self.model_config.max_model_len = max_model_len
        self.scheduler_config.max_model_len = max_model_len
3709
3710

    def try_verify_and_update_config(self):
3711
3712
3713
        if self.model_config is None:
            return

3714
3715
3716
3717
3718
        # Avoid running try_verify_and_update_config multiple times
        if getattr(self.model_config, "config_updated", False):
            return
        self.model_config.config_updated = True

3719
        architecture = self.model_config.architecture
3720
3721
3722
        if architecture is None:
            return

3723
3724
        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP, HybridAttentionMambaModelConfig)
3725
3726
3727
        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_config(self)
3728

3729
3730
3731
        if self.model_config.is_hybrid:
            HybridAttentionMambaModelConfig.verify_and_update_config(self)

3732
        if self.model_config.convert_type == "classify":
3733
3734
3735
3736
            # Maybe convert ForCausalLM into ForSequenceClassification model.
            from vllm.model_executor.models.adapters import (
                SequenceClassificationConfig)
            SequenceClassificationConfig.verify_and_update_config(self)
3737

3738
    def __str__(self):
3739
        return (
3740
3741
3742
3743
3744
            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}, "
3745
            f"revision={self.model_config.revision}, "
3746
            f"tokenizer_revision={self.model_config.tokenizer_revision}, "
3747
3748
            f"trust_remote_code={self.model_config.trust_remote_code}, "
            f"dtype={self.model_config.dtype}, "
3749
3750
            f"max_seq_len={self.model_config.max_model_len}, "
            f"download_dir={self.load_config.download_dir!r}, "
3751
            f"load_format={self.load_config.load_format}, "
3752
3753
            f"tensor_parallel_size={self.parallel_config.tensor_parallel_size}, "  # noqa
            f"pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, "  # noqa
3754
            f"data_parallel_size={self.parallel_config.data_parallel_size}, "  # noqa
3755
3756
3757
3758
            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}, "
3759
            f"device_config={self.device_config.device}, "
3760
3761
3762
3763
3764
3765
3766
            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}, "
3767
3768
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
3769
3770
3771


_current_vllm_config: Optional[VllmConfig] = None
3772
_current_prefix: Optional[str] = None
3773
3774
3775


@contextmanager
3776
3777
3778
def set_current_vllm_config(vllm_config: VllmConfig,
                            check_compile=False,
                            prefix: Optional[str] = None):
3779
    """
3780
    Temporarily set the current vLLM config.
3781
    Used during model initialization.
3782
    We save the current vLLM config in a global variable,
3783
    so that all modules can access it, e.g. custom ops
3784
    can access the vLLM config to determine how to dispatch.
3785
    """
3786
    global _current_vllm_config, _current_prefix
3787
    old_vllm_config = _current_vllm_config
3788
    old_prefix = _current_prefix
3789
3790
3791
3792
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
3793
        _current_prefix = prefix
3794
        yield
3795
3796
3797
    except Exception:
        raise
    else:
3798
3799
3800
3801
        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)
3802
3803
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
3804
3805
3806
3807
3808
3809
3810
3811
3812
            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"
3813
                " if you want it to be supported.",
3814
                vllm_config.model_config.model)
3815
    finally:
3816
        _current_vllm_config = old_vllm_config
3817
        _current_prefix = old_prefix
3818
3819
3820
3821
3822
3823
3824
3825
        # 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
3826
3827
3828
3829
3830
3831
3832


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.
3833
        logger.warning("Current vLLM config is not set.")
3834
3835
3836
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
3837
3838


3839
3840
3841
3842
3843
3844
3845
3846
3847
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


3848
3849
3850
3851
3852
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
3853
    a max of 16 on a 64-bit system).
3854
3855
3856
3857
3858

    Args:
        text (str): The text to check

    Returns:
3859
        result (bool): `True` if a match is found, `False` otherwise.
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
    """
    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}")
3873
3874
3875
3876
3877


T = TypeVar("T")


3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
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

3897
    return {
3898
3899
3900
        layer_name: forward_context[layer_name]
        for layer_name in layer_names
        if isinstance(forward_context[layer_name], layer_type)
3901
    }
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931


@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:
3932
        return self.min_energy_split_window_size is not None
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950


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)