"examples/backends/vllm/launch/xpu/agg_multimodal_xpu.sh" did not exist on "550bf98ce2569867911ad533b03ce4d09c308955"
config.py 225 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import ast
5
import copy
6
import enum
7
import hashlib
8
import inspect
9
import json
10
import textwrap
Robert Shaw's avatar
Robert Shaw committed
11
import uuid
12
import warnings
13
from collections import Counter
14
from collections.abc import Mapping
15
from contextlib import contextmanager
16
17
from dataclasses import (MISSING, Field, asdict, field, fields, is_dataclass,
                         replace)
18
from functools import cached_property, lru_cache
19
from importlib.util import find_spec
20
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Literal, Optional,
21
                    Protocol, TypeVar, Union, cast, get_args)
22

23
import regex as re
24
import torch
25
26
27
from pydantic import (ConfigDict, SkipValidation, TypeAdapter, field_validator,
                      model_validator)
from pydantic.dataclasses import dataclass
28
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
29
from torch.distributed import ProcessGroup, ReduceOp
30
from typing_extensions import Self, assert_never, runtime_checkable
31

32
import vllm.envs as envs
33
from vllm import version
34
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
35
from vllm.logger import init_logger
36
from vllm.model_executor.layers.quantization import QuantizationMethods
37
from vllm.platforms import current_platform
38
39
40
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
41
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
42
43
    maybe_override_with_speculators_target_model, try_get_generation_config,
    try_get_safetensors_metadata, try_get_tokenizer_config, uses_mrope)
44
from vllm.transformers_utils.s3_utils import S3Model
45
from vllm.transformers_utils.utils import is_s3, maybe_model_redirect
46
47
# yapf conflicts with isort for this block
# yapf: disable
48
49
50
from vllm.utils import (DEFAULT_MAX_NUM_BATCHED_TOKENS,
                        MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                        POOLING_MODEL_MAX_NUM_BATCHED_TOKENS, GiB_bytes,
51
                        LayerBlockType, LazyLoader, common_broadcastable_dtype,
52
53
54
                        cuda_device_count_stateless, get_cpu_memory,
                        get_open_port, is_torch_equal_or_newer, random_uuid,
                        resolve_obj_by_qualname)
55

56
57
# yapf: enable

58
if TYPE_CHECKING:
59
    from _typeshed import DataclassInstance
60
    from ray.runtime_env import RuntimeEnv
61
    from ray.util.placement_group import PlacementGroup
62
    from transformers.configuration_utils import PretrainedConfig
63

64
65
    import vllm.model_executor.layers.quantization as me_quant
    import vllm.model_executor.models as me_models
66
    from vllm.executor.executor_base import ExecutorBase
67
    from vllm.model_executor.layers.quantization import QuantizationMethods
68
69
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
70
    from vllm.model_executor.model_loader import LoadFormats
71
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
72

73
    ConfigType = type[DataclassInstance]
74
    HfOverrides = Union[dict, Callable[[type], type]]
75
else:
76
    DataclassInstance = Any
77
    PlacementGroup = Any
78
    RuntimeEnv = Any
79
    PretrainedConfig = Any
80
    ExecutorBase = Any
81
    QuantizationConfig = Any
82
    QuantizationMethods = Any
83
    BaseModelLoader = Any
84
    LoadFormats = Any
85
    TensorizerConfig = Any
86
    ConfigType = type
87
88
89
90
91
92
    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")
93

94
logger = init_logger(__name__)
95
DataclassInstanceT = TypeVar("DataclassInstanceT", bound=DataclassInstance)
96
97
ConfigT = TypeVar("ConfigT", bound=ConfigType)

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
183
184
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"


185
186
187
188
189
190
191
192
193
194
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
195

196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
        Can be removed when Python 3.9 support is dropped.
        """
        iterator = iter(iterable)
        a = next(iterator, None)

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

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

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

    out = {}

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

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

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

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

            out[target.id] = doc

    return out


237
def config(cls: ConfigT) -> ConfigT:
238
239
240
    """
    A decorator that ensures all fields in a dataclass have default values
    and that each field has a docstring.
241
242
243
244
245
246

    If a `ConfigT` is used as a CLI argument itself, the default value provided
    by `get_kwargs` will be the result parsing a JSON string as the kwargs
    (i.e. `ConfigT(**json.loads(cli_arg))`). However, if a particular `ConfigT`
    requires custom construction from CLI (i.e. `CompilationConfig`), it can
    have a `from_cli` method, which will be called instead.
247

248
249
250
    Config validation is performed by the tools/validate_config.py
    script, which is invoked during the pre-commit checks.
    """
251
252
253
    return cls


254
def get_field(cls: ConfigType, name: str) -> Field:
255
256
257
258
259
260
261
    """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__}.")
262
    named_field: Field = cls_fields[name]
263
264
265
266
267
268
269
270
    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.")


271
272
273
274
def is_init_field(cls: ConfigType, name: str) -> bool:
    return next(f for f in fields(cls) if f.name == name).init


275
276
TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
277
278
LogprobsMode = Literal["raw_logprobs", "raw_logits", "processed_logprobs",
                       "processed_logits"]
279
280
281


@config
282
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
283
class ModelConfig:
284
285
    """Configuration for the model."""

286
    model: str = "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
372
373
374
375
376
377
378
379
    """Method used to quantize the weights. If `None`, we first check the
    `quantization_config` attribute in the model config file. If that is
    `None`, we assume the model weights are not quantized and use `dtype` to
    determine the data type of the weights."""
    enforce_eager: bool = False
    """Whether to always use eager-mode PyTorch. If True, we will disable CUDA
    graph and always execute the model in eager mode. If False, we will use
    CUDA graph and eager execution in hybrid for maximal performance and
    flexibility."""
    max_seq_len_to_capture: int = 8192
    """Maximum sequence len covered by CUDA graphs. When a sequence has context
    length larger than this, we fall back to eager mode. Additionally for
    encoder-decoder models, if the sequence length of the encoder input is
    larger than this, we fall back to the eager mode."""
    max_logprobs: int = 20
    """Maximum number of log probabilities to return when `logprobs` is
    specified in `SamplingParams`. The default value comes the default for the
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
383
384
385
386
387
388
    logprobs_mode: LogprobsMode = "raw_logprobs"
    """Indicates the content returned in the logprobs and prompt_logprobs.
    Supported mode:
    1) raw_logprobs, 2) processed_logprobs, 3) raw_logits, 4) processed_logits.
    Raw means the values before applying logit processors, like bad words.
    Processed means the values after applying such processors.
    """
389
390
391
392
393
394
395
396
397
398
399
400
401
402
    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."""
403
404
405
406
    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."""
407
408
409
410
411
412
413
414
415
416
417
    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."""
418
    interleave_mm_strings: bool = False
419
    """Enable fully interleaved support for multimodal prompts, while using
420
    --chat-template-content-format=string. Defaults to False."""
421
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
422
423
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
424
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
425
426
427
428
429
430
431
432
433
434
435
436
437
438
    use_async_output_proc: bool = True
    """Whether to use async output processor."""
    config_format: Union[str, ConfigFormat] = ConfigFormat.AUTO.value
    """The format of the model config to load:\n
    - "auto" will try to load the config in hf format if available else it
    will try to load in mistral format.\n
    - "hf" will load the config in hf format.\n
    - "mistral" will load the config in mistral format."""
    hf_token: Optional[Union[bool, str]] = None
    """The token to use as HTTP bearer authorization for remote files . If
    `True`, will use the token generated when running `huggingface-cli login`
    (stored in `~/.huggingface`)."""
    hf_overrides: HfOverrides = field(default_factory=dict)
    """If a dictionary, contains arguments to be forwarded to the Hugging Face
439
    config. If a callable, it is called to update the HuggingFace config."""
440
441
442
443
444
    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}`.
445
    """
446
    disable_mm_preprocessor_cache: bool = False
447
    """If `True`, disable caching of the multi-modal processor."""
448
449
450
451
    override_neuron_config: dict[str, Any] = field(default_factory=dict)
    """Initialize non-default neuron config or override default neuron config
    that are specific to Neuron devices, this argument will be used to
    configure the neuron config that can not be gathered from the vllm
452
    arguments. e.g. `{"cast_logits_dtype": "bfloat16"}`."""
453
454
455
456
457
458
    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}`.
459
    """
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
    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
475
    `--generation-config vllm`, only the override parameters are used."""
476
477
478
479
480
481
482
483
484
    enable_sleep_mode: bool = False
    """Enable sleep mode for the engine (only cuda platform is supported)."""
    model_impl: Union[str, ModelImpl] = ModelImpl.AUTO.value
    """Which implementation of the model to use:\n
    - "auto" will try to use the vLLM implementation, if it exists, and fall
    back to the Transformers implementation if no vLLM implementation is
    available.\n
    - "vllm" will use the vLLM model implementation.\n
    - "transformers" will use the Transformers model implementation."""
485
486
    override_attention_dtype: Optional[str] = None
    """Override dtype for attention"""
487

488
489
490
491
492
493
494
495
496
497
498
499
    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.
        """
500
        factors: list[Any] = []
501
502
503
504
505
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
506
507
508
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
509
        factors.append(self.trust_remote_code)
510
511
512
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
513
514
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
515
516
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
517
518
        str_factors = str(factors)
        assert_hashable(str_factors)
519
520
        return hashlib.sha256(str(factors).encode()).hexdigest()

521
    def __post_init__(self) -> None:
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
        # 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)

540
541
542
543
544
545
546
547
548
        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)

549
550
551
        # 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)
552
553
554
555
556
557
558
559
560
561
562
563
        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):
564
            hf_overrides_kw = {}
565
            hf_overrides_fn = self.hf_overrides
566
        else:
567
            hf_overrides_kw = self.hf_overrides
568
            hf_overrides_fn = None
569

570
571
        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
572
            hf_overrides_kw.update(hf_override)
573
            hf_overrides_str = json.dumps(hf_overrides_kw)
574
575
576
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
577
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
578
579
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
580
            hf_overrides_kw.update(hf_override)
581
            hf_overrides_str = json.dumps(hf_overrides_kw)
582
583
584
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
585
586
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

587
        self.maybe_pull_model_tokenizer_for_s3(self.model, self.tokenizer)
588

589
590
591
592
        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 "
593
594
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
595
596
                "for instructions on how to install it.")

597
598
        from vllm.platforms import current_platform

599
600
601
602
603
604
        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)

605
606
607
608
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
609

610
611
612
        if isinstance(self.config_format, str):
            self.config_format = ConfigFormat(self.config_format)

613
        hf_config = get_config(self.hf_config_path or self.model,
614
615
616
617
618
619
                               self.trust_remote_code,
                               self.revision,
                               self.code_revision,
                               self.config_format,
                               hf_overrides_kw=hf_overrides_kw,
                               hf_overrides_fn=hf_overrides_fn)
620

621
        self.hf_config = hf_config
622
        self.hf_text_config = get_hf_text_config(self.hf_config)
623
624
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
625
        self.encoder_config = self._get_encoder_config()
626
        self.hf_image_processor_config = get_hf_image_processor_config(
627
            self.model, hf_token=self.hf_token, revision=self.revision)
628

629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
        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
686
            else:
687
688
689
690
691
692
693
694
                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}"
695
696
            warnings.warn(msg, DeprecationWarning, stacklevel=2)

697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
        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.")
716

717
718
719
720
721
722
        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)
723
724
        self._model_info = model_info
        self._architecture = arch
725
        logger.info("Resolved architecture: %s", arch)
726
727
728
729
730
731
732
733
734
735

        self.pooler_config = self._init_pooler_config()

        self.dtype = _get_and_verify_dtype(
            self.model,
            self.hf_config,
            self.dtype,
            is_pooling_model=self.runner_type == "pooling",
            revision=self.revision,
        )
736

737
        # Workaround for Gemma 2 which uses interleaved sliding window
738
739
        # attention, but it's not specified in its config.
        # TODO: remove this when Gemma 2 config updated in HuggingFace.
740
741
742
        if self.hf_text_config.model_type == "gemma2":
            self.hf_text_config.sliding_window_pattern = 2

743
744
745
746
747
        # TODO: remove this when Gemma 3n config updated in HuggingFace.
        if self.hf_text_config.model_type == "gemma3n_text":
            # 4 sliding window attention followed by 1 full attention
            self.hf_text_config.sliding_window_pattern = "LLLLG"

748
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
749
750
        sliding_window_pattern = getattr(self.hf_text_config,
                                         "sliding_window_pattern", None)
751
752
        has_interleaved_attention = sliding_window_pattern is not None or (
            isinstance(sliding_window, list))
753

754
        if not self.disable_sliding_window and has_interleaved_attention:
755
756
            if not envs.VLLM_USE_V1 and (backend := envs.VLLM_ATTENTION_BACKEND
                                         ) in ("XFORMERS", "FLASHINFER"):
757
758
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
759

760
                logger.warning_once(
761
762
763
764
765
                    "%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).",  # noqa: E501
                    self.hf_text_config.model_type,
                    backend,
                    sliding_window_len_min,
                )
766
767
768
769
770
771
772
773
                self.disable_sliding_window = True
            else:
                # for a model with interleaved attention,
                # the scheduler and the model treat it as full attention
                # (i.e., not dropping any tokens outside the window).
                # only the attention layer itself is aware of the sliding
                # window, and use the window size to compute the attention.
                self.hf_text_config.interleaved_sliding_window = sliding_window
774
775
776
777

                if hasattr(self.hf_text_config, "sliding_window"):
                    delattr(self.hf_text_config, "sliding_window")

778
                sliding_window = None
Woosuk Kwon's avatar
Woosuk Kwon committed
779

780
        self.original_max_model_len = self.max_model_len
781
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
782
        self.multimodal_config = self._init_multimodal_config()
783

784
785
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
786

787
788
789
        if (not current_platform.is_neuron() and self.override_neuron_config):
            raise ValueError(
                "`override_neuron_config` is only supported on Neuron.")
790

791
792
793
        # Avoid running try_verify_and_update_config multiple times
        self.config_updated = False

794
        self._verify_quantization()
795
        self._verify_cuda_graph()
796
        self._verify_bnb_config()
797

798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
    @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

814
815
816
    def _get_transformers_backend_cls(self) -> str:
        """Determine which Transformers backend class will be used if
        `model_impl` is set to `transformers` or `auto`."""
817
818
        if getattr(self, "runner_type", self.runner) == "pooling":
            return "TransformersModel"
819
820
821
822
        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"
823
824
825
826
827
        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()
828

829
830
    @property
    def registry(self):
831
        return me_models.ModelRegistry
832
833
834

    @property
    def architectures(self) -> list[str]:
835
        return getattr(self.hf_config, "architectures", [])
836

837
838
    @property
    def architecture(self) -> str:
839
        """The architecture vllm actually used."""
840
841
        return self._architecture

842
843
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
844
        """Pull model/tokenizer from S3 to temporary directory when needed.
845

846
        Args:
847
848
            model: Model name or path
            tokenizer: Tokenizer name or path
849
        """
850
851
852
853
854
855
856
857
858
859
860
861
        if not (is_s3(model) or is_s3(tokenizer)):
            return

        if is_s3(model):
            s3_model = S3Model()
            s3_model.pull_files(model,
                                allow_pattern=["*.model", "*.py", "*.json"])
            self.model_weights = model
            self.model = s3_model.dir

            # If tokenizer is same as model, download to same directory
            if model == tokenizer:
862
863
864
865
866
                s3_model.pull_files(model,
                                    ignore_pattern=[
                                        "*.pt", "*.safetensors", "*.bin",
                                        "*.tensors"
                                    ])
867
868
869
870
871
872
873
                self.tokenizer = s3_model.dir
                return

        # Only download tokenizer if needed and not already handled
        if is_s3(tokenizer):
            s3_tokenizer = S3Model()
            s3_tokenizer.pull_files(
874
875
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors"])
876
            self.tokenizer = s3_tokenizer.dir
877

878
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
879
        if self._model_info.supports_multimodal:
880
            return MultiModalConfig(
881
                limit_per_prompt=self.limit_mm_per_prompt,
882
                media_io_kwargs=self.media_io_kwargs,
883
884
                mm_processor_kwargs=self.mm_processor_kwargs,
                disable_mm_preprocessor_cache=self.
885
886
                disable_mm_preprocessor_cache,
                interleave_mm_strings=self.interleave_mm_strings)
887
888

        return None
889

890
891
892
893
894
895
    def set_disable_mm_preprocessor_cache(self, value: bool) -> None:
        mm_config = self.get_multimodal_config()

        self.disable_mm_preprocessor_cache = value
        mm_config.disable_mm_preprocessor_cache = value

896
897
898
899
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

900
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
901
        if self.runner_type == "pooling":
902
903
904
905
906
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
907
908
909
910
911

            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():
912
913
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
914

915
            return pooler_config
916

917
918
        return None

919
    def _verify_tokenizer_mode(self) -> None:
920
921
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
922
923
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
924
                f"one of {get_args(TokenizerMode)}.")
925
        self.tokenizer_mode = tokenizer_mode
926

927
928
929
930
931
932
933
934
935
936
    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"

937
        for arch in architectures:
938
939
940
941
942
943
944
945
946
947
948
949
            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"
950

951
    def _get_runner_type(
952
        self,
953
        architectures: list[str],
954
955
956
957
958
959
960
        runner: RunnerOption,
    ) -> RunnerType:
        if runner != "auto":
            return runner

        runner_type = self._get_default_runner_type(architectures)

961
962
963
964
965
966
        # 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)
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996

        return runner_type

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

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

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

        # This is to handle Sentence Transformers models that use *ForCausalLM
        # and also multi-modal pooling models which are not defined as
        # Sentence Transformers models
        if runner_type == "pooling":
997
998
            return "embed"

999
1000
1001
1002
1003
1004
1005
1006
1007
1008
        return "none"

    def _get_convert_type(
        self,
        architectures: list[str],
        runner_type: RunnerType,
        convert: ConvertOption,
    ) -> ConvertType:
        if convert != "auto":
            return convert
1009

1010
1011
        convert_type = self._get_default_convert_type(architectures,
                                                      runner_type)
1012

1013
1014
1015
1016
1017
1018
        # 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)
1019
1020

        return convert_type
1021

1022
    def _get_supported_generation_tasks(
1023
        self,
1024
1025
        architectures: list[str],
        convert_type: ConvertType,
1026
1027
1028
    ) -> list[_ResolvedTask]:
        registry = self.registry

1029
        if registry.is_transcription_only_model(architectures, self):
1030
1031
            return ["transcription"]

1032
        # TODO: Use get_supported_generation_tasks once V0 is removed
1033
        supported_tasks = list[_ResolvedTask]()
1034
1035
        if (registry.is_text_generation_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["generate"]):
1036
1037
            supported_tasks.append("generate")

1038
1039
        if registry.is_transcription_model(architectures, self):
            supported_tasks.append("transcription")
1040
1041

        return supported_tasks
1042

1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
    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"

1060
1061
    def _get_supported_pooling_tasks(
        self,
1062
1063
        architectures: list[str],
        convert_type: ConvertType,
1064
    ) -> list[_ResolvedTask]:
1065
        registry = self.registry
1066

1067
        # TODO: Use get_supported_pooling_tasks once V0 is removed
1068
        supported_tasks = list[_ResolvedTask]()
1069
1070
        if (registry.is_pooling_model(architectures, self)
                or convert_type in _RUNNER_CONVERTS["pooling"]):
1071
            supported_tasks.append("encode")
1072

1073
1074
1075
            extra_task = (self._get_default_pooling_task(architectures)
                          if convert_type == "none" else convert_type)
            supported_tasks.append(extra_task)
1076
1077
1078
1079
1080

        return supported_tasks

    def _get_supported_tasks(
        self,
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
        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"]
1093

1094
        assert_never(runner_type)
1095

1096
1097
1098
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
1099
            # compressed-tensors uses a "compression_config" key
1100
            quant_cfg = getattr(self.hf_config, "compression_config", None)
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115

        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}")

1116
1117
        return quant_cfg

1118
    def _verify_quantization(self) -> None:
1119
        supported_quantization = me_quant.QUANTIZATION_METHODS
1120
        optimized_quantization_methods = [
1121
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
1122
            "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
1123
            "quark", "modelopt_fp4", "bitblas", "gptq_bitblas", "inc"
1124
        ]
1125
        if self.quantization is not None:
1126
1127
            self.quantization = cast(me_quant.QuantizationMethods,
                                     self.quantization)
1128
1129

        # Parse quantization method from the HF model config, if available.
1130
1131
        quant_cfg = self._parse_quant_hf_config()

1132
        if quant_cfg is not None:
1133
            # Use the community standard 'quant_method'
1134
            quant_method = quant_cfg.get("quant_method", "").lower()
1135
1136

            # Normalize library names
1137
1138
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
1139

1140
            quant_cfg["quant_method"] = quant_method
1141

1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
            # Quantization methods which are overrides (i.e. they have a
            # `override_quantization_method` method) must be checked in order
            # of preference (this is particularly important for GPTQ).
            overrides = [
                "marlin",
                "bitblas",
                "gptq_marlin_24",
                "gptq_marlin",
                "gptq_bitblas",
                "awq_marlin",
                "ipex",
                "moe_wna16",
1154
1155
                "modelopt",
                "modelopt_fp4",
1156
1157
1158
1159
1160
1161
1162
1163
1164
            ]
            quantization_methods = [
                q for q in supported_quantization if q not in overrides
            ]
            # Any custom overrides will be in quantization_methods so we place
            # them at the start of the list so custom overrides have preference
            # over the built in ones.
            quantization_methods = quantization_methods + overrides

1165
            # Detect which checkpoint is it
1166
            for name in quantization_methods:
1167
                method = me_quant.get_quantization_config(name)
1168
1169
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
1170
1171
1172
1173
                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.
1174
                    if (name in get_args(me_quant.QuantizationMethods)
1175
1176
1177
1178
1179
1180
                            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.")
1181
1182
1183
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
1184

1185
            # Verify quantization configurations.
1186
            if self.quantization is None:
1187
1188
                self.quantization = quant_method
            elif self.quantization != quant_method:
1189
1190
                raise ValueError(
                    "Quantization method specified in the model config "
1191
                    f"({quant_method}) does not match the quantization "
1192
1193
1194
1195
1196
1197
1198
1199
                    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}.")
1200
            from vllm.platforms import current_platform
1201
            current_platform.verify_quantization(self.quantization)
1202
            if self.quantization not in optimized_quantization_methods:
1203
                logger.warning(
1204
                    "%s quantization is not fully "
1205
                    "optimized yet. The speed can be slower than "
1206
                    "non-quantized models.", self.quantization)
1207

1208
    def _verify_cuda_graph(self) -> None:
1209
1210
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
1211
        # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
1212
        ROCM_UNSUPPORTED_MODELS = ['mllama']
1213
1214
1215
1216
1217
1218
        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()):
1219
1220
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1221
                "to eager mode.", self.hf_config.model_type)
1222
            self.enforce_eager = True
1223

1224
1225
    def _verify_bnb_config(self) -> None:
        """
1226
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
1227
        yet support CUDA graph.
1228
        # TODO Remove this when bitsandbytes supports.
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
        """
        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(
1243
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1244
                "fallback to the eager mode.")
1245

1246
1247
            self.enforce_eager = True

1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
    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.")

1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
    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

1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
    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

1292
        # Reminder: Please update docs/features/compatibility_matrix.md
1293
        # If the feature combo become valid
1294
        from vllm.platforms import current_platform
1295
        if not current_platform.is_async_output_supported(self.enforce_eager):
1296
1297
1298
1299
1300
1301
1302
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

1303
        # Async postprocessor is not necessary for pooling models
1304
        # since there is no token generation
1305
        if self.runner_type == "pooling":
1306
1307
            self.use_async_output_proc = False

1308
        # Reminder: Please update docs/features/compatibility_matrix.md
1309
        # If the feature combo become valid
1310
1311
1312
        if speculative_config:
            self.use_async_output_proc = False

1313
1314
1315
1316
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
1317
1318
1319
1320
1321
1322

        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.")

1323
1324
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
1325
1326
1327
1328
1329
1330
1331
        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}).")

1332
        if parallel_config.enable_expert_parallel:
1333
1334
            self._verify_with_expert_parallelism()

1335
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1336
        if pipeline_parallel_size > 1:
1337
1338
            if not self.registry.is_pp_supported_model(self.architectures,
                                                       self):
1339
1340
1341
1342
1343
1344
                raise NotImplementedError(
                    "Pipeline parallelism is not supported for this model. "
                    "Supported models implement the `SupportsPP` interface.")

            if self.use_async_output_proc:
                self.use_async_output_proc = False
1345

1346
1347
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
1348
        """Get the sliding window size, or None if disabled."""
1349
1350
1351
1352

        # Some models, like Qwen2 and Qwen1.5, use `use_sliding_window` in
        # addition to sliding window size. We check if that field is present
        # and if it's False, return None.
1353
1354
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
1355
            return None
1356
        return getattr(self.hf_text_config, "sliding_window", None)
1357

1358
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
1359
1360
1361
1362
1363
1364
1365
1366
        """Get the sliding window size, or None if disabled.
        """
        # If user disables sliding window, return None.
        if self.disable_sliding_window:
            return None
        # Otherwise get the value from the hf config.
        return self.get_hf_config_sliding_window()

1367
    def get_vocab_size(self) -> int:
1368
        return getattr(self.hf_text_config, "vocab_size", 0)
1369

1370
    def get_hidden_size(self) -> int:
1371
        return getattr(self.hf_text_config, "hidden_size", 0)
1372

1373
1374
    @property
    def is_deepseek_mla(self) -> bool:
1375
1376
1377
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
bigmoyan's avatar
bigmoyan committed
1378
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp', 'kimi_k2'):
1379
1380
1381
1382
1383
1384
1385
1386
            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
1387

1388
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1389
        # TODO remove hard code
1390
        if self.is_deepseek_mla:
1391
1392
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1393
            if self.use_mla:
1394
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1395
1396
1397
1398
1399
            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
1400

1401
1402
1403
1404
1405
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1406
1407
1408
        if self.is_attention_free:
            return 0

1409
1410
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1411
            return self.hf_text_config.head_dim
1412

1413
        # FIXME(woosuk): This may not be true for all models.
1414
1415
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1416

1417
1418
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1419
        # For GPTBigCode & Falcon:
1420
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1421
1422
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1423
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1424
        new_decoder_arch_falcon = (
1425
            self.hf_config.model_type in falcon_model_types
1426
            and getattr(self.hf_config, "new_decoder_architecture", False))
1427
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1428
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1429
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1430
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1431
            return 1
1432

1433
        # For DBRX and MPT
1434
1435
1436
1437
1438
        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":
1439
1440
1441
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1442
1443
1444
1445
1446
1447
1448
1449
        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")

1450
1451
1452
        if self.is_attention_free:
            return 0

1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
        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:
1463
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1464
1465
1466
1467
1468
            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.
1469
        return self.hf_text_config.num_attention_heads
1470
1471
1472

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1473
1474
1475
1476
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1477
1478
1479
1480
1481
1482
1483
        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)
1484

1485
1486
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1487
1488
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1489

1490
    def get_layers_start_end_indices(
1491
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1492
        from vllm.distributed.utils import get_pp_indices
1493
        if (self.hf_text_config.model_type == "deepseek_mtp"
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1494
1495
                or self.hf_config.model_type == "mimo_mtp"
                or self.hf_config.model_type == "glm4_moe_mtp"):
1496
1497
1498
1499
1500
            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)
1501
1502
1503
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1504
1505
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1506
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1507

1508
1509
1510
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
        start, end = self.get_layers_start_end_indices(parallel_config)
        return end - start
Mor Zusman's avatar
Mor Zusman committed
1511

1512
1513
1514
1515
1516
1517
1518
1519
    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
1520
1521
1522
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
        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
1533
1534
1535
1536
        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])
1537
        else:
1538
            # Hybrid model Jamba
1539
1540
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
            if layers_block_type_value is not None:
                if hasattr(self.hf_text_config,
                           "model_type") and (self.hf_text_config.model_type
                                              == "zamba2"):
                    if attn_block_type:
                        return sum(t == "hybrid"
                                   for t in layers_block_type_value[start:end])
                    else:
                        return self.get_num_layers(parallel_config)
                return sum(t == block_type.value
                           for t in layers_block_type_value[start:end])

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

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

            return sum(t == 1 for t in attn_type_list[start:end])
Mor Zusman's avatar
Mor Zusman committed
1566

1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
    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

1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
    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

1590
    def try_get_generation_config(self) -> dict[str, Any]:
1591
1592
1593
        """
        This method attempts to retrieve the non-default values of the
        generation config for this model.
1594

1595
1596
1597
1598
1599
1600
1601
1602
        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"}:
1603
            config = try_get_generation_config(
1604
                self.hf_config_path or self.model,
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
                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()

1619
    def get_diff_sampling_param(self) -> dict[str, Any]:
1620
        """
1621
1622
1623
1624
1625
1626
1627
1628
1629
        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"`
1630
1631

        Returns:
1632
            A dictionary containing the non-default sampling parameters.
1633
        """
1634
        if self.generation_config == "vllm":
1635
1636
1637
1638
1639
1640
1641
            config = {}
        else:
            config = self.try_get_generation_config()

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

1642
1643
1644
1645
1646
1647
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1648
            "max_new_tokens",
1649
1650
1651
1652
1653
1654
        ]
        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
            }
1655
1656
1657
1658
1659
            # 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")
1660
1661
        else:
            diff_sampling_param = {}
1662
1663
1664
1665
1666
1667
1668

        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`.")
1669
1670
        return diff_sampling_param

1671
    @property
1672
    def is_encoder_decoder(self) -> bool:
1673
        """Extract the HF encoder/decoder model flag."""
1674
        """
1675
        For Mllama, VLLM overrides HF's is_encoder_decoder flag and sets it to
1676
        True to enable cross-attention
1677
        Neuron needs all multimodal data to be in the decoder and does not
1678
1679
1680
1681
1682
1683
        need to explicitly enable cross-attention
        """
        if (current_platform.is_neuron()
                and self.hf_config.model_type == "mllama"):
            return False

1684
1685
1686
1687
1688
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1689

1690
1691
1692
1693
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
    @property
    def processor_return_mm_hashes(self) -> bool:
        """Whether the multi-modal processor should output hashes."""
        mm_config = self.multimodal_config
        if mm_config is None:
            return False

        return not mm_config.disable_mm_preprocessor_cache

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

        return not mm_config.disable_mm_preprocessor_cache

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

        return envs.VLLM_MM_INPUT_CACHE_GIB

1719
1720
    @property
    def is_cross_encoder(self) -> bool:
1721
1722
        return (self._model_info.supports_cross_encoding
                or self.convert_type == "classify")
1723

1724
    @property
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
    def is_pp_supported(self) -> bool:
        return self._model_info.supports_pp

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

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

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

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

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

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

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

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

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

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

1771
    def get_and_verify_max_len(self, max_model_len: int):
1772
1773
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
1774
        tokenizer_config = None
1775
1776
        if (self.runner_type == "pooling" and getattr(
                self.hf_config, "position_embedding_type", "") == "absolute"):
1777
1778
1779
1780
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
                revision=self.tokenizer_revision)
1781
1782
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
1783
            tokenizer_config=tokenizer_config,
1784
1785
1786
1787
1788
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
            sliding_window_len=self.get_hf_config_sliding_window(),
            spec_target_max_model_len=self.spec_target_max_model_len,
            encoder_config=self.encoder_config)
1789
        logger.info("Using max model len %s", max_model_len)
1790
1791
        return max_model_len

1792

1793
BlockSize = Literal[1, 8, 16, 32, 64, 128]
1794
CacheDType = Literal["auto", "fp8", "fp8_e4m3", "fp8_e5m2", "fp8_inc"]
1795
PrefixCachingHashAlgo = Literal["builtin", "sha256", "sha256_cbor_64bit"]
1796
1797
1798
1799


@config
@dataclass
1800
class CacheConfig:
1801
    """Configuration for the KV cache."""
1802

1803
    block_size: SkipValidation[BlockSize] = None  # type: ignore
1804
1805
1806
    """Size of a contiguous cache block in number of tokens. This is ignored on
    neuron devices and set to `--max-model-len`. On CUDA devices, only block
    sizes up to 32 are supported. On HPU devices, block size defaults to 128.
1807
1808

    This config has no static default. If left unspecified by the user, it will
1809
    be set in `Platform.check_and_update_config()` based on the current
1810
    platform."""
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
    gpu_memory_utilization: float = 0.9
    """The fraction of GPU memory to be used for the model executor, which can
    range from 0 to 1. For example, a value of 0.5 would imply 50% GPU memory
    utilization. If unspecified, will use the default value of 0.9. This is a
    per-instance limit, and only applies to the current vLLM instance. It does
    not matter if you have another vLLM instance running on the same GPU. For
    example, if you have two vLLM instances running on the same GPU, you can
    set the GPU memory utilization to 0.5 for each instance."""
    swap_space: float = 4
    """Size of the CPU swap space per GPU (in GiB)."""
    cache_dtype: CacheDType = "auto"
    """Data type for kv cache storage. If "auto", will use model data type.
    CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. ROCm (AMD GPU) supports
1824
    fp8 (=fp8_e4m3). Intel Gaudi (HPU) supports fp8 (using fp8_inc)."""
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
    is_attention_free: bool = False
    """Whether the model is attention-free. This is primarily set in
    `ModelConfig` and that value should be manually duplicated here."""
    num_gpu_blocks_override: Optional[int] = None
    """Number of GPU blocks to use. This overrides the profiled `num_gpu_blocks`
    if specified. Does nothing if `None`. Used for testing preemption."""
    sliding_window: Optional[int] = None
    """Sliding window size for the KV cache. This is primarily set in
    `ModelConfig` and that value should be manually duplicated here."""
    enable_prefix_caching: Optional[bool] = None
    """Whether to enable prefix caching. Disabled by default for V0. Enabled by
    default for V1."""
    prefix_caching_hash_algo: PrefixCachingHashAlgo = "builtin"
    """Set the hash algorithm for prefix caching:\n
    - "builtin" is Python's built-in hash.\n
1840
1841
    - "sha256" is collision resistant but with certain overheads.
    This option uses Pickle for object serialization before hashing.\n
1842
1843
    - "sha256_cbor_64bit" provides a reproducible, cross-language compatible
    hash. It serializes objects using canonical CBOR and hashes them with
1844
1845
    SHA-256. The resulting hash consists of the lower 64 bits of the SHA-256
    digest."""
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
    cpu_offload_gb: float = 0
    """The space in GiB to offload to CPU, per GPU. Default is 0, which means
    no offloading. Intuitively, this argument can be seen as a virtual way to
    increase the GPU memory size. For example, if you have one 24 GB GPU and
    set this to 10, virtually you can think of it as a 34 GB GPU. Then you can
    load a 13B model with BF16 weight, which requires at least 26GB GPU memory.
    Note that this requires fast CPU-GPU interconnect, as part of the model is
    loaded from CPU memory to GPU memory on the fly in each model forward pass.
    """
    calculate_kv_scales: bool = False
    """This enables dynamic calculation of `k_scale` and `v_scale` when
    kv_cache_dtype is fp8. If `False`, the scales will be loaded from the model
    checkpoint if available. Otherwise, the scales will default to 1.0."""
1859
1860
    cpu_kvcache_space_bytes: Optional[int] = None
    """(CPU backend only) CPU key-value cache space."""
1861
1862
1863
    mamba_page_size_padded: Optional[int] = None
    """ Optional override for mamba page size; used by hybrid mamba/attention
    models to ensure exact alignment with attention page size."""
1864
1865
1866
1867
1868
1869

    # Will be set after profiling.
    num_gpu_blocks: Optional[int] = field(default=None, init=False)
    """The number of blocks to allocate for GPU memory."""
    num_cpu_blocks: Optional[int] = field(default=None, init=False)
    """The number of blocks to allocate for CPU memory."""
1870

1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
    kv_sharing_fast_prefill: bool = False
    """This feature is work in progress and no prefill optimization takes place
    with this flag enabled currently.

    In some KV sharing setups, e.g. YOCO (https://arxiv.org/abs/2405.05254),
    some layers can skip tokens corresponding to prefill. This flag enables
    attention metadata for eligible layers to be overriden with metadata
    necessary for implementating this optimization in some models (e.g. Gemma3n)
    """

1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
    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.
        """
1893
        factors: list[Any] = []
1894
1895
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
1896
1897
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1898
1899
        return hash_str

1900
1901
1902
    def __post_init__(self) -> None:
        self.swap_space_bytes = self.swap_space * GiB_bytes

1903
        self._verify_cache_dtype()
1904
        self._verify_prefix_caching()
1905

1906
    def metrics_info(self):
1907
1908
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1909
1910
        return {key: str(value) for key, value in self.__dict__.items()}

1911
1912
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
1913
1914
1915
1916
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1917
1918
1919
1920
1921
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1922
1923
1924
1925
1926
        if self.kv_sharing_fast_prefill:
            logger.warning_once(
                "--kv-sharing-fast-prefill is currently work in progress "
                "and not functional yet (i.e. no prefill savings)")

1927
1928
        return self

1929
1930
1931
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1932
        elif self.cache_dtype in get_args(CacheDType):
1933
            logger.info(
1934
1935
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1936
                "Meanwhile, it may cause accuracy drop without a proper "
1937
                "scaling factor.")
1938
1939
1940
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1941
1942
1943
1944
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

1945
        if self.sliding_window is not None and not envs.VLLM_USE_V1:
1946
1947
1948
1949
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

1950
1951
        if (self.enable_prefix_caching and self.prefix_caching_hash_algo
                not in get_args(PrefixCachingHashAlgo)):
1952
1953
            raise ValueError(
                "Unknown prefix caching hash algorithm: "
1954
1955
                f"{self.prefix_caching_hash_algo}. Must be one of "
                f"{get_args(PrefixCachingHashAlgo)}.")
1956

1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
        total_cpu_memory = get_cpu_memory()
        # FIXME(woosuk): Here, it is assumed that the GPUs in a tensor parallel
        # group are in the same node. However, the GPUs may span multiple nodes.
        num_gpus_per_node = parallel_config.tensor_parallel_size
        cpu_memory_usage = self.swap_space_bytes * num_gpus_per_node

1967
1968
1969
        msg = (f"{cpu_memory_usage / GiB_bytes:.2f} GiB out of the "
               f"{total_cpu_memory / GiB_bytes:.2f} GiB total CPU memory "
               "is allocated for the swap space.")
1970
1971
1972
        if cpu_memory_usage > 0.7 * total_cpu_memory:
            raise ValueError("Too large swap space. " + msg)
        elif cpu_memory_usage > 0.4 * total_cpu_memory:
1973
            logger.warning("Possibly too large swap space. %s", msg)
1974

1975

1976
@config
1977
1978
@dataclass
class LoadConfig:
1979
1980
    """Configuration for loading the model weights."""

1981
    load_format: Union[str, LoadFormats] = "auto"
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
    """The format of the model weights to load:\n
    - "auto" will try to load the weights in the safetensors format and fall
    back to the pytorch bin format if safetensors format is not available.\n
    - "pt" will load the weights in the pytorch bin format.\n
    - "safetensors" will load the weights in the safetensors format.\n
    - "npcache" will load the weights in pytorch format and store a numpy cache
    to speed up the loading.\n
    - "dummy" will initialize the weights with random values, which is mainly
    for profiling.\n
    - "tensorizer" will use CoreWeave's tensorizer library for fast weight
    loading. See the Tensorize vLLM Model script in the Examples section for
    more information.\n
    - "runai_streamer" will load the Safetensors weights using Run:ai Model
    Streamer.\n
    - "bitsandbytes" will load the weights using bitsandbytes quantization.\n
    - "sharded_state" will load weights from pre-sharded checkpoint files,
    supporting efficient loading of tensor-parallel models.\n
    - "gguf" will load weights from GGUF format files (details specified in
    https://github.com/ggml-org/ggml/blob/master/docs/gguf.md).\n
    - "mistral" will load weights from consolidated safetensors files used by
2002
2003
    Mistral models.
    - Other custom values can be supported via plugins."""
2004
    download_dir: Optional[str] = None
2005
2006
    """Directory to download and load the weights, default to the default
    cache directory of Hugging Face."""
2007
2008
    model_loader_extra_config: Union[dict, TensorizerConfig] = field(
        default_factory=dict)
2009
    """Extra config for model loader. This will be passed to the model loader
2010
    corresponding to the chosen load_format."""
2011
2012
2013
    device: Optional[str] = None
    """Device to which model weights will be loaded, default to
    device_config.device"""
2014
    ignore_patterns: Optional[Union[list[str], str]] = None
2015
2016
    """The list of patterns to ignore when loading the model. Default to
    "original/**/*" to avoid repeated loading of llama's checkpoints."""
2017
    use_tqdm_on_load: bool = True
2018
2019
    """Whether to enable tqdm for showing progress bar when loading model
    weights."""
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
    pt_load_map_location: Union[str, dict[str, str]] = "cpu"
    """
    pt_load_map_location: the map location for loading pytorch checkpoint, to
    support loading checkpoints can only be loaded on certain devices like
    "cuda", this is equivalent to {"": "cuda"}. Another supported format is
    mapping from different devices like from GPU 1 to GPU 0:
    {"cuda:1": "cuda:0"}. Note that when passed from command line, the strings
    in dictionary needs to be double quoted for json parsing. For more details,
    see original doc for `map_location` in https://pytorch.org/docs/stable/generated/torch.load.html
    """
2030

2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
    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.
2045
        factors: list[Any] = []
2046
2047
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2048
2049
        return hash_str

2050
    def __post_init__(self):
2051
        self.load_format = self.load_format.lower()
2052
2053
2054
2055
2056
2057
2058
        if self.ignore_patterns is not None and len(self.ignore_patterns) > 0:
            logger.info(
                "Ignoring the following patterns when downloading weights: %s",
                self.ignore_patterns)
        else:
            self.ignore_patterns = ["original/**/*"]

2059

2060
2061
2062
DistributedExecutorBackend = Literal["ray", "mp", "uni", "external_launcher"]


2063
@config
2064
@dataclass
2065
class ParallelConfig:
2066
    """Configuration for the distributed execution."""
2067

2068
2069
2070
2071
2072
2073
2074
    pipeline_parallel_size: int = 1
    """Number of pipeline parallel groups."""
    tensor_parallel_size: int = 1
    """Number of tensor parallel groups."""
    data_parallel_size: int = 1
    """Number of data parallel groups. MoE layers will be sharded according to
    the product of the tensor parallel size and data parallel size."""
2075
2076
    data_parallel_size_local: int = 1
    """Number of local data parallel groups."""
2077
2078
    data_parallel_rank: int = 0
    """Rank of the data parallel group."""
2079
2080
2081
    data_parallel_rank_local: Optional[int] = None
    """Local rank of the data parallel group,
    set only in SPMD mode."""
2082
    data_parallel_master_ip: str = "127.0.0.1"
2083
    """IP of the data parallel master."""
2084
2085
    data_parallel_rpc_port: int = 29550
    """Port for data parallel messaging."""
2086
2087
    data_parallel_master_port: int = 29500
    """Port of the data parallel master."""
Rui Qiao's avatar
Rui Qiao committed
2088
2089
    data_parallel_backend: str = "mp"
    """Backend to use for data parallel, either "mp" or "ray"."""
2090
2091
    data_parallel_external_lb: bool = False
    """Whether to use "external" DP LB mode. Applies only to online serving
2092
2093
2094
2095
2096
2097
2098
2099
    and when data_parallel_size > 0. This is useful for a "one-pod-per-rank"
    wide-EP setup in Kuberentes. Set implicitly when --data-parallel-rank
    is provided explicitly to vllm serve."""
    data_parallel_hybrid_lb: bool = False
    """Whether to use "hybrid" DP LB mode. Applies only to online serving
    and when data_parallel_size > 0. Enables running an AsyncLLM
    and API server on a "per-node" basis where vLLM load balances
    between local data parallel ranks, but an external LB balances
2100
    between vLLM nodes/replicas. Set explicitly in conjunction with
2101
    --data-parallel-start-rank."""
2102
2103
    enable_expert_parallel: bool = False
    """Use expert parallelism instead of tensor parallelism for MoE layers."""
2104
2105
2106
2107
2108
2109
2110
2111
2112
    enable_eplb: bool = False
    """Enable expert parallelism load balancing for MoE layers."""
    num_redundant_experts: int = 0
    """Number of redundant experts to use for expert parallelism."""
    eplb_window_size: int = 1000
    """Window size for expert load recording."""
    eplb_step_interval: int = 3000
    """
    Interval for rearranging experts in expert parallelism.
2113

2114
2115
2116
2117
2118
2119
2120
2121
2122
    Note that if this is greater than the EPLB window size, only the metrics
    of the last `eplb_window_size` steps will be used for rearranging experts.
    """
    eplb_log_balancedness: bool = False
    """
    Log the balancedness each step of expert parallelism.
    This is turned off by default since it will cause communication overhead.
    """

2123
    max_parallel_loading_workers: Optional[int] = None
2124
    """Maximum number of parallel loading workers when loading model
2125
2126
    sequentially in multiple batches. To avoid RAM OOM when using tensor
    parallel and large models."""
2127
2128

    disable_custom_all_reduce: bool = False
2129
    """Disable the custom all-reduce kernel and fall back to NCCL."""
2130
2131

    ray_workers_use_nsight: bool = False
2132
    """Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler."""
2133

2134
2135
2136
    ray_runtime_env: Optional["RuntimeEnv"] = None
    """Ray runtime environment to pass to distributed workers."""

2137
    placement_group: Optional["PlacementGroup"] = None
2138
    """ray distributed model workers placement group."""
2139

2140
    distributed_executor_backend: Optional[Union[DistributedExecutorBackend,
2141
                                                 type["ExecutorBase"]]] = None
2142
2143
2144
2145
2146
2147
    """Backend to use for distributed model
    workers, either "ray" or "mp" (multiprocessing). If the product
    of pipeline_parallel_size and tensor_parallel_size is less than
    or equal to the number of GPUs available, "mp" will be used to
    keep processing on a single host. Otherwise, this will default
    to "ray" if Ray is installed and fail otherwise. Note that tpu
2148
    only support Ray for distributed inference."""
2149
2150

    worker_cls: str = "auto"
2151
2152
    """The full name of the worker class to use. If "auto", the worker class
    will be determined based on the platform."""
2153
    sd_worker_cls: str = "auto"
Ning Xie's avatar
Ning Xie committed
2154
    """The full name of the worker class to use for speculative decoding.
2155
    If "auto", the worker class will be determined based on the platform."""
2156
    worker_extension_cls: str = ""
2157
2158
2159
2160
    """The full name of the worker extension class to use. The worker extension
    class is dynamically inherited by the worker class. This is used to inject
    new attributes and methods to the worker class for use in collective_rpc
    calls."""
2161
2162

    world_size: int = field(init=False)
2163
    """world_size is TPxPP, it affects the number of workers we create."""
2164
2165

    rank: int = 0
2166
    """Global rank in distributed setup."""
2167

2168
    enable_multimodal_encoder_data_parallel: bool = False
2169
    """ Use data parallelism instead of tensor parallelism for vision encoder.
2170
2171
    Only support LLama4 for now"""

2172
2173
2174
2175
2176
2177
    @property
    def world_size_across_dp(self) -> int:
        """world_size_across_dp is TPxPPxDP, it is the size of the world
        including data parallelism."""
        return self.world_size * self.data_parallel_size

2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
    def get_next_dp_init_port(self) -> int:
        """
        We might need to initialize process groups in multiple
        processes that is related to data parallelism,
        e.g. both in the worker and in the engine, which
        can live in different processes. To avoid port conflicts, we
        increment the port number each time we need to initialize a
        new process group related to data parallelism.
        """
        answer = self.data_parallel_master_port
        self.data_parallel_master_port += 1
        return answer

    def stateless_init_dp_group(self) -> "ProcessGroup":
2192
2193
2194
2195
2196
2197
2198
2199
2200
        # NOTE: In high-concurrency scenarios multiple processes
        # can pick the same (currently free) port through a race
        # condition when calling `get_open_port()`. When the first
        # process binds the port the others will subsequently fail
        # with `torch.distributed.DistNetworkError: EADDRINUSE`.
        # To make the initialization more robust we retry a few times
        # with a fresh port whenever this specific error is observed.
        from torch.distributed import DistNetworkError

2201
2202
2203
        from vllm.distributed.utils import (
            stateless_init_torch_distributed_process_group)

2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
        max_retries = 5
        last_exc: Optional[Exception] = None
        for _ in range(max_retries):
            try:
                # use gloo since the engine process might not have cuda device
                return stateless_init_torch_distributed_process_group(
                    self.data_parallel_master_ip,
                    self.get_next_dp_init_port(),
                    self.data_parallel_rank,
                    self.data_parallel_size,
                    backend="gloo")
            except DistNetworkError as e:
                # We only want to retry when the root cause is EADDRINUSE.
                if "EADDRINUSE" in str(e):
                    logger.warning(
                        "Address already in use. Retrying with a new port.")
                    last_exc = e
                    continue  # try again with a new port
                raise e

        # If we get here all retries have failed.
        assert last_exc is not None
        raise last_exc
2227
2228
2229

    @staticmethod
    def has_unfinished_dp(dp_group: "ProcessGroup",
youkaichao's avatar
youkaichao committed
2230
                          has_unfinished: bool) -> bool:
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
        tensor = torch.tensor([has_unfinished],
                              dtype=torch.int32,
                              device="cpu")
        # dp rank 0: has_unfinished_seqs=True
        # dp rank 1: has_unfinished_seqs=False
        # aggregated: has_unfinished_seqs=True
        # so this is an OR operation, i.e. MAX in integers
        torch.distributed.all_reduce(tensor, op=ReduceOp.MAX, group=dp_group)
        aggregated_has_unfinished = bool(tensor.item())
        return aggregated_has_unfinished

2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
    @staticmethod
    def sync_kv_cache_memory_size(dp_group: "ProcessGroup",
                                  kv_cache_memory: int) -> int:
        if kv_cache_memory == -1:
            kv_cache_memory = torch.iinfo(torch.int64).max
        tensor = torch.tensor([kv_cache_memory],
                              dtype=torch.int64,
                              device="cpu")
        # we cannot use broadcast for stateless dp group since it depends
        # on global rank
        torch.distributed.all_reduce(tensor, op=ReduceOp.MIN, group=dp_group)
        return tensor.item()

2255
2256
2257
2258
2259
2260
2261
2262
    def compute_hash(self):
        """
        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.
        """
2263
        factors: list[Any] = []
2264
2265
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
2266
        factors.append(self.enable_expert_parallel)
2267
2268
        factors.append(self.data_parallel_size)
        factors.append(envs.VLLM_ALL2ALL_BACKEND)
2269
2270
        return hashlib.sha256(str(factors).encode()).hexdigest()

2271
2272
2273
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
2274

2275
2276
2277
2278
2279
2280
        if self.data_parallel_size_local > self.data_parallel_size:
            raise ValueError(
                f"data_parallel_size_local ({self.data_parallel_size_local}) "
                f"must be <= data_parallel_size ({self.data_parallel_size})")

        if self.data_parallel_size > 1 or self.data_parallel_size_local == 0:
2281
2282
            # Data parallel was specified in the engine args.
            self.data_parallel_master_port = get_open_port()
2283
2284
2285
2286
2287

            if not (0 <= self.data_parallel_rank < self.data_parallel_size):
                raise ValueError(
                    f"data_parallel_rank ({self.data_parallel_rank})"
                    f" must be in the range [0, {self.data_parallel_size})")
2288
2289
2290
2291
2292
2293
2294
2295
        else:
            # Otherwise fall back to env vars (e.g. for offline SPMD case).
            self.data_parallel_size = envs.VLLM_DP_SIZE
            self.data_parallel_rank = envs.VLLM_DP_RANK
            self.data_parallel_rank_local = envs.VLLM_DP_RANK_LOCAL
            self.data_parallel_master_ip = envs.VLLM_DP_MASTER_IP
            self.data_parallel_master_port = envs.VLLM_DP_MASTER_PORT

2296
2297
2298
2299
            if self.data_parallel_external_lb:
                raise ValueError("data_parallel_external_lb can only "
                                 "be set when data_parallel_size > 1")

2300
2301
2302
2303
2304
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

2305
2306
2307
2308
2309
2310
2311
2312
2313
        if self.enable_eplb:
            if not current_platform.is_cuda():
                raise ValueError(
                    "Expert parallelism load balancing is only supported on "
                    "CUDA devices now.")
            if self.num_redundant_experts < 0:
                raise ValueError(
                    "num_redundant_experts must be non-negative, but got "
                    f"{self.num_redundant_experts}.")
2314
2315
2316
2317
2318
2319
2320
2321
2322
            if not self.enable_expert_parallel:
                raise ValueError(
                    "enable_expert_parallel must be True to use EPLB.")
            if self.tensor_parallel_size * self.data_parallel_size <= 1:
                raise ValueError(
                    "EPLB requires tensor_parallel_size or data_parallel_size "
                    f"to be greater than 1, but got "
                    f"TP={self.tensor_parallel_size},DP={self.data_parallel_size}."
                )
2323
2324
2325
2326
2327
        else:
            if self.num_redundant_experts != 0:
                raise ValueError(
                    "num_redundant_experts should be used with EPLB."
                    f"{self.num_redundant_experts}.")
2328
        if self.distributed_executor_backend is None and self.world_size > 1:
2329
2330
2331
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

2332
            from vllm.executor import ray_utils
2333
            backend: DistributedExecutorBackend = "mp"
2334
            ray_found = ray_utils.ray_is_available()
2335
2336
            if current_platform.is_neuron():
                # neuron uses single process to control multiple devices
2337
2338
                backend = "uni"
            elif current_platform.is_tpu() and envs.VLLM_XLA_USE_SPMD:
2339
2340
2341
                backend = "uni"
            elif (current_platform.is_cuda()
                  and cuda_device_count_stateless() < self.world_size):
2342
                if not ray_found:
2343
2344
                    raise ValueError("Unable to load Ray: "
                                     f"{ray_utils.ray_import_err}. Ray is "
2345
2346
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
2347
                                     "ray`.")
2348
                backend = "ray"
Rui Qiao's avatar
Rui Qiao committed
2349
2350
2351
2352
            elif self.data_parallel_backend == "ray":
                logger.info("Using ray distributed inference because "
                            "data_parallel_backend is ray")
                backend = "ray"
2353
            elif ray_found:
2354
                if self.placement_group:
2355
                    backend = "ray"
2356
2357
2358
2359
2360
2361
                else:
                    from ray import is_initialized as ray_is_initialized
                    if ray_is_initialized():
                        from ray.util import get_current_placement_group
                        if get_current_placement_group():
                            backend = "ray"
2362
            self.distributed_executor_backend = backend
2363
2364
            logger.debug("Defaulting to use %s for distributed inference",
                         backend)
2365

2366
2367
2368
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

2369
2370
2371
2372
2373
2374
    @property
    def use_ray(self) -> bool:
        return self.distributed_executor_backend == "ray" or (
            isinstance(self.distributed_executor_backend, type)
            and self.distributed_executor_backend.uses_ray)

2375
2376
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2377
2378
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
2379
        from vllm.platforms import current_platform
2380
        if self.distributed_executor_backend not in (
2381
2382
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
2383
2384
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
2385
            raise ValueError(
2386
2387
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
2388
2389
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
2390
        if self.use_ray:
2391
2392
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
2393
2394

        if not current_platform.use_custom_allreduce():
2395
            self.disable_custom_all_reduce = True
Aaron Pham's avatar
Aaron Pham committed
2396
            logger.debug(
2397
                "Disabled the custom all-reduce kernel because it is not "
2398
                "supported on current platform.")
2399
        if self.ray_workers_use_nsight and not self.use_ray:
2400
2401
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
2402

2403
        return self
2404

2405

2406
PreemptionMode = Literal["swap", "recompute"]
2407
2408
2409
2410
SchedulerPolicy = Literal["fcfs", "priority"]


@config
2411
@dataclass
2412
class SchedulerConfig:
2413
    """Scheduler configuration."""
2414

2415
2416
    runner_type: RunnerType = "generate"
    """The runner type to launch for the model."""
2417

2418
    max_num_batched_tokens: SkipValidation[int] = None  # type: ignore
2419
    """Maximum number of tokens to be processed in a single iteration.
2420

2421
2422
    This config has no static default. If left unspecified by the user, it will
    be set in `EngineArgs.create_engine_config` based on the usage context."""
2423

2424
    max_num_seqs: SkipValidation[int] = None  # type: ignore
2425
    """Maximum number of sequences to be processed in a single iteration.
2426

2427
2428
    This config has no static default. If left unspecified by the user, it will
    be set in `EngineArgs.create_engine_config` based on the usage context."""
2429

2430
    max_model_len: SkipValidation[int] = None  # type: ignore
2431
2432
2433
    """Maximum length of a sequence (including prompt and generated text). This
    is primarily set in `ModelConfig` and that value should be manually
    duplicated here."""
2434

2435
    max_num_partial_prefills: int = 1
2436
2437
    """For chunked prefill, the maximum number of sequences that can be
    partially prefilled concurrently."""
2438
2439

    max_long_partial_prefills: int = 1
2440
2441
2442
2443
    """For chunked prefill, the maximum number of prompts longer than
    long_prefill_token_threshold that will be prefilled concurrently. Setting
    this less than max_num_partial_prefills will allow shorter prompts to jump
    the queue in front of longer prompts in some cases, improving latency."""
2444
2445

    long_prefill_token_threshold: int = 0
2446
2447
    """For chunked prefill, a request is considered long if the prompt is
    longer than this number of tokens."""
2448

2449
    num_lookahead_slots: int = 0
2450
2451
2452
2453
2454
2455
2456
    """The number of slots to allocate per sequence per
    step, beyond the known token ids. This is used in speculative
    decoding to store KV activations of tokens which may or may not be
    accepted.

    NOTE: This will be replaced by speculative config in the future; it is
    present to enable correctness tests until then."""
2457

2458
2459
2460
2461
    cuda_graph_sizes: list[int] = field(default_factory=list)
    """Cuda graph capture sizes
    1. if none provided, then default set to [min(max_num_seqs * 2, 512)]
    2. if one value is provided, then the capture list would follow the
2462
    pattern: [1, 2, 4] + [i for i in range(8, cuda_graph_sizes + 1, 8)]
2463
    3. more than one value (e.g. 1 2 128) is provided, then the capture list
2464
    will follow the provided list."""
2465

2466
    delay_factor: float = 0.0
2467
2468
    """Apply a delay (of delay factor multiplied by previous
    prompt latency) before scheduling next prompt."""
2469

2470
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2471
2472
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""
2473
2474

    is_multimodal_model: bool = False
2475
2476
2477
2478
2479
    """True if the model is multimodal."""

    # TODO (ywang96): Make this configurable.
    max_num_encoder_input_tokens: int = field(init=False)
    """Multimodal encoder compute budget, only used in V1.
2480

2481
2482
2483
2484
2485
2486
2487
2488
2489
    NOTE: This is not currently configurable. It will be overridden by
    max_num_batched_tokens in case max multimodal embedding size is larger."""

    # TODO (ywang96): Make this configurable.
    encoder_cache_size: int = field(init=False)
    """Multimodal encoder cache size, only used in V1.

    NOTE: This is not currently configurable. It will be overridden by
    max_num_batched_tokens in case max multimodal embedding size is larger."""
2490

2491
    preemption_mode: Optional[PreemptionMode] = None
2492
2493
2494
2495
2496
2497
    """Whether to perform preemption by swapping or
    recomputation. If not specified, we determine the mode as follows:
    We use recomputation by default since it incurs lower overhead than
    swapping. However, when the sequence group has multiple sequences
    (e.g., beam search), recomputation is not currently supported. In
    such a case, we use swapping instead."""
2498
2499

    num_scheduler_steps: int = 1
2500
    """Maximum number of forward steps per scheduler call."""
2501

2502
2503
    multi_step_stream_outputs: bool = True
    """If False, then multi-step will stream outputs at the end of all steps"""
2504
2505

    send_delta_data: bool = False
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
    """Private API. If used, scheduler sends delta data to
    workers instead of an entire data. It should be enabled only
    when SPMD worker architecture is enabled. I.e.,
    VLLM_USE_RAY_SPMD_WORKER=1"""

    policy: SchedulerPolicy = "fcfs"
    """The scheduling policy to use:\n
    - "fcfs" means first come first served, i.e. requests are handled in order
    of arrival.\n
    - "priority" means requests are handled based on given priority (lower
    value means earlier handling) and time of arrival deciding any ties)."""
2517
2518

    chunked_prefill_enabled: bool = field(init=False)
2519
    """True if chunked prefill is enabled."""
2520

2521
    disable_chunked_mm_input: bool = False
2522
2523
2524
2525
2526
2527
    """If set to true and chunked prefill is enabled, we do not want to
    partially schedule a multimodal item. Only used in V1
    This ensures that if a request has a mixed prompt
    (like text tokens TTTT followed by image tokens IIIIIIIIII) where only
    some image tokens can be scheduled (like TTTTIIIII, leaving IIIII),
    it will be scheduled as TTTT in one step and IIIIIIIIII in the next."""
2528

2529
2530
    # scheduler class or path. "vllm.core.scheduler.Scheduler" (default)
    # or "mod.custom_class".
2531
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
2532
2533
2534
    """The scheduler class to use. "vllm.core.scheduler.Scheduler" is the
    default scheduler. Can be a class directly or the path to a class of form
    "mod.custom_class"."""
2535

2536
2537
2538
2539
2540
2541
    disable_hybrid_kv_cache_manager: bool = False
    """If set to True, KV cache manager will allocate the same size of KV cache
    for all attention layers even if there are multiple type of attention layers
    like full attention and sliding window attention.
    """

2542
2543
2544
2545
2546
2547
2548
    async_scheduling: bool = False
    """EXPERIMENTAL: If set to True, perform async scheduling. This may help
    reduce the CPU overheads, leading to better latency and throughput. However,
    async scheduling is currently not supported with some features such as
    structured outputs, speculative decoding, and pipeline parallelism.
    """

2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
    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.
2563
        factors: list[Any] = []
2564
2565
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2566
2567
        return hash_str

2568
    def __post_init__(self) -> None:
2569
2570
2571
2572
2573
2574
        if self.max_model_len is None:
            self.max_model_len = 8192

        if self.max_num_seqs is None:
            self.max_num_seqs = 128

2575
2576
2577
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
2578
2579
2580
2581
                    # Multi-step Chunked-Prefill doesn't allow prompt-chunking
                    # for now. Have max_num_batched_tokens set to max_model_len
                    # so we don't reject sequences on account of a short
                    # max_num_batched_tokens.
2582
                    self.max_num_batched_tokens = max(
2583
                        self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2584
                else:
2585
                    self.max_num_batched_tokens = (
2586
                        DEFAULT_MAX_NUM_BATCHED_TOKENS)
2587
            else:
2588
                # If max_model_len is too short, use
2589
                # DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
2590
                # for higher throughput.
2591
                self.max_num_batched_tokens = max(
2592
                    self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2593

2594
2595
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
2596
2597
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2598
                    POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
2599
                )
2600
            if self.is_multimodal_model:
2601
                # The value needs to be at least the number of multimodal tokens
2602
2603
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2604
                    MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
2605
2606
                )

2607
2608
2609
2610
2611
2612
2613
            # When using default settings,
            # Ensure max_num_batched_tokens does not exceed model limit.
            # Some models (e.g., Whisper) have embeddings tied to max length.
            self.max_num_batched_tokens = min(
                self.max_num_seqs * self.max_model_len,
                self.max_num_batched_tokens)

2614
2615
2616
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

2617
        if self.enable_chunked_prefill:
2618
2619
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
2620
                self.max_num_batched_tokens)
2621

2622
        self.chunked_prefill_enabled = self.enable_chunked_prefill
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
        if self.max_num_partial_prefills > 1:
            if self.long_prefill_token_threshold == 0:
                self.long_prefill_token_threshold = int(self.max_model_len *
                                                        0.04)

            logger.info(
                "Concurrent partial prefills enabled with "
                "max_num_partial_prefills=%d, max_long_partial_prefills=%d, "
                "long_prefill_token_threshold=%d",
                self.max_num_partial_prefills, self.max_long_partial_prefills,
                self.long_prefill_token_threshold)

2635
2636
2637
2638
2639
2640
2641
        # NOTE: Default set cuda_graph_sizes to [min(max_num_seqs * 2, 512)].
        # This avoids OOM in tight memory scenarios with small max_num_seqs,
        # and prevents capture of many large graphs (>512) that would greatly
        # increase startup time with limited performance benefit.
        if not self.cuda_graph_sizes:
            self.cuda_graph_sizes = [min(self.max_num_seqs * 2, 512)]

2642
2643
2644
2645
        if self.async_scheduling:
            self.scheduler_cls = (
                "vllm.v1.core.sched.async_scheduler.AsyncScheduler")

2646
2647
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2648
2649
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
2650
2651
2652
2653
2654
2655
2656
            raise ValueError(
                f"max_num_batched_tokens ({self.max_num_batched_tokens}) is "
                f"smaller than max_model_len ({self.max_model_len}). "
                "This effectively limits the maximum sequence length to "
                "max_num_batched_tokens and makes vLLM reject longer "
                "sequences. Please increase max_num_batched_tokens or "
                "decrease max_model_len.")
2657

2658
2659
2660
2661
2662
        if self.max_num_batched_tokens < self.max_num_seqs:
            raise ValueError(
                f"max_num_batched_tokens ({self.max_num_batched_tokens}) must "
                "be greater than or equal to max_num_seqs "
                f"({self.max_num_seqs}).")
2663

2664
2665
        if self.max_num_batched_tokens > self.max_num_seqs * self.max_model_len:
            logger.warning(
2666
                "max_num_batched_tokens (%d) exceeds max_num_seqs "
2667
2668
2669
2670
                "* max_model_len (%d). This may lead to unexpected behavior.",
                self.max_num_batched_tokens,
                self.max_num_seqs * self.max_model_len)

2671
2672
2673
2674
2675
2676
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

2677
2678
2679
2680
2681
2682
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
        if self.max_num_partial_prefills < 1:
            raise ValueError(
                f"max_num_partial_prefills ({self.max_num_partial_prefills}) "
                "must be greater than or equal to 1.")
        elif self.max_num_partial_prefills > 1:
            if not self.chunked_prefill_enabled:
                raise ValueError("Chunked prefill must be enabled to set "
                                 "max_num_partial_prefills > 1.")

            if self.long_prefill_token_threshold > self.max_model_len:
                raise ValueError(
                    "long_prefill_token_threshold "
                    f"({self.long_prefill_token_threshold}) cannot be greater "
                    f"than the max_model_len ({self.max_model_len}).")

        if (self.max_long_partial_prefills
                < 1) or (self.max_long_partial_prefills
                         > self.max_num_partial_prefills):
            raise ValueError(
                f"max_long_partial_prefills ({self.max_long_partial_prefills}) "
                "must be greater than or equal to 1 and less than or equal to "
                f"max_num_partial_prefills ({self.max_num_partial_prefills}).")

2706
2707
        return self

2708
2709
2710
2711
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

2712

2713
Device = Literal["auto", "cuda", "neuron", "cpu", "tpu", "xpu"]
2714
2715
2716


@config
2717
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
2718
class DeviceConfig:
2719
2720
    """Configuration for the device to use for vLLM execution."""

2721
    device: SkipValidation[Optional[Union[Device, torch.device]]] = "auto"
2722
    """Device type for vLLM execution.
2723
2724
2725
    This parameter is deprecated and will be
    removed in a future release.
    It will now be set automatically based
2726
    on the current platform."""
2727
2728
2729
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
2730

2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
    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.
2746
        factors: list[Any] = []
2747
2748
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2749
2750
        return hash_str

2751
2752
    def __post_init__(self):
        if self.device == "auto":
2753
            # Automated device type detection
2754
            from vllm.platforms import current_platform
2755
            self.device_type = current_platform.device_type
2756
            if not self.device_type:
2757
2758
2759
2760
                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.")
2761
2762
        else:
            # Device type is assigned explicitly
2763
2764
2765
2766
            if isinstance(self.device, str):
                self.device_type = self.device
            elif isinstance(self.device, torch.device):
                self.device_type = self.device.type
2767
2768

        # Some device types require processing inputs on CPU
2769
        if self.device_type in ["neuron"]:
2770
            self.device = torch.device("cpu")
2771
2772
        elif self.device_type in ["tpu"]:
            self.device = None
2773
2774
2775
2776
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

2777

2778
2779
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
                            "mlp_speculator", "draft_model", "deepseek_mtp"]
2780
2781
2782


@config
2783
@dataclass
2784
class SpeculativeConfig:
2785
    """Configuration for speculative decoding."""
2786

2787
    # General speculative decoding control
2788
    num_speculative_tokens: SkipValidation[int] = None  # type: ignore
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
    """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."""
2802
    draft_tensor_parallel_size: Optional[int] = None
2803
2804
    """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."""
2805
    disable_logprobs: bool = True
2806
2807
2808
    """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."""
2809

2810
    # Draft model configuration
2811
    quantization: Optional[me_quant.QuantizationMethods] = None
2812
2813
2814
    """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."""
2815
    max_model_len: Optional[int] = None
2816
2817
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
2818
    revision: Optional[str] = None
2819
2820
2821
    """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."""
2822
    code_revision: Optional[str] = None
2823
2824
2825
    """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."""
2826

2827
    # Advanced control
2828
    disable_by_batch_size: Optional[int] = None
2829
2830
2831
2832
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
2833
    prompt_lookup_max: Optional[int] = None
2834
2835
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
2836
    prompt_lookup_min: Optional[int] = None
2837
2838
2839
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

2840
    speculative_token_tree: Optional[str] = None
2841
    """Specifies the tree structure for speculative token generation.
2842
    """
2843
    # required configuration params passed from engine
2844
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2845
    """The configuration of the target model."""
2846
2847
    target_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2848
    """The parallel configuration for the target model."""
2849
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2850
2851
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
2852
    disable_log_stats: SkipValidation[bool] = None  # type: ignore
2853
2854
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
2855
2856

    # params generated in the post-init stage
2857
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2858
    """The configuration of the draft model initialized internal."""
2859
2860
    draft_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2861
    """The parallel configuration for the draft model initialized internal."""
2862

2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
    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.
        """
2875
        factors: list[Any] = []
2876
2877
2878
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2879
2880
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2881
2882
        return hash_str

2883
2884
2885
2886
2887
    @classmethod
    def from_dict(cls, dict_value: dict) -> "SpeculativeConfig":
        """Parse the CLI value for the speculative config."""
        return cls(**dict_value)

2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
    @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"]
            })
2898
2899
2900
2901
2902
2903
2904
2905
2906

        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
2907
2908
2909
2910
2911
2912
2913
2914
2915

        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"]
            })
2916

2917
2918
        return hf_config

2919
    def __post_init__(self):
2920

2921
2922
2923
2924
2925
2926
2927
        # 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.
2928
2929
2930
2931

        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
2932
            if self.target_model_config and \
2933
2934
2935
2936
                (self.target_model_config.hf_text_config.model_type \
                        == "deepseek_v3" or
                    self.target_model_config.hf_text_config.model_type \
                        == "mimo"):
2937
2938
2939
2940
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2941
            else:
2942
2943
2944
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2945
2946
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2947
2948
2949
2950
2951
2952
2953
        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"
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
            # 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
2968
            if self.prompt_lookup_min < 1:
2969
2970
2971
2972
2973
                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")
2974
            if self.prompt_lookup_min > self.prompt_lookup_max:
2975
2976
2977
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2978

2979
2980
2981
            # 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.
2982
2983
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2984
        else:
2985
2986
2987
2988
2989
2990
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
2991
                    runner="draft",
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
                    tokenizer=self.target_model_config.tokenizer,
                    tokenizer_mode=self.target_model_config.tokenizer_mode,
                    trust_remote_code=self.target_model_config.
                    trust_remote_code,
                    allowed_local_media_path=self.target_model_config.
                    allowed_local_media_path,
                    dtype=self.target_model_config.dtype,
                    seed=self.target_model_config.seed,
                    revision=self.revision,
                    code_revision=self.code_revision,
                    tokenizer_revision=self.target_model_config.
                    tokenizer_revision,
                    spec_target_max_model_len=self.target_model_config.
                    max_model_len,
                    quantization=self.quantization,
                    enforce_eager=self.target_model_config.enforce_eager,
                    max_seq_len_to_capture=self.target_model_config.
                    max_seq_len_to_capture,
                    max_logprobs=self.target_model_config.max_logprobs,
                    hf_overrides=SpeculativeConfig.hf_config_override,
                )
3013

3014
                # Automatically detect the method
3015
                if self.method in ('eagle', 'eagle3'):
3016
                    pass
3017
3018
                elif "eagle-" in self.draft_model_config.model.lower() or \
                        "eagle3-" in self.draft_model_config.model.lower():
3019
3020
3021
3022
3023
3024
                    self.method = "eagle"
                elif self.draft_model_config.hf_config.model_type == "medusa":
                    self.method = "medusa"
                elif (self.draft_model_config.hf_config.model_type ==
                      "mlp_speculator"):
                    self.method = "mlp_speculator"
3025
                elif (self.draft_model_config.hf_config.model_type
Yuxuan Zhang's avatar
Yuxuan Zhang committed
3026
                      in ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp")):
Jiayi Yao's avatar
Jiayi Yao committed
3027
3028
3029
3030
3031
3032
3033
                    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."
                            )
3034
                else:
3035
                    self.method = "draft_model"
3036
3037
3038
3039
3040
                    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.")
3041
3042

                # Replace hf_config for EAGLE draft_model
3043
                if self.method in ("eagle", "eagle3"):
3044
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
3045
                        raise ValueError(
3046
3047
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
3048

3049
3050
                    from vllm.transformers_utils.configs import (
                        SpeculatorsConfig)
3051
3052
                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
3053

3054
                    if isinstance(self.draft_model_config.hf_config,
3055
                                  (EAGLEConfig, SpeculatorsConfig)):
3056
3057
3058
                        pass
                    else:
                        eagle_config = EAGLEConfig(
3059
                            self.draft_model_config.hf_config,
3060
3061
                            method=self.method,
                            model_type="eagle")
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
                        self.draft_model_config.hf_config = eagle_config

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

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

3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
                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)))

3096
3097
3098
3099
3100
3101
                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
                )
3102

3103
3104
3105
3106
3107
3108
                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,
                    ))
3109

3110
3111
3112
3113
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
3114

3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
    @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,
        )

3150
    @staticmethod
3151
    def _verify_and_get_draft_tp(
3152
3153
3154
3155
3156
3157
            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.
3158
        """
3159
3160
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
3161
        if speculative_draft_tensor_parallel_size is None:
3162
3163
3164
3165
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
3166
3167
3168
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
3169
3170
3171
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
3172
3173
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
3174
            raise ValueError(
3175
                f"{speculative_draft_tensor_parallel_size=} cannot be "
3176
                f"other value than 1 or target model tensor_parallel_size")
3177
        return speculative_draft_tensor_parallel_size
3178

3179
3180
3181
3182
3183
3184
3185
3186
3187
    @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.
        """
3188
3189
3190
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
3191
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
3192
3193
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
            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

3205
3206
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
3207
3208
3209
3210
3211
3212
        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.")

3213
3214
3215
3216
3217
3218
3219
        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)
3220
3221
3222
3223
3224
3225

        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=}")
3226

3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
        from vllm.transformers_utils.configs import SpeculatorsConfig

        eagle3_target_supported = ["llama"]
        if self.draft_model_config and isinstance(
                self.draft_model_config.hf_config, SpeculatorsConfig):
            eagle3_target_supported.append("qwen")

        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):
3238
            raise ValueError(
3239
                f"Eagle3 is only supported for {eagle3_target_supported} models. "  # noqa: E501
3240
3241
                f"Got {self.target_model_config.hf_text_config.model_type=}")

3242
3243
        return self

3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
    @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

3254
    def use_eagle(self) -> bool:
Jiayi Yao's avatar
Jiayi Yao committed
3255
        return self.method in ("eagle", "eagle3", "deepseek_mtp")
3256

3257
    def __repr__(self) -> str:
3258
3259
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
3260
        num_spec_tokens = self.num_speculative_tokens
3261
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
3262
3263


3264
3265
3266
3267
LoRADType = Literal["auto", "float16", "bfloat16"]


@config
3268
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3269
class LoRAConfig:
3270
3271
3272
3273
3274
3275
    """Configuration for LoRA."""

    max_lora_rank: int = 16
    """Max LoRA rank."""
    max_loras: int = 1
    """Max number of LoRAs in a single batch."""
3276
    fully_sharded_loras: bool = False
3277
3278
3279
3280
    """By default, only half of the LoRA computation is sharded with tensor
    parallelism. Enabling this will use the fully sharded layers. At high
    sequence length, max rank or tensor parallel size, this is likely faster.
    """
3281
    max_cpu_loras: Optional[int] = None
3282
3283
3284
3285
    """Maximum number of LoRAs to store in CPU memory. Must be >= than
    `max_loras`."""
    lora_dtype: Union[torch.dtype, LoRADType] = "auto"
    """Data type for LoRA. If auto, will default to base model dtype."""
3286
    lora_extra_vocab_size: int = 256
3287
3288
    """Maximum size of extra vocabulary that can be present in a LoRA adapter
    (added to the base model vocabulary)."""
3289
3290
    lora_vocab_padding_size: ClassVar[int] = current_platform\
        .get_lora_vocab_padding_size()
3291

3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
    default_mm_loras: Optional[dict[str, str]] = None
    """Dictionary mapping specific modalities to LoRA model paths; this field
    is only applicable to multimodal models and should be leveraged when a
    model always expects a LoRA to be active when a given modality is present.
    Note that currently, if a request provides multiple additional
    modalities, each of which have their own LoRA, we do NOT apply
    default_mm_loras because we currently only support one lora adapter
    per prompt. When run in offline mode, the lora IDs for n modalities
    will be automatically assigned to 1-n with the names of the modalities
    in alphabetic order."""
3302
    bias_enabled: bool = False
3303
    """Enable bias for LoRA adapters."""
3304

3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
    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.
        """
3317
        factors: list[Any] = []
3318
3319
3320
3321
3322
        factors.append(self.max_lora_rank)
        factors.append(self.max_loras)
        factors.append(self.fully_sharded_loras)
        factors.append(self.lora_dtype)
        factors.append(self.lora_extra_vocab_size)
3323
        factors.append(self.lora_vocab_padding_size)
3324
        factors.append(self.bias_enabled)
3325
3326
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3327
3328
        return hash_str

3329
    def __post_init__(self):
3330
        # Setting the maximum rank to 512 should be able to satisfy the vast
3331
        # majority of applications.
3332
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
3333
        possible_lora_extra_vocab_size = (256, 512)
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
        if self.max_lora_rank not in possible_max_ranks:
            raise ValueError(
                f"max_lora_rank ({self.max_lora_rank}) must be one of "
                f"{possible_max_ranks}.")
        if self.lora_extra_vocab_size not in possible_lora_extra_vocab_size:
            raise ValueError(
                f"lora_extra_vocab_size ({self.lora_extra_vocab_size}) "
                f"must be one of {possible_lora_extra_vocab_size}.")
        if self.max_loras < 1:
            raise ValueError(f"max_loras ({self.max_loras}) must be >= 1.")
        if self.max_cpu_loras is None:
            self.max_cpu_loras = self.max_loras
        elif self.max_cpu_loras < self.max_loras:
            raise ValueError(
                f"max_cpu_loras ({self.max_cpu_loras}) must be >= "
zspo's avatar
zspo committed
3349
                f"max_loras ({self.max_loras})")
3350

3351
    def verify_with_cache_config(self, cache_config: CacheConfig):
3352
3353
3354
        if cache_config.cpu_offload_gb > 0 and not envs.VLLM_USE_V1:
            raise ValueError(
                "V0 LoRA does not support CPU offload, please use V1.")
3355

3356
3357
3358
3359
3360
3361
3362
    def verify_with_model_config(self, model_config: ModelConfig):
        if self.lora_dtype in (None, "auto"):
            self.lora_dtype = model_config.dtype
        elif isinstance(self.lora_dtype, str):
            self.lora_dtype = getattr(torch, self.lora_dtype)


3363
@config
3364
@dataclass
3365
class MultiModalConfig:
3366
3367
    """Controls the behavior of multimodal models."""

3368
3369
    limit_per_prompt: dict[str, int] = \
        cast(dict[str, int], get_field(ModelConfig, "limit_mm_per_prompt"))
3370
    """
3371
    The maximum number of input items allowed per prompt for each modality.
3372
    Defaults to 1 (V0) or 999 (V1) for each modality.
3373
3374

    For example, to allow up to 16 images and 2 videos per prompt:
3375
    `{"image": 16, "video": 2}`
3376
3377
    """

3378
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
3379
3380
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
3381
3382
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """

3383
3384
3385
    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
3386
    `transformers.AutoProcessor.from_pretrained`.
3387
3388
3389
3390

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

    For example, for Phi-3-Vision:
3391
    `{"num_crops": 4}`.
3392
3393
3394
3395
    """

    disable_mm_preprocessor_cache: bool = False
    """
3396
    If `True`, disable caching of the multi-modal processor.
3397
3398
    """

3399
3400
3401
3402
3403
    interleave_mm_strings: bool = False
    """
    Enable fully interleaved support for multimodal prompts.
    """

3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
    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.
3418
        factors: list[Any] = []
3419
3420
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3421
3422
        return hash_str

3423
3424
3425
3426
3427
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
3428
3429
3430
3431
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
3432

3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
    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)
3443

3444

3445
@config
3446
3447
@dataclass
class PoolerConfig:
3448
    """Controls the behavior of output pooling in pooling models."""
3449
3450

    pooling_type: Optional[str] = None
3451
    """
3452
    The pooling method of the pooling model. This should be a key in
3453
    [`vllm.model_executor.layers.pooler.PoolingType`][].
3454
3455
    """

3456
    ## for embeddings models
3457
3458
    normalize: Optional[bool] = None
    """
3459
3460
3461
3462
3463
3464
    Whether to normalize the embeddings outputs. 
    """
    dimensions: Optional[int] = None
    """
    Reduce the dimensions of embeddings if model 
    support matryoshka representation.
3465
3466
    """

3467
3468
    ## for classification models
    activation: Optional[bool] = None
3469
    """
3470
    Whether to apply activation function to the classification outputs. 
3471
3472
    """

3473
3474
3475
3476
3477
    ## for reward models
    softmax: Optional[bool] = None
    """
    Whether to apply softmax to the reward outputs. 
    """
3478
3479
    step_tag_id: Optional[int] = None
    """
3480
    If set, only the score corresponding to the ``step_tag_id`` in the
3481
3482
3483
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """
3484
    returned_token_ids: Optional[list[int]] = None
3485
    """
3486
3487
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
3488
3489
3490
    ``math-shepherd-mistral-7b-prm`` model.
    """

3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
    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.
3505
        factors: list[Any] = []
3506
3507
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3508
3509
        return hash_str

3510

3511
3512
3513
3514
3515
3516
3517
3518
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

3519
3520
3521
3522
3523
3524
3525
# model_type -> reason
_FLOAT16_NOT_SUPPORTED_MODELS = {
    "gemma2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "gemma3": "Numerical instability. Please use bfloat16 or float32 instead.",
    "plamo2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "glm4": "Numerical instability. Please use bfloat16 or float32 instead.",
}
3526

3527

3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
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,
3546
    config: PretrainedConfig,
3547
3548
3549
    *,
    revision: Optional[str],
):
3550
3551
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
3552
    config_dtype = getattr(config, "torch_dtype", None)
3553

3554
    # Fallbacks for multi-modal models if the root config
3555
    # does not define torch_dtype
3556
3557
    if config_dtype is None:
        config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
3558
3559
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)
3560
3561
    if config_dtype is None and hasattr(config, "encoder_config"):
        config_dtype = getattr(config.encoder_config, "torch_dtype", None)
3562

3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
    # 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)

3578
3579
3580
    if config_dtype is None:
        config_dtype = torch.float32

3581
    return config_dtype
3582

Shinichi Hemmi's avatar
Shinichi Hemmi committed
3583

3584
3585
3586
3587
3588
3589
3590
def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform
3591

3592
3593
3594
3595
    supported_dtypes = [
        dtype for dtype in current_platform.supported_dtypes
        if _is_valid_dtype(model_type, dtype)
    ]
3596

3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
    if is_pooling_model and torch.float16 in supported_dtypes:
        preferred_dtype = torch.float16
    else:
        preferred_dtype = supported_dtypes[0]

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

    if config_dtype in supported_dtypes:
        return config_dtype

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

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

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

    return preferred_dtype


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

    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            # Set default dtype from model config
            torch_dtype = _resolve_auto_dtype(
                model_type,
                config_dtype,
                is_pooling_model=is_pooling_model,
            )
3650
        else:
3651
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
3652
                raise ValueError(f"Unknown dtype: {dtype!r}")
3653
3654
3655
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
3656
    else:
3657
        raise ValueError(f"Unknown dtype: {dtype}")
3658

3659
3660
    _check_valid_dtype(model_type, torch_dtype)

3661
3662
3663
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
3664
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
3665
3666
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
3667
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
3668
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
3669
            # Casting between float16 and bfloat16 is allowed with a warning.
3670
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
3671
3672

    return torch_dtype
3673
3674
3675
3676


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
3677
    tokenizer_config: Optional[dict],
3678
    max_model_len: Optional[int],
3679
    disable_sliding_window: bool,
3680
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
3681
    spec_target_max_model_len: Optional[int] = None,
3682
    encoder_config: Optional[Any] = None,
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
) -> 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",
3693
3694
        # ChatGLM2
        "seq_length",
3695
3696
        # Command-R
        "model_max_length",
3697
3698
        # Whisper
        "max_target_positions",
3699
3700
3701
3702
3703
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
3704
    # Choose the smallest "max_length" from the possible keys
3705
    max_len_key = None
3706
    for key in possible_keys:
3707
3708
3709
3710
3711
        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
3712
3713
3714
3715
    # 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
3716
3717
3718
3719

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
    if disable_sliding_window and sliding_window_len is not None:
3720
3721

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
3722
        max_len_key = "sliding_window" \
3723
3724
3725
            if sliding_window_len_min < derived_max_model_len else max_len_key
        derived_max_model_len = min(derived_max_model_len,
                                    sliding_window_len_min)
3726

3727
3728
3729
3730
3731
3732
3733
    # 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)

3734
3735
    # If none of the keys were found in the config, use a default and
    # log a warning.
3736
    if derived_max_model_len == float("inf"):
3737
3738
3739
3740
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

3741
3742
3743
3744
3745
        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

3746
3747
3748
3749
        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: "
3750
            "%s. Assuming the model's maximum length is %d.", possible_keys,
3751
            default_max_len)
3752
        derived_max_model_len = default_max_len
3753

3754
    rope_scaling = getattr(hf_config, "rope_scaling", None)
3755
3756
3757
    # 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:
3758
3759
3760
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770

        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.")

3771
3772
3773
3774
            # 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)

3775
3776
3777
3778
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
3779

3780
3781
3782
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

3783
3784
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
3785
    if max_model_len is None:
3786
        max_model_len = int(derived_max_model_len)
3787
3788
3789
3790
3791
3792
3793
3794
        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)
3795
    elif max_model_len > derived_max_model_len:
3796
3797
3798
3799
3800
        # 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:
3801
3802
3803
3804
3805
3806
3807
            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.")
3808
        else:
3809
            msg = (
3810
                f"User-specified max_model_len ({max_model_len}) is greater "
3811
3812
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
3813
                f"{model_max_length} in model's config.json). This may lead "
3814
3815
3816
3817
3818
3819
3820
3821
3822
                "to incorrect model outputs or CUDA errors.")
            if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
                logger.warning(
                    "%s Make sure the value is correct and within the "
                    "model context size.", msg)
            else:
                raise ValueError(
                    f"{msg} To allow overriding this maximum, set "
                    "the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1")
3823
    return int(max_model_len)
3824
3825


3826
def get_min_sliding_window(
3827
        sliding_window: Union[int, list[Optional[int]]]) -> int:
3828
3829
3830
3831
3832
3833
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


3834
def get_served_model_name(model: str,
3835
                          served_model_name: Optional[Union[str, list[str]]]):
3836
    """
3837
3838
3839
3840
    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
3841
3842
3843
3844
3845
3846
3847
3848
3849
    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


3850
GuidedDecodingBackend = Literal["auto", "xgrammar", "guidance", "outlines"]
3851
3852
3853


@config
3854
3855
@dataclass
class DecodingConfig:
3856
    """Dataclass which contains the decoding strategy of the engine."""
3857

3858
    backend: GuidedDecodingBackend = "auto"
3859
3860
3861
3862
    """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."""
3863

3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
    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`."""

3876
    reasoning_backend: str = ""
3877
    """Select the reasoning parser depending on the model that you're using.
3878
    This is used to parse the reasoning content into OpenAI API format."""
3879

3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
    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.
3894
        factors: list[Any] = []
3895
3896
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3897
3898
        return hash_str

3899
    def __post_init__(self):
3900
3901
3902
3903
3904
3905
3906
3907
        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.")

3908

3909
3910
3911
3912
DetailedTraceModules = Literal["model", "worker", "all"]


@config
3913
3914
@dataclass
class ObservabilityConfig:
3915
    """Configuration for observability - metrics and tracing."""
3916

3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
    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)
3932

3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
    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))
3958

3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
    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.
3973
        factors: list[Any] = []
3974
3975
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3976
3977
        return hash_str

3978
    def __post_init__(self):
3979
3980
3981
3982
3983
        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()

3984
        from vllm.tracing import is_otel_available, otel_import_error_traceback
3985
3986
3987
3988
3989
        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}")
3990

3991
3992
3993
3994
3995
3996
    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(","))

3997

3998
3999
4000
4001
4002
4003
4004
4005
KVProducer = Literal["kv_producer", "kv_both"]
KVConsumer = Literal["kv_consumer", "kv_both"]
KVRole = Literal[KVProducer, KVConsumer]


@config
@dataclass
class KVTransferConfig:
4006
4007
4008
    """Configuration for distributed KV cache transfer."""

    kv_connector: Optional[str] = None
4009
4010
    """The KV connector for vLLM to transmit KV caches between vLLM instances.
    """
4011

4012
    engine_id: Optional[str] = None
Robert Shaw's avatar
Robert Shaw committed
4013
4014
    """The engine id for KV transfers."""

4015
    kv_buffer_device: Optional[str] = "cuda"
4016
4017
    """The device used by kv connector to buffer the KV cache.
    Currently only support 'cuda'."""
4018
4019

    kv_buffer_size: float = 1e9
4020
4021
    """The buffer size for TorchDistributedConnector. Measured in number of
    bytes. Recommended value: 1e9 (about 1GB)."""
4022

4023
4024
    kv_role: Optional[KVRole] = None
    """Whether this vLLM instance produces, consumes KV cache, or both. Choices
Robert Shaw's avatar
Robert Shaw committed
4025
    are 'kv_producer', 'kv_consumer', and 'kv_both'."""
4026
4027

    kv_rank: Optional[int] = None
4028
4029
4030
    """The rank of this vLLM instance in the KV cache transfer. Typical value:
    0 for prefill instance, 1 for decode instance.
    Currently only 1P1D is supported."""
4031
4032

    kv_parallel_size: int = 1
4033
4034
    """The number of parallel instances for KV cache transfer. For
    PyNcclConnector, this should be 2."""
4035
4036

    kv_ip: str = "127.0.0.1"
4037
    """The KV connector ip, used to build distributed connection."""
4038
4039

    kv_port: int = 14579
4040
    """The KV connector port, used to build distributed connection."""
4041

4042
4043
    kv_connector_extra_config: dict[str, Any] = field(default_factory=dict)
    """any extra config that the connector may need."""
4044

4045
4046
4047
4048
    kv_connector_module_path: Optional[str] = None
    """The Python module path to dynamically load the KV connector from.
    Only supported in V1."""

4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
    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.
4063
        factors: list[Any] = []
4064
4065
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
4066
4067
        return hash_str

4068
    def __post_init__(self) -> None:
4069
4070
4071
        if self.engine_id is None:
            self.engine_id = str(uuid.uuid4())

4072
4073
4074
        if self.kv_role is not None and self.kv_role not in get_args(KVRole):
            raise ValueError(f"Unsupported kv_role: {self.kv_role}. "
                             f"Supported roles are {get_args(KVRole)}")
4075
4076
4077

        if self.kv_connector is not None and self.kv_role is None:
            raise ValueError("Please specify kv_disagg_role when kv_connector "
4078
                             f"is set, supported roles are {get_args(KVRole)}")
4079
4080
4081
4082

    @property
    def is_kv_transfer_instance(self) -> bool:
        return self.kv_connector is not None and \
4083
            self.kv_role in get_args(KVRole)
4084
4085
4086
4087

    @property
    def is_kv_producer(self) -> bool:
        return self.kv_connector is not None and \
4088
            self.kv_role in get_args(KVProducer)
4089
4090
4091
4092

    @property
    def is_kv_consumer(self) -> bool:
        return self.kv_connector is not None and \
4093
            self.kv_role in get_args(KVConsumer)
4094

4095
4096
4097
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

4098

4099
4100
4101
@config
@dataclass
class KVEventsConfig:
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
    """Configuration for KV event publishing."""

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

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

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

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

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

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

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

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


4141
4142
4143
4144
4145
4146
4147
4148
class CompilationLevel:
    # constants for the levels of the compilation process
    NO_COMPILATION = 0
    DYNAMO_AS_IS = 1
    DYNAMO_ONCE = 2
    PIECEWISE = 3


4149
4150
4151
4152
4153
4154
4155
4156
4157
@config
@dataclass
class PassConfig:
    """Configuration for custom Inductor passes.

    This is separate from general `CompilationConfig` so that inductor passes
    don't all have access to full configuration - that would create a cycle as
    the `PassManager` is set as a property of config."""

4158
    enable_fusion: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
4159
4160
4161
    """Whether to enable the custom fusion (RMSNorm/SiluMul+quant) pass."""
    enable_attn_fusion: bool = False
    """Whether to enable the custom attention+quant fusion pass."""
4162
    enable_noop: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
4163
4164
4165
    """Whether to enable the custom no-op elimination pass."""
    enable_sequence_parallelism: bool = False
    """Whether to enable sequence parallelism."""
4166
4167
    enable_async_tp: bool = False
    """Whether to enable async TP."""
4168
4169
    enable_fi_allreduce_fusion: bool = False
    """Whether to enable flashinfer allreduce fusion."""
4170
    fi_allreduce_fusion_max_token_num: int = 16384
4171
    """Max number of tokens to used in flashinfer allreduce fusion."""
4172

4173
4174
    # TODO(luka) better pass enabling system.

4175
4176
4177
4178
    def uuid(self):
        """
        Produces a hash unique to the pass configuration.
        Any new fields that affect compilation should be added to the hash.
4179
        Any future fields that don't affect compilation should be excluded.
4180
        """
4181
        return InductorPass.hash_dict(asdict(self))
4182
4183

    def __post_init__(self) -> None:
4184
4185
4186
4187
4188
4189
4190
4191
4192
        if not self.enable_noop:
            if self.enable_fusion:
                logger.warning_once(
                    "Fusion enabled but reshape elimination disabled. "
                    "RMSNorm/SiluMul + quant (fp8) fusion might not work")
            if self.enable_attn_fusion:
                logger.warning_once(
                    "Fusion enabled but reshape elimination disabled. "
                    "Attention + quant (fp8) fusion might not work")
4193
4194
4195
4196
4197
4198
4199


@config
@dataclass
class CompilationConfig:
    """Configuration for compilation. It has three parts:

4200
    - Top-level Compilation control:
4201
4202
4203
4204
4205
4206
        - [`level`][vllm.config.CompilationConfig.level]
        - [`debug_dump_path`][vllm.config.CompilationConfig.debug_dump_path]
        - [`cache_dir`][vllm.config.CompilationConfig.cache_dir]
        - [`backend`][vllm.config.CompilationConfig.backend]
        - [`custom_ops`][vllm.config.CompilationConfig.custom_ops]
        - [`splitting_ops`][vllm.config.CompilationConfig.splitting_ops]
4207
    - CudaGraph capture:
4208
4209
4210
4211
4212
4213
4214
4215
        - [`use_cudagraph`][vllm.config.CompilationConfig.use_cudagraph]
        - [`cudagraph_capture_sizes`]
        [vllm.config.CompilationConfig.cudagraph_capture_sizes]
        - [`cudagraph_num_of_warmups`]
        [vllm.config.CompilationConfig.cudagraph_num_of_warmups]
        - [`cudagraph_copy_inputs`]
        [vllm.config.CompilationConfig.cudagraph_copy_inputs]
        - [`full_cuda_graph`][vllm.config.CompilationConfig.full_cuda_graph]
4216
    - Inductor compilation:
4217
4218
4219
4220
4221
        - [`use_inductor`][vllm.config.CompilationConfig.use_inductor]
        - [`compile_sizes`][vllm.config.CompilationConfig.compile_sizes]
        - [`inductor_compile_config`]
        [vllm.config.CompilationConfig.inductor_compile_config]
        - [`inductor_passes`][vllm.config.CompilationConfig.inductor_passes]
4222
        - custom inductor passes
4223

4224
4225
4226
4227
4228
4229
4230
4231
4232
    Why we have different sizes for cudagraph and inductor:
    - cudagraph: a cudagraph captured for a specific size can only be used
        for the same size. We need to capture all the sizes we want to use.
    - inductor: a graph compiled by inductor for a general shape can be used
        for different sizes. Inductor can also compile for specific sizes,
        where it can have more information to optimize the graph with fully
        static shapes. However, we find the general shape compilation is
        sufficient for most cases. It might be beneficial to compile for
        certain small batchsizes, where inductor is good at optimizing.
4233
4234
    """
    # Top-level Compilation control
4235
    level: Optional[int] = None
4236
4237
    """The level of compilation:

4238
4239
    - None: If None, we will select the default compilation level.
      For V1 engine this is 3, for V0 engine this is 0.
4240
4241
4242
4243
    - 0: no compilation.
    - 1: dynamo as is.
    - 2: dynamo once.
    - 3: piecewise compilation."""
4244
    debug_dump_path: str = ""
4245
    """The path to dump the debug information."""
4246
    cache_dir: str = ""
4247
4248
4249
    """The directory to store the compiled graph, to accelerate Inductor
    compilation. By default, it will use model-related information to generate
    a cache directory."""
4250
    backend: str = ""
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
    """The backend for compilation. It needs to be a string:

    - "" (empty string): use the default backend.
    - "eager"/"openxla"/...: use the specified backend registered in PyTorch.
    - "full.module.name": a qualified name which can be used to import the

    backend function.
    We use string to avoid serialization issues when using compilation in a
    distributed setting. When the compilation level is 1 or 2, the backend is
    used for the compilation directly (it sees the whole graph). When the
    compilation level is 3, the backend is used for the piecewise compilation
    (it sees a part of the graph)."""
    custom_ops: list[str] = field(default_factory=list)
    """Fine-grained control over which custom ops to enable/disable. Use 'all'
    to enable all, 'none' to disable all. Also specify a list of custom op
    names to enable (prefixed with a '+'), or disable (prefixed with a '-').
    Examples:

    - 'all,-op1' to enable all except op1
    - 'none,+op1,+op2' to enable only op1 and op2

    By default, all custom ops are enabled when running without Inductor and
4273
4274
    disabled when running with Inductor: level>=PIECEWISE and use_inductor=True.
    Inductor generates (fused) Triton kernels for disabled custom ops."""
4275
4276
4277
4278
4279
    splitting_ops: list[str] = field(default_factory=list)
    """A list of ops to split the full graph into subgraphs, used in piecewise
    compilation."""

    # Inductor capture
4280
    use_inductor: bool = True
4281
4282
    """Whether to use inductor compilation:

4283
4284
4285
4286
4287
    - False: inductor compilation is not used. graph runs in eager
        (custom_ops enabled by default).
    - True: inductor compilation is used (custom_ops disabled by default).
        One graph for symbolic shape and one graph per size in compile_sizes
        are compiled using configurations in inductor_compile_config.
4288

4289
    This setting is ignored if level<PIECEWISE."""
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
    compile_sizes: Optional[list[Union[int, str]]] = None
    """Sizes to compile for inductor. In addition
    to integers, it also supports "cudagraph_capture_sizes" to
    specify the sizes for cudagraph capture."""
    inductor_compile_config: dict = field(default_factory=dict)
    """Additional configurations for inductor.
    - None: use default configurations."""
    inductor_passes: dict[str, str] = field(default_factory=dict)
    """Additional passes for inductor. It is a dictionary
    from pass name to pass function qualified name. We use function
    name because the config uses JSON format. If we pass the config
    from Python, functions can also be passed directly via Python object
    constructor, e.g. `CompilationConfig(inductor_passes={"a": func})`."""

    # CudaGraph compilation
4305
    use_cudagraph: bool = field(default_factory=lambda: envs.VLLM_USE_V1)
4306
4307
4308
4309
4310
    """Whether to use cudagraph inside compilation.
    - False: cudagraph inside compilation is not used.
    - True: cudagraph inside compilation is used. It requires
        that all input buffers have fixed addresses, and all
        splitting ops write their outputs to input buffers.
4311
4312
    In the vLLM V1 Engine, this flag only applies for
    CompilationLevel.PIECEWISE (aka -O3).
4313
4314
4315
4316
    Note that this is orthogonal to the cudagraph capture logic
    outside of compilation.
    TODO: move outside cudagraph logic into compilation.
    torch.compile will handle cudagraph capture logic in the future."""
4317
    cudagraph_num_of_warmups: int = 0
4318
4319
4320
4321
    """Number of warmup runs for cudagraph.
    It means the first several runs will be treated as warmup runs.
    Only after that, the execution will be recorded, and the recorded
    cudagraph will be used for subsequent runs."""
4322
    cudagraph_capture_sizes: Optional[list[int]] = None
4323
4324
4325
    """Sizes to capture cudagraph.
    - None (default): capture sizes are inferred from vllm config.
    - list[int]: capture sizes are specified as given."""
4326
    cudagraph_copy_inputs: bool = False
4327
4328
4329
4330
4331
    """Whether to copy input tensors for
    cudagraph. If the caller can guarantee that the same input buffers
    are always used, it can set this to False. Otherwise, it should
    set this to True, and the compiler will copy the input to an
    internally managed buffer. Default is False."""
4332
    full_cuda_graph: bool = False
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
    """whether to use a full cuda graph for the entire forward pass rather than
    splitting certain operations such as attention into subgraphs. Thus this
    flag cannot be used together with splitting_ops. This may provide
    performance benefits for smaller models."""

    pass_config: PassConfig = field(default_factory=PassConfig)
    """Custom inductor passes, see PassConfig for more details"""

    max_capture_size: int = field(default=None, init=False)  # type: ignore
    """not configurable, computed after init"""
    local_cache_dir: str = field(default=None, init=False)  # type: ignore
    """local cache dir for each rank"""
    bs_to_padded_graph_size: list[int] = field(
        default=None,  # type: ignore
        init=False)
    """optimization:
    Intuitively, bs_to_padded_graph_size should be dict[int, int].
    since we know all keys are in a range [0, max_capture_size],
    we can optimize it to list[int] for better lookup performance."""
4352

4353
    # keep track of enabled and disabled custom ops
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
    enabled_custom_ops: Counter[str] = field(default_factory=Counter,
                                             init=False)
    """custom ops that are enabled"""
    disabled_custom_ops: Counter[str] = field(default_factory=Counter,
                                              init=False)
    """custom ops that are disabled"""
    traced_files: set[str] = field(default_factory=set, init=False)
    """files that are traced for compilation"""
    compilation_time: float = field(default=0.0, init=False)
    """time taken for compilation"""

    static_forward_context: dict[str, Any] = field(default_factory=dict,
                                                   init=False)
    """Per-model forward context
    Map from layer name to layer objects that need to be accessed outside
    model code, e.g., Attention, FusedMOE when dp_size>1."""
4370

4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
    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.
        """
4383
        factors: list[Any] = []
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
        factors.append(self.level)
        factors.append(self.backend)
        factors.append(self.custom_ops)
        factors.append(self.splitting_ops)
        factors.append(self.use_inductor)
        factors.append(self.inductor_compile_config)
        factors.append(self.inductor_passes)
        factors.append(self.pass_config.uuid())
        return hashlib.sha256(str(factors).encode()).hexdigest()

4394
4395
    def __repr__(self) -> str:
        exclude = {
4396
4397
4398
4399
4400
4401
4402
4403
4404
            "static_forward_context": True,
            "enabled_custom_ops": True,
            "disabled_custom_ops": True,
            "compilation_time": True,
            "bs_to_padded_graph_size": True,
            "traced_files": True,
            "inductor_compile_config": {
                "post_grad_custom_post_pass": True,
            },
4405
        }
4406
4407
4408
4409
4410
4411
4412
4413
4414

        # exclude default attr in pass_config
        pass_config_exclude = {}
        for attr, default_val in vars(PassConfig()).items():
            if getattr(self.pass_config, attr) == default_val:
                pass_config_exclude[attr] = True
        if pass_config_exclude:
            exclude["pass_config"] = pass_config_exclude

4415
4416
4417
4418
        # The cast to string is necessary because Pydantic is mocked in docs
        # builds and sphinx-argparse doesn't know the return type of decode()
        return str(
            TypeAdapter(CompilationConfig).dump_json(
4419
4420
4421
                self,
                exclude=exclude,  # type: ignore[arg-type]
                exclude_unset=True).decode())
4422
4423
4424

    __str__ = __repr__

4425
4426
    @classmethod
    def from_cli(cls, cli_value: str) -> "CompilationConfig":
4427
4428
4429
        """Parse the CLI value for the compilation config.
        -O1, -O2, -O3, etc. is handled in FlexibleArgumentParser.
        """
4430
        return TypeAdapter(CompilationConfig).validate_json(cli_value)
4431

4432
    def __post_init__(self) -> None:
4433
4434
4435
4436
        count_none = self.custom_ops.count("none")
        count_all = self.custom_ops.count("all")
        assert count_none + count_all <= 1, "Can only specify 'none' or 'all'"

Michael Goin's avatar
Michael Goin committed
4437
4438
4439
4440
4441
4442
4443
4444
        # TODO(zou3519/luka): There are 2 issues with auto-functionalization V2:
        # 1. A bug in PyTorch, fixed in 2.7:
        #    https://github.com/pytorch/pytorch/issues/147924
        # 2. Custom passes (fusion) rely on auto-functionalization V1 and don't
        #    work with V2. Addressing this will take extra engineering effort
        #    and it is not yet a priority. RFC here:
        #    https://github.com/vllm-project/vllm/issues/14703

4445
        if is_torch_equal_or_newer("2.6"):
Michael Goin's avatar
Michael Goin committed
4446
4447
4448
4449
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

4450
4451
4452
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
4453
4454
4455
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
4456
4457
4458
4459
4460
4461
4462
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
4463
4464
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
4465

4466
4467
        if isinstance(self.pass_config, dict):
            self.pass_config = PassConfig(**self.pass_config)
4468

4469
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
        if self.level == CompilationLevel.NO_COMPILATION:
            raise ValueError("No compilation level is set.")

        from torch._dynamo.backends.registry import list_backends
        torch_backends = list_backends(exclude_tags=tuple())
        if self.level in [
                CompilationLevel.DYNAMO_AS_IS, CompilationLevel.DYNAMO_ONCE
        ]:
            if self.backend == "":
                return "eager"
            if self.backend in torch_backends:
                return self.backend
            return resolve_obj_by_qualname(self.backend)

        # TODO: pass user-specified backend to piecewise compilation
        # merge with the config use_inductor
        assert self.level == CompilationLevel.PIECEWISE
4487

4488
        from vllm.compilation.backends import VllmBackend
4489
        return VllmBackend(vllm_config)
4490

4491
    def init_with_cudagraph_sizes(self,
4492
                                  cudagraph_capture_sizes: list[int]) -> None:
4493
        """To complete the initialization of config,
4494
4495
        we need to know the cudagraph sizes."""

4496
        if self.cudagraph_capture_sizes is None:
4497
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
4498
        else:
4499
            # de-duplicate the sizes provided by the config
4500
4501
4502
4503
4504
4505
            dedup_sizes = list(set(self.cudagraph_capture_sizes))
            if len(dedup_sizes) < len(self.cudagraph_capture_sizes):
                logger.info(("cudagraph sizes specified by model runner"
                             " %s is overridden by config %s"),
                            cudagraph_capture_sizes, dedup_sizes)
            self.cudagraph_capture_sizes = dedup_sizes
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520

        computed_compile_sizes = []
        if self.compile_sizes is not None:
            # de-duplicate the sizes provided by the config
            self.compile_sizes = list(set(self.compile_sizes))
            for x in self.compile_sizes:
                if isinstance(x, str):
                    assert x == "cudagraph_capture_sizes", \
                    "Unrecognized size type in compile_sizes, " \
                    f"expect 'cudagraph_capture_sizes', got {x}"
                    computed_compile_sizes.extend(self.cudagraph_capture_sizes)
                else:
                    assert isinstance(x, int)
                    computed_compile_sizes.append(x)
        self.compile_sizes = computed_compile_sizes  # type: ignore
4521

4522
        # sort to make sure cudagraph capture sizes are in descending order
4523
4524
4525
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
4526

4527
4528
4529
4530
        # pre-compute the mapping from batch size to padded graph size
        self.bs_to_padded_graph_size = [
            0 for i in range(self.max_capture_size + 1)
        ]
4531
4532
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
4533
4534
4535
4536
4537
4538
4539
            for bs in range(start, end):
                if bs == start:
                    self.bs_to_padded_graph_size[bs] = start
                else:
                    self.bs_to_padded_graph_size[bs] = end
        self.bs_to_padded_graph_size[
            self.max_capture_size] = self.max_capture_size
4540

4541
4542
    def set_splitting_ops_for_v1(self):
        # NOTE: this function needs to be called
4543
4544
4545
4546
4547
        if self.splitting_ops and self.full_cuda_graph:
            raise ValueError("full_cuda_graph cannot be used together with "
                             "splitting_ops, as Full CUDA graph will override "
                             f"the splitting_ops: {self.splitting_ops}")

4548
        if not self.splitting_ops:
4549
            self.splitting_ops = [] if self.full_cuda_graph else [
4550
4551
                "vllm.unified_attention",
                "vllm.unified_attention_with_output",
4552
                "vllm.mamba_mixer2",
4553
4554
            ]

4555

4556
@config
4557
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
4558
4559
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
4560
4561
4562
    simplifies passing around the distinct configurations in the codebase.
    """

4563
4564
4565
    # TODO: use default_factory once default constructing ModelConfig doesn't
    # try to download a model
    model_config: ModelConfig = None  # type: ignore
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
    """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."""
4577
    lora_config: Optional[LoRAConfig] = None
4578
4579
4580
    """LoRA configuration."""
    speculative_config: Optional[SpeculativeConfig] = None
    """Speculative decoding configuration."""
4581
    decoding_config: DecodingConfig = field(default_factory=DecodingConfig)
4582
    """Decoding configuration."""
4583
    observability_config: Optional[ObservabilityConfig] = None
4584
    """Observability configuration."""
4585
    quant_config: Optional[QuantizationConfig] = None
4586
4587
4588
    """Quantization configuration."""
    compilation_config: CompilationConfig = field(
        default_factory=CompilationConfig)
4589
    """`torch.compile` and cudagraph capture configuration for the model.
4590

4591
4592
    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}'`).
4593
    Currently, -O <n> and -O=<n> are supported as well but this will likely be
4594
    removed in favor of clearer -O<n> syntax in the future.
4595
4596
4597

    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
4598
    production, also default in V1.
4599
4600
4601
4602
4603
4604

    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."""
4605
    kv_events_config: Optional[KVEventsConfig] = None
4606
    """The configurations for event publishing."""
4607
    # some opaque config, only used to provide additional information
4608
4609
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
4610
4611
4612
4613
    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."""
4614
    instance_id: str = ""
4615
    """The ID of the vLLM instance."""
4616

4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
    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.
        """
4629
        factors: list[Any] = []
4630
4631

        # summarize vllm config
4632
        vllm_factors: list[Any] = []
4633
4634
        from vllm import __version__
        vllm_factors.append(__version__)
4635
        vllm_factors.append(envs.VLLM_USE_V1)
4636
4637
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
4638
4639
        else:
            vllm_factors.append("None")
4640
4641
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
4642
4643
        else:
            vllm_factors.append("None")
4644
4645
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
4646
4647
        else:
            vllm_factors.append("None")
4648
4649
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
4650
4651
        else:
            vllm_factors.append("None")
4652
4653
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
4654
4655
        else:
            vllm_factors.append("None")
4656
4657
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
4658
4659
        else:
            vllm_factors.append("None")
4660
4661
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
4662
4663
4664
4665
4666
            # 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))
4667
4668
        else:
            vllm_factors.append("None")
4669
4670
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
4671
4672
        else:
            vllm_factors.append("None")
4673
4674
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
4675
4676
        else:
            vllm_factors.append("None")
4677
4678
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
4679
4680
        else:
            vllm_factors.append("None")
4681
4682
4683
4684
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
4685
4686
        else:
            vllm_factors.append("None")
4687
4688
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
4689
4690
4691
        else:
            vllm_factors.append("None")
        if self.additional_config:
4692
4693
4694
4695
4696
4697
4698
4699
            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)
4700
4701
        else:
            vllm_factors.append("None")
4702
4703
        factors.append(vllm_factors)

4704
4705
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
4706
4707
        return hash_str

4708
4709
4710
4711
4712
4713
    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]
4714

4715
4716
4717
4718
4719
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
4720
        from vllm.platforms import current_platform
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
        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
4743

4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
    @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)

4755
4756
4757
4758
4759
4760
4761
4762
4763
    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

4764
4765
4766
4767
4768
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

4769
4770
4771
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
4772
4773
4774

        self.try_verify_and_update_config()

4775
4776
4777
4778
4779
        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)
4780
4781
            self.model_config.verify_dual_chunk_attention_config(
                self.load_config)
4782

4783
        self.cache_config.verify_with_parallel_config(self.parallel_config)
4784

4785
        if self.lora_config is not None:
4786
            self.lora_config.verify_with_cache_config(self.cache_config)
4787
            self.lora_config.verify_with_model_config(self.model_config)
4788

4789
        if self.quant_config is None and self.model_config is not None:
4790
4791
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
4792

4793
        from vllm.platforms import current_platform
4794
        if self.model_config is not None and \
4795
4796
4797
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
4798
            logger.warning_once(
4799
4800
4801
4802
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
        # 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
            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

4819
4820
4821
4822
4823
        # 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
4824
4825
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
4826
4827
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
4828
4829
            # By default, V1 uses piecewise CUDA graphs. If full_cuda_graph
            # is set to True, full CUDA graphs will be used.
4830
            self.compilation_config.cudagraph_num_of_warmups = 1
4831
            self.compilation_config.set_splitting_ops_for_v1()
4832

4833
        self._set_cudagraph_sizes()
4834

4835
        if self.cache_config.cpu_offload_gb > 0 and \
4836
4837
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
4838
            logger.warning(
4839
                "CPU offload is not supported with `torch.compile` in v0 yet."
4840
4841
4842
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4843
4844
4845
4846
4847
4848
        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`.")
4849
4850
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4851
4852
        if self.compilation_config.full_cuda_graph and \
            not self.model_config.disable_cascade_attn:
4853
4854
            logger.info("full_cuda_graph is not supported with "
                        "cascade attention. Disabling cascade attention.")
4855
            self.model_config.disable_cascade_attn = True
4856

4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
        disable_chunked_prefill_reasons: list[str] = []

        if self.model_config and self.model_config.pooler_config:
            pooling_type = self.model_config.pooler_config.pooling_type
            if pooling_type is None or pooling_type.lower() != "last":
                disable_chunked_prefill_reasons.append(
                    "Only \"last\" pooling supports chunked "
                    "prefill and prefix caching; disabling both.")

        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
            self.scheduler_config.max_num_batched_tokens = max(
                self.scheduler_config.max_model_len,
                DEFAULT_MAX_NUM_BATCHED_TOKENS)

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

4878
        if (self.kv_events_config is not None
4879
4880
4881
4882
4883
                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.")
4884
4885
        if (self.kv_events_config is not None
                and self.kv_events_config.publisher != "null"
4886
4887
4888
4889
4890
                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.")
4891
4892
        current_platform.check_and_update_config(self)

4893
4894
4895
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

4896
4897
4898
4899
4900
4901
4902
        if (envs.VLLM_USE_V1
                and not self.scheduler_config.disable_hybrid_kv_cache_manager):
            # logger should only print warning message for hybrid models. As we
            # can't know whether the model is hybrid or not now, so we don't log
            # warning message here and will log it later.
            if not (current_platform.is_cuda() or current_platform.is_rocm()):
                # Hybrid KV cache manager is not supported on non-GPU platforms.
4903
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4904
4905
            if self.kv_transfer_config is not None:
                # Hybrid KV cache manager is not compatible with KV transfer.
4906
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4907
4908
            if self.kv_events_config is not None:
                # Hybrid KV cache manager is not compatible with KV events.
4909
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4910
            if self.model_config is not None and \
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
                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
4928

4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
    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
        ]

4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
    def _set_cudagraph_sizes(self):
        """
        cudagraph batchsize padding logic:

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

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

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

4965
4966
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
4967
4968

        During runtime, if batchsize is larger than
4969
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4970
4971
        no cudagraph will be used.
        If the batch size is no larger than
4972
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
        we can quickly find the padded graph size for a given batch size by
        looking up `vllm_config.compilation_config.bs_to_padded_graph_size`.
        """

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

                possible_sizes = [1, 2, 4] + [8 * i for i in range(1, 1025)]
4985
4986
4987
4988
4989
                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)

4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
                # 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:
5011
5012
5013
5014
5015
5016
5017
5018
                cuda_graph_sizes = self.scheduler_config.cuda_graph_sizes
                if len(cuda_graph_sizes) == 1:
                    batch_size_capture_list = [1, 2, 4] + [
                        i for i in range(8, cuda_graph_sizes[0] + 1, 8)
                    ]
                elif len(cuda_graph_sizes) > 1:
                    batch_size_capture_list = sorted(cuda_graph_sizes)
                else:
Cyrus Leung's avatar
Cyrus Leung committed
5019
                    raise TypeError(f"Invalid value for {cuda_graph_sizes=}.")
5020
5021
5022
5023
                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)
5024
5025
5026
5027
5028
                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
                ]
5029
5030
5031
5032

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

5033
    def recalculate_max_model_len(self, max_model_len: int):
5034
        # Can only be called in try_verify_and_update_config
5035
        model_config = self.model_config
5036
        max_model_len = model_config.get_and_verify_max_len(max_model_len)
5037
5038
        self.model_config.max_model_len = max_model_len
        self.scheduler_config.max_model_len = max_model_len
5039
5040

    def try_verify_and_update_config(self):
5041
5042
5043
        if self.model_config is None:
            return

5044
5045
5046
5047
5048
        # Avoid running try_verify_and_update_config multiple times
        if getattr(self.model_config, "config_updated", False):
            return
        self.model_config.config_updated = True

5049
        architecture = self.model_config.architecture
5050
5051
5052
        if architecture is None:
            return

5053
5054
        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP, HybridAttentionMambaModelConfig)
5055
5056
5057
        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_config(self)
5058

5059
5060
5061
        if self.model_config.is_hybrid:
            HybridAttentionMambaModelConfig.verify_and_update_config(self)

5062
        if self.model_config.convert_type == "classify":
5063
5064
5065
5066
5067
            # Maybe convert ForCausalLM into ForSequenceClassification model.
            from vllm.model_executor.models.adapters import (
                SequenceClassificationConfig)
            SequenceClassificationConfig.verify_and_update_config(self)

5068
    def __str__(self):
5069
        return (
5070
5071
5072
5073
5074
            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}, "
5075
            f"revision={self.model_config.revision}, "
5076
5077
            f"override_neuron_config={self.model_config.override_neuron_config}, "  # noqa
            f"tokenizer_revision={self.model_config.tokenizer_revision}, "
5078
5079
            f"trust_remote_code={self.model_config.trust_remote_code}, "
            f"dtype={self.model_config.dtype}, "
5080
5081
            f"max_seq_len={self.model_config.max_model_len}, "
            f"download_dir={self.load_config.download_dir!r}, "
5082
            f"load_format={self.load_config.load_format}, "
5083
5084
            f"tensor_parallel_size={self.parallel_config.tensor_parallel_size}, "  # noqa
            f"pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, "  # noqa
5085
5086
5087
5088
            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}, "
5089
            f"device_config={self.device_config.device}, "
5090
5091
5092
5093
5094
5095
5096
5097
5098
            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"num_scheduler_steps={self.scheduler_config.num_scheduler_steps}, "
            f"multi_step_stream_outputs={self.scheduler_config.multi_step_stream_outputs}, "  # noqa
            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}, "
5099
5100
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
5101
5102
5103


_current_vllm_config: Optional[VllmConfig] = None
5104
_current_prefix: Optional[str] = None
5105
5106
5107


@contextmanager
5108
5109
5110
def set_current_vllm_config(vllm_config: VllmConfig,
                            check_compile=False,
                            prefix: Optional[str] = None):
5111
    """
5112
    Temporarily set the current vLLM config.
5113
    Used during model initialization.
5114
    We save the current vLLM config in a global variable,
5115
    so that all modules can access it, e.g. custom ops
5116
    can access the vLLM config to determine how to dispatch.
5117
    """
5118
    global _current_vllm_config, _current_prefix
5119
    old_vllm_config = _current_vllm_config
5120
    old_prefix = _current_prefix
5121
5122
5123
5124
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
5125
        _current_prefix = prefix
5126
        yield
5127
5128
5129
    except Exception:
        raise
    else:
5130
5131
5132
5133
        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)
5134
5135
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
5136
5137
5138
5139
5140
5141
5142
5143
5144
            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"
5145
                " if you want it to be supported.",
5146
                vllm_config.model_config.model)
5147
    finally:
5148
        _current_vllm_config = old_vllm_config
5149
        _current_prefix = old_prefix
5150
5151
5152
5153
5154
5155
5156
5157
        # 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
5158
5159
5160
5161
5162
5163
5164


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.
5165
        logger.warning("Current vLLM config is not set.")
5166
5167
5168
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
5169
5170


5171
5172
5173
5174
5175
5176
5177
5178
5179
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


5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
def contains_object_print(text):
    """
    Check if the text looks like a printed Python object, e.g.
    contains any substring matching the pattern: "at 0xFFFFFFF>"
    We match against 0x followed by 2-16 hex chars (there's
    a max of 16 on a 64 bit system).

    Args:
        text (str): The text to check

    Returns:
5191
        result (bool): `True` if a match is found, `False` otherwise.
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
    """
    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}")
5205
5206
5207
5208
5209


T = TypeVar("T")


5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
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

5229
    return {
5230
5231
5232
        layer_name: forward_context[layer_name]
        for layer_name in layer_names
        if isinstance(forward_context[layer_name], layer_type)
5233
    }
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263


@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:
5264
        return self.min_energy_split_window_size is not None
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282


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)