config.py 214 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 contextlib import contextmanager
15
16
from dataclasses import (MISSING, Field, asdict, field, fields, is_dataclass,
                         replace)
17
from functools import cached_property
18
from importlib.util import find_spec
19
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Literal, Optional,
20
                    Protocol, TypeVar, Union, cast, get_args)
21

22
import regex as re
23
import torch
24
25
26
from pydantic import (ConfigDict, SkipValidation, TypeAdapter, field_validator,
                      model_validator)
from pydantic.dataclasses import dataclass
27
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
28
from torch.distributed import ProcessGroup, ReduceOp
29
from typing_extensions import Self, runtime_checkable
30

31
import vllm.envs as envs
32
from vllm import version
33
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
34
from vllm.logger import init_logger
35
from vllm.model_executor.layers.quantization import QuantizationMethods
36
from vllm.platforms import current_platform
37
38
39
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
40
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
41
42
    try_get_generation_config, try_get_safetensors_metadata,
    try_get_tokenizer_config, uses_mrope)
43
from vllm.transformers_utils.s3_utils import S3Model
44
from vllm.transformers_utils.utils import is_s3, maybe_model_redirect
45
46
# yapf conflicts with isort for this block
# yapf: disable
47
48
49
from vllm.utils import (DEFAULT_MAX_NUM_BATCHED_TOKENS,
                        MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                        POOLING_MODEL_MAX_NUM_BATCHED_TOKENS, GiB_bytes,
50
                        LayerBlockType, LazyLoader, common_broadcastable_dtype,
51
52
53
                        cuda_device_count_stateless, get_cpu_memory,
                        get_open_port, is_torch_equal_or_newer, random_uuid,
                        resolve_obj_by_qualname)
54

55
56
# yapf: enable

57
if TYPE_CHECKING:
58
    from _typeshed import DataclassInstance
59
    from ray.util.placement_group import PlacementGroup
60
    from transformers.configuration_utils import PretrainedConfig
61

62
63
    import vllm.model_executor.layers.quantization as me_quant
    import vllm.model_executor.models as me_models
64
    from vllm.executor.executor_base import ExecutorBase
65
    from vllm.model_executor.layers.quantization import QuantizationMethods
66
67
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
68
    from vllm.model_executor.model_loader import BaseModelLoader
69
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
70

71
    ConfigType = type[DataclassInstance]
72
    HfOverrides = Union[dict, Callable[[type], type]]
73
else:
74
    DataclassInstance = Any
75
    PlacementGroup = Any
76
    PretrainedConfig = Any
77
    ExecutorBase = Any
78
    QuantizationConfig = Any
79
    QuantizationMethods = Any
80
81
    BaseModelLoader = Any
    TensorizerConfig = Any
82
    ConfigType = type
83
84
85
86
87
88
    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")
89

90
logger = init_logger(__name__)
91
DataclassInstanceT = TypeVar("DataclassInstanceT", bound=DataclassInstance)
92
93
ConfigT = TypeVar("ConfigT", bound=ConfigType)

94
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
95
                     "score", "reward", "transcription", "draft"]
96

97
_ResolvedTask = Literal["generate", "transcription", "encode", "embed",
98
                        "classify", "reward", "draft"]
99

100
RunnerOption = Literal["auto", "generate", "pooling", "draft"]
101

102
RunnerType = Literal["generate", "pooling", "draft"]
103

104
105
_RUNNER_TASKS: dict[RunnerType, list[_ResolvedTask]] = {
    "generate": ["generate", "transcription"],
106
    "pooling": ["encode", "embed", "classify", "reward"],
107
    "draft": [],
108
}
109

110

111
@runtime_checkable
112
113
114
115
116
117
class SupportsHash(Protocol):

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


118
119
class SupportsMetricsInfo(Protocol):

120
    def metrics_info(self) -> dict[str, str]:
121
122
123
        ...


124
125
126
127
128
129
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"


130
131
132
133
134
135
136
137
138
139
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
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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
        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


182
def config(cls: ConfigT) -> ConfigT:
183
184
185
    """
    A decorator that ensures all fields in a dataclass have default values
    and that each field has a docstring.
186
187
188
189
190
191

    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.
192

193
194
195
    Config validation is performed by the tools/validate_config.py
    script, which is invoked during the pre-commit checks.
    """
196
197
198
    return cls


199
def get_field(cls: ConfigType, name: str) -> Field:
200
201
202
203
204
205
206
    """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__}.")
207
    named_field: Field = cls_fields[name]
208
209
210
211
212
213
214
215
    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.")


216
217
218
219
def is_init_field(cls: ConfigType, name: str) -> bool:
    return next(f for f in fields(cls) if f.name == name).init


220
221
TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
222
223
LogprobsMode = Literal["raw_logprobs", "raw_logits", "processed_logprobs",
                       "processed_logits"]
224
225
226


@config
227
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
228
class ModelConfig:
229
230
    """Configuration for the model."""

231
    model: str = "Qwen/Qwen3-0.6B"
232
233
234
    """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."""
235
236
237
238
239
240
241
242
    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."""
    task: TaskOption = "auto"
    """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."""
243
    tokenizer: SkipValidation[str] = None  # type: ignore
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
    """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
265
266
    """Random seed for reproducibility. Initialized to None in V0, but
    initialized to 0 in V1."""
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
    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)
282
    """RoPE scaling configuration. For example,
283
284
285
286
287
288
289
290
    `{"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."""
291
    max_model_len: SkipValidation[int] = None  # type: ignore
292
293
    """Model context length (prompt and output). If unspecified, will be
    automatically derived from the model config.
294

295
296
297
298
299
300
    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
301
    """Specify the maximum length for spec decoding draft models."""
302
    quantization: SkipValidation[Optional[QuantizationMethods]] = None
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
    """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
    OpenAI Chat Completions API."""
321
322
323
324
325
326
327
    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.
    """
328
329
330
331
332
333
334
335
336
337
338
339
340
341
    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."""
342
343
344
345
    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."""
346
347
348
349
350
351
352
353
354
355
356
    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."""
357
    interleave_mm_strings: bool = False
358
    """Enable fully interleaved support for multimodal prompts, while using
359
    --chat-template-content-format=string. Defaults to False."""
360
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
361
362
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
363
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
364
365
366
367
368
369
370
371
372
373
374
375
376
377
    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
378
    config. If a callable, it is called to update the HuggingFace config."""
379
380
381
382
383
    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}`.
384
    """
385
386
387
388
389
390
391
    disable_mm_preprocessor_cache: bool = False
    """If `True`, disable caching of the multi-modal preprocessor/mapper (not
    recommended)."""
    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
392
    arguments. e.g. `{"cast_logits_dtype": "bfloat16"}`."""
393
394
395
396
397
398
    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}`.
399
    """
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
    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
415
    `--generation-config vllm`, only the override parameters are used."""
416
417
418
419
420
421
422
423
424
    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."""
425
426
    override_attention_dtype: Optional[str] = None
    """Override dtype for attention"""
427

428
429
430
431
432
433
434
435
436
437
438
439
    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.
        """
440
        factors: list[Any] = []
441
442
443
444
445
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
446
447
448
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
449
        factors.append(self.trust_remote_code)
450
451
452
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
453
454
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
455
456
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
457
458
        str_factors = str(factors)
        assert_hashable(str_factors)
459
460
        return hashlib.sha256(str(factors).encode()).hexdigest()

461
    def __post_init__(self) -> None:
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
        # 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)

480
481
482
        # 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)
483
484
485
486
487
488
489
490
491
492
493
494
        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):
495
            hf_overrides_kw = {}
496
            hf_overrides_fn = self.hf_overrides
497
        else:
498
            hf_overrides_kw = self.hf_overrides
499
            hf_overrides_fn = None
500

501
502
        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
503
            hf_overrides_kw.update(hf_override)
504
            hf_overrides_str = json.dumps(hf_overrides_kw)
505
506
507
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
508
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
509
510
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
511
            hf_overrides_kw.update(hf_override)
512
            hf_overrides_str = json.dumps(hf_overrides_kw)
513
514
515
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
516
517
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

518
        self.maybe_pull_model_tokenizer_for_s3(self.model, self.tokenizer)
519

520
521
522
523
        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 "
524
525
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
526
527
                "for instructions on how to install it.")

528
529
        from vllm.platforms import current_platform

530
531
532
533
534
535
        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)

536
537
538
539
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
540

541
542
543
        if isinstance(self.config_format, str):
            self.config_format = ConfigFormat(self.config_format)

544
        hf_config = get_config(self.hf_config_path or self.model,
545
546
547
548
549
550
                               self.trust_remote_code,
                               self.revision,
                               self.code_revision,
                               self.config_format,
                               hf_overrides_kw=hf_overrides_kw,
                               hf_overrides_fn=hf_overrides_fn)
551
552
        self.hf_config = hf_config

553
        self.hf_text_config = get_hf_text_config(self.hf_config)
554
555
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
556
        self.encoder_config = self._get_encoder_config()
557
        self.hf_image_processor_config = get_hf_image_processor_config(
558
            self.model, hf_token=self.hf_token, revision=self.revision)
559

560
561
562
        # For pooling models, self.task is used to indicate the
        # user-selected task
        if self.task == "score":
563
            if self._is_classify_task(self.architectures):
564
565
566
567
568
569
570
571
572
573
                self.task = "classify"
            else:
                self.task = "embed"
        elif self.task == "embedding":
            msg = ("The 'embedding' task has been renamed to 'embed', please "
                   "use the new name. The old name will be removed in v1.0.")
            warnings.warn(msg, DeprecationWarning, stacklevel=2)

            self.task = "embed"

574
575
576
577
        model_info, arch = self.registry.inspect_model_cls(self.architectures)
        self._model_info = model_info
        self._architecture = arch

578
579
580
581
582
583
584
585
586
587
588
589
590
        all_supported_tasks = self._get_supported_tasks(self.task)
        logger.debug("Tasks supported by runner type: %s", all_supported_tasks)
        supported_runner_types = self._get_supported_runner_types(
            all_supported_tasks)
        runner_type = self._resolve_runner(self.runner, self.task,
                                           supported_runner_types,
                                           all_supported_tasks)

        logger.debug("Selected runner type: %s", runner_type)
        # For pooling models, self.task is used to indicate the
        # user-selected task
        if runner_type == "pooling" and self.task == "auto":
            selected_task = all_supported_tasks[runner_type][-1]
591
            assert selected_task != "encode"
592
593
594
595
596
597
598
            self.task = selected_task
        self.supported_runner_types = supported_runner_types
        self.runner_type = runner_type
        self.supported_tasks = all_supported_tasks[runner_type]

        if self.runner_type in ("draft",
                                "generate") and self.task != "transcription":
599
600
601
602
603
604
605
606
607
608
609
610
611
            self.truncation_side = "left"
        else:
            self.truncation_side = "right"

        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,
        )
612

613
614
615
616
617
618
        # Workaround for Gemma 2 which uses interleaved sliding window
        # attention, but it's not specified in its config. TODO: remove this
        # when Gemma 2 is fixed in Transformers.
        if self.hf_text_config.model_type == "gemma2":
            self.hf_text_config.sliding_window_pattern = 2

619
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
620
621
        sliding_window_pattern = getattr(self.hf_text_config,
                                         "sliding_window_pattern", None)
622
623
        has_interleaved_attention = sliding_window_pattern is not None or (
            isinstance(sliding_window, list))
624

625
        if not self.disable_sliding_window and has_interleaved_attention:
626
627
            if (backend :=
                    envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"):
628
629
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
630

631
                logger.warning_once(
632
633
634
635
636
                    "%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,
                )
637
638
639
640
641
642
643
644
                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
645
646
647
648

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

649
                sliding_window = None
Woosuk Kwon's avatar
Woosuk Kwon committed
650

651
        self.original_max_model_len = self.max_model_len
652
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
653
        self.multimodal_config = self._init_multimodal_config()
654
655
        self.model_supports_multimodal_raw_input = (
            self.registry.supports_multimodal_raw_input(self.architectures))
656
657
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
658

659
        self.is_attention_free = self._init_attention_free()
660
        self.is_hybrid = self._init_is_hybrid()
661
        self.has_noops = self._init_has_noops()
662
663
        self.has_inner_state = self._init_has_inner_state()

664
665
666
        if (not current_platform.is_neuron() and self.override_neuron_config):
            raise ValueError(
                "`override_neuron_config` is only supported on Neuron.")
667

668
        self._verify_quantization()
669
        self._verify_cuda_graph()
670
        self._verify_bnb_config()
671

672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
    @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

688
689
690
691
692
693
694
695
696
697
    def _get_transformers_backend_cls(self) -> str:
        """Determine which Transformers backend class will be used if
        `model_impl` is set to `transformers` or `auto`."""
        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"
        else:
            return "TransformersForCausalLM"

698
699
    @property
    def registry(self):
700
        return me_models.ModelRegistry
701
702
703

    @property
    def architectures(self) -> list[str]:
704
        # architectures in the model config.
705
706
707
708
709
710
711
712
713
714
715
716
717
        architectures = getattr(self.hf_config, "architectures", [])
        # The registry assumes that it can always inspect the vLLM model class
        # for a given architecture. This assumption breaks down for the
        # Transformers backend, which may use a different class depending on
        # the model type. To work around this, we add the correct Transformers
        # backend class to the architectures list. We must do this here because
        # we need access to the `hf_config` to determine the backend class.
        transformers_backend_cls = self._get_transformers_backend_cls()
        if (self.model_impl != ModelImpl.VLLM.value
                and all(arch != transformers_backend_cls
                        for arch in architectures)):
            architectures.append(transformers_backend_cls)
        return architectures
718

719
720
721
722
723
724
    @property
    def architecture(self) -> str:
        # The architecture vllm actually used.
        return self._architecture

    @property
725
    def model_info(self):
726
727
        return self._model_info

728
729
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
730
        """Pull model/tokenizer from S3 to temporary directory when needed.
731

732
        Args:
733
734
            model: Model name or path
            tokenizer: Tokenizer name or path
735
        """
736
737
738
739
740
741
742
743
744
745
746
747
        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:
748
749
750
751
752
                s3_model.pull_files(model,
                                    ignore_pattern=[
                                        "*.pt", "*.safetensors", "*.bin",
                                        "*.tensors"
                                    ])
753
754
755
756
757
758
759
                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(
760
761
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors"])
762
            self.tokenizer = s3_tokenizer.dir
763

764
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
765
        if self.registry.is_multimodal_model(self.architectures):
766
            return MultiModalConfig(
767
                limit_per_prompt=self.limit_mm_per_prompt,
768
                media_io_kwargs=self.media_io_kwargs,
769
770
                mm_processor_kwargs=self.mm_processor_kwargs,
                disable_mm_preprocessor_cache=self.
771
772
                disable_mm_preprocessor_cache,
                interleave_mm_strings=self.interleave_mm_strings)
773

774
        if self.limit_mm_per_prompt:
775
776
            raise ValueError("`limit_mm_per_prompt` is only supported for "
                             "multimodal models.")
777
        if self.mm_processor_kwargs:
778
779
            raise ValueError("`mm_processor_kwargs` is only supported for "
                             "multimodal models.")
780
        if self.disable_mm_preprocessor_cache:
781
782
            raise ValueError("`disable_mm_preprocessor_cache` is only "
                             "supported for multimodal models.")
783
784
785
        if self.interleave_mm_strings:
            raise ValueError("`interleave_mm_strings` is only "
                             "supported for multimodal models.")
786
787

        return None
788

789
790
791
792
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

793
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
794
        if self.runner_type == "pooling":
795
796
797
798
799
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
800
801
802
803
804

            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():
805
806
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
807

808
            if self.is_matryoshka:
809
810
811
                if pooler_config.normalize is None:
                    pooler_config.normalize = True
                elif not pooler_config.normalize:
812
813
814
815
816
                    raise ValueError(
                        "`normalize` must be enabled (set to True) "
                        "for models that are compatible with "
                        "Matryoshka Representation.")

817
            return pooler_config
818

819
820
        return None

821
    def _init_attention_free(self) -> bool:
822
        return self.registry.is_attention_free_model(self.architectures)
823

824
    def _init_is_hybrid(self) -> bool:
825
        return self.registry.is_hybrid_model(self.architectures)
826

827
828
829
830
    def _init_has_noops(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return self.registry.is_noops_model(architectures)

831
    def _init_has_inner_state(self) -> bool:
832
        return self.registry.model_has_inner_state(self.architectures)
833

834
    def _verify_tokenizer_mode(self) -> None:
835
836
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
837
838
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
839
                f"one of {get_args(TokenizerMode)}.")
840
        self.tokenizer_mode = tokenizer_mode
841

842
843
844
845
846
847
    def _is_classify_task(self, architectures: list[str]):
        for arch in architectures:
            if arch.endswith("ForSequenceClassification"):
                return True
        return self.registry.is_cross_encoder_model(architectures)

848
    def _get_preferred_pooling_task(
849
        self,
850
        architectures: list[str],
851
    ) -> _ResolvedTask:
852
853
854
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
855
        if self.registry.is_transcription_model(architectures):
856
            return "transcription"
857

858
        suffix_to_preferred_task: list[tuple[str, _ResolvedTask]] = [
859
860
861
862
863
864
            # Other models follow this pattern
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]

        for suffix, pref_task in suffix_to_preferred_task:
865
            if self.architecture.endswith(suffix):
866
867
                return pref_task

868
        return "embed"
869

870
    def _get_supported_generation_tasks(
871
        self,
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
        task_option: TaskOption,
    ) -> list[_ResolvedTask]:
        registry = self.registry
        architectures = self.architectures

        if registry.is_transcription_only_model(architectures):
            return ["transcription"]

        supported_tasks = list[_ResolvedTask]()
        if registry.is_text_generation_model(architectures):
            supported_tasks.append("generate")

            if registry.is_transcription_model(architectures):
                supported_tasks.append("transcription")

        return supported_tasks
888

889
890
891
892
    def _get_supported_pooling_tasks(
        self,
        task_option: TaskOption,
    ) -> list[_ResolvedTask]:
893
894
        registry = self.registry
        architectures = self.architectures
895

896
897
        supported_tasks = list[_ResolvedTask]()
        if registry.is_pooling_model(architectures):
898
            supported_tasks.append("encode")
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915

            # For now, users must specify the task (other than "pooling")
            # to use for pooling models
            if task_option == "auto":
                preferred_task = self._get_preferred_pooling_task(
                    architectures)

                supported_tasks.append(preferred_task)
            elif task_option in _RUNNER_TASKS["pooling"]:
                supported_tasks.append(cast(_ResolvedTask, task_option))

        return supported_tasks

    def _get_supported_tasks(
        self,
        task_option: TaskOption,
    ) -> dict[RunnerType, list[_ResolvedTask]]:
916
917
918
919
920
921
922
923
        if self._is_classify_task(self.architectures):
            return {"generate": [], "pooling": ["classify"], "draft": []}
        else:
            return {
                "generate": self._get_supported_generation_tasks(task_option),
                "pooling": self._get_supported_pooling_tasks(task_option),
                "draft": ["draft"]
            }
924

925
926
927
928
929
930
931
932
933
    def _get_supported_runner_types(
        self,
        supported_tasks: dict[RunnerType, list[_ResolvedTask]],
    ) -> set[RunnerType]:
        return {
            runner
            for runner, runner_tasks in supported_tasks.items()
            if len(runner_tasks) > 0
        }
934

935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
    def _resolve_runner(
        self,
        runner_option: RunnerOption,
        task_option: TaskOption,
        supported_runner_types: set[RunnerType],
        supported_tasks: dict[RunnerType, list[_ResolvedTask]],
    ) -> RunnerType:
        if not supported_runner_types:
            raise ValueError("This model does not support any model runners!")

        if runner_option != "auto":
            if runner_option not in supported_runner_types:
                raise ValueError(
                    f"This model does not support runner={runner_option!r}. "
                    f"Available runners: {supported_runner_types}")
950

951
            return runner_option
952

953
954
955
956
        if task_option != "auto":
            for runner, runner_tasks in supported_tasks.items():
                if task_option in runner_tasks:
                    return runner
957
            else:
958
959
960
961
962
963
964
                task_runner: RunnerType = next(
                    runner for runner, tasks in _RUNNER_TASKS.items()
                    if task_option in tasks)
                raise ValueError(
                    f"This model does not support task={task_option!r}. "
                    f"Available tasks for runner={task_runner!r}: "
                    f"{supported_tasks[task_runner]}")
965

966
967
968
969
970
        if "classify" in supported_tasks.get("pooling", []):
            # When multiple pooling tasks are present, default to
            # pooling (eg cross-encoder) for non-standard architectures.
            return "pooling"

971
972
973
974
975
976
977
978
        suffix_to_preferred_runner: list[tuple[str, RunnerType]] = [
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("EmbeddingModel", "pooling"),
            ("RewardModel", "pooling"),
        ]
979

980
        for suffix, pref_runner in suffix_to_preferred_runner:
981
982
            if self.architecture.endswith(
                    suffix) and pref_runner in supported_runner_types:
983
                return pref_runner
984

985
986
987
988
        if "generate" in supported_runner_types:
            return "generate"
        if "pooling" in supported_runner_types:
            return "pooling"
989

990
        raise AssertionError("This line should not be reached")
991

992
993
994
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
995
            # compressed-tensors uses a "compression_config" key
996
            quant_cfg = getattr(self.hf_config, "compression_config", None)
997
998
        return quant_cfg

999
    def _verify_quantization(self) -> None:
1000
        supported_quantization = me_quant.QUANTIZATION_METHODS
1001
        optimized_quantization_methods = [
1002
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
1003
            "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
1004
            "quark", "modelopt_fp4", "bitblas", "gptq_bitblas", "inc"
1005
        ]
1006
        if self.quantization is not None:
1007
1008
            self.quantization = cast(me_quant.QuantizationMethods,
                                     self.quantization)
1009
1010

        # Parse quantization method from the HF model config, if available.
1011
1012
        quant_cfg = self._parse_quant_hf_config()

1013
        if quant_cfg is not None:
1014
            # Use the community standard 'quant_method'
1015
            quant_method = quant_cfg.get("quant_method", "").lower()
1016
1017

            # Normalize library names
1018
1019
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
1020

1021
            quant_cfg["quant_method"] = quant_method
1022

1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
            # 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",
1035
1036
                "modelopt",
                "modelopt_fp4",
1037
1038
1039
1040
1041
1042
1043
1044
1045
            ]
            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

1046
            # Detect which checkpoint is it
1047
            for name in quantization_methods:
1048
                method = me_quant.get_quantization_config(name)
1049
1050
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
1051
1052
1053
1054
                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.
1055
                    if (name in get_args(me_quant.QuantizationMethods)
1056
1057
1058
1059
1060
1061
                            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.")
1062
1063
1064
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
1065

1066
            # Verify quantization configurations.
1067
            if self.quantization is None:
1068
1069
                self.quantization = quant_method
            elif self.quantization != quant_method:
1070
1071
                raise ValueError(
                    "Quantization method specified in the model config "
1072
                    f"({quant_method}) does not match the quantization "
1073
1074
1075
1076
1077
1078
1079
1080
                    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}.")
1081
            from vllm.platforms import current_platform
1082
            current_platform.verify_quantization(self.quantization)
1083
            if self.quantization not in optimized_quantization_methods:
1084
                logger.warning(
1085
                    "%s quantization is not fully "
1086
                    "optimized yet. The speed can be slower than "
1087
                    "non-quantized models.", self.quantization)
1088

1089
    def _verify_cuda_graph(self) -> None:
1090
1091
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
1092
        # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
1093
        ROCM_UNSUPPORTED_MODELS = ['mllama']
1094
1095
1096
1097
1098
1099
        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()):
1100
1101
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1102
                "to eager mode.", self.hf_config.model_type)
1103
            self.enforce_eager = True
1104

1105
1106
    def _verify_bnb_config(self) -> None:
        """
1107
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
1108
        yet support CUDA graph.
1109
        # TODO Remove this when bitsandbytes supports.
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
        """
        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(
1124
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1125
                "fallback to the eager mode.")
1126

1127
1128
            self.enforce_eager = True

1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
    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.")

1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
    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

1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
    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

1173
        # Reminder: Please update docs/features/compatibility_matrix.md
1174
        # If the feature combo become valid
1175
        from vllm.platforms import current_platform
1176
        if not current_platform.is_async_output_supported(self.enforce_eager):
1177
1178
1179
1180
1181
1182
1183
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

1184
        # Async postprocessor is not necessary for pooling models
1185
        # since there is no token generation
1186
        if self.runner_type == "pooling":
1187
1188
            self.use_async_output_proc = False

1189
        # Reminder: Please update docs/features/compatibility_matrix.md
1190
        # If the feature combo become valid
1191
1192
1193
        if speculative_config:
            self.use_async_output_proc = False

1194
1195
1196
1197
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
1198
1199
1200
1201
1202
1203

        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.")

1204
1205
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
1206
1207
1208
1209
1210
1211
1212
        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}).")

1213
        if parallel_config.enable_expert_parallel:
1214
1215
            self._verify_with_expert_parallelism()

1216
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1217
        if pipeline_parallel_size > 1:
1218
            if not self.registry.is_pp_supported_model(self.architectures):
1219
1220
1221
1222
1223
1224
                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
1225

1226
1227
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
1228
        """Get the sliding window size, or None if disabled."""
1229
1230
1231
1232

        # 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.
1233
1234
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
1235
            return None
1236
        return getattr(self.hf_text_config, "sliding_window", None)
1237

1238
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
1239
1240
1241
1242
1243
1244
1245
1246
        """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()

1247
    def get_vocab_size(self) -> int:
1248
        return getattr(self.hf_text_config, "vocab_size", 0)
1249

1250
    def get_hidden_size(self) -> int:
1251
        return getattr(self.hf_text_config, "hidden_size", 0)
1252

1253
1254
    @property
    def is_deepseek_mla(self) -> bool:
1255
1256
1257
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
bigmoyan's avatar
bigmoyan committed
1258
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp', 'kimi_k2'):
1259
1260
1261
1262
1263
1264
1265
1266
            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
1267

1268
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1269
        # TODO remove hard code
1270
        if self.is_deepseek_mla:
1271
1272
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1273
            if self.use_mla:
1274
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1275
1276
1277
1278
1279
            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
1280

1281
1282
1283
1284
1285
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1286
1287
1288
        if self.is_attention_free:
            return 0

1289
1290
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1291
            return self.hf_text_config.head_dim
1292

1293
        # FIXME(woosuk): This may not be true for all models.
1294
1295
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1296

1297
1298
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1299
        # For GPTBigCode & Falcon:
1300
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1301
1302
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1303
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1304
        new_decoder_arch_falcon = (
1305
            self.hf_config.model_type in falcon_model_types
1306
            and getattr(self.hf_config, "new_decoder_architecture", False))
1307
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1308
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1309
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1310
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1311
            return 1
1312

1313
        # For DBRX and MPT
1314
1315
1316
1317
1318
        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":
1319
1320
1321
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1322
1323
1324
1325
1326
1327
1328
1329
        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")

1330
1331
1332
        if self.is_attention_free:
            return 0

1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
        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:
1343
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1344
1345
1346
1347
1348
            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.
1349
        return self.hf_text_config.num_attention_heads
1350
1351
1352

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1353
1354
1355
1356
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1357
1358
1359
1360
1361
1362
1363
        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)
1364

1365
1366
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1367
1368
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1369

1370
    def get_layers_start_end_indices(
1371
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1372
        from vllm.distributed.utils import get_pp_indices
1373
        if (self.hf_text_config.model_type == "deepseek_mtp"
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1374
1375
                or self.hf_config.model_type == "mimo_mtp"
                or self.hf_config.model_type == "glm4_moe_mtp"):
1376
1377
1378
1379
1380
            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)
1381
1382
1383
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1384
1385
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1386
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1387

1388
1389
1390
    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
1391

1392
1393
1394
1395
1396
1397
1398
1399
    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
1400
1401
1402
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
        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
1413
1414
1415
1416
        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])
1417
        else:
1418
            # Hybrid model Jamba
1419
1420
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
            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
1446

1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
    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

1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
    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

1470
    def try_get_generation_config(self) -> dict[str, Any]:
1471
        if self.generation_config in ("auto", "vllm"):
1472
            config = try_get_generation_config(
1473
                self.hf_config_path or self.model,
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
                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()

1488
    def get_diff_sampling_param(self) -> dict[str, Any]:
1489
        """
1490
        This method returns a dictionary containing the parameters
1491
1492
        that differ from the default sampling parameters. If
        `generation_config` is `"vllm"`, an empty dictionary is returned.
1493
1494

        Returns:
1495
            dict[str, Any]: A dictionary with the differing sampling
1496
            parameters, if `generation_config` is `"vllm"` an empty dictionary.
1497
        """
1498
        if self.generation_config == "vllm":
1499
1500
1501
1502
1503
1504
1505
            config = {}
        else:
            config = self.try_get_generation_config()

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

1506
1507
1508
1509
1510
1511
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1512
            "max_new_tokens",
1513
1514
1515
1516
1517
1518
        ]
        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
            }
1519
1520
1521
1522
1523
            # 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")
1524
1525
        else:
            diff_sampling_param = {}
1526
1527
1528
1529
1530
1531
1532

        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`.")
1533
1534
        return diff_sampling_param

1535
    @property
1536
    def is_encoder_decoder(self) -> bool:
1537
        """Extract the HF encoder/decoder model flag."""
1538
        """
1539
        For Mllama, VLLM overrides HF's is_encoder_decoder flag and sets it to
1540
        True to enable cross-attention
1541
        Neuron needs all multimodal data to be in the decoder and does not
1542
1543
1544
1545
1546
1547
        need to explicitly enable cross-attention
        """
        if (current_platform.is_neuron()
                and self.hf_config.model_type == "mllama"):
            return False

1548
1549
1550
1551
1552
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1553

1554
1555
1556
1557
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1558
1559
    @property
    def is_cross_encoder(self) -> bool:
1560
        return self.task == "classify"
1561

1562
1563
    @property
    def use_mla(self) -> bool:
1564
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1565

1566
1567
1568
    @property
    def is_v1_compatible(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
1569
        return me_models.ModelRegistry.is_v1_compatible(architectures)
1570

1571
1572
    @property
    def is_matryoshka(self) -> bool:
1573
        return (bool(getattr(self.hf_config, "matryoshka_dimensions", None))
1574
1575
                or getattr(self.hf_config, "is_matryoshka", False))

1576
1577
1578
1579
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)

1580
1581
1582
1583
1584
1585
    @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)

1586
    def get_and_verify_max_len(self, max_model_len: int):
1587
1588
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
1589
        tokenizer_config = None
1590
1591
        if (self.runner_type == "pooling" and getattr(
                self.hf_config, "position_embedding_type", "") == "absolute"):
1592
1593
1594
1595
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
                revision=self.tokenizer_revision)
1596
1597
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
1598
            tokenizer_config=tokenizer_config,
1599
1600
1601
1602
1603
            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)
1604
        logger.info("Using max model len %s", max_model_len)
1605
1606
        return max_model_len

1607

1608
BlockSize = Literal[1, 8, 16, 32, 64, 128]
1609
CacheDType = Literal["auto", "fp8", "fp8_e4m3", "fp8_e5m2", "fp8_inc"]
1610
PrefixCachingHashAlgo = Literal["builtin", "sha256", "sha256_cbor_64bit"]
1611
1612
1613
1614


@config
@dataclass
1615
class CacheConfig:
1616
    """Configuration for the KV cache."""
1617

1618
    block_size: SkipValidation[BlockSize] = None  # type: ignore
1619
1620
1621
    """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.
1622
1623

    This config has no static default. If left unspecified by the user, it will
1624
    be set in `Platform.check_and_update_config()` based on the current
1625
    platform."""
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
    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
1639
    fp8 (=fp8_e4m3). Intel Gaudi (HPU) supports fp8 (using fp8_inc)."""
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
    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
1655
1656
1657
1658
1659
1660
    - "sha256" is collision resistant but with certain overheads.
    This option uses Pickle for object serialization before hashing.\n
    - "sha256_cbor_64bit" provides a reproducible, cross-language compatible 
    hash. It serializes objects using canonical CBOR and hashes them with 
    SHA-256. The resulting hash consists of the lower 64 bits of the SHA-256
    digest."""
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
    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."""
1674
1675
    cpu_kvcache_space_bytes: Optional[int] = None
    """(CPU backend only) CPU key-value cache space."""
1676
1677
1678
    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."""
1679
1680
1681
1682
1683
1684

    # 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."""
1685

1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
    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.
        """
1698
        factors: list[Any] = []
1699
1700
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
1701
1702
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1703
1704
        return hash_str

1705
1706
1707
    def __post_init__(self) -> None:
        self.swap_space_bytes = self.swap_space * GiB_bytes

1708
        self._verify_cache_dtype()
1709
        self._verify_prefix_caching()
1710

1711
    def metrics_info(self):
1712
1713
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1714
1715
        return {key: str(value) for key, value in self.__dict__.items()}

1716
1717
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
1718
1719
1720
1721
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1722
1723
1724
1725
1726
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1727
1728
        return self

1729
1730
1731
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1732
        elif self.cache_dtype in get_args(CacheDType):
1733
            logger.info(
1734
1735
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1736
                "Meanwhile, it may cause accuracy drop without a proper "
1737
                "scaling factor.")
1738
1739
1740
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1741
1742
1743
1744
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

1745
        if self.sliding_window is not None and not envs.VLLM_USE_V1:
1746
1747
1748
1749
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

1750
1751
        if (self.enable_prefix_caching and self.prefix_caching_hash_algo
                not in get_args(PrefixCachingHashAlgo)):
1752
1753
            raise ValueError(
                "Unknown prefix caching hash algorithm: "
1754
1755
                f"{self.prefix_caching_hash_algo}. Must be one of "
                f"{get_args(PrefixCachingHashAlgo)}.")
1756

1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
    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

1767
1768
1769
        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.")
1770
1771
1772
        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:
1773
            logger.warning("Possibly too large swap space. %s", msg)
1774

1775

1776
1777
1778
1779
1780
1781
1782
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
1783
    SHARDED_STATE = "sharded_state"
1784
    GGUF = "gguf"
1785
    BITSANDBYTES = "bitsandbytes"
1786
    MISTRAL = "mistral"
1787
    RUNAI_STREAMER = "runai_streamer"
1788
    RUNAI_STREAMER_SHARDED = "runai_streamer_sharded"
1789
    FASTSAFETENSORS = "fastsafetensors"
1790
1791


1792
@config
1793
1794
@dataclass
class LoadConfig:
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
    """Configuration for loading the model weights."""

    load_format: Union[str, LoadFormat,
                       "BaseModelLoader"] = LoadFormat.AUTO.value
    """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
    Mistral models."""
1820
    download_dir: Optional[str] = None
1821
1822
    """Directory to download and load the weights, default to the default
    cache directory of Hugging Face."""
1823
1824
    model_loader_extra_config: Union[dict, TensorizerConfig] = field(
        default_factory=dict)
1825
    """Extra config for model loader. This will be passed to the model loader
1826
    corresponding to the chosen load_format."""
1827
1828
1829
    device: Optional[str] = None
    """Device to which model weights will be loaded, default to
    device_config.device"""
1830
    ignore_patterns: Optional[Union[list[str], str]] = None
1831
1832
    """The list of patterns to ignore when loading the model. Default to
    "original/**/*" to avoid repeated loading of llama's checkpoints."""
1833
    use_tqdm_on_load: bool = True
1834
1835
    """Whether to enable tqdm for showing progress bar when loading model
    weights."""
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
    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
    """
1846

1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
    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.
1861
        factors: list[Any] = []
1862
1863
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1864
1865
        return hash_str

1866
    def __post_init__(self):
1867
1868
1869
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1870

1871
1872
1873
1874
1875
1876
1877
        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/**/*"]

1878

1879
1880
1881
DistributedExecutorBackend = Literal["ray", "mp", "uni", "external_launcher"]


1882
@config
1883
@dataclass
1884
class ParallelConfig:
1885
    """Configuration for the distributed execution."""
1886

1887
1888
1889
1890
1891
1892
1893
    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."""
1894
1895
    data_parallel_size_local: int = 1
    """Number of local data parallel groups."""
1896
1897
    data_parallel_rank: int = 0
    """Rank of the data parallel group."""
1898
1899
1900
    data_parallel_rank_local: Optional[int] = None
    """Local rank of the data parallel group,
    set only in SPMD mode."""
1901
    data_parallel_master_ip: str = "127.0.0.1"
1902
    """IP of the data parallel master."""
1903
1904
    data_parallel_rpc_port: int = 29550
    """Port for data parallel messaging."""
1905
1906
    data_parallel_master_port: int = 29500
    """Port of the data parallel master."""
Rui Qiao's avatar
Rui Qiao committed
1907
1908
    data_parallel_backend: str = "mp"
    """Backend to use for data parallel, either "mp" or "ray"."""
1909
1910
1911
1912
    data_parallel_external_lb: bool = False
    """Whether to use "external" DP LB mode. Applies only to online serving
    and when data_parallel_size > 0. Set implicitly when
    data_parallel_rank is provided explicitly to vllm serve."""
1913
1914
    enable_expert_parallel: bool = False
    """Use expert parallelism instead of tensor parallelism for MoE layers."""
1915
1916
1917
1918
1919
1920
1921
1922
1923
    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.
1924

1925
1926
1927
1928
1929
1930
1931
1932
1933
    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.
    """

1934
    max_parallel_loading_workers: Optional[int] = None
1935
    """Maximum number of parallel loading workers when loading model
1936
1937
    sequentially in multiple batches. To avoid RAM OOM when using tensor
    parallel and large models."""
1938
1939

    disable_custom_all_reduce: bool = False
1940
    """Disable the custom all-reduce kernel and fall back to NCCL."""
1941
1942

    ray_workers_use_nsight: bool = False
1943
    """Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler."""
1944
1945

    placement_group: Optional["PlacementGroup"] = None
1946
    """ray distributed model workers placement group."""
1947

1948
    distributed_executor_backend: Optional[Union[DistributedExecutorBackend,
1949
                                                 type["ExecutorBase"]]] = None
1950
1951
1952
1953
1954
1955
    """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
1956
    only support Ray for distributed inference."""
1957
1958

    worker_cls: str = "auto"
1959
1960
    """The full name of the worker class to use. If "auto", the worker class
    will be determined based on the platform."""
1961
    sd_worker_cls: str = "auto"
Ning Xie's avatar
Ning Xie committed
1962
    """The full name of the worker class to use for speculative decoding.
1963
    If "auto", the worker class will be determined based on the platform."""
1964
    worker_extension_cls: str = ""
1965
1966
1967
1968
    """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."""
1969
1970

    world_size: int = field(init=False)
1971
    """world_size is TPxPP, it affects the number of workers we create."""
1972
1973

    rank: int = 0
1974
    """Global rank in distributed setup."""
1975

1976
    enable_multimodal_encoder_data_parallel: bool = False
1977
    """ Use data parallelism instead of tensor parallelism for vision encoder.
1978
1979
    Only support LLama4 for now"""

1980
1981
1982
1983
1984
1985
    @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

1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
    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":
2000
2001
2002
2003
2004
2005
2006
2007
2008
        # 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

2009
2010
2011
        from vllm.distributed.utils import (
            stateless_init_torch_distributed_process_group)

2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
        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
2035
2036
2037

    @staticmethod
    def has_unfinished_dp(dp_group: "ProcessGroup",
youkaichao's avatar
youkaichao committed
2038
                          has_unfinished: bool) -> bool:
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
        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

2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
    @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()

2063
2064
2065
2066
2067
2068
2069
2070
    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.
        """
2071
        factors: list[Any] = []
2072
2073
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
2074
        factors.append(self.enable_expert_parallel)
2075
2076
        factors.append(self.data_parallel_size)
        factors.append(envs.VLLM_ALL2ALL_BACKEND)
2077
2078
        return hashlib.sha256(str(factors).encode()).hexdigest()

2079
2080
2081
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
2082

2083
2084
2085
2086
2087
2088
        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:
2089
2090
            # Data parallel was specified in the engine args.
            self.data_parallel_master_port = get_open_port()
2091
2092
2093
2094
2095

            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})")
2096
2097
2098
2099
2100
2101
2102
2103
        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

2104
2105
2106
2107
            if self.data_parallel_external_lb:
                raise ValueError("data_parallel_external_lb can only "
                                 "be set when data_parallel_size > 1")

2108
2109
2110
2111
2112
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

2113
2114
2115
2116
2117
2118
2119
2120
2121
        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}.")
2122
2123
2124
2125
2126
2127
2128
2129
2130
            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}."
                )
2131
2132
2133
2134
2135
        else:
            if self.num_redundant_experts != 0:
                raise ValueError(
                    "num_redundant_experts should be used with EPLB."
                    f"{self.num_redundant_experts}.")
2136
        if self.distributed_executor_backend is None and self.world_size > 1:
2137
2138
2139
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

2140
            from vllm.executor import ray_utils
2141
            backend: DistributedExecutorBackend = "mp"
2142
            ray_found = ray_utils.ray_is_available()
2143
2144
            if current_platform.is_neuron():
                # neuron uses single process to control multiple devices
2145
2146
                backend = "uni"
            elif current_platform.is_tpu() and envs.VLLM_XLA_USE_SPMD:
2147
2148
2149
                backend = "uni"
            elif (current_platform.is_cuda()
                  and cuda_device_count_stateless() < self.world_size):
2150
                if not ray_found:
2151
2152
                    raise ValueError("Unable to load Ray: "
                                     f"{ray_utils.ray_import_err}. Ray is "
2153
2154
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
2155
                                     "ray`.")
2156
                backend = "ray"
Rui Qiao's avatar
Rui Qiao committed
2157
2158
2159
2160
            elif self.data_parallel_backend == "ray":
                logger.info("Using ray distributed inference because "
                            "data_parallel_backend is ray")
                backend = "ray"
2161
            elif ray_found:
2162
                if self.placement_group:
2163
                    backend = "ray"
2164
2165
2166
2167
2168
2169
                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"
2170
            self.distributed_executor_backend = backend
2171
2172
            logger.debug("Defaulting to use %s for distributed inference",
                         backend)
2173

2174
2175
2176
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

2177
2178
2179
2180
2181
2182
    @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)

2183
2184
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2185
2186
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
2187
        from vllm.platforms import current_platform
2188
        if self.distributed_executor_backend not in (
2189
2190
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
2191
2192
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
2193
            raise ValueError(
2194
2195
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
2196
2197
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
2198
        if self.use_ray:
2199
2200
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
2201
2202

        if not current_platform.use_custom_allreduce():
2203
            self.disable_custom_all_reduce = True
Aaron Pham's avatar
Aaron Pham committed
2204
            logger.debug(
2205
                "Disabled the custom all-reduce kernel because it is not "
2206
                "supported on current platform.")
2207
        if self.ray_workers_use_nsight and not self.use_ray:
2208
2209
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
2210

2211
        return self
2212

2213

2214
PreemptionMode = Literal["swap", "recompute"]
2215
2216
2217
2218
SchedulerPolicy = Literal["fcfs", "priority"]


@config
2219
@dataclass
2220
class SchedulerConfig:
2221
    """Scheduler configuration."""
2222

2223
2224
    runner_type: RunnerType = "generate"
    """The runner type to launch for the model."""
2225

2226
    max_num_batched_tokens: SkipValidation[int] = None  # type: ignore
2227
    """Maximum number of tokens to be processed in a single iteration.
2228

2229
2230
    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."""
2231

2232
    max_num_seqs: SkipValidation[int] = None  # type: ignore
2233
    """Maximum number of sequences to be processed in a single iteration.
2234

2235
2236
    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."""
2237

2238
    max_model_len: SkipValidation[int] = None  # type: ignore
2239
2240
2241
    """Maximum length of a sequence (including prompt and generated text). This
    is primarily set in `ModelConfig` and that value should be manually
    duplicated here."""
2242

2243
    max_num_partial_prefills: int = 1
2244
2245
    """For chunked prefill, the maximum number of sequences that can be
    partially prefilled concurrently."""
2246
2247

    max_long_partial_prefills: int = 1
2248
2249
2250
2251
    """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."""
2252
2253

    long_prefill_token_threshold: int = 0
2254
2255
    """For chunked prefill, a request is considered long if the prompt is
    longer than this number of tokens."""
2256

2257
    num_lookahead_slots: int = 0
2258
2259
2260
2261
2262
2263
2264
    """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."""
2265

2266
2267
2268
2269
    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
2270
    pattern: [1, 2, 4] + [i for i in range(8, cuda_graph_sizes + 1, 8)]
2271
    3. more than one value (e.g. 1 2 128) is provided, then the capture list
2272
    will follow the provided list."""
2273

2274
    delay_factor: float = 0.0
2275
2276
    """Apply a delay (of delay factor multiplied by previous
    prompt latency) before scheduling next prompt."""
2277

2278
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2279
2280
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""
2281
2282

    is_multimodal_model: bool = False
2283
2284
2285
2286
2287
    """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.
2288

2289
2290
2291
2292
2293
2294
2295
2296
2297
    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."""
2298

2299
    preemption_mode: Optional[PreemptionMode] = None
2300
2301
2302
2303
2304
2305
    """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."""
2306
2307

    num_scheduler_steps: int = 1
2308
    """Maximum number of forward steps per scheduler call."""
2309

2310
2311
    multi_step_stream_outputs: bool = True
    """If False, then multi-step will stream outputs at the end of all steps"""
2312
2313

    send_delta_data: bool = False
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
    """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)."""
2325
2326

    chunked_prefill_enabled: bool = field(init=False)
2327
    """True if chunked prefill is enabled."""
2328

2329
    disable_chunked_mm_input: bool = False
2330
2331
2332
2333
2334
2335
    """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."""
2336

2337
2338
    # scheduler class or path. "vllm.core.scheduler.Scheduler" (default)
    # or "mod.custom_class".
2339
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
2340
2341
2342
    """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"."""
2343

2344
2345
2346
2347
2348
2349
    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.
    """

2350
2351
2352
2353
2354
2355
2356
    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.
    """

2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
    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.
2371
        factors: list[Any] = []
2372
2373
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2374
2375
        return hash_str

2376
    def __post_init__(self) -> None:
2377
2378
2379
2380
2381
2382
        if self.max_model_len is None:
            self.max_model_len = 8192

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

2383
2384
2385
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
2386
2387
2388
2389
                    # 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.
2390
                    self.max_num_batched_tokens = max(
2391
                        self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2392
                else:
2393
                    self.max_num_batched_tokens = (
2394
                        DEFAULT_MAX_NUM_BATCHED_TOKENS)
2395
            else:
2396
                # If max_model_len is too short, use
2397
                # DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
2398
                # for higher throughput.
2399
                self.max_num_batched_tokens = max(
2400
                    self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2401

2402
2403
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
2404
2405
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2406
                    POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
2407
                )
2408
            if self.is_multimodal_model:
2409
                # The value needs to be at least the number of multimodal tokens
2410
2411
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2412
                    MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
2413
2414
                )

2415
2416
2417
2418
2419
2420
2421
            # 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)

2422
2423
2424
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

2425
        if self.enable_chunked_prefill:
2426
2427
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
2428
                self.max_num_batched_tokens)
2429

2430
        self.chunked_prefill_enabled = self.enable_chunked_prefill
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
        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)

2443
2444
2445
2446
2447
2448
2449
        # 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)]

2450
2451
2452
2453
        if self.async_scheduling:
            self.scheduler_cls = (
                "vllm.v1.core.sched.async_scheduler.AsyncScheduler")

2454
2455
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2456
2457
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
2458
2459
2460
2461
2462
2463
2464
            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.")
2465

2466
2467
2468
2469
2470
        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}).")
2471

2472
2473
        if self.max_num_batched_tokens > self.max_num_seqs * self.max_model_len:
            logger.warning(
2474
                "max_num_batched_tokens (%d) exceeds max_num_seqs "
2475
2476
2477
2478
                "* max_model_len (%d). This may lead to unexpected behavior.",
                self.max_num_batched_tokens,
                self.max_num_seqs * self.max_model_len)

2479
2480
2481
2482
2483
2484
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

2485
2486
2487
2488
2489
2490
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
        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}).")

2514
2515
        return self

2516
2517
2518
2519
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

2520

2521
Device = Literal["auto", "cuda", "neuron", "cpu", "tpu", "xpu"]
2522
2523
2524


@config
2525
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
2526
class DeviceConfig:
2527
2528
    """Configuration for the device to use for vLLM execution."""

2529
    device: SkipValidation[Optional[Union[Device, torch.device]]] = "auto"
2530
    """Device type for vLLM execution.
2531
2532
2533
    This parameter is deprecated and will be
    removed in a future release.
    It will now be set automatically based
2534
    on the current platform."""
2535
2536
2537
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
2538

2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
    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.
2554
        factors: list[Any] = []
2555
2556
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2557
2558
        return hash_str

2559
2560
    def __post_init__(self):
        if self.device == "auto":
2561
            # Automated device type detection
2562
            from vllm.platforms import current_platform
2563
            self.device_type = current_platform.device_type
2564
            if not self.device_type:
2565
2566
2567
2568
                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.")
2569
2570
        else:
            # Device type is assigned explicitly
2571
2572
2573
2574
            if isinstance(self.device, str):
                self.device_type = self.device
            elif isinstance(self.device, torch.device):
                self.device_type = self.device.type
2575
2576

        # Some device types require processing inputs on CPU
2577
        if self.device_type in ["neuron"]:
2578
            self.device = torch.device("cpu")
2579
2580
        elif self.device_type in ["tpu"]:
            self.device = None
2581
2582
2583
2584
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

2585

2586
2587
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
                            "mlp_speculator", "draft_model", "deepseek_mtp"]
2588
2589
2590


@config
2591
@dataclass
2592
class SpeculativeConfig:
2593
    """Configuration for speculative decoding."""
2594

2595
    # General speculative decoding control
2596
    num_speculative_tokens: SkipValidation[int] = None  # type: ignore
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
    """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."""
2610
    draft_tensor_parallel_size: Optional[int] = None
2611
2612
    """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."""
2613
    disable_logprobs: bool = True
2614
2615
2616
    """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."""
2617

2618
    # Draft model configuration
2619
    quantization: Optional[me_quant.QuantizationMethods] = None
2620
2621
2622
    """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."""
2623
    max_model_len: Optional[int] = None
2624
2625
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
2626
    revision: Optional[str] = None
2627
2628
2629
    """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."""
2630
    code_revision: Optional[str] = None
2631
2632
2633
    """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."""
2634

2635
    # Advanced control
2636
    disable_by_batch_size: Optional[int] = None
2637
2638
2639
2640
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
2641
    prompt_lookup_max: Optional[int] = None
2642
2643
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
2644
    prompt_lookup_min: Optional[int] = None
2645
2646
2647
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

2648
    speculative_token_tree: Optional[str] = None
2649
    """Specifies the tree structure for speculative token generation.
2650
    """
2651
    # required configuration params passed from engine
2652
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2653
    """The configuration of the target model."""
2654
2655
    target_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2656
    """The parallel configuration for the target model."""
2657
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2658
2659
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
2660
    disable_log_stats: SkipValidation[bool] = None  # type: ignore
2661
2662
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
2663
2664

    # params generated in the post-init stage
2665
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2666
    """The configuration of the draft model initialized internal."""
2667
2668
    draft_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2669
    """The parallel configuration for the draft model initialized internal."""
2670

2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
    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.
        """
2683
        factors: list[Any] = []
2684
2685
2686
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2687
2688
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2689
2690
        return hash_str

2691
2692
2693
2694
2695
    @classmethod
    def from_dict(cls, dict_value: dict) -> "SpeculativeConfig":
        """Parse the CLI value for the speculative config."""
        return cls(**dict_value)

2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
    @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"]
            })
2706
2707
2708
2709
2710
2711
2712
2713
2714

        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
2715
2716
2717
2718
2719
2720
2721
2722
2723

        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"]
            })
2724

2725
2726
        return hf_config

2727
    def __post_init__(self):
2728

2729
2730
2731
2732
2733
2734
2735
        # 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.
2736
2737
2738
2739

        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
2740
            if self.target_model_config and \
2741
2742
2743
2744
                (self.target_model_config.hf_text_config.model_type \
                        == "deepseek_v3" or
                    self.target_model_config.hf_text_config.model_type \
                        == "mimo"):
2745
2746
2747
2748
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2749
            else:
2750
2751
2752
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2753
2754
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2755
2756
2757
2758
2759
2760
2761
        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"
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
            # 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
2776
            if self.prompt_lookup_min < 1:
2777
2778
2779
2780
2781
                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")
2782
            if self.prompt_lookup_min > self.prompt_lookup_max:
2783
2784
2785
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2786

2787
2788
2789
            # 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.
2790
2791
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2792
        else:
2793
2794
2795
2796
2797
2798
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
2799
                    runner="draft",
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
                    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,
                )
2821

2822
                # Automatically detect the method
2823
                if self.method in ('eagle', 'eagle3'):
2824
                    pass
2825
2826
                elif "eagle-" in self.draft_model_config.model.lower() or \
                        "eagle3-" in self.draft_model_config.model.lower():
2827
2828
2829
2830
2831
2832
                    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"
2833
                elif (self.draft_model_config.hf_config.model_type
Yuxuan Zhang's avatar
Yuxuan Zhang committed
2834
                      in ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp")):
Jiayi Yao's avatar
Jiayi Yao committed
2835
2836
2837
2838
2839
2840
2841
                    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."
                            )
2842
                else:
2843
                    self.method = "draft_model"
2844
2845
2846
2847
2848
                    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.")
2849
2850

                # Replace hf_config for EAGLE draft_model
2851
                if self.method in ("eagle", "eagle3"):
2852
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2853
                        raise ValueError(
2854
2855
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2856
2857
2858
2859

                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
                    if isinstance(self.draft_model_config.hf_config,
2860
                                  EAGLEConfig):
2861
2862
2863
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2864
                            self.draft_model_config.hf_config,
2865
2866
                            method=self.method,
                            model_type="eagle")
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
                        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=}")

                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
                )
2894

2895
2896
2897
2898
2899
2900
                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,
                    ))
2901

2902
2903
2904
2905
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2906

2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
    @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,
        )

2942
    @staticmethod
2943
    def _verify_and_get_draft_tp(
2944
2945
2946
2947
2948
2949
            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.
2950
        """
2951
2952
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2953
        if speculative_draft_tensor_parallel_size is None:
2954
2955
2956
2957
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2958
2959
2960
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2961
2962
2963
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2964
2965
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2966
            raise ValueError(
2967
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2968
                f"other value than 1 or target model tensor_parallel_size")
2969
        return speculative_draft_tensor_parallel_size
2970

2971
2972
2973
2974
2975
2976
2977
2978
2979
    @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.
        """
2980
2981
2982
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2983
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2984
2985
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
            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

2997
2998
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2999
3000
3001
3002
3003
3004
        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.")

3005
3006
3007
3008
3009
3010
3011
        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)
3012
3013
3014
3015
3016
3017

        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=}")
3018

3019
3020
3021
3022
3023
3024
        if self.method == "eagle3" and self.target_model_config and \
            "llama" not in self.target_model_config.hf_text_config.model_type:
            raise ValueError(
                "Eagle3 is only supported for Llama models. "
                f"Got {self.target_model_config.hf_text_config.model_type=}")

3025
3026
        return self

3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
    @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

3037
    def use_eagle(self) -> bool:
Jiayi Yao's avatar
Jiayi Yao committed
3038
        return self.method in ("eagle", "eagle3", "deepseek_mtp")
3039

3040
    def __repr__(self) -> str:
3041
3042
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
3043
        num_spec_tokens = self.num_speculative_tokens
3044
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
3045
3046


3047
3048
3049
3050
LoRADType = Literal["auto", "float16", "bfloat16"]


@config
3051
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3052
class LoRAConfig:
3053
3054
3055
3056
3057
3058
    """Configuration for LoRA."""

    max_lora_rank: int = 16
    """Max LoRA rank."""
    max_loras: int = 1
    """Max number of LoRAs in a single batch."""
3059
    fully_sharded_loras: bool = False
3060
3061
3062
3063
    """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.
    """
3064
    max_cpu_loras: Optional[int] = None
3065
3066
3067
3068
    """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."""
3069
    lora_extra_vocab_size: int = 256
3070
3071
    """Maximum size of extra vocabulary that can be present in a LoRA adapter
    (added to the base model vocabulary)."""
3072
3073
    lora_vocab_padding_size: ClassVar[int] = current_platform\
        .get_lora_vocab_padding_size()
3074

3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
    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."""
3085
    bias_enabled: bool = False
3086
    """Enable bias for LoRA adapters."""
3087

3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
    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.
        """
3100
        factors: list[Any] = []
3101
3102
3103
3104
3105
        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)
3106
        factors.append(self.lora_vocab_padding_size)
3107
        factors.append(self.bias_enabled)
3108
3109
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3110
3111
        return hash_str

3112
    def __post_init__(self):
3113
        # Setting the maximum rank to 512 should be able to satisfy the vast
3114
        # majority of applications.
3115
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
3116
        possible_lora_extra_vocab_size = (256, 512)
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
        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
3132
                f"max_loras ({self.max_loras})")
3133

3134
    def verify_with_cache_config(self, cache_config: CacheConfig):
3135
3136
3137
        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.")
3138

3139
3140
3141
3142
3143
3144
3145
    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)


3146
@config
3147
@dataclass
3148
class MultiModalConfig:
3149
3150
    """Controls the behavior of multimodal models."""

3151
3152
    limit_per_prompt: dict[str, int] = \
        cast(dict[str, int], get_field(ModelConfig, "limit_mm_per_prompt"))
3153
    """
3154
    The maximum number of input items allowed per prompt for each modality.
3155
    Defaults to 1 (V0) or 999 (V1) for each modality.
3156
3157

    For example, to allow up to 16 images and 2 videos per prompt:
3158
    `{"images": 16, "videos": 2}`
3159
3160
    """

3161
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
3162
3163
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
3164
3165
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """

3166
3167
3168
    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
3169
    `transformers.AutoProcessor.from_pretrained`.
3170
3171
3172
3173

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

    For example, for Phi-3-Vision:
3174
    `{"num_crops": 4}`.
3175
3176
3177
3178
    """

    disable_mm_preprocessor_cache: bool = False
    """
3179
    If `True`, disable caching of the processed multi-modal inputs.
3180
3181
    """

3182
3183
3184
3185
3186
    interleave_mm_strings: bool = False
    """
    Enable fully interleaved support for multimodal prompts.
    """

3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
    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.
3201
        factors: list[Any] = []
3202
3203
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3204
3205
        return hash_str

3206
3207
3208
3209
3210
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
3211
3212
3213
3214
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
3215

3216
    # TODO: Add configs to init vision tower or not.
3217

3218

3219
@config
3220
3221
@dataclass
class PoolerConfig:
3222
    """Controls the behavior of output pooling in pooling models."""
3223
3224

    pooling_type: Optional[str] = None
3225
    """
3226
    The pooling method of the pooling model. This should be a key in
3227
    [`vllm.model_executor.layers.pooler.PoolingType`][].
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
    """

    normalize: Optional[bool] = None
    """
    Whether to normalize the pooled outputs. Usually, this should be set to
    ``True`` for embedding outputs.
    """

    softmax: Optional[bool] = None
    """
    Whether to apply softmax to the pooled outputs. Usually, this should be set
    to ``True`` for classification outputs.
    """

    step_tag_id: Optional[int] = None
    """
3244
    If set, only the score corresponding to the ``step_tag_id`` in the
3245
3246
3247
3248
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

3249
    returned_token_ids: Optional[list[int]] = None
3250
    """
3251
3252
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
3253
3254
3255
    ``math-shepherd-mistral-7b-prm`` model.
    """

3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
    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.
3270
        factors: list[Any] = []
3271
3272
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3273
3274
        return hash_str

3275

3276
3277
3278
3279
3280
3281
3282
3283
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

3284
3285
3286
3287
3288
3289
3290
# 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.",
}
3291

3292

3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
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,
3311
    config: PretrainedConfig,
3312
3313
3314
    *,
    revision: Optional[str],
):
3315
3316
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
3317
    config_dtype = getattr(config, "torch_dtype", None)
3318

3319
    # Fallbacks for multi-modal models if the root config
3320
    # does not define torch_dtype
3321
3322
    if config_dtype is None:
        config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
3323
3324
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)
3325
3326
    if config_dtype is None and hasattr(config, "encoder_config"):
        config_dtype = getattr(config.encoder_config, "torch_dtype", None)
3327

3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
    # 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)

3343
3344
3345
    if config_dtype is None:
        config_dtype = torch.float32

3346
    return config_dtype
3347

Shinichi Hemmi's avatar
Shinichi Hemmi committed
3348

3349
3350
3351
3352
3353
3354
3355
def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform
3356

3357
3358
3359
3360
    supported_dtypes = [
        dtype for dtype in current_platform.supported_dtypes
        if _is_valid_dtype(model_type, dtype)
    ]
3361

3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
    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,
            )
3415
        else:
3416
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
3417
                raise ValueError(f"Unknown dtype: {dtype!r}")
3418
3419
3420
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
3421
    else:
3422
        raise ValueError(f"Unknown dtype: {dtype}")
3423

3424
3425
    _check_valid_dtype(model_type, torch_dtype)

3426
3427
3428
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
3429
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
3430
3431
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
3432
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
3433
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
3434
            # Casting between float16 and bfloat16 is allowed with a warning.
3435
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
3436
3437

    return torch_dtype
3438
3439
3440
3441


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
3442
    tokenizer_config: Optional[dict],
3443
    max_model_len: Optional[int],
3444
    disable_sliding_window: bool,
3445
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
3446
    spec_target_max_model_len: Optional[int] = None,
3447
    encoder_config: Optional[Any] = None,
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
) -> 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",
3458
3459
        # ChatGLM2
        "seq_length",
3460
3461
        # Command-R
        "model_max_length",
3462
3463
        # Whisper
        "max_target_positions",
3464
3465
3466
3467
3468
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
3469
    # Choose the smallest "max_length" from the possible keys
3470
    max_len_key = None
3471
    for key in possible_keys:
3472
3473
3474
3475
3476
        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
3477
3478
3479
3480
    # 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
3481
3482
3483
3484

    # 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:
3485
3486

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
3487
        max_len_key = "sliding_window" \
3488
3489
3490
            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)
3491

3492
3493
3494
3495
3496
3497
3498
    # 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)

3499
3500
    # If none of the keys were found in the config, use a default and
    # log a warning.
3501
    if derived_max_model_len == float("inf"):
3502
3503
3504
3505
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

3506
3507
3508
3509
3510
        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

3511
3512
3513
3514
        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: "
3515
            "%s. Assuming the model's maximum length is %d.", possible_keys,
3516
            default_max_len)
3517
        derived_max_model_len = default_max_len
3518

3519
    rope_scaling = getattr(hf_config, "rope_scaling", None)
3520
3521
3522
    # 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:
3523
3524
3525
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535

        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.")

3536
3537
3538
3539
            # 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)

3540
3541
3542
3543
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
3544

3545
3546
3547
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

3548
3549
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
3550
    if max_model_len is None:
3551
        max_model_len = int(derived_max_model_len)
3552
3553
3554
3555
3556
3557
3558
3559
        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)
3560
    elif max_model_len > derived_max_model_len:
3561
3562
3563
3564
3565
        # 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:
3566
3567
3568
3569
3570
3571
3572
            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.")
3573
        else:
3574
            msg = (
3575
                f"User-specified max_model_len ({max_model_len}) is greater "
3576
3577
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
3578
                f"{model_max_length} in model's config.json). This may lead "
3579
3580
3581
3582
3583
3584
3585
3586
3587
                "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")
3588
    return int(max_model_len)
3589
3590


3591
def get_min_sliding_window(
3592
        sliding_window: Union[int, list[Optional[int]]]) -> int:
3593
3594
3595
3596
3597
3598
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


3599
def get_served_model_name(model: str,
3600
                          served_model_name: Optional[Union[str, list[str]]]):
3601
    """
3602
3603
3604
3605
    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
3606
3607
3608
3609
3610
3611
3612
3613
3614
    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


3615
GuidedDecodingBackendV0 = Literal["auto", "outlines", "lm-format-enforcer",
3616
                                  "xgrammar", "guidance"]
3617
3618

GuidedDecodingBackendV1 = Literal["auto", "xgrammar", "guidance", "outlines"]
3619
3620
GuidedDecodingBackend = Literal[GuidedDecodingBackendV0,
                                GuidedDecodingBackendV1]
3621
3622
3623


@config
3624
3625
@dataclass
class DecodingConfig:
3626
    """Dataclass which contains the decoding strategy of the engine."""
3627

3628
    backend: GuidedDecodingBackend = "auto" if envs.VLLM_USE_V1 else "xgrammar"
3629
3630
3631
3632
    """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."""
3633

3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
    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`."""

3646
    reasoning_backend: str = ""
3647
    """Select the reasoning parser depending on the model that you're using.
3648
    This is used to parse the reasoning content into OpenAI API format."""
3649

3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
    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.
3664
        factors: list[Any] = []
3665
3666
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3667
3668
        return hash_str

3669
    def __post_init__(self):
3670
        if envs.VLLM_USE_V1:
3671
            valid_guided_backends = get_args(GuidedDecodingBackendV1)
3672
        else:
3673
            valid_guided_backends = get_args(GuidedDecodingBackendV0)
3674
3675
        if self.backend not in valid_guided_backends:
            raise ValueError(f"Invalid backend '{self.backend}',"
3676
                             f" must be one of {valid_guided_backends}")
3677
3678
3679
3680
3681
3682
3683
3684
        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.")

3685

3686
3687
3688
3689
DetailedTraceModules = Literal["model", "worker", "all"]


@config
3690
3691
@dataclass
class ObservabilityConfig:
3692
    """Configuration for observability - metrics and tracing."""
3693

3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
    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)
3709

3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
    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))
3735

3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
    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.
3750
        factors: list[Any] = []
3751
3752
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3753
3754
        return hash_str

3755
    def __post_init__(self):
3756
3757
3758
3759
3760
        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()

3761
        from vllm.tracing import is_otel_available, otel_import_error_traceback
3762
3763
3764
3765
3766
        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}")
3767

3768
3769
3770
3771
3772
3773
    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(","))

3774

3775
3776
3777
3778
3779
3780
3781
3782
KVProducer = Literal["kv_producer", "kv_both"]
KVConsumer = Literal["kv_consumer", "kv_both"]
KVRole = Literal[KVProducer, KVConsumer]


@config
@dataclass
class KVTransferConfig:
3783
3784
3785
    """Configuration for distributed KV cache transfer."""

    kv_connector: Optional[str] = None
3786
3787
    """The KV connector for vLLM to transmit KV caches between vLLM instances.
    """
3788

3789
    engine_id: Optional[str] = None
Robert Shaw's avatar
Robert Shaw committed
3790
3791
    """The engine id for KV transfers."""

3792
    kv_buffer_device: Optional[str] = "cuda"
3793
3794
    """The device used by kv connector to buffer the KV cache.
    Currently only support 'cuda'."""
3795
3796

    kv_buffer_size: float = 1e9
3797
3798
    """The buffer size for TorchDistributedConnector. Measured in number of
    bytes. Recommended value: 1e9 (about 1GB)."""
3799

3800
3801
    kv_role: Optional[KVRole] = None
    """Whether this vLLM instance produces, consumes KV cache, or both. Choices
Robert Shaw's avatar
Robert Shaw committed
3802
    are 'kv_producer', 'kv_consumer', and 'kv_both'."""
3803
3804

    kv_rank: Optional[int] = None
3805
3806
3807
    """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."""
3808
3809

    kv_parallel_size: int = 1
3810
3811
    """The number of parallel instances for KV cache transfer. For
    PyNcclConnector, this should be 2."""
3812
3813

    kv_ip: str = "127.0.0.1"
3814
    """The KV connector ip, used to build distributed connection."""
3815
3816

    kv_port: int = 14579
3817
    """The KV connector port, used to build distributed connection."""
3818

3819
3820
    kv_connector_extra_config: dict[str, Any] = field(default_factory=dict)
    """any extra config that the connector may need."""
3821

3822
3823
3824
3825
    kv_connector_module_path: Optional[str] = None
    """The Python module path to dynamically load the KV connector from.
    Only supported in V1."""

3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
    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.
3840
        factors: list[Any] = []
3841
3842
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3843
3844
        return hash_str

3845
    def __post_init__(self) -> None:
3846
3847
3848
        if self.engine_id is None:
            self.engine_id = str(uuid.uuid4())

3849
3850
3851
        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)}")
3852
3853
3854

        if self.kv_connector is not None and self.kv_role is None:
            raise ValueError("Please specify kv_disagg_role when kv_connector "
3855
                             f"is set, supported roles are {get_args(KVRole)}")
3856
3857
3858
3859

    @property
    def is_kv_transfer_instance(self) -> bool:
        return self.kv_connector is not None and \
3860
            self.kv_role in get_args(KVRole)
3861
3862
3863
3864

    @property
    def is_kv_producer(self) -> bool:
        return self.kv_connector is not None and \
3865
            self.kv_role in get_args(KVProducer)
3866
3867
3868
3869

    @property
    def is_kv_consumer(self) -> bool:
        return self.kv_connector is not None and \
3870
            self.kv_role in get_args(KVConsumer)
3871

3872
3873
3874
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

3875

3876
3877
3878
@config
@dataclass
class KVEventsConfig:
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
    """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.
    """


3918
3919
3920
3921
3922
3923
3924
3925
class CompilationLevel:
    # constants for the levels of the compilation process
    NO_COMPILATION = 0
    DYNAMO_AS_IS = 1
    DYNAMO_ONCE = 2
    PIECEWISE = 3


3926
3927
3928
3929
3930
3931
3932
3933
3934
@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."""

3935
    enable_fusion: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
3936
3937
3938
    """Whether to enable the custom fusion (RMSNorm/SiluMul+quant) pass."""
    enable_attn_fusion: bool = False
    """Whether to enable the custom attention+quant fusion pass."""
3939
    enable_noop: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
3940
3941
3942
    """Whether to enable the custom no-op elimination pass."""
    enable_sequence_parallelism: bool = False
    """Whether to enable sequence parallelism."""
3943
3944
    enable_async_tp: bool = False
    """Whether to enable async TP."""
3945
3946
3947
3948
    enable_fi_allreduce_fusion: bool = False
    """Whether to enable flashinfer allreduce fusion."""
    fi_allreduce_fusion_max_token_num: int = 1024
    """Max number of tokens to used in flashinfer allreduce fusion."""
3949

3950
3951
    # TODO(luka) better pass enabling system.

3952
3953
3954
3955
    def uuid(self):
        """
        Produces a hash unique to the pass configuration.
        Any new fields that affect compilation should be added to the hash.
3956
        Any future fields that don't affect compilation should be excluded.
3957
        """
3958
        return InductorPass.hash_dict(asdict(self))
3959
3960

    def __post_init__(self) -> None:
3961
3962
3963
3964
3965
3966
3967
3968
3969
        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")
3970
3971
3972
3973
3974
3975
3976


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

3977
    - Top-level Compilation control:
3978
3979
3980
3981
3982
3983
        - [`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]
3984
    - CudaGraph capture:
3985
3986
3987
3988
3989
3990
3991
3992
        - [`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]
3993
    - Inductor compilation:
3994
3995
3996
3997
3998
        - [`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]
3999
        - custom inductor passes
4000

4001
4002
4003
4004
4005
4006
4007
4008
4009
    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.
4010
4011
    """
    # Top-level Compilation control
4012
    level: int = 0
4013
4014
4015
4016
4017
4018
    """The level of compilation:

    - 0: no compilation.
    - 1: dynamo as is.
    - 2: dynamo once.
    - 3: piecewise compilation."""
4019
    debug_dump_path: str = ""
4020
    """The path to dump the debug information."""
4021
    cache_dir: str = ""
4022
4023
4024
    """The directory to store the compiled graph, to accelerate Inductor
    compilation. By default, it will use model-related information to generate
    a cache directory."""
4025
    backend: str = ""
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
    """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
4048
4049
    disabled when running with Inductor: level>=PIECEWISE and use_inductor=True.
    Inductor generates (fused) Triton kernels for disabled custom ops."""
4050
4051
4052
4053
4054
    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
4055
    use_inductor: bool = True
4056
4057
    """Whether to use inductor compilation:

4058
4059
4060
4061
4062
    - 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.
4063

4064
    This setting is ignored if level<PIECEWISE."""
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
    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
4080
    use_cudagraph: bool = field(default_factory=lambda: envs.VLLM_USE_V1)
4081
4082
4083
4084
4085
    """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.
4086
4087
    In the vLLM V1 Engine, this flag only applies for
    CompilationLevel.PIECEWISE (aka -O3).
4088
4089
4090
4091
    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."""
4092
    cudagraph_num_of_warmups: int = 0
4093
4094
4095
4096
    """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."""
4097
    cudagraph_capture_sizes: Optional[list[int]] = None
4098
4099
4100
    """Sizes to capture cudagraph.
    - None (default): capture sizes are inferred from vllm config.
    - list[int]: capture sizes are specified as given."""
4101
    cudagraph_copy_inputs: bool = False
4102
4103
4104
4105
4106
    """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."""
4107
    full_cuda_graph: bool = False
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
    """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."""
4127

4128
    # keep track of enabled and disabled custom ops
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
    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."""
4145

4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
    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.
        """
4158
        factors: list[Any] = []
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
        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()

4169
4170
    def __repr__(self) -> str:
        exclude = {
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
            "static_forward_context": True,
            "enabled_custom_ops": True,
            "disabled_custom_ops": True,
            "compilation_time": True,
            "bs_to_padded_graph_size": True,
            "pass_config": True,
            "traced_files": True,
            "inductor_compile_config": {
                "post_grad_custom_post_pass": True,
            },
4181
        }
4182
4183
4184
4185
        # 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(
4186
4187
4188
                self,
                exclude=exclude,  # type: ignore[arg-type]
                exclude_unset=True).decode())
4189
4190
4191

    __str__ = __repr__

4192
4193
    @classmethod
    def from_cli(cls, cli_value: str) -> "CompilationConfig":
4194
4195
4196
        """Parse the CLI value for the compilation config.
        -O1, -O2, -O3, etc. is handled in FlexibleArgumentParser.
        """
4197
        return TypeAdapter(CompilationConfig).validate_json(cli_value)
4198

4199
    def __post_init__(self) -> None:
4200
4201
4202
4203
        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
4204
4205
4206
4207
4208
4209
4210
4211
        # 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

4212
        if is_torch_equal_or_newer("2.6"):
Michael Goin's avatar
Michael Goin committed
4213
4214
4215
4216
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

4217
4218
4219
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
4220
4221
4222
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
4223
4224
4225
4226
4227
4228
4229
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
4230
4231
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
4232

4233
4234
        if isinstance(self.pass_config, dict):
            self.pass_config = PassConfig(**self.pass_config)
4235

4236
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
        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
4254

4255
        from vllm.compilation.backends import VllmBackend
4256
        return VllmBackend(vllm_config)
4257

4258
    def init_with_cudagraph_sizes(self,
4259
                                  cudagraph_capture_sizes: list[int]) -> None:
4260
        """To complete the initialization of config,
4261
4262
        we need to know the cudagraph sizes."""

4263
        if self.cudagraph_capture_sizes is None:
4264
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
4265
        else:
4266
            # de-duplicate the sizes provided by the config
4267
4268
4269
4270
4271
4272
            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
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287

        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
4288

4289
        # sort to make sure cudagraph capture sizes are in descending order
4290
4291
4292
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
4293

4294
4295
4296
4297
        # 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)
        ]
4298
4299
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
4300
4301
4302
4303
4304
4305
4306
            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
4307

4308
4309
    def set_splitting_ops_for_v1(self):
        # NOTE: this function needs to be called
4310
4311
4312
4313
4314
        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}")

4315
        if not self.splitting_ops:
4316
            self.splitting_ops = [] if self.full_cuda_graph else [
4317
4318
                "vllm.unified_attention",
                "vllm.unified_attention_with_output",
4319
                "vllm.mamba_mixer2",
4320
4321
            ]

4322

4323
@config
4324
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
4325
4326
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
4327
4328
4329
    simplifies passing around the distinct configurations in the codebase.
    """

4330
4331
4332
    # TODO: use default_factory once default constructing ModelConfig doesn't
    # try to download a model
    model_config: ModelConfig = None  # type: ignore
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
    """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."""
4344
    lora_config: Optional[LoRAConfig] = None
4345
4346
4347
    """LoRA configuration."""
    speculative_config: Optional[SpeculativeConfig] = None
    """Speculative decoding configuration."""
4348
    decoding_config: DecodingConfig = field(default_factory=DecodingConfig)
4349
    """Decoding configuration."""
4350
    observability_config: Optional[ObservabilityConfig] = None
4351
    """Observability configuration."""
4352
    quant_config: Optional[QuantizationConfig] = None
4353
4354
4355
    """Quantization configuration."""
    compilation_config: CompilationConfig = field(
        default_factory=CompilationConfig)
4356
    """`torch.compile` and cudagraph capture configuration for the model.
4357

4358
4359
    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}'`).
4360
    Currently, -O <n> and -O=<n> are supported as well but this will likely be
4361
    removed in favor of clearer -O<n> syntax in the future.
4362
4363
4364

    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
4365
    production, also default in V1.
4366
4367
4368
4369
4370
4371

    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."""
4372
    kv_events_config: Optional[KVEventsConfig] = None
4373
    """The configurations for event publishing."""
4374
    # some opaque config, only used to provide additional information
4375
4376
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
4377
4378
4379
4380
    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."""
4381
    instance_id: str = ""
4382
    """The ID of the vLLM instance."""
4383

4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
    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.
        """
4396
        factors: list[Any] = []
4397
4398

        # summarize vllm config
4399
        vllm_factors: list[Any] = []
4400
4401
        from vllm import __version__
        vllm_factors.append(__version__)
4402
        vllm_factors.append(envs.VLLM_USE_V1)
4403
4404
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
4405
4406
        else:
            vllm_factors.append("None")
4407
4408
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
4409
4410
        else:
            vllm_factors.append("None")
4411
4412
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
4413
4414
        else:
            vllm_factors.append("None")
4415
4416
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
4417
4418
        else:
            vllm_factors.append("None")
4419
4420
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
4421
4422
        else:
            vllm_factors.append("None")
4423
4424
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
4425
4426
        else:
            vllm_factors.append("None")
4427
4428
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
4429
4430
4431
4432
4433
            # 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))
4434
4435
        else:
            vllm_factors.append("None")
4436
4437
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
4438
4439
        else:
            vllm_factors.append("None")
4440
4441
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
4442
4443
        else:
            vllm_factors.append("None")
4444
4445
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
4446
4447
        else:
            vllm_factors.append("None")
4448
4449
4450
4451
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
4452
4453
        else:
            vllm_factors.append("None")
4454
4455
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
4456
4457
4458
        else:
            vllm_factors.append("None")
        if self.additional_config:
4459
4460
4461
4462
4463
4464
4465
4466
            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)
4467
4468
        else:
            vllm_factors.append("None")
4469
4470
        factors.append(vllm_factors)

4471
4472
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
4473
4474
        return hash_str

4475
4476
4477
4478
4479
4480
    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]
4481

4482
4483
4484
4485
4486
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
4487
        from vllm.platforms import current_platform
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
        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
4510

4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
    @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)

4522
4523
4524
4525
4526
4527
4528
4529
4530
    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

4531
4532
4533
4534
4535
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

4536
4537
4538
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
4539
4540
4541

        self.try_verify_and_update_config()

4542
4543
4544
4545
4546
        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)
4547
4548
            self.model_config.verify_dual_chunk_attention_config(
                self.load_config)
4549

4550
        self.cache_config.verify_with_parallel_config(self.parallel_config)
4551

4552
        if self.lora_config is not None:
4553
            self.lora_config.verify_with_cache_config(self.cache_config)
4554
            self.lora_config.verify_with_model_config(self.model_config)
4555

4556
        if self.quant_config is None and self.model_config is not None:
4557
4558
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
4559

4560
        from vllm.platforms import current_platform
4561
        if self.model_config is not None and \
4562
4563
4564
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
4565
            logger.warning_once(
4566
4567
4568
4569
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

4570
4571
4572
4573
4574
        # 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
4575
4576
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
4577
4578
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
4579
4580
            # By default, V1 uses piecewise CUDA graphs. If full_cuda_graph
            # is set to True, full CUDA graphs will be used.
4581
            self.compilation_config.cudagraph_num_of_warmups = 1
4582
            self.compilation_config.level = CompilationLevel.PIECEWISE
4583
            self.compilation_config.set_splitting_ops_for_v1()
4584

4585
        self._set_cudagraph_sizes()
4586

4587
        if self.cache_config.cpu_offload_gb > 0 and \
4588
4589
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
4590
            logger.warning(
4591
                "CPU offload is not supported with `torch.compile` in v0 yet."
4592
4593
4594
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4595
4596
4597
4598
4599
4600
        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`.")
4601
4602
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4603
4604
        if self.compilation_config.full_cuda_graph and \
            not self.model_config.disable_cascade_attn:
4605
4606
            logger.info("full_cuda_graph is not supported with "
                        "cascade attention. Disabling cascade attention.")
4607
            self.model_config.disable_cascade_attn = True
4608

4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
        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

4630
        if (self.kv_events_config is not None
4631
4632
4633
4634
4635
                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.")
4636
4637
        if (self.kv_events_config is not None
                and self.kv_events_config.publisher != "null"
4638
4639
4640
4641
4642
                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.")
4643
4644
        current_platform.check_and_update_config(self)

4645
4646
4647
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

4648
4649
4650
4651
4652
4653
4654
        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.
4655
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4656
4657
            if self.kv_transfer_config is not None:
                # Hybrid KV cache manager is not compatible with KV transfer.
4658
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4659
4660
            if self.kv_events_config is not None:
                # Hybrid KV cache manager is not compatible with KV events.
4661
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4662
4663
4664
4665
4666
4667
4668
            if self.model_config is not None and \
                self.model_config.attention_chunk_size is not None and \
                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
4669

4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
    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
        ]

4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
    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.

4706
4707
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
4708
4709

        During runtime, if batchsize is larger than
4710
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4711
4712
        no cudagraph will be used.
        If the batch size is no larger than
4713
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
        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)]
4726
4727
4728
4729
4730
                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)

4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
                # 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:
4752
4753
4754
4755
4756
4757
4758
4759
                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
4760
                    raise TypeError(f"Invalid value for {cuda_graph_sizes=}.")
4761
4762
4763
4764
                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)
4765
4766
4767
4768
4769
                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
                ]
4770
4771
4772
4773

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

4774
    def recalculate_max_model_len(self, max_model_len: int):
4775
        # Can only be called in try_verify_and_update_config
4776
        model_config = self.model_config
4777
        max_model_len = model_config.get_and_verify_max_len(max_model_len)
4778
4779
        self.model_config.max_model_len = max_model_len
        self.scheduler_config.max_model_len = max_model_len
4780
4781
4782
4783
4784
4785

    def try_verify_and_update_config(self):
        architecture = getattr(self.model_config, "architecture", None)
        if architecture is None:
            return

4786
4787
        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP, HybridAttentionMambaModelConfig)
4788
4789
4790
        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_config(self)
4791

4792
4793
4794
        if self.model_config.is_hybrid:
            HybridAttentionMambaModelConfig.verify_and_update_config(self)

4795
4796
4797
4798
4799
4800
        if self.model_config.task == "classify":
            # Maybe convert ForCausalLM into ForSequenceClassification model.
            from vllm.model_executor.models.adapters import (
                SequenceClassificationConfig)
            SequenceClassificationConfig.verify_and_update_config(self)

4801
    def __str__(self):
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
        return (
            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}, "
            f"revision={self.model_config.revision}, "
            f"override_neuron_config={self.model_config.override_neuron_config},"
            f" tokenizer_revision={self.model_config.tokenizer_revision}, "
            f"trust_remote_code={self.model_config.trust_remote_code}, "
            f"dtype={self.model_config.dtype}, "
            f"max_seq_len={self.model_config.max_model_len},"
            f" download_dir={self.load_config.download_dir!r}, "
            f"load_format={self.load_config.load_format}, "
            f"tensor_parallel_size={self.parallel_config.tensor_parallel_size},"
            f" pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, "  # noqa
            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}, "
            f" device_config={self.device_config.device}, "
            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}, "
4832
4833
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
4834
4835
4836


_current_vllm_config: Optional[VllmConfig] = None
4837
_current_prefix: Optional[str] = None
4838
4839
4840


@contextmanager
4841
4842
4843
def set_current_vllm_config(vllm_config: VllmConfig,
                            check_compile=False,
                            prefix: Optional[str] = None):
4844
    """
4845
    Temporarily set the current vLLM config.
4846
    Used during model initialization.
4847
    We save the current vLLM config in a global variable,
4848
    so that all modules can access it, e.g. custom ops
4849
    can access the vLLM config to determine how to dispatch.
4850
    """
4851
    global _current_vllm_config, _current_prefix
4852
    old_vllm_config = _current_vllm_config
4853
    old_prefix = _current_prefix
4854
4855
4856
4857
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
4858
        _current_prefix = prefix
4859
        yield
4860
4861
4862
    except Exception:
        raise
    else:
4863
4864
4865
4866
        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)
4867
4868
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
4869
4870
4871
4872
4873
4874
4875
4876
4877
            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"
4878
                " if you want it to be supported.",
4879
                vllm_config.model_config.model)
4880
    finally:
4881
        _current_vllm_config = old_vllm_config
4882
        _current_prefix = old_prefix
4883
4884
4885
4886
4887
4888
4889


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.
4890
        logger.warning("Current vLLM config is not set.")
4891
4892
4893
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
4894
4895


4896
4897
4898
4899
4900
4901
4902
4903
4904
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


4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
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:
4916
        result (bool): `True` if a match is found, `False` otherwise.
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
    """
    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}")
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942


T = TypeVar("T")


def get_layers_from_vllm_config(vllm_config: VllmConfig,
                                layer_type: type[T]) -> dict[str, T]:
    return {
        layer_name: layer
        for layer_name, layer in
        vllm_config.compilation_config.static_forward_context.items()
        if isinstance(layer, layer_type)
    }
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972


@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:
4973
        return self.min_energy_split_window_size is not None
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991


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)