config.py 216 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
222
223
224
TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]


@config
225
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
226
class ModelConfig:
227
228
    """Configuration for the model."""

229
    model: str = "Qwen/Qwen3-0.6B"
230
231
232
    """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."""
233
234
235
236
237
238
239
240
    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."""
241
    tokenizer: SkipValidation[str] = None  # type: ignore
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
    """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
263
264
    """Random seed for reproducibility. Initialized to None in V0, but
    initialized to 0 in V1."""
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
    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)
280
    """RoPE scaling configuration. For example,
281
282
283
284
285
286
287
288
    `{"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."""
289
    max_model_len: SkipValidation[int] = None  # type: ignore
290
291
    """Model context length (prompt and output). If unspecified, will be
    automatically derived from the model config.
292

293
294
295
296
297
298
    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
299
    """Specify the maximum length for spec decoding draft models."""
300
    quantization: SkipValidation[Optional[QuantizationMethods]] = None
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
    """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."""
    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."""
333
334
335
336
    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."""
337
338
339
340
341
342
343
344
345
346
347
    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."""
348
    interleave_mm_strings: bool = False
349
    """Enable fully interleaved support for multimodal prompts, while using
350
    --chat-template-content-format=string. Defaults to False."""
351
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
352
353
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
354
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
355
356
357
358
359
360
361
362
363
364
365
366
367
368
    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
369
    config. If a callable, it is called to update the HuggingFace config."""
370
371
372
373
374
    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}`.
375
    """
376
377
378
379
380
381
382
    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
383
    arguments. e.g. `{"cast_logits_dtype": "bfloat16"}`."""
384
385
386
387
388
389
    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}`.
390
    """
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
    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
406
    `--generation-config vllm`, only the override parameters are used."""
407
408
409
410
411
412
413
414
415
    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."""
416
417
    override_attention_dtype: Optional[str] = None
    """Override dtype for attention"""
418

419
420
421
422
423
424
425
426
427
428
429
430
    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.
        """
431
        factors: list[Any] = []
432
433
434
435
436
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
437
438
439
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
440
        factors.append(self.trust_remote_code)
441
442
443
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
444
445
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
446
447
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
448
449
        str_factors = str(factors)
        assert_hashable(str_factors)
450
451
        return hashlib.sha256(str(factors).encode()).hexdigest()

452
    def __post_init__(self) -> None:
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
        # 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)

471
472
473
        # 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)
474
475
476
477
478
479
480
481
482
483
484
485
        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):
486
            hf_overrides_kw = {}
487
            hf_overrides_fn = self.hf_overrides
488
        else:
489
            hf_overrides_kw = self.hf_overrides
490
            hf_overrides_fn = None
491

492
493
        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
494
            hf_overrides_kw.update(hf_override)
495
            hf_overrides_str = json.dumps(hf_overrides_kw)
496
497
498
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
499
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
500
501
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
502
            hf_overrides_kw.update(hf_override)
503
            hf_overrides_str = json.dumps(hf_overrides_kw)
504
505
506
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
507
508
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

509
        self.maybe_pull_model_tokenizer_for_s3(self.model, self.tokenizer)
510

511
512
513
514
        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 "
515
516
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
517
518
                "for instructions on how to install it.")

519
520
        from vllm.platforms import current_platform

521
522
523
524
525
526
        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)

527
528
529
530
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
531

532
533
534
        if isinstance(self.config_format, str):
            self.config_format = ConfigFormat(self.config_format)

535
        hf_config = get_config(self.hf_config_path or self.model,
536
537
538
539
540
541
                               self.trust_remote_code,
                               self.revision,
                               self.code_revision,
                               self.config_format,
                               hf_overrides_kw=hf_overrides_kw,
                               hf_overrides_fn=hf_overrides_fn)
542
543
        self.hf_config = hf_config

544
        self.hf_text_config = get_hf_text_config(self.hf_config)
545
546
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
547
        self.encoder_config = self._get_encoder_config()
548
        self.hf_image_processor_config = get_hf_image_processor_config(
549
            self.model, hf_token=self.hf_token, revision=self.revision)
550

551
552
553
        # For pooling models, self.task is used to indicate the
        # user-selected task
        if self.task == "score":
554
            if self._is_classify_task(self.architectures):
555
556
557
558
559
560
561
562
563
564
                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"

565
566
567
568
        model_info, arch = self.registry.inspect_model_cls(self.architectures)
        self._model_info = model_info
        self._architecture = arch

569
570
571
572
573
574
575
576
577
578
579
580
581
        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]
582
            assert selected_task != "encode"
583
584
585
586
587
588
589
            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":
590
591
592
593
594
595
596
597
598
599
600
601
602
            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,
        )
603

604
605
606
607
608
609
        # 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

610
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
611
612
        sliding_window_pattern = getattr(self.hf_text_config,
                                         "sliding_window_pattern", None)
613
614
        has_interleaved_attention = sliding_window_pattern is not None or (
            isinstance(sliding_window, list))
615

616
        if not self.disable_sliding_window and has_interleaved_attention:
617
618
            if (backend :=
                    envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"):
619
620
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
621

622
                logger.warning_once(
623
624
625
626
627
                    "%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,
                )
628
629
630
631
632
633
634
635
                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
636
637
638
639

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

640
                sliding_window = None
Woosuk Kwon's avatar
Woosuk Kwon committed
641

642
        self.original_max_model_len = self.max_model_len
643
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
644
        self.multimodal_config = self._init_multimodal_config()
645
646
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
647

648
        self.is_attention_free = self._init_attention_free()
649
        self.is_hybrid = self._init_is_hybrid()
650
        self.has_noops = self._init_has_noops()
651
652
        self.has_inner_state = self._init_has_inner_state()

653
654
655
        if (not current_platform.is_neuron() and self.override_neuron_config):
            raise ValueError(
                "`override_neuron_config` is only supported on Neuron.")
656

657
        self._verify_quantization()
658
        self._verify_cuda_graph()
659
        self._verify_bnb_config()
660

661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
    @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

677
678
679
680
681
682
683
684
685
686
    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"

687
688
    @property
    def registry(self):
689
        return me_models.ModelRegistry
690
691
692

    @property
    def architectures(self) -> list[str]:
693
        # architectures in the model config.
694
695
696
697
698
699
700
701
702
703
704
705
706
        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
707

708
709
710
711
712
713
    @property
    def architecture(self) -> str:
        # The architecture vllm actually used.
        return self._architecture

    @property
714
    def model_info(self):
715
716
        return self._model_info

717
718
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
719
        """Pull model/tokenizer from S3 to temporary directory when needed.
720

721
        Args:
722
723
            model: Model name or path
            tokenizer: Tokenizer name or path
724
        """
725
726
727
728
729
730
731
732
733
734
735
736
        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:
737
738
739
740
741
                s3_model.pull_files(model,
                                    ignore_pattern=[
                                        "*.pt", "*.safetensors", "*.bin",
                                        "*.tensors"
                                    ])
742
743
744
745
746
747
748
                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(
749
750
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors"])
751
            self.tokenizer = s3_tokenizer.dir
752

753
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
754
        if self.registry.is_multimodal_model(self.architectures):
755
            return MultiModalConfig(
756
                limit_per_prompt=self.limit_mm_per_prompt,
757
                media_io_kwargs=self.media_io_kwargs,
758
759
                mm_processor_kwargs=self.mm_processor_kwargs,
                disable_mm_preprocessor_cache=self.
760
761
                disable_mm_preprocessor_cache,
                interleave_mm_strings=self.interleave_mm_strings)
762

763
        if self.limit_mm_per_prompt:
764
765
            raise ValueError("`limit_mm_per_prompt` is only supported for "
                             "multimodal models.")
766
        if self.mm_processor_kwargs:
767
768
            raise ValueError("`mm_processor_kwargs` is only supported for "
                             "multimodal models.")
769
        if self.disable_mm_preprocessor_cache:
770
771
            raise ValueError("`disable_mm_preprocessor_cache` is only "
                             "supported for multimodal models.")
772
773
774
        if self.interleave_mm_strings:
            raise ValueError("`interleave_mm_strings` is only "
                             "supported for multimodal models.")
775
776

        return None
777

778
779
780
781
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

782
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
783
        if self.runner_type == "pooling":
784
785
786
787
788
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
789
790
791
792
793

            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():
794
795
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
796

797
            if self.is_matryoshka:
798
799
800
                if pooler_config.normalize is None:
                    pooler_config.normalize = True
                elif not pooler_config.normalize:
801
802
803
804
805
                    raise ValueError(
                        "`normalize` must be enabled (set to True) "
                        "for models that are compatible with "
                        "Matryoshka Representation.")

806
            return pooler_config
807

808
809
        return None

810
    def _init_attention_free(self) -> bool:
811
        return self.registry.is_attention_free_model(self.architectures)
812

813
    def _init_is_hybrid(self) -> bool:
814
        return self.registry.is_hybrid_model(self.architectures)
815

816
817
818
819
    def _init_has_noops(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return self.registry.is_noops_model(architectures)

820
    def _init_has_inner_state(self) -> bool:
821
        return self.registry.model_has_inner_state(self.architectures)
822

823
    def _verify_tokenizer_mode(self) -> None:
824
825
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
826
827
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
828
                f"one of {get_args(TokenizerMode)}.")
829
        self.tokenizer_mode = tokenizer_mode
830

831
832
833
834
835
836
    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)

837
    def _get_preferred_pooling_task(
838
        self,
839
        architectures: list[str],
840
    ) -> _ResolvedTask:
841
842
843
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
844
        if self.registry.is_transcription_model(architectures):
845
            return "transcription"
846

847
        suffix_to_preferred_task: list[tuple[str, _ResolvedTask]] = [
848
849
850
851
852
853
            # Other models follow this pattern
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]

        for suffix, pref_task in suffix_to_preferred_task:
854
            if self.architecture.endswith(suffix):
855
856
                return pref_task

857
        return "embed"
858

859
    def _get_supported_generation_tasks(
860
        self,
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
        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
877

878
879
880
881
    def _get_supported_pooling_tasks(
        self,
        task_option: TaskOption,
    ) -> list[_ResolvedTask]:
882
883
        registry = self.registry
        architectures = self.architectures
884

885
886
        supported_tasks = list[_ResolvedTask]()
        if registry.is_pooling_model(architectures):
887
            supported_tasks.append("encode")
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904

            # 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]]:
905
906
907
908
909
910
911
912
        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"]
            }
913

914
915
916
917
918
919
920
921
922
    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
        }
923

924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
    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}")
939

940
            return runner_option
941

942
943
944
945
        if task_option != "auto":
            for runner, runner_tasks in supported_tasks.items():
                if task_option in runner_tasks:
                    return runner
946
            else:
947
948
949
950
951
952
953
                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]}")
954

955
956
957
958
959
        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"

960
961
962
963
964
965
966
967
        suffix_to_preferred_runner: list[tuple[str, RunnerType]] = [
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("EmbeddingModel", "pooling"),
            ("RewardModel", "pooling"),
        ]
968

969
        for suffix, pref_runner in suffix_to_preferred_runner:
970
971
            if self.architecture.endswith(
                    suffix) and pref_runner in supported_runner_types:
972
                return pref_runner
973

974
975
976
977
        if "generate" in supported_runner_types:
            return "generate"
        if "pooling" in supported_runner_types:
            return "pooling"
978

979
        raise AssertionError("This line should not be reached")
980

981
982
983
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
984
            # compressed-tensors uses a "compression_config" key
985
            quant_cfg = getattr(self.hf_config, "compression_config", None)
986
987
        return quant_cfg

988
    def _verify_quantization(self) -> None:
989
        supported_quantization = me_quant.QUANTIZATION_METHODS
990
        optimized_quantization_methods = [
991
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
992
            "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
993
            "quark", "modelopt_fp4", "bitblas", "gptq_bitblas", "inc"
994
        ]
995
        if self.quantization is not None:
996
997
            self.quantization = cast(me_quant.QuantizationMethods,
                                     self.quantization)
998
999

        # Parse quantization method from the HF model config, if available.
1000
1001
        quant_cfg = self._parse_quant_hf_config()

1002
        if quant_cfg is not None:
1003
            # Use the community standard 'quant_method'
1004
            quant_method = quant_cfg.get("quant_method", "").lower()
1005
1006

            # Normalize library names
1007
1008
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
1009

1010
            quant_cfg["quant_method"] = quant_method
1011

1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
            # 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",
1024
1025
                "modelopt",
                "modelopt_fp4",
1026
1027
1028
1029
1030
1031
1032
1033
1034
            ]
            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

1035
            # Detect which checkpoint is it
1036
            for name in quantization_methods:
1037
                method = me_quant.get_quantization_config(name)
1038
1039
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
1040
1041
1042
1043
                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.
1044
                    if (name in get_args(me_quant.QuantizationMethods)
1045
1046
1047
1048
1049
1050
                            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.")
1051
1052
1053
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
1054

1055
            # Verify quantization configurations.
1056
            if self.quantization is None:
1057
1058
                self.quantization = quant_method
            elif self.quantization != quant_method:
1059
1060
                raise ValueError(
                    "Quantization method specified in the model config "
1061
                    f"({quant_method}) does not match the quantization "
1062
1063
1064
1065
1066
1067
1068
1069
                    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}.")
1070
            from vllm.platforms import current_platform
1071
            current_platform.verify_quantization(self.quantization)
1072
            if self.quantization not in optimized_quantization_methods:
1073
                logger.warning(
1074
                    "%s quantization is not fully "
1075
                    "optimized yet. The speed can be slower than "
1076
                    "non-quantized models.", self.quantization)
1077

1078
    def _verify_cuda_graph(self) -> None:
1079
1080
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
1081
        # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
1082
        ROCM_UNSUPPORTED_MODELS = ['mllama']
1083
1084
1085
1086
1087
1088
        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()):
1089
1090
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1091
                "to eager mode.", self.hf_config.model_type)
1092
            self.enforce_eager = True
1093

1094
1095
    def _verify_bnb_config(self) -> None:
        """
1096
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
1097
        yet support CUDA graph.
1098
        # TODO Remove this when bitsandbytes supports.
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
        """
        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(
1113
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1114
                "fallback to the eager mode.")
1115

1116
1117
            self.enforce_eager = True

1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
    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.")

1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
    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

1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
    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

1162
        # Reminder: Please update docs/features/compatibility_matrix.md
1163
        # If the feature combo become valid
1164
        from vllm.platforms import current_platform
1165
        if not current_platform.is_async_output_supported(self.enforce_eager):
1166
1167
1168
1169
1170
1171
1172
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

1173
        # Async postprocessor is not necessary for pooling models
1174
        # since there is no token generation
1175
        if self.runner_type == "pooling":
1176
1177
            self.use_async_output_proc = False

1178
        # Reminder: Please update docs/features/compatibility_matrix.md
1179
        # If the feature combo become valid
1180
1181
1182
        if speculative_config:
            self.use_async_output_proc = False

1183
1184
1185
1186
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
1187
1188
1189
1190
1191
1192

        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.")

1193
1194
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
1195
1196
1197
1198
1199
1200
1201
        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}).")

1202
        if parallel_config.enable_expert_parallel:
1203
1204
            self._verify_with_expert_parallelism()

1205
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1206
        if pipeline_parallel_size > 1:
1207
            if not self.registry.is_pp_supported_model(self.architectures):
1208
1209
1210
1211
1212
1213
                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
1214

1215
1216
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
1217
        """Get the sliding window size, or None if disabled."""
1218
1219
1220
1221

        # 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.
1222
1223
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
1224
            return None
1225
        return getattr(self.hf_text_config, "sliding_window", None)
1226

1227
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
1228
1229
1230
1231
1232
1233
1234
1235
        """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()

1236
    def get_vocab_size(self) -> int:
1237
        return self.hf_text_config.vocab_size
1238

1239
    def get_hidden_size(self) -> int:
1240
        return self.hf_text_config.hidden_size
1241

1242
1243
    @property
    def is_deepseek_mla(self) -> bool:
1244
1245
1246
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
bigmoyan's avatar
bigmoyan committed
1247
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp', 'kimi_k2'):
1248
1249
1250
1251
1252
1253
1254
1255
            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
1256

1257
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1258
        # TODO remove hard code
1259
        if self.is_deepseek_mla:
1260
1261
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1262
            if self.use_mla:
1263
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1264
1265
1266
1267
1268
            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
1269

1270
1271
1272
1273
1274
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1275
1276
1277
        if self.is_attention_free:
            return 0

1278
1279
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1280
            return self.hf_text_config.head_dim
1281

1282
        # FIXME(woosuk): This may not be true for all models.
1283
1284
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1285

1286
1287
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1288
        # For GPTBigCode & Falcon:
1289
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1290
1291
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1292
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1293
        new_decoder_arch_falcon = (
1294
            self.hf_config.model_type in falcon_model_types
1295
            and getattr(self.hf_config, "new_decoder_architecture", False))
1296
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1297
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1298
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1299
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1300
            return 1
1301

1302
        # For DBRX and MPT
1303
1304
1305
1306
1307
        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":
1308
1309
1310
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1311
1312
1313
1314
1315
1316
1317
1318
        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")

1319
1320
1321
        if self.is_attention_free:
            return 0

1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
        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:
1332
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1333
1334
1335
1336
1337
            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.
1338
        return self.hf_text_config.num_attention_heads
1339
1340
1341

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1342
1343
1344
1345
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1346
1347
1348
1349
1350
1351
1352
        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)
1353

1354
1355
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1356
1357
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1358

1359
    def get_layers_start_end_indices(
1360
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1361
        from vllm.distributed.utils import get_pp_indices
1362
        if (self.hf_text_config.model_type == "deepseek_mtp"
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1363
1364
                or self.hf_config.model_type == "mimo_mtp"
                or self.hf_config.model_type == "glm4_moe_mtp"):
1365
1366
1367
1368
1369
            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)
1370
1371
1372
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1373
1374
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1375
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1376

1377
1378
1379
    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
1380

1381
1382
1383
1384
1385
1386
1387
1388
    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
1389
1390
1391
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
        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
1402
1403
1404
1405
        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])
1406
        else:
1407
            # Hybrid model Jamba
1408
1409
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
            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
1435

1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
    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

1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
    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

1459
    def try_get_generation_config(self) -> dict[str, Any]:
1460
        if self.generation_config in ("auto", "vllm"):
1461
            config = try_get_generation_config(
1462
                self.hf_config_path or self.model,
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
                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()

1477
    def get_diff_sampling_param(self) -> dict[str, Any]:
1478
        """
1479
        This method returns a dictionary containing the parameters
1480
1481
        that differ from the default sampling parameters. If
        `generation_config` is `"vllm"`, an empty dictionary is returned.
1482
1483

        Returns:
1484
            dict[str, Any]: A dictionary with the differing sampling
1485
            parameters, if `generation_config` is `"vllm"` an empty dictionary.
1486
        """
1487
        if self.generation_config == "vllm":
1488
1489
1490
1491
1492
1493
1494
            config = {}
        else:
            config = self.try_get_generation_config()

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

1495
1496
1497
1498
1499
1500
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1501
            "max_new_tokens",
1502
1503
1504
1505
1506
1507
        ]
        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
            }
1508
1509
1510
1511
1512
            # 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")
1513
1514
        else:
            diff_sampling_param = {}
1515
1516
1517
1518
1519
1520
1521

        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`.")
1522
1523
        return diff_sampling_param

1524
    @property
1525
    def is_encoder_decoder(self) -> bool:
1526
        """Extract the HF encoder/decoder model flag."""
1527
        """
1528
        For Mllama, VLLM overrides HF's is_encoder_decoder flag and sets it to
1529
        True to enable cross-attention
1530
        Neuron needs all multimodal data to be in the decoder and does not
1531
1532
1533
1534
1535
1536
        need to explicitly enable cross-attention
        """
        if (current_platform.is_neuron()
                and self.hf_config.model_type == "mllama"):
            return False

1537
1538
1539
1540
1541
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1542

1543
1544
1545
1546
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1547
1548
    @property
    def is_cross_encoder(self) -> bool:
1549
        return self.task == "classify"
1550

1551
1552
    @property
    def use_mla(self) -> bool:
1553
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1554

1555
1556
1557
    @property
    def is_v1_compatible(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
1558
        return me_models.ModelRegistry.is_v1_compatible(architectures)
1559

1560
1561
    @property
    def is_matryoshka(self) -> bool:
1562
        return (bool(getattr(self.hf_config, "matryoshka_dimensions", None))
1563
1564
                or getattr(self.hf_config, "is_matryoshka", False))

1565
1566
1567
1568
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)

1569
1570
1571
1572
1573
1574
    @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)

1575
    def get_and_verify_max_len(self, max_model_len: int):
1576
1577
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
1578
        tokenizer_config = None
1579
1580
        if (self.runner_type == "pooling" and getattr(
                self.hf_config, "position_embedding_type", "") == "absolute"):
1581
1582
1583
1584
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
                revision=self.tokenizer_revision)
1585
1586
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
1587
            tokenizer_config=tokenizer_config,
1588
1589
1590
1591
1592
            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)
1593
        logger.info("Using max model len %s", max_model_len)
1594
1595
        return max_model_len

1596

1597
BlockSize = Literal[1, 8, 16, 32, 64, 128]
1598
CacheDType = Literal["auto", "fp8", "fp8_e4m3", "fp8_e5m2", "fp8_inc"]
1599
PrefixCachingHashAlgo = Literal["builtin", "sha256", "sha256_cbor_64bit"]
1600
1601
1602
1603


@config
@dataclass
1604
class CacheConfig:
1605
    """Configuration for the KV cache."""
1606

1607
    block_size: SkipValidation[BlockSize] = None  # type: ignore
1608
1609
1610
    """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.
1611
1612

    This config has no static default. If left unspecified by the user, it will
1613
    be set in `Platform.check_and_update_config()` based on the current
1614
    platform."""
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
    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
1628
    fp8 (=fp8_e4m3). Intel Gaudi (HPU) supports fp8 (using fp8_inc)."""
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
    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
1644
1645
1646
1647
1648
1649
    - "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."""
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
    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."""
1663
1664
    cpu_kvcache_space_bytes: Optional[int] = None
    """(CPU backend only) CPU key-value cache space."""
1665
1666
1667
    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."""
1668
1669
1670
1671
1672
1673

    # 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."""
1674

1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
    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.
        """
1687
        factors: list[Any] = []
1688
1689
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
1690
1691
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1692
1693
        return hash_str

1694
1695
1696
    def __post_init__(self) -> None:
        self.swap_space_bytes = self.swap_space * GiB_bytes

1697
        self._verify_cache_dtype()
1698
        self._verify_prefix_caching()
1699

1700
    def metrics_info(self):
1701
1702
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1703
1704
        return {key: str(value) for key, value in self.__dict__.items()}

1705
1706
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
1707
1708
1709
1710
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1711
1712
1713
1714
1715
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1716
1717
        return self

1718
1719
1720
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1721
        elif self.cache_dtype in get_args(CacheDType):
1722
            logger.info(
1723
1724
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1725
                "Meanwhile, it may cause accuracy drop without a proper "
1726
                "scaling factor.")
1727
1728
1729
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1730
1731
1732
1733
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

1734
        if self.sliding_window is not None and not envs.VLLM_USE_V1:
1735
1736
1737
1738
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

1739
1740
        if (self.enable_prefix_caching and self.prefix_caching_hash_algo
                not in get_args(PrefixCachingHashAlgo)):
1741
1742
            raise ValueError(
                "Unknown prefix caching hash algorithm: "
1743
1744
                f"{self.prefix_caching_hash_algo}. Must be one of "
                f"{get_args(PrefixCachingHashAlgo)}.")
1745

1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
    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

1756
1757
1758
        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.")
1759
1760
1761
        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:
1762
            logger.warning("Possibly too large swap space. %s", msg)
1763

1764

1765
1766
1767
1768
1769
1770
1771
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
1772
    SHARDED_STATE = "sharded_state"
1773
    GGUF = "gguf"
1774
    BITSANDBYTES = "bitsandbytes"
1775
    MISTRAL = "mistral"
1776
    RUNAI_STREAMER = "runai_streamer"
1777
    RUNAI_STREAMER_SHARDED = "runai_streamer_sharded"
1778
    FASTSAFETENSORS = "fastsafetensors"
1779
1780


1781
@config
1782
1783
@dataclass
class LoadConfig:
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
    """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."""
1809
    download_dir: Optional[str] = None
1810
1811
    """Directory to download and load the weights, default to the default
    cache directory of Hugging Face."""
1812
1813
    model_loader_extra_config: Union[dict, TensorizerConfig] = field(
        default_factory=dict)
1814
    """Extra config for model loader. This will be passed to the model loader
1815
    corresponding to the chosen load_format."""
1816
1817
1818
    device: Optional[str] = None
    """Device to which model weights will be loaded, default to
    device_config.device"""
1819
    ignore_patterns: Optional[Union[list[str], str]] = None
1820
1821
    """The list of patterns to ignore when loading the model. Default to
    "original/**/*" to avoid repeated loading of llama's checkpoints."""
1822
    use_tqdm_on_load: bool = True
1823
1824
    """Whether to enable tqdm for showing progress bar when loading model
    weights."""
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
    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
    """
1835

1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
    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.
1850
        factors: list[Any] = []
1851
1852
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1853
1854
        return hash_str

1855
    def __post_init__(self):
1856
1857
1858
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1859

1860
1861
1862
1863
1864
1865
1866
        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/**/*"]

1867

1868
1869
1870
DistributedExecutorBackend = Literal["ray", "mp", "uni", "external_launcher"]


1871
@config
1872
@dataclass
1873
class ParallelConfig:
1874
    """Configuration for the distributed execution."""
1875

1876
1877
1878
1879
1880
1881
1882
    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."""
1883
1884
    data_parallel_size_local: int = 1
    """Number of local data parallel groups."""
1885
1886
    data_parallel_rank: int = 0
    """Rank of the data parallel group."""
1887
1888
1889
    data_parallel_rank_local: Optional[int] = None
    """Local rank of the data parallel group,
    set only in SPMD mode."""
1890
    data_parallel_master_ip: str = "127.0.0.1"
1891
    """IP of the data parallel master."""
1892
1893
    data_parallel_rpc_port: int = 29550
    """Port for data parallel messaging."""
1894
1895
    data_parallel_master_port: int = 29500
    """Port of the data parallel master."""
Rui Qiao's avatar
Rui Qiao committed
1896
1897
    data_parallel_backend: str = "mp"
    """Backend to use for data parallel, either "mp" or "ray"."""
1898
1899
1900
1901
    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."""
1902
1903
    enable_expert_parallel: bool = False
    """Use expert parallelism instead of tensor parallelism for MoE layers."""
1904
1905
1906
1907
1908
1909
1910
1911
1912
    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.
1913

1914
1915
1916
1917
1918
1919
1920
1921
1922
    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.
    """

1923
    max_parallel_loading_workers: Optional[int] = None
1924
    """Maximum number of parallel loading workers when loading model
1925
1926
    sequentially in multiple batches. To avoid RAM OOM when using tensor
    parallel and large models."""
1927
1928

    disable_custom_all_reduce: bool = False
1929
    """Disable the custom all-reduce kernel and fall back to NCCL."""
1930
1931

    ray_workers_use_nsight: bool = False
1932
    """Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler."""
1933
1934

    placement_group: Optional["PlacementGroup"] = None
1935
    """ray distributed model workers placement group."""
1936

1937
    distributed_executor_backend: Optional[Union[DistributedExecutorBackend,
1938
                                                 type["ExecutorBase"]]] = None
1939
1940
1941
1942
1943
1944
    """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
1945
    only support Ray for distributed inference."""
1946
1947

    worker_cls: str = "auto"
1948
1949
    """The full name of the worker class to use. If "auto", the worker class
    will be determined based on the platform."""
1950
    sd_worker_cls: str = "auto"
Ning Xie's avatar
Ning Xie committed
1951
    """The full name of the worker class to use for speculative decoding.
1952
    If "auto", the worker class will be determined based on the platform."""
1953
    worker_extension_cls: str = ""
1954
1955
1956
1957
    """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."""
1958
1959

    world_size: int = field(init=False)
1960
    """world_size is TPxPP, it affects the number of workers we create."""
1961
1962

    rank: int = 0
1963
    """Global rank in distributed setup."""
1964

1965
    enable_multimodal_encoder_data_parallel: bool = False
1966
    """ Use data parallelism instead of tensor parallelism for vision encoder.
1967
1968
    Only support LLama4 for now"""

1969
1970
1971
1972
1973
1974
    @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

1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
    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":
1989
1990
1991
1992
1993
1994
1995
1996
1997
        # 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

1998
1999
2000
        from vllm.distributed.utils import (
            stateless_init_torch_distributed_process_group)

2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
        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
2024
2025
2026

    @staticmethod
    def has_unfinished_dp(dp_group: "ProcessGroup",
youkaichao's avatar
youkaichao committed
2027
                          has_unfinished: bool) -> bool:
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
        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

2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
    @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()

2052
2053
2054
2055
2056
2057
2058
2059
    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.
        """
2060
        factors: list[Any] = []
2061
2062
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
2063
        factors.append(self.enable_expert_parallel)
2064
2065
        factors.append(self.data_parallel_size)
        factors.append(envs.VLLM_ALL2ALL_BACKEND)
2066
2067
        return hashlib.sha256(str(factors).encode()).hexdigest()

2068
2069
2070
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
2071

2072
2073
2074
2075
2076
2077
        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:
2078
2079
            # Data parallel was specified in the engine args.
            self.data_parallel_master_port = get_open_port()
2080
2081
2082
2083
2084

            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})")
2085
2086
2087
2088
2089
2090
2091
2092
        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

2093
2094
2095
2096
            if self.data_parallel_external_lb:
                raise ValueError("data_parallel_external_lb can only "
                                 "be set when data_parallel_size > 1")

2097
2098
2099
2100
2101
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

2102
2103
2104
2105
2106
2107
2108
2109
2110
        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}.")
2111
2112
2113
2114
2115
2116
2117
2118
2119
            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}."
                )
2120
2121
2122
2123
2124
        else:
            if self.num_redundant_experts != 0:
                raise ValueError(
                    "num_redundant_experts should be used with EPLB."
                    f"{self.num_redundant_experts}.")
2125
        if self.distributed_executor_backend is None and self.world_size > 1:
2126
2127
2128
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

2129
            from vllm.executor import ray_utils
2130
            backend: DistributedExecutorBackend = "mp"
2131
            ray_found = ray_utils.ray_is_available()
2132
2133
            if current_platform.is_neuron():
                # neuron uses single process to control multiple devices
2134
2135
                backend = "uni"
            elif current_platform.is_tpu() and envs.VLLM_XLA_USE_SPMD:
2136
2137
2138
                backend = "uni"
            elif (current_platform.is_cuda()
                  and cuda_device_count_stateless() < self.world_size):
2139
                if not ray_found:
2140
2141
                    raise ValueError("Unable to load Ray: "
                                     f"{ray_utils.ray_import_err}. Ray is "
2142
2143
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
2144
                                     "ray`.")
2145
                backend = "ray"
Rui Qiao's avatar
Rui Qiao committed
2146
2147
2148
2149
            elif self.data_parallel_backend == "ray":
                logger.info("Using ray distributed inference because "
                            "data_parallel_backend is ray")
                backend = "ray"
2150
            elif ray_found:
2151
                if self.placement_group:
2152
                    backend = "ray"
2153
2154
2155
2156
2157
2158
                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"
2159
            self.distributed_executor_backend = backend
2160
2161
            logger.debug("Defaulting to use %s for distributed inference",
                         backend)
2162

2163
2164
2165
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

2166
2167
2168
2169
2170
2171
    @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)

2172
2173
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2174
2175
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
2176
        from vllm.platforms import current_platform
2177
        if self.distributed_executor_backend not in (
2178
2179
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
2180
2181
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
2182
            raise ValueError(
2183
2184
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
2185
2186
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
2187
        if self.use_ray:
2188
2189
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
2190
2191

        if not current_platform.use_custom_allreduce():
2192
            self.disable_custom_all_reduce = True
Aaron Pham's avatar
Aaron Pham committed
2193
            logger.debug(
2194
                "Disabled the custom all-reduce kernel because it is not "
2195
                "supported on current platform.")
2196
        if self.ray_workers_use_nsight and not self.use_ray:
2197
2198
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
2199

2200
        return self
2201

2202

2203
PreemptionMode = Literal["swap", "recompute"]
2204
2205
2206
2207
SchedulerPolicy = Literal["fcfs", "priority"]


@config
2208
@dataclass
2209
class SchedulerConfig:
2210
    """Scheduler configuration."""
2211

2212
2213
    runner_type: RunnerType = "generate"
    """The runner type to launch for the model."""
2214

2215
    max_num_batched_tokens: SkipValidation[int] = None  # type: ignore
2216
    """Maximum number of tokens to be processed in a single iteration.
2217

2218
2219
    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."""
2220

2221
    max_num_seqs: SkipValidation[int] = None  # type: ignore
2222
    """Maximum number of sequences to be processed in a single iteration.
2223

2224
2225
    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."""
2226

2227
    max_model_len: SkipValidation[int] = None  # type: ignore
2228
2229
2230
    """Maximum length of a sequence (including prompt and generated text). This
    is primarily set in `ModelConfig` and that value should be manually
    duplicated here."""
2231

2232
    max_num_partial_prefills: int = 1
2233
2234
    """For chunked prefill, the maximum number of sequences that can be
    partially prefilled concurrently."""
2235
2236

    max_long_partial_prefills: int = 1
2237
2238
2239
2240
    """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."""
2241
2242

    long_prefill_token_threshold: int = 0
2243
2244
    """For chunked prefill, a request is considered long if the prompt is
    longer than this number of tokens."""
2245

2246
    num_lookahead_slots: int = 0
2247
2248
2249
2250
2251
2252
2253
    """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."""
2254

2255
2256
2257
2258
    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
2259
    pattern: [1, 2, 4] + [i for i in range(8, cuda_graph_sizes + 1, 8)]
2260
    3. more than one value (e.g. 1 2 128) is provided, then the capture list
2261
    will follow the provided list."""
2262

2263
    delay_factor: float = 0.0
2264
2265
    """Apply a delay (of delay factor multiplied by previous
    prompt latency) before scheduling next prompt."""
2266

2267
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2268
2269
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""
2270
2271

    is_multimodal_model: bool = False
2272
2273
2274
2275
2276
    """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.
2277

2278
2279
2280
2281
2282
2283
2284
2285
2286
    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."""
2287

2288
    preemption_mode: Optional[PreemptionMode] = None
2289
2290
2291
2292
2293
2294
    """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."""
2295
2296

    num_scheduler_steps: int = 1
2297
    """Maximum number of forward steps per scheduler call."""
2298

2299
2300
    multi_step_stream_outputs: bool = True
    """If False, then multi-step will stream outputs at the end of all steps"""
2301
2302

    send_delta_data: bool = False
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
    """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)."""
2314
2315

    chunked_prefill_enabled: bool = field(init=False)
2316
    """True if chunked prefill is enabled."""
2317

2318
    disable_chunked_mm_input: bool = False
2319
2320
2321
2322
2323
2324
    """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."""
2325

2326
2327
    # scheduler class or path. "vllm.core.scheduler.Scheduler" (default)
    # or "mod.custom_class".
2328
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
2329
2330
2331
    """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"."""
2332

2333
2334
2335
2336
2337
2338
    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.
    """

2339
2340
2341
2342
2343
2344
2345
    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.
    """

2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
    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.
2360
        factors: list[Any] = []
2361
2362
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2363
2364
        return hash_str

2365
    def __post_init__(self) -> None:
2366
2367
2368
2369
2370
2371
        if self.max_model_len is None:
            self.max_model_len = 8192

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

2372
2373
2374
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
2375
2376
2377
2378
                    # 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.
2379
                    self.max_num_batched_tokens = max(
2380
                        self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2381
                else:
2382
                    self.max_num_batched_tokens = (
2383
                        DEFAULT_MAX_NUM_BATCHED_TOKENS)
2384
            else:
2385
                # If max_model_len is too short, use
2386
                # DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
2387
                # for higher throughput.
2388
                self.max_num_batched_tokens = max(
2389
                    self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2390

2391
2392
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
2393
2394
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2395
                    POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
2396
                )
2397
            if self.is_multimodal_model:
2398
                # The value needs to be at least the number of multimodal tokens
2399
2400
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2401
                    MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
2402
2403
                )

2404
2405
2406
2407
2408
2409
2410
            # 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)

2411
2412
2413
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

2414
        if self.enable_chunked_prefill:
2415
2416
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
2417
                self.max_num_batched_tokens)
2418

2419
        self.chunked_prefill_enabled = self.enable_chunked_prefill
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
        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)

2432
2433
2434
2435
2436
2437
2438
        # 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)]

2439
2440
2441
2442
        if self.async_scheduling:
            self.scheduler_cls = (
                "vllm.v1.core.sched.async_scheduler.AsyncScheduler")

2443
2444
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2445
2446
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
2447
2448
2449
2450
2451
2452
2453
            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.")
2454

2455
2456
2457
2458
2459
        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}).")
2460

2461
2462
        if self.max_num_batched_tokens > self.max_num_seqs * self.max_model_len:
            logger.warning(
2463
                "max_num_batched_tokens (%d) exceeds max_num_seqs "
2464
2465
2466
2467
                "* max_model_len (%d). This may lead to unexpected behavior.",
                self.max_num_batched_tokens,
                self.max_num_seqs * self.max_model_len)

2468
2469
2470
2471
2472
2473
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

2474
2475
2476
2477
2478
2479
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
        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}).")

2503
2504
        return self

2505
2506
2507
2508
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

2509

2510
Device = Literal["auto", "cuda", "neuron", "cpu", "tpu", "xpu"]
2511
2512
2513


@config
2514
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
2515
class DeviceConfig:
2516
2517
    """Configuration for the device to use for vLLM execution."""

2518
    device: SkipValidation[Optional[Union[Device, torch.device]]] = "auto"
2519
    """Device type for vLLM execution.
2520
2521
2522
    This parameter is deprecated and will be
    removed in a future release.
    It will now be set automatically based
2523
    on the current platform."""
2524
2525
2526
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
2527

2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
    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.
2543
        factors: list[Any] = []
2544
2545
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2546
2547
        return hash_str

2548
2549
    def __post_init__(self):
        if self.device == "auto":
2550
            # Automated device type detection
2551
            from vllm.platforms import current_platform
2552
            self.device_type = current_platform.device_type
2553
            if not self.device_type:
2554
2555
2556
2557
                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.")
2558
2559
        else:
            # Device type is assigned explicitly
2560
2561
2562
2563
            if isinstance(self.device, str):
                self.device_type = self.device
            elif isinstance(self.device, torch.device):
                self.device_type = self.device.type
2564
2565

        # Some device types require processing inputs on CPU
2566
        if self.device_type in ["neuron"]:
2567
            self.device = torch.device("cpu")
2568
2569
        elif self.device_type in ["tpu"]:
            self.device = None
2570
2571
2572
2573
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

2574

2575
2576
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
                            "mlp_speculator", "draft_model", "deepseek_mtp"]
2577
2578
2579


@config
2580
@dataclass
2581
class SpeculativeConfig:
2582
    """Configuration for speculative decoding."""
2583

2584
    # General speculative decoding control
2585
    num_speculative_tokens: SkipValidation[int] = None  # type: ignore
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
    """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."""
2599
    draft_tensor_parallel_size: Optional[int] = None
2600
2601
    """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."""
2602
    disable_logprobs: bool = True
2603
2604
2605
    """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."""
2606

2607
    # Draft model configuration
2608
    quantization: Optional[me_quant.QuantizationMethods] = None
2609
2610
2611
    """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."""
2612
    max_model_len: Optional[int] = None
2613
2614
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
2615
    revision: Optional[str] = None
2616
2617
2618
    """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."""
2619
    code_revision: Optional[str] = None
2620
2621
2622
    """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."""
2623

2624
    # Advanced control
2625
    disable_by_batch_size: Optional[int] = None
2626
2627
2628
2629
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
2630
    prompt_lookup_max: Optional[int] = None
2631
2632
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
2633
    prompt_lookup_min: Optional[int] = None
2634
2635
2636
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

2637
    speculative_token_tree: Optional[str] = None
2638
    """Specifies the tree structure for speculative token generation.
2639
    """
2640
    # required configuration params passed from engine
2641
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2642
    """The configuration of the target model."""
2643
2644
    target_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2645
    """The parallel configuration for the target model."""
2646
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2647
2648
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
2649
    disable_log_stats: SkipValidation[bool] = None  # type: ignore
2650
2651
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
2652
2653

    # params generated in the post-init stage
2654
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2655
    """The configuration of the draft model initialized internal."""
2656
2657
    draft_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2658
    """The parallel configuration for the draft model initialized internal."""
2659

2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
    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.
        """
2672
        factors: list[Any] = []
2673
2674
2675
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2676
2677
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2678
2679
        return hash_str

2680
2681
2682
2683
2684
    @classmethod
    def from_dict(cls, dict_value: dict) -> "SpeculativeConfig":
        """Parse the CLI value for the speculative config."""
        return cls(**dict_value)

2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
    @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"]
            })
2695
2696
2697
2698
2699
2700
2701
2702
2703

        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
2704
2705
2706
2707
2708
2709
2710
2711
2712

        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"]
            })
2713

2714
2715
        return hf_config

2716
    def __post_init__(self):
2717

2718
2719
2720
2721
2722
2723
2724
        # 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.
2725
2726
2727
2728

        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
2729
            if self.target_model_config and \
2730
2731
2732
2733
                (self.target_model_config.hf_text_config.model_type \
                        == "deepseek_v3" or
                    self.target_model_config.hf_text_config.model_type \
                        == "mimo"):
2734
2735
2736
2737
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2738
            else:
2739
2740
2741
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2742
2743
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2744
2745
2746
2747
2748
2749
2750
        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"
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
            # 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
2765
            if self.prompt_lookup_min < 1:
2766
2767
2768
2769
2770
                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")
2771
            if self.prompt_lookup_min > self.prompt_lookup_max:
2772
2773
2774
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2775

2776
2777
2778
            # 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.
2779
2780
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2781
        else:
2782
2783
2784
2785
2786
2787
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
2788
                    runner="draft",
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
                    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,
                )
2810

2811
                # Automatically detect the method
2812
                if self.method in ('eagle', 'eagle3'):
2813
                    pass
2814
2815
                elif "eagle-" in self.draft_model_config.model.lower() or \
                        "eagle3-" in self.draft_model_config.model.lower():
2816
2817
2818
2819
2820
2821
                    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"
2822
                elif (self.draft_model_config.hf_config.model_type
Yuxuan Zhang's avatar
Yuxuan Zhang committed
2823
                      in ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp")):
Jiayi Yao's avatar
Jiayi Yao committed
2824
2825
2826
2827
2828
2829
2830
                    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."
                            )
2831
                else:
2832
                    self.method = "draft_model"
2833
2834
2835
2836
2837
                    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.")
2838
2839

                # Replace hf_config for EAGLE draft_model
2840
                if self.method in ("eagle", "eagle3"):
2841
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2842
                        raise ValueError(
2843
2844
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2845
2846
2847
2848

                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
                    if isinstance(self.draft_model_config.hf_config,
2849
                                  EAGLEConfig):
2850
2851
2852
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2853
                            self.draft_model_config.hf_config,
2854
2855
                            method=self.method,
                            model_type="eagle")
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
                        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
                )
2883

2884
2885
2886
2887
2888
2889
                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,
                    ))
2890

2891
2892
2893
2894
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2895

2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
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
    @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,
        )

2931
    @staticmethod
2932
    def _verify_and_get_draft_tp(
2933
2934
2935
2936
2937
2938
            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.
2939
        """
2940
2941
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2942
        if speculative_draft_tensor_parallel_size is None:
2943
2944
2945
2946
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2947
2948
2949
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2950
2951
2952
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2953
2954
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2955
            raise ValueError(
2956
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2957
                f"other value than 1 or target model tensor_parallel_size")
2958
        return speculative_draft_tensor_parallel_size
2959

2960
2961
2962
2963
2964
2965
2966
2967
2968
    @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.
        """
2969
2970
2971
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2972
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2973
2974
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
            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

2986
2987
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2988
2989
2990
2991
2992
2993
        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.")

2994
2995
2996
2997
2998
2999
3000
        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)
3001
3002
3003
3004
3005
3006

        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=}")
3007

3008
3009
3010
3011
3012
3013
        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=}")

3014
3015
        return self

3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
    @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

3026
    def use_eagle(self) -> bool:
Jiayi Yao's avatar
Jiayi Yao committed
3027
        return self.method in ("eagle", "eagle3", "deepseek_mtp")
3028

3029
    def __repr__(self) -> str:
3030
3031
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
3032
        num_spec_tokens = self.num_speculative_tokens
3033
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
3034
3035


3036
3037
3038
3039
LoRADType = Literal["auto", "float16", "bfloat16"]


@config
3040
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3041
class LoRAConfig:
3042
3043
3044
3045
3046
3047
    """Configuration for LoRA."""

    max_lora_rank: int = 16
    """Max LoRA rank."""
    max_loras: int = 1
    """Max number of LoRAs in a single batch."""
3048
    fully_sharded_loras: bool = False
3049
3050
3051
3052
    """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.
    """
3053
    max_cpu_loras: Optional[int] = None
3054
3055
3056
3057
    """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."""
3058
    lora_extra_vocab_size: int = 256
3059
3060
    """Maximum size of extra vocabulary that can be present in a LoRA adapter
    (added to the base model vocabulary)."""
3061
3062
    lora_vocab_padding_size: ClassVar[int] = current_platform\
        .get_lora_vocab_padding_size()
3063

3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
    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."""
3074
    bias_enabled: bool = False
3075
    """Enable bias for LoRA adapters."""
3076

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

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
3089
        factors: list[Any] = []
3090
3091
3092
3093
3094
        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)
3095
        factors.append(self.lora_vocab_padding_size)
3096
        factors.append(self.bias_enabled)
3097
3098
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3099
3100
        return hash_str

3101
    def __post_init__(self):
3102
        # Setting the maximum rank to 512 should be able to satisfy the vast
3103
        # majority of applications.
3104
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
3105
        possible_lora_extra_vocab_size = (256, 512)
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
        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
3121
                f"max_loras ({self.max_loras})")
3122

3123
    def verify_with_cache_config(self, cache_config: CacheConfig):
3124
3125
3126
        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.")
3127

3128
3129
3130
3131
3132
3133
3134
    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)


3135
@config
3136
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3137
class PromptAdapterConfig:
3138
3139
    """Configuration for PromptAdapters."""

3140
3141
3142
3143
    max_prompt_adapters: int = 1
    """Max number of PromptAdapters in a batch."""
    max_prompt_adapter_token: int = 0
    """Max number of PromptAdapters tokens."""
3144
    max_cpu_prompt_adapters: Optional[int] = None
3145
3146
3147
3148
3149
    """Maximum number of PromptAdapters to store in CPU memory. Must be >= than
    `max_prompt_adapters`."""
    prompt_adapter_dtype: Union[torch.dtype, str] = "auto"
    """Data type for PromptAdapter. If auto, will default to base model dtype.
    """
3150

3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
    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.
3165
        factors: list[Any] = []
3166
3167
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3168
3169
        return hash_str

3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
    def __post_init__(self):

        if self.max_prompt_adapters < 1:
            raise ValueError(f"max_prompt_adapters "
                             f"({self.max_prompt_adapters}) must be >= 1.")
        if self.max_prompt_adapter_token == 0:
            raise ValueError("max_prompt_adapter_token must be set.")
        if self.max_cpu_prompt_adapters is None:
            self.max_cpu_prompt_adapters = self.max_prompt_adapters

    def verify_with_model_config(self, model_config: ModelConfig):
3181
        if self.prompt_adapter_dtype == "auto":
3182
3183
3184
3185
3186
3187
            self.prompt_adapter_dtype = model_config.dtype
        elif isinstance(self.prompt_adapter_dtype, str):
            self.prompt_adapter_dtype = getattr(torch,
                                                self.prompt_adapter_dtype)


3188
@config
3189
@dataclass
3190
class MultiModalConfig:
3191
3192
    """Controls the behavior of multimodal models."""

3193
3194
    limit_per_prompt: dict[str, int] = \
        cast(dict[str, int], get_field(ModelConfig, "limit_mm_per_prompt"))
3195
    """
3196
    The maximum number of input items allowed per prompt for each modality.
3197
    Defaults to 1 (V0) or 999 (V1) for each modality.
3198
3199

    For example, to allow up to 16 images and 2 videos per prompt:
3200
    `{"images": 16, "videos": 2}`
3201
3202
    """

3203
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
3204
3205
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
3206
3207
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """

3208
3209
3210
    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
3211
    `transformers.AutoProcessor.from_pretrained`.
3212
3213
3214
3215

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

    For example, for Phi-3-Vision:
3216
    `{"num_crops": 4}`.
3217
3218
3219
3220
    """

    disable_mm_preprocessor_cache: bool = False
    """
3221
    If `True`, disable caching of the processed multi-modal inputs.
3222
3223
    """

3224
3225
3226
3227
3228
    interleave_mm_strings: bool = False
    """
    Enable fully interleaved support for multimodal prompts.
    """

3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
    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.
3243
        factors: list[Any] = []
3244
3245
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3246
3247
        return hash_str

3248
3249
3250
3251
3252
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
3253
3254
3255
3256
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
3257

3258
    # TODO: Add configs to init vision tower or not.
3259

3260

3261
@config
3262
3263
@dataclass
class PoolerConfig:
3264
    """Controls the behavior of output pooling in pooling models."""
3265
3266

    pooling_type: Optional[str] = None
3267
    """
3268
    The pooling method of the pooling model. This should be a key in
3269
    [`vllm.model_executor.layers.pooler.PoolingType`][].
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
    """

    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
    """
3286
    If set, only the score corresponding to the ``step_tag_id`` in the
3287
3288
3289
3290
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

3291
    returned_token_ids: Optional[list[int]] = None
3292
    """
3293
3294
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
3295
3296
3297
    ``math-shepherd-mistral-7b-prm`` model.
    """

3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
    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.
3312
        factors: list[Any] = []
3313
3314
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3315
3316
        return hash_str

3317

3318
3319
3320
3321
3322
3323
3324
3325
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

3326
3327
3328
3329
3330
3331
3332
# 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.",
}
3333

3334

3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
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,
3353
    config: PretrainedConfig,
3354
3355
3356
    *,
    revision: Optional[str],
):
3357
3358
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
3359
    config_dtype = getattr(config, "torch_dtype", None)
3360

3361
    # Fallbacks for multi-modal models if the root config
3362
    # does not define torch_dtype
3363
3364
    if config_dtype is None:
        config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
3365
3366
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)
3367
3368
    if config_dtype is None and hasattr(config, "encoder_config"):
        config_dtype = getattr(config.encoder_config, "torch_dtype", None)
3369

3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
    # 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)

3385
3386
3387
    if config_dtype is None:
        config_dtype = torch.float32

3388
    return config_dtype
3389

Shinichi Hemmi's avatar
Shinichi Hemmi committed
3390

3391
3392
3393
3394
3395
3396
3397
def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform
3398

3399
3400
3401
3402
    supported_dtypes = [
        dtype for dtype in current_platform.supported_dtypes
        if _is_valid_dtype(model_type, dtype)
    ]
3403

3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
    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,
            )
3457
        else:
3458
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
3459
                raise ValueError(f"Unknown dtype: {dtype!r}")
3460
3461
3462
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
3463
    else:
3464
        raise ValueError(f"Unknown dtype: {dtype}")
3465

3466
3467
    _check_valid_dtype(model_type, torch_dtype)

3468
3469
3470
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
3471
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
3472
3473
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
3474
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
3475
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
3476
            # Casting between float16 and bfloat16 is allowed with a warning.
3477
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
3478
3479

    return torch_dtype
3480
3481
3482
3483


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
3484
    tokenizer_config: Optional[dict],
3485
    max_model_len: Optional[int],
3486
    disable_sliding_window: bool,
3487
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
3488
    spec_target_max_model_len: Optional[int] = None,
3489
    encoder_config: Optional[Any] = None,
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
) -> 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",
3500
3501
        # ChatGLM2
        "seq_length",
3502
3503
        # Command-R
        "model_max_length",
3504
3505
        # Whisper
        "max_target_positions",
3506
3507
3508
3509
3510
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
3511
    # Choose the smallest "max_length" from the possible keys
3512
    max_len_key = None
3513
    for key in possible_keys:
3514
3515
3516
3517
3518
        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
3519
3520
3521
3522
    # 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
3523
3524
3525
3526

    # 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:
3527
3528

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
3529
        max_len_key = "sliding_window" \
3530
3531
3532
            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)
3533

3534
3535
3536
3537
3538
3539
3540
    # 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)

3541
3542
    # If none of the keys were found in the config, use a default and
    # log a warning.
3543
    if derived_max_model_len == float("inf"):
3544
3545
3546
3547
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

3548
3549
3550
3551
3552
        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

3553
3554
3555
3556
        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: "
3557
            "%s. Assuming the model's maximum length is %d.", possible_keys,
3558
            default_max_len)
3559
        derived_max_model_len = default_max_len
3560

3561
    rope_scaling = getattr(hf_config, "rope_scaling", None)
3562
3563
3564
    # 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:
3565
3566
3567
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577

        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.")

3578
3579
3580
3581
            # 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)

3582
3583
3584
3585
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
3586

3587
3588
3589
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

3590
3591
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
3592
    if max_model_len is None:
3593
        max_model_len = int(derived_max_model_len)
3594
3595
3596
3597
3598
3599
3600
3601
        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)
3602
    elif max_model_len > derived_max_model_len:
3603
3604
3605
3606
3607
        # 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:
3608
3609
3610
3611
3612
3613
3614
            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.")
3615
        else:
3616
            msg = (
3617
                f"User-specified max_model_len ({max_model_len}) is greater "
3618
3619
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
3620
                f"{model_max_length} in model's config.json). This may lead "
3621
3622
3623
3624
3625
3626
3627
3628
3629
                "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")
3630
    return int(max_model_len)
3631
3632


3633
def get_min_sliding_window(
3634
        sliding_window: Union[int, list[Optional[int]]]) -> int:
3635
3636
3637
3638
3639
3640
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


3641
def get_served_model_name(model: str,
3642
                          served_model_name: Optional[Union[str, list[str]]]):
3643
    """
3644
3645
3646
3647
    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
3648
3649
3650
3651
3652
3653
3654
3655
3656
    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


3657
GuidedDecodingBackendV0 = Literal["auto", "outlines", "lm-format-enforcer",
3658
                                  "xgrammar", "guidance"]
3659
3660

GuidedDecodingBackendV1 = Literal["auto", "xgrammar", "guidance", "outlines"]
3661
3662
GuidedDecodingBackend = Literal[GuidedDecodingBackendV0,
                                GuidedDecodingBackendV1]
3663
3664
3665


@config
3666
3667
@dataclass
class DecodingConfig:
3668
    """Dataclass which contains the decoding strategy of the engine."""
3669

3670
    backend: GuidedDecodingBackend = "auto" if envs.VLLM_USE_V1 else "xgrammar"
3671
3672
3673
3674
    """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."""
3675

3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
    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`."""

3688
    reasoning_backend: str = ""
3689
    """Select the reasoning parser depending on the model that you're using.
3690
    This is used to parse the reasoning content into OpenAI API format."""
3691

3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
    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.
3706
        factors: list[Any] = []
3707
3708
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3709
3710
        return hash_str

3711
    def __post_init__(self):
3712
        if envs.VLLM_USE_V1:
3713
            valid_guided_backends = get_args(GuidedDecodingBackendV1)
3714
        else:
3715
            valid_guided_backends = get_args(GuidedDecodingBackendV0)
3716
3717
        if self.backend not in valid_guided_backends:
            raise ValueError(f"Invalid backend '{self.backend}',"
3718
                             f" must be one of {valid_guided_backends}")
3719
3720
3721
3722
3723
3724
3725
3726
        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.")

3727

3728
3729
3730
3731
DetailedTraceModules = Literal["model", "worker", "all"]


@config
3732
3733
@dataclass
class ObservabilityConfig:
3734
    """Configuration for observability - metrics and tracing."""
3735

3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
    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)
3751

3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
    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))
3777

3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
    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.
3792
        factors: list[Any] = []
3793
3794
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3795
3796
        return hash_str

3797
    def __post_init__(self):
3798
3799
3800
3801
3802
        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()

3803
        from vllm.tracing import is_otel_available, otel_import_error_traceback
3804
3805
3806
3807
3808
        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}")
3809

3810
3811
3812
3813
3814
3815
    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(","))

3816

3817
3818
3819
3820
3821
3822
3823
3824
KVProducer = Literal["kv_producer", "kv_both"]
KVConsumer = Literal["kv_consumer", "kv_both"]
KVRole = Literal[KVProducer, KVConsumer]


@config
@dataclass
class KVTransferConfig:
3825
3826
3827
    """Configuration for distributed KV cache transfer."""

    kv_connector: Optional[str] = None
3828
3829
    """The KV connector for vLLM to transmit KV caches between vLLM instances.
    """
3830

3831
    engine_id: Optional[str] = None
Robert Shaw's avatar
Robert Shaw committed
3832
3833
    """The engine id for KV transfers."""

3834
    kv_buffer_device: Optional[str] = "cuda"
3835
3836
    """The device used by kv connector to buffer the KV cache.
    Currently only support 'cuda'."""
3837
3838

    kv_buffer_size: float = 1e9
3839
3840
    """The buffer size for TorchDistributedConnector. Measured in number of
    bytes. Recommended value: 1e9 (about 1GB)."""
3841

3842
3843
    kv_role: Optional[KVRole] = None
    """Whether this vLLM instance produces, consumes KV cache, or both. Choices
Robert Shaw's avatar
Robert Shaw committed
3844
    are 'kv_producer', 'kv_consumer', and 'kv_both'."""
3845
3846

    kv_rank: Optional[int] = None
3847
3848
3849
    """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."""
3850
3851

    kv_parallel_size: int = 1
3852
3853
    """The number of parallel instances for KV cache transfer. For
    PyNcclConnector, this should be 2."""
3854
3855

    kv_ip: str = "127.0.0.1"
3856
    """The KV connector ip, used to build distributed connection."""
3857
3858

    kv_port: int = 14579
3859
    """The KV connector port, used to build distributed connection."""
3860

3861
3862
    kv_connector_extra_config: dict[str, Any] = field(default_factory=dict)
    """any extra config that the connector may need."""
3863

3864
3865
3866
3867
    kv_connector_module_path: Optional[str] = None
    """The Python module path to dynamically load the KV connector from.
    Only supported in V1."""

3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
    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.
3882
        factors: list[Any] = []
3883
3884
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3885
3886
        return hash_str

3887
    def __post_init__(self) -> None:
3888
3889
3890
        if self.engine_id is None:
            self.engine_id = str(uuid.uuid4())

3891
3892
3893
        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)}")
3894
3895
3896

        if self.kv_connector is not None and self.kv_role is None:
            raise ValueError("Please specify kv_disagg_role when kv_connector "
3897
                             f"is set, supported roles are {get_args(KVRole)}")
3898
3899
3900
3901

    @property
    def is_kv_transfer_instance(self) -> bool:
        return self.kv_connector is not None and \
3902
            self.kv_role in get_args(KVRole)
3903
3904
3905
3906

    @property
    def is_kv_producer(self) -> bool:
        return self.kv_connector is not None and \
3907
            self.kv_role in get_args(KVProducer)
3908
3909
3910
3911

    @property
    def is_kv_consumer(self) -> bool:
        return self.kv_connector is not None and \
3912
            self.kv_role in get_args(KVConsumer)
3913

3914
3915
3916
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

3917

3918
3919
3920
@config
@dataclass
class KVEventsConfig:
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
    """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.
    """


3960
3961
3962
3963
3964
3965
3966
3967
class CompilationLevel:
    # constants for the levels of the compilation process
    NO_COMPILATION = 0
    DYNAMO_AS_IS = 1
    DYNAMO_ONCE = 2
    PIECEWISE = 3


3968
3969
3970
3971
3972
3973
3974
3975
3976
@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."""

3977
    enable_fusion: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
3978
3979
3980
    """Whether to enable the custom fusion (RMSNorm/SiluMul+quant) pass."""
    enable_attn_fusion: bool = False
    """Whether to enable the custom attention+quant fusion pass."""
3981
    enable_noop: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
3982
3983
3984
    """Whether to enable the custom no-op elimination pass."""
    enable_sequence_parallelism: bool = False
    """Whether to enable sequence parallelism."""
3985
3986
    enable_async_tp: bool = False
    """Whether to enable async TP."""
3987
3988
3989
3990
    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."""
3991

3992
3993
    # TODO(luka) better pass enabling system.

3994
3995
3996
3997
    def uuid(self):
        """
        Produces a hash unique to the pass configuration.
        Any new fields that affect compilation should be added to the hash.
3998
        Any future fields that don't affect compilation should be excluded.
3999
        """
4000
        return InductorPass.hash_dict(asdict(self))
4001
4002

    def __post_init__(self) -> None:
4003
4004
4005
4006
4007
4008
4009
4010
4011
        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")
4012
4013
4014
4015
4016
4017
4018


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

4019
    - Top-level Compilation control:
4020
4021
4022
4023
4024
4025
        - [`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]
4026
    - CudaGraph capture:
4027
4028
4029
4030
4031
4032
4033
4034
        - [`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]
4035
    - Inductor compilation:
4036
4037
4038
4039
4040
        - [`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]
4041
        - custom inductor passes
4042

4043
4044
4045
4046
4047
4048
4049
4050
4051
    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.
4052
4053
    """
    # Top-level Compilation control
4054
    level: int = 0
4055
4056
4057
4058
4059
4060
    """The level of compilation:

    - 0: no compilation.
    - 1: dynamo as is.
    - 2: dynamo once.
    - 3: piecewise compilation."""
4061
    debug_dump_path: str = ""
4062
    """The path to dump the debug information."""
4063
    cache_dir: str = ""
4064
4065
4066
    """The directory to store the compiled graph, to accelerate Inductor
    compilation. By default, it will use model-related information to generate
    a cache directory."""
4067
    backend: str = ""
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
    """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
4090
4091
    disabled when running with Inductor: level>=PIECEWISE and use_inductor=True.
    Inductor generates (fused) Triton kernels for disabled custom ops."""
4092
4093
4094
4095
4096
    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
4097
    use_inductor: bool = True
4098
4099
    """Whether to use inductor compilation:

4100
4101
4102
4103
4104
    - 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.
4105

4106
    This setting is ignored if level<PIECEWISE."""
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
    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
4122
    use_cudagraph: bool = field(default_factory=lambda: envs.VLLM_USE_V1)
4123
4124
4125
4126
4127
    """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.
4128
4129
    In the vLLM V1 Engine, this flag only applies for
    CompilationLevel.PIECEWISE (aka -O3).
4130
4131
4132
4133
    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."""
4134
    cudagraph_num_of_warmups: int = 0
4135
4136
4137
4138
    """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."""
4139
    cudagraph_capture_sizes: Optional[list[int]] = None
4140
4141
4142
    """Sizes to capture cudagraph.
    - None (default): capture sizes are inferred from vllm config.
    - list[int]: capture sizes are specified as given."""
4143
    cudagraph_copy_inputs: bool = False
4144
4145
4146
4147
4148
    """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."""
4149
    full_cuda_graph: bool = False
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
    """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."""
4169

4170
    # keep track of enabled and disabled custom ops
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
    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."""
4187

4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
    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.
        """
4200
        factors: list[Any] = []
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
        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()

4211
4212
    def __repr__(self) -> str:
        exclude = {
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
            "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,
            },
4223
        }
4224
4225
4226
4227
        # 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(
4228
4229
4230
                self,
                exclude=exclude,  # type: ignore[arg-type]
                exclude_unset=True).decode())
4231
4232
4233

    __str__ = __repr__

4234
4235
    @classmethod
    def from_cli(cls, cli_value: str) -> "CompilationConfig":
4236
4237
4238
        """Parse the CLI value for the compilation config.
        -O1, -O2, -O3, etc. is handled in FlexibleArgumentParser.
        """
4239
        return TypeAdapter(CompilationConfig).validate_json(cli_value)
4240

4241
    def __post_init__(self) -> None:
4242
4243
4244
4245
        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
4246
4247
4248
4249
4250
4251
4252
4253
        # 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

4254
        if is_torch_equal_or_newer("2.6"):
Michael Goin's avatar
Michael Goin committed
4255
4256
4257
4258
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

4259
4260
4261
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
4262
4263
4264
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
4265
4266
4267
4268
4269
4270
4271
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
4272
4273
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
4274

4275
4276
        if isinstance(self.pass_config, dict):
            self.pass_config = PassConfig(**self.pass_config)
4277

4278
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
        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
4296

4297
        from vllm.compilation.backends import VllmBackend
4298
        return VllmBackend(vllm_config)
4299

4300
    def init_with_cudagraph_sizes(self,
4301
                                  cudagraph_capture_sizes: list[int]) -> None:
4302
        """To complete the initialization of config,
4303
4304
        we need to know the cudagraph sizes."""

4305
        if self.cudagraph_capture_sizes is None:
4306
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
4307
        else:
4308
            # de-duplicate the sizes provided by the config
4309
4310
4311
4312
4313
4314
            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
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329

        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
4330

4331
        # sort to make sure cudagraph capture sizes are in descending order
4332
4333
4334
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
4335

4336
4337
4338
4339
        # 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)
        ]
4340
4341
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
4342
4343
4344
4345
4346
4347
4348
            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
4349

4350
4351
    def set_splitting_ops_for_v1(self):
        # NOTE: this function needs to be called
4352
4353
4354
4355
4356
        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}")

4357
        if not self.splitting_ops:
4358
            self.splitting_ops = [] if self.full_cuda_graph else [
4359
4360
                "vllm.unified_attention",
                "vllm.unified_attention_with_output",
4361
                "vllm.mamba_mixer2",
4362
4363
            ]

4364

4365
@config
4366
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
4367
4368
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
4369
4370
4371
    simplifies passing around the distinct configurations in the codebase.
    """

4372
4373
4374
    # TODO: use default_factory once default constructing ModelConfig doesn't
    # try to download a model
    model_config: ModelConfig = None  # type: ignore
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
    """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."""
4386
    lora_config: Optional[LoRAConfig] = None
4387
4388
4389
    """LoRA configuration."""
    speculative_config: Optional[SpeculativeConfig] = None
    """Speculative decoding configuration."""
4390
    decoding_config: DecodingConfig = field(default_factory=DecodingConfig)
4391
    """Decoding configuration."""
4392
    observability_config: Optional[ObservabilityConfig] = None
4393
    """Observability configuration."""
4394
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
4395
    """Prompt adapter configuration."""
4396
    quant_config: Optional[QuantizationConfig] = None
4397
4398
4399
    """Quantization configuration."""
    compilation_config: CompilationConfig = field(
        default_factory=CompilationConfig)
4400
    """`torch.compile` and cudagraph capture configuration for the model.
4401

4402
4403
    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}'`).
4404
    Currently, -O <n> and -O=<n> are supported as well but this will likely be
4405
    removed in favor of clearer -O<n> syntax in the future.
4406
4407
4408

    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
4409
    production, also default in V1.
4410
4411
4412
4413
4414
4415

    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."""
4416
    kv_events_config: Optional[KVEventsConfig] = None
4417
    """The configurations for event publishing."""
4418
    # some opaque config, only used to provide additional information
4419
4420
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
4421
4422
4423
4424
    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."""
4425
    instance_id: str = ""
4426
    """The ID of the vLLM instance."""
4427

4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
    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.
        """
4440
        factors: list[Any] = []
4441
4442

        # summarize vllm config
4443
        vllm_factors: list[Any] = []
4444
4445
        from vllm import __version__
        vllm_factors.append(__version__)
4446
        vllm_factors.append(envs.VLLM_USE_V1)
4447
4448
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
4449
4450
        else:
            vllm_factors.append("None")
4451
4452
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
4453
4454
        else:
            vllm_factors.append("None")
4455
4456
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
4457
4458
        else:
            vllm_factors.append("None")
4459
4460
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
4461
4462
        else:
            vllm_factors.append("None")
4463
4464
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
4465
4466
        else:
            vllm_factors.append("None")
4467
4468
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
4469
4470
        else:
            vllm_factors.append("None")
4471
4472
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
4473
4474
4475
4476
4477
            # 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))
4478
4479
        else:
            vllm_factors.append("None")
4480
4481
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
4482
4483
        else:
            vllm_factors.append("None")
4484
4485
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
4486
4487
        else:
            vllm_factors.append("None")
4488
4489
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
4490
4491
        else:
            vllm_factors.append("None")
4492
4493
        if self.prompt_adapter_config:
            vllm_factors.append(self.prompt_adapter_config.compute_hash())
4494
4495
        else:
            vllm_factors.append("None")
4496
4497
4498
4499
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
4500
4501
        else:
            vllm_factors.append("None")
4502
4503
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
4504
4505
4506
        else:
            vllm_factors.append("None")
        if self.additional_config:
4507
4508
4509
4510
4511
4512
4513
4514
            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)
4515
4516
        else:
            vllm_factors.append("None")
4517
4518
        factors.append(vllm_factors)

4519
4520
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
4521
4522
        return hash_str

4523
4524
4525
4526
4527
4528
    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]
4529

4530
4531
4532
4533
4534
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
4535
        from vllm.platforms import current_platform
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
        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
4558

4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
    @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)

4570
4571
4572
4573
4574
4575
4576
4577
4578
    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

4579
4580
4581
4582
4583
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

4584
4585
4586
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
4587
4588
4589

        self.try_verify_and_update_config()

4590
4591
4592
4593
4594
        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)
4595
4596
            self.model_config.verify_dual_chunk_attention_config(
                self.load_config)
4597

4598
        self.cache_config.verify_with_parallel_config(self.parallel_config)
4599

4600
        if self.lora_config is not None:
4601
            self.lora_config.verify_with_cache_config(self.cache_config)
4602
            self.lora_config.verify_with_model_config(self.model_config)
4603
        if self.prompt_adapter_config is not None:
4604
4605
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
4606

4607
        if self.quant_config is None and self.model_config is not None:
4608
4609
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
4610

4611
        from vllm.platforms import current_platform
4612
        if self.model_config is not None and \
4613
4614
4615
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
4616
            logger.warning_once(
4617
4618
4619
4620
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

4621
4622
4623
4624
4625
        # 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
4626
4627
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
4628
4629
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
4630
4631
            # By default, V1 uses piecewise CUDA graphs. If full_cuda_graph
            # is set to True, full CUDA graphs will be used.
4632
            self.compilation_config.cudagraph_num_of_warmups = 1
4633
            self.compilation_config.level = CompilationLevel.PIECEWISE
4634
            self.compilation_config.set_splitting_ops_for_v1()
4635

4636
        self._set_cudagraph_sizes()
4637

4638
        if self.cache_config.cpu_offload_gb > 0 and \
4639
4640
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
4641
            logger.warning(
4642
                "CPU offload is not supported with `torch.compile` in v0 yet."
4643
4644
4645
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4646
4647
4648
4649
4650
4651
        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`.")
4652
4653
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4654
4655
        if self.compilation_config.full_cuda_graph and \
            not self.model_config.disable_cascade_attn:
4656
4657
            logger.info("full_cuda_graph is not supported with "
                        "cascade attention. Disabling cascade attention.")
4658
            self.model_config.disable_cascade_attn = True
4659

4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
        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

4681
        if (self.kv_events_config is not None
4682
4683
4684
4685
4686
                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.")
4687
4688
        if (self.kv_events_config is not None
                and self.kv_events_config.publisher != "null"
4689
4690
4691
4692
4693
                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.")
4694
4695
        current_platform.check_and_update_config(self)

4696
4697
4698
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

4699
4700
4701
4702
4703
4704
4705
        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.
4706
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4707
4708
            if self.kv_transfer_config is not None:
                # Hybrid KV cache manager is not compatible with KV transfer.
4709
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4710
4711
            if self.kv_events_config is not None:
                # Hybrid KV cache manager is not compatible with KV events.
4712
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4713
4714
4715
4716
4717
4718
4719
            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
4720

4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
    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
        ]

4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
    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.

4757
4758
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
4759
4760

        During runtime, if batchsize is larger than
4761
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4762
4763
        no cudagraph will be used.
        If the batch size is no larger than
4764
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
        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)]
4777
4778
4779
4780
4781
                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)

4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
                # 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:
4803
4804
4805
4806
4807
4808
4809
4810
                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
4811
                    raise TypeError(f"Invalid value for {cuda_graph_sizes=}.")
4812
4813
4814
4815
                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)
4816
4817
4818
4819
4820
                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
                ]
4821
4822
4823
4824

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

4825
    def recalculate_max_model_len(self, max_model_len: int):
4826
        # Can only be called in try_verify_and_update_config
4827
        model_config = self.model_config
4828
        max_model_len = model_config.get_and_verify_max_len(max_model_len)
4829
4830
        self.model_config.max_model_len = max_model_len
        self.scheduler_config.max_model_len = max_model_len
4831
4832
4833
4834
4835
4836

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

4837
4838
        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP, HybridAttentionMambaModelConfig)
4839
4840
4841
        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_config(self)
4842

4843
4844
4845
        if self.model_config.is_hybrid:
            HybridAttentionMambaModelConfig.verify_and_update_config(self)

4846
4847
4848
4849
4850
4851
        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)

4852
    def __str__(self):
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
        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}, "
4883
4884
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
4885
4886
4887


_current_vllm_config: Optional[VllmConfig] = None
4888
_current_prefix: Optional[str] = None
4889
4890
4891


@contextmanager
4892
4893
4894
def set_current_vllm_config(vllm_config: VllmConfig,
                            check_compile=False,
                            prefix: Optional[str] = None):
4895
    """
4896
    Temporarily set the current vLLM config.
4897
    Used during model initialization.
4898
    We save the current vLLM config in a global variable,
4899
    so that all modules can access it, e.g. custom ops
4900
    can access the vLLM config to determine how to dispatch.
4901
    """
4902
    global _current_vllm_config, _current_prefix
4903
    old_vllm_config = _current_vllm_config
4904
    old_prefix = _current_prefix
4905
4906
4907
4908
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
4909
        _current_prefix = prefix
4910
        yield
4911
4912
4913
    except Exception:
        raise
    else:
4914
4915
4916
4917
        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)
4918
4919
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
4920
4921
4922
4923
4924
4925
4926
4927
4928
            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"
4929
                " if you want it to be supported.",
4930
                vllm_config.model_config.model)
4931
    finally:
4932
        _current_vllm_config = old_vllm_config
4933
        _current_prefix = old_prefix
4934
4935
4936
4937
4938
4939
4940


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.
4941
        logger.warning("Current vLLM config is not set.")
4942
4943
4944
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
4945
4946


4947
4948
4949
4950
4951
4952
4953
4954
4955
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


4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
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:
4967
        result (bool): `True` if a match is found, `False` otherwise.
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
    """
    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}")
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993


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)
    }
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023


@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:
5024
        return self.min_energy_split_window_size is not None
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042


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)