config.py 215 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
98
_ResolvedTask = Literal["generate", "transcription", "pooling", "embed",
                        "classify", "reward", "draft"]
99

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

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

104
105
106
107
_RUNNER_TASKS: dict[RunnerType, list[_ResolvedTask]] = {
    "generate": ["generate", "transcription"],
    "pooling": ["pooling", "embed", "classify", "reward"],
    "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
349
350
    interleave_mm_strings: bool = False
    """Enable fully interleaved support for multimodal prompts, while using 
    --chat-template-content-format=string. Defaults to False."""
351
352
353
354
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
    """Additional args passed to process media inputs, keyed by modalities. 
    For example, to set num_frames for video, set 
    `--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
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
        # For pooling models, self.task is used to indicate the
        # user-selected task
        if self.task == "score":
            if self.registry.is_cross_encoder_model(self.architectures):
                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"

        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]
            assert selected_task != "pooling"
            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":
586
587
588
589
            self.truncation_side = "left"
        else:
            self.truncation_side = "right"

590
591
592
593
        model_info, arch = self.registry.inspect_model_cls(self.architectures)
        self._model_info = model_info
        self._architecture = arch

594
595
596
597
598
599
600
601
602
        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
    @property
    def registry(self):
679
        return me_models.ModelRegistry
680
681
682

    @property
    def architectures(self) -> list[str]:
683
        # architectures in the model config.
684
685
        return getattr(self.hf_config, "architectures", [])

686
687
688
689
690
691
    @property
    def architecture(self) -> str:
        # The architecture vllm actually used.
        return self._architecture

    @property
692
    def model_info(self):
693
694
        return self._model_info

695
696
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
697
        """Pull model/tokenizer from S3 to temporary directory when needed.
698

699
        Args:
700
701
            model: Model name or path
            tokenizer: Tokenizer name or path
702
        """
703
704
705
706
707
708
709
710
711
712
713
714
        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:
715
716
717
718
719
                s3_model.pull_files(model,
                                    ignore_pattern=[
                                        "*.pt", "*.safetensors", "*.bin",
                                        "*.tensors"
                                    ])
720
721
722
723
724
725
726
                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(
727
728
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors"])
729
            self.tokenizer = s3_tokenizer.dir
730

731
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
732
        if self.registry.is_multimodal_model(self.architectures):
733
            return MultiModalConfig(
734
                limit_per_prompt=self.limit_mm_per_prompt,
735
                media_io_kwargs=self.media_io_kwargs,
736
737
                mm_processor_kwargs=self.mm_processor_kwargs,
                disable_mm_preprocessor_cache=self.
738
739
                disable_mm_preprocessor_cache,
                interleave_mm_strings=self.interleave_mm_strings)
740

741
        if self.limit_mm_per_prompt:
742
743
            raise ValueError("`limit_mm_per_prompt` is only supported for "
                             "multimodal models.")
744
        if self.mm_processor_kwargs:
745
746
            raise ValueError("`mm_processor_kwargs` is only supported for "
                             "multimodal models.")
747
        if self.disable_mm_preprocessor_cache:
748
749
            raise ValueError("`disable_mm_preprocessor_cache` is only "
                             "supported for multimodal models.")
750
751
752
        if self.interleave_mm_strings:
            raise ValueError("`interleave_mm_strings` is only "
                             "supported for multimodal models.")
753
754

        return None
755

756
757
758
759
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

760
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
761
        if self.runner_type == "pooling":
762
763
764
765
766
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
767
768
769
770
771

            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():
772
773
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
774

775
            if self.is_matryoshka:
776
777
778
                if pooler_config.normalize is None:
                    pooler_config.normalize = True
                elif not pooler_config.normalize:
779
780
781
782
783
                    raise ValueError(
                        "`normalize` must be enabled (set to True) "
                        "for models that are compatible with "
                        "Matryoshka Representation.")

784
            return pooler_config
785

786
787
        return None

788
    def _init_attention_free(self) -> bool:
789
        return self.registry.is_attention_free_model(self.architectures)
790

791
    def _init_is_hybrid(self) -> bool:
792
        return self.registry.is_hybrid_model(self.architectures)
793

794
795
796
797
    def _init_has_noops(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return self.registry.is_noops_model(architectures)

798
    def _init_has_inner_state(self) -> bool:
799
        return self.registry.model_has_inner_state(self.architectures)
800

801
    def _verify_tokenizer_mode(self) -> None:
802
803
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
804
805
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
806
                f"one of {get_args(TokenizerMode)}.")
807
        self.tokenizer_mode = tokenizer_mode
808

809
    def _get_preferred_pooling_task(
810
        self,
811
        architectures: list[str],
812
    ) -> _ResolvedTask:
813
814
815
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
816
        if self.registry.is_cross_encoder_model(architectures):
817
            return "classify"
818
        if self.registry.is_transcription_model(architectures):
819
            return "transcription"
820

821
        suffix_to_preferred_task: list[tuple[str, _ResolvedTask]] = [
822
823
824
825
826
            # Other models follow this pattern
            ("ForSequenceClassification", "classify"),
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]
827
        _, arch = self.registry.inspect_model_cls(architectures)
828
829

        for suffix, pref_task in suffix_to_preferred_task:
830
            if arch.endswith(suffix):
831
832
                return pref_task

833
        return "embed"
834

835
    def _get_supported_generation_tasks(
836
        self,
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
        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
853

854
855
856
857
    def _get_supported_pooling_tasks(
        self,
        task_option: TaskOption,
    ) -> list[_ResolvedTask]:
858
859
        registry = self.registry
        architectures = self.architectures
860

861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
        supported_tasks = list[_ResolvedTask]()
        if registry.is_pooling_model(architectures):
            supported_tasks.append("pooling")

            # 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]]:
        return {
            "generate": self._get_supported_generation_tasks(task_option),
            "pooling": self._get_supported_pooling_tasks(task_option),
            "draft": ["draft"]
885
        }
886

887
888
889
890
891
892
893
894
895
    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
        }
896

897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
    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}")
912

913
            return runner_option
914

915
916
917
918
        if task_option != "auto":
            for runner, runner_tasks in supported_tasks.items():
                if task_option in runner_tasks:
                    return runner
919
            else:
920
921
922
923
924
925
926
                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]}")
927

928
929
930
931
932
933
934
935
936
937
        suffix_to_preferred_runner: list[tuple[str, RunnerType]] = [
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("ForSequenceClassification", "pooling"),
            ("EmbeddingModel", "pooling"),
            ("RewardModel", "pooling"),
        ]
        _, arch = self.registry.inspect_model_cls(self.architectures)
938

939
940
941
        for suffix, pref_runner in suffix_to_preferred_runner:
            if arch.endswith(suffix) and pref_runner in supported_runner_types:
                return pref_runner
942

943
944
945
946
947
948
949
950
        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"
        if "generate" in supported_runner_types:
            return "generate"
        if "pooling" in supported_runner_types:
            return "pooling"
951

952
        raise AssertionError("This line should not be reached")
953

954
955
956
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
957
            # compressed-tensors uses a "compression_config" key
958
            quant_cfg = getattr(self.hf_config, "compression_config", None)
959
960
        return quant_cfg

961
    def _verify_quantization(self) -> None:
962
        supported_quantization = me_quant.QUANTIZATION_METHODS
963
        optimized_quantization_methods = [
964
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
965
            "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
966
            "quark", "modelopt_fp4", "bitblas", "gptq_bitblas", "inc"
967
        ]
968
        if self.quantization is not None:
969
970
            self.quantization = cast(me_quant.QuantizationMethods,
                                     self.quantization)
971
972

        # Parse quantization method from the HF model config, if available.
973
974
        quant_cfg = self._parse_quant_hf_config()

975
976
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
977
978
979
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
            quant_cfg["quant_method"] = quant_method
980

981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
            # 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",
            ]
            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

1002
            # Detect which checkpoint is it
1003
            for name in quantization_methods:
1004
                method = me_quant.get_quantization_config(name)
1005
1006
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
1007
1008
1009
1010
                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.
1011
                    if (name in get_args(me_quant.QuantizationMethods)
1012
1013
1014
1015
1016
1017
                            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.")
1018
1019
1020
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
1021

1022
            # Verify quantization configurations.
1023
            if self.quantization is None:
1024
1025
                self.quantization = quant_method
            elif self.quantization != quant_method:
1026
1027
                raise ValueError(
                    "Quantization method specified in the model config "
1028
                    f"({quant_method}) does not match the quantization "
1029
1030
1031
1032
1033
1034
1035
1036
                    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}.")
1037
            from vllm.platforms import current_platform
1038
            current_platform.verify_quantization(self.quantization)
1039
            if self.quantization not in optimized_quantization_methods:
1040
                logger.warning(
1041
                    "%s quantization is not fully "
1042
                    "optimized yet. The speed can be slower than "
1043
                    "non-quantized models.", self.quantization)
1044

1045
    def _verify_cuda_graph(self) -> None:
1046
1047
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
1048
        # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
1049
        ROCM_UNSUPPORTED_MODELS = ['mllama']
1050
1051
1052
1053
1054
1055
        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()):
1056
1057
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1058
                "to eager mode.", self.hf_config.model_type)
1059
            self.enforce_eager = True
1060

1061
1062
    def _verify_bnb_config(self) -> None:
        """
1063
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
1064
        yet support CUDA graph.
1065
        # TODO Remove this when bitsandbytes supports.
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
        """
        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(
1080
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1081
                "fallback to the eager mode.")
1082

1083
1084
            self.enforce_eager = True

1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
    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.")

1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
    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

1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
    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

1129
        # Reminder: Please update docs/features/compatibility_matrix.md
1130
        # If the feature combo become valid
1131
        from vllm.platforms import current_platform
1132
        if not current_platform.is_async_output_supported(self.enforce_eager):
1133
1134
1135
1136
1137
1138
1139
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

1140
        # Async postprocessor is not necessary for pooling models
1141
        # since there is no token generation
1142
        if self.runner_type == "pooling":
1143
1144
            self.use_async_output_proc = False

1145
        # Reminder: Please update docs/features/compatibility_matrix.md
1146
        # If the feature combo become valid
1147
1148
1149
        if speculative_config:
            self.use_async_output_proc = False

1150
1151
1152
1153
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
1154
1155
1156
1157
1158
1159

        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.")

1160
1161
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
1162
1163
1164
1165
1166
1167
1168
        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}).")

1169
        if parallel_config.enable_expert_parallel:
1170
1171
            self._verify_with_expert_parallelism()

1172
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1173
        if pipeline_parallel_size > 1:
1174
            if not self.registry.is_pp_supported_model(self.architectures):
1175
1176
1177
1178
1179
1180
                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
1181

1182
1183
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
1184
        """Get the sliding window size, or None if disabled."""
1185
1186
1187
1188

        # 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.
1189
1190
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
1191
            return None
1192
        return getattr(self.hf_text_config, "sliding_window", None)
1193

1194
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
1195
1196
1197
1198
1199
1200
1201
1202
        """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()

1203
    def get_vocab_size(self) -> int:
1204
        return self.hf_text_config.vocab_size
1205

1206
    def get_hidden_size(self) -> int:
1207
        return self.hf_text_config.hidden_size
1208

1209
1210
    @property
    def is_deepseek_mla(self) -> bool:
1211
1212
1213
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
bigmoyan's avatar
bigmoyan committed
1214
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp', 'kimi_k2'):
1215
1216
1217
1218
1219
1220
1221
1222
            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
1223

1224
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1225
        # TODO remove hard code
1226
        if self.is_deepseek_mla:
1227
1228
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1229
            if self.use_mla:
1230
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1231
1232
1233
1234
1235
            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
1236

1237
1238
1239
1240
1241
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1242
1243
1244
        if self.is_attention_free:
            return 0

1245
1246
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1247
            return self.hf_text_config.head_dim
1248

1249
        # FIXME(woosuk): This may not be true for all models.
1250
1251
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1252

1253
1254
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1255
        # For GPTBigCode & Falcon:
1256
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1257
1258
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1259
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1260
        new_decoder_arch_falcon = (
1261
            self.hf_config.model_type in falcon_model_types
1262
            and getattr(self.hf_config, "new_decoder_architecture", False))
1263
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1264
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1265
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1266
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1267
            return 1
1268

1269
        # For DBRX and MPT
1270
1271
1272
1273
1274
        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":
1275
1276
1277
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1278
1279
1280
1281
1282
1283
1284
1285
        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")

1286
1287
1288
        if self.is_attention_free:
            return 0

1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
        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:
1299
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1300
1301
1302
1303
1304
            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.
1305
        return self.hf_text_config.num_attention_heads
1306
1307
1308

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1309
1310
1311
1312
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1313
1314
1315
1316
1317
1318
1319
        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)
1320

1321
1322
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1323
1324
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1325

1326
    def get_layers_start_end_indices(
1327
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1328
        from vllm.distributed.utils import get_pp_indices
1329
1330
        if (self.hf_text_config.model_type == "deepseek_mtp"
                or self.hf_config.model_type == "mimo_mtp"):
1331
1332
1333
1334
1335
            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)
1336
1337
1338
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1339
1340
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1341
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1342

1343
1344
1345
    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
1346

1347
1348
1349
1350
1351
1352
1353
1354
    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
1355
1356
1357
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
        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
1368
1369
1370
1371
        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])
1372
        else:
1373
            # Hybrid model Jamba
1374
1375
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
            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
1401

1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
    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

1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
    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

1425
    def try_get_generation_config(self) -> dict[str, Any]:
1426
        if self.generation_config in ("auto", "vllm"):
1427
            config = try_get_generation_config(
1428
                self.hf_config_path or self.model,
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
                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()

1443
    def get_diff_sampling_param(self) -> dict[str, Any]:
1444
        """
1445
        This method returns a dictionary containing the parameters
1446
1447
        that differ from the default sampling parameters. If
        `generation_config` is `"vllm"`, an empty dictionary is returned.
1448
1449

        Returns:
1450
            dict[str, Any]: A dictionary with the differing sampling
1451
            parameters, if `generation_config` is `"vllm"` an empty dictionary.
1452
        """
1453
        if self.generation_config == "vllm":
1454
1455
1456
1457
1458
1459
1460
            config = {}
        else:
            config = self.try_get_generation_config()

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

1461
1462
1463
1464
1465
1466
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1467
            "max_new_tokens",
1468
1469
1470
1471
1472
1473
        ]
        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
            }
1474
1475
1476
1477
1478
            # 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")
1479
1480
        else:
            diff_sampling_param = {}
1481
1482
1483
1484
1485
1486
1487

        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`.")
1488
1489
        return diff_sampling_param

1490
    @property
1491
    def is_encoder_decoder(self) -> bool:
1492
        """Extract the HF encoder/decoder model flag."""
1493
        """
1494
        For Mllama, VLLM overrides HF's is_encoder_decoder flag and sets it to
1495
        True to enable cross-attention
1496
        Neuron needs all multimodal data to be in the decoder and does not
1497
1498
1499
1500
1501
1502
        need to explicitly enable cross-attention
        """
        if (current_platform.is_neuron()
                and self.hf_config.model_type == "mllama"):
            return False

1503
1504
1505
1506
1507
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1508

1509
1510
1511
1512
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1513
1514
    @property
    def is_cross_encoder(self) -> bool:
1515
        return self.task == "classify"
1516

1517
1518
    @property
    def use_mla(self) -> bool:
1519
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1520

1521
1522
1523
    @property
    def is_v1_compatible(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
1524
        return me_models.ModelRegistry.is_v1_compatible(architectures)
1525

1526
1527
1528
1529
1530
    @property
    def is_matryoshka(self) -> bool:
        return (hasattr(self.hf_config, "matryoshka_dimensions")
                or getattr(self.hf_config, "is_matryoshka", False))

1531
1532
1533
1534
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)

1535
1536
1537
1538
1539
1540
    @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)

1541
    def get_and_verify_max_len(self, max_model_len: int):
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
        # For pooling models, the tokenizer's `model_max_length` is often a
        # reliable source for the maximum sequence length. However, for
        # generative models, this can be incorrect and unduly limit the
        # context window (e.g., DeepSeek-R1). Therefore, we only consider
        # tokenizer_config for pooling models.
        tokenizer_config = None
        if self.runner_type == "pooling":
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
                revision=self.tokenizer_revision)
1553
1554
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
1555
            tokenizer_config=tokenizer_config,
1556
1557
1558
1559
1560
            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)
1561
        logger.info("Using max model len %s", max_model_len)
1562
1563
        return max_model_len

1564

1565
BlockSize = Literal[1, 8, 16, 32, 64, 128]
1566
CacheDType = Literal["auto", "fp8", "fp8_e4m3", "fp8_e5m2", "fp8_inc"]
1567
PrefixCachingHashAlgo = Literal["builtin", "sha256", "sha256_cbor_64bit"]
1568
1569
1570
1571


@config
@dataclass
1572
class CacheConfig:
1573
    """Configuration for the KV cache."""
1574

1575
    block_size: SkipValidation[BlockSize] = None  # type: ignore
1576
1577
1578
    """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.
1579
1580

    This config has no static default. If left unspecified by the user, it will
1581
    be set in `Platform.check_and_update_config()` based on the current
1582
    platform."""
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
    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
1596
    fp8 (=fp8_e4m3). Intel Gaudi (HPU) supports fp8 (using fp8_inc)."""
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
    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
1612
1613
1614
1615
1616
1617
    - "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."""
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
    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."""
1631
1632
    cpu_kvcache_space_bytes: Optional[int] = None
    """(CPU backend only) CPU key-value cache space."""
1633
1634
1635
    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."""
1636
1637
1638
1639
1640
1641

    # 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."""
1642

1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
    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.
        """
1655
        factors: list[Any] = []
1656
1657
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
1658
1659
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1660
1661
        return hash_str

1662
1663
1664
    def __post_init__(self) -> None:
        self.swap_space_bytes = self.swap_space * GiB_bytes

1665
        self._verify_cache_dtype()
1666
        self._verify_prefix_caching()
1667

1668
    def metrics_info(self):
1669
1670
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1671
1672
        return {key: str(value) for key, value in self.__dict__.items()}

1673
1674
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
1675
1676
1677
1678
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1679
1680
1681
1682
1683
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1684
1685
        return self

1686
1687
1688
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1689
        elif self.cache_dtype in get_args(CacheDType):
1690
            logger.info(
1691
1692
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1693
                "Meanwhile, it may cause accuracy drop without a proper "
1694
                "scaling factor.")
1695
1696
1697
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1698
1699
1700
1701
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

1702
        if self.sliding_window is not None and not envs.VLLM_USE_V1:
1703
1704
1705
1706
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

1707
1708
        if (self.enable_prefix_caching and self.prefix_caching_hash_algo
                not in get_args(PrefixCachingHashAlgo)):
1709
1710
            raise ValueError(
                "Unknown prefix caching hash algorithm: "
1711
1712
                f"{self.prefix_caching_hash_algo}. Must be one of "
                f"{get_args(PrefixCachingHashAlgo)}.")
1713

1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
    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

1724
1725
1726
        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.")
1727
1728
1729
        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:
1730
            logger.warning("Possibly too large swap space. %s", msg)
1731

1732

1733
1734
1735
1736
1737
1738
1739
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
1740
    SHARDED_STATE = "sharded_state"
1741
    GGUF = "gguf"
1742
    BITSANDBYTES = "bitsandbytes"
1743
    MISTRAL = "mistral"
1744
    RUNAI_STREAMER = "runai_streamer"
1745
    RUNAI_STREAMER_SHARDED = "runai_streamer_sharded"
1746
    FASTSAFETENSORS = "fastsafetensors"
1747
1748


1749
@config
1750
1751
@dataclass
class LoadConfig:
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
    """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."""
1777
    download_dir: Optional[str] = None
1778
1779
    """Directory to download and load the weights, default to the default
    cache directory of Hugging Face."""
1780
1781
    model_loader_extra_config: Union[dict, TensorizerConfig] = field(
        default_factory=dict)
1782
    """Extra config for model loader. This will be passed to the model loader
1783
    corresponding to the chosen load_format."""
1784
1785
1786
    device: Optional[str] = None
    """Device to which model weights will be loaded, default to
    device_config.device"""
1787
    ignore_patterns: Optional[Union[list[str], str]] = None
1788
1789
    """The list of patterns to ignore when loading the model. Default to
    "original/**/*" to avoid repeated loading of llama's checkpoints."""
1790
    use_tqdm_on_load: bool = True
1791
1792
    """Whether to enable tqdm for showing progress bar when loading model
    weights."""
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
    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
    """
1803

1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
    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.
1818
        factors: list[Any] = []
1819
1820
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1821
1822
        return hash_str

1823
    def __post_init__(self):
1824
1825
1826
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1827

1828
1829
1830
1831
1832
1833
1834
        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/**/*"]

1835

1836
1837
1838
DistributedExecutorBackend = Literal["ray", "mp", "uni", "external_launcher"]


1839
@config
1840
@dataclass
1841
class ParallelConfig:
1842
    """Configuration for the distributed execution."""
1843

1844
1845
1846
1847
1848
1849
1850
    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."""
1851
1852
    data_parallel_size_local: int = 1
    """Number of local data parallel groups."""
1853
1854
    data_parallel_rank: int = 0
    """Rank of the data parallel group."""
1855
1856
1857
    data_parallel_rank_local: Optional[int] = None
    """Local rank of the data parallel group,
    set only in SPMD mode."""
1858
    data_parallel_master_ip: str = "127.0.0.1"
1859
    """IP of the data parallel master."""
1860
1861
    data_parallel_rpc_port: int = 29550
    """Port for data parallel messaging."""
1862
1863
    data_parallel_master_port: int = 29500
    """Port of the data parallel master."""
Rui Qiao's avatar
Rui Qiao committed
1864
1865
    data_parallel_backend: str = "mp"
    """Backend to use for data parallel, either "mp" or "ray"."""
1866
1867
1868
1869
    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."""
1870
1871
    enable_expert_parallel: bool = False
    """Use expert parallelism instead of tensor parallelism for MoE layers."""
1872
1873
1874
1875
1876
1877
1878
1879
1880
    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.
1881

1882
1883
1884
1885
1886
1887
1888
1889
1890
    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.
    """

1891
    max_parallel_loading_workers: Optional[int] = None
1892
    """Maximum number of parallel loading workers when loading model
1893
1894
    sequentially in multiple batches. To avoid RAM OOM when using tensor
    parallel and large models."""
1895
1896

    disable_custom_all_reduce: bool = False
1897
    """Disable the custom all-reduce kernel and fall back to NCCL."""
1898
1899

    ray_workers_use_nsight: bool = False
1900
    """Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler."""
1901
1902

    placement_group: Optional["PlacementGroup"] = None
1903
    """ray distributed model workers placement group."""
1904

1905
    distributed_executor_backend: Optional[Union[DistributedExecutorBackend,
1906
                                                 type["ExecutorBase"]]] = None
1907
1908
1909
1910
1911
1912
    """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
1913
    only support Ray for distributed inference."""
1914
1915

    worker_cls: str = "auto"
1916
1917
    """The full name of the worker class to use. If "auto", the worker class
    will be determined based on the platform."""
1918
    sd_worker_cls: str = "auto"
Ning Xie's avatar
Ning Xie committed
1919
    """The full name of the worker class to use for speculative decoding.
1920
    If "auto", the worker class will be determined based on the platform."""
1921
    worker_extension_cls: str = ""
1922
1923
1924
1925
    """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."""
1926
1927

    world_size: int = field(init=False)
1928
    """world_size is TPxPP, it affects the number of workers we create."""
1929
1930

    rank: int = 0
1931
    """Global rank in distributed setup."""
1932

1933
    enable_multimodal_encoder_data_parallel: bool = False
1934
    """ Use data parallelism instead of tensor parallelism for vision encoder.
1935
1936
    Only support LLama4 for now"""

1937
1938
1939
1940
1941
1942
    @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

1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
    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":
1957
1958
1959
1960
1961
1962
1963
1964
1965
        # 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

1966
1967
1968
        from vllm.distributed.utils import (
            stateless_init_torch_distributed_process_group)

1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
        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
1992
1993
1994

    @staticmethod
    def has_unfinished_dp(dp_group: "ProcessGroup",
youkaichao's avatar
youkaichao committed
1995
                          has_unfinished: bool) -> bool:
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
        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

2007
2008
2009
2010
2011
2012
2013
2014
    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.
        """
2015
        factors: list[Any] = []
2016
2017
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
2018
        factors.append(self.enable_expert_parallel)
2019
2020
        factors.append(self.data_parallel_size)
        factors.append(envs.VLLM_ALL2ALL_BACKEND)
2021
2022
        return hashlib.sha256(str(factors).encode()).hexdigest()

2023
2024
2025
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
2026

2027
2028
2029
2030
2031
2032
        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:
2033
2034
            # Data parallel was specified in the engine args.
            self.data_parallel_master_port = get_open_port()
2035
2036
2037
2038
2039

            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})")
2040
2041
2042
2043
2044
2045
2046
2047
        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

2048
2049
2050
2051
            if self.data_parallel_external_lb:
                raise ValueError("data_parallel_external_lb can only "
                                 "be set when data_parallel_size > 1")

2052
2053
2054
2055
2056
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
        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}.")
        else:
            if self.num_redundant_experts != 0:
                raise ValueError(
                    "num_redundant_experts should be used with EPLB."
                    f"{self.num_redundant_experts}.")
2071
        if self.distributed_executor_backend is None and self.world_size > 1:
2072
2073
2074
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

2075
            from vllm.executor import ray_utils
2076
            backend: DistributedExecutorBackend = "mp"
2077
            ray_found = ray_utils.ray_is_available()
2078
2079
            if current_platform.is_neuron():
                # neuron uses single process to control multiple devices
2080
2081
                backend = "uni"
            elif current_platform.is_tpu() and envs.VLLM_XLA_USE_SPMD:
2082
2083
2084
                backend = "uni"
            elif (current_platform.is_cuda()
                  and cuda_device_count_stateless() < self.world_size):
2085
2086
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
2087
2088
2089
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
2090
                backend = "ray"
Rui Qiao's avatar
Rui Qiao committed
2091
2092
2093
2094
            elif self.data_parallel_backend == "ray":
                logger.info("Using ray distributed inference because "
                            "data_parallel_backend is ray")
                backend = "ray"
2095
            elif ray_found:
2096
                if self.placement_group:
2097
                    backend = "ray"
2098
2099
2100
2101
2102
2103
                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"
2104
            self.distributed_executor_backend = backend
2105
2106
            logger.debug("Defaulting to use %s for distributed inference",
                         backend)
2107

2108
2109
2110
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

2111
2112
2113
2114
2115
2116
    @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)

2117
2118
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2119
2120
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
2121
        from vllm.platforms import current_platform
2122
        if self.distributed_executor_backend not in (
2123
2124
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
2125
2126
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
2127
            raise ValueError(
2128
2129
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
2130
2131
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
2132
        if self.use_ray:
2133
2134
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
2135
2136

        if not current_platform.use_custom_allreduce():
2137
            self.disable_custom_all_reduce = True
Aaron Pham's avatar
Aaron Pham committed
2138
            logger.debug(
2139
                "Disabled the custom all-reduce kernel because it is not "
2140
                "supported on current platform.")
2141
        if self.ray_workers_use_nsight and not self.use_ray:
2142
2143
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
2144

2145
        return self
2146

2147

2148
PreemptionMode = Literal["swap", "recompute"]
2149
2150
2151
2152
SchedulerPolicy = Literal["fcfs", "priority"]


@config
2153
@dataclass
2154
class SchedulerConfig:
2155
    """Scheduler configuration."""
2156

2157
2158
    runner_type: RunnerType = "generate"
    """The runner type to launch for the model."""
2159

2160
    max_num_batched_tokens: SkipValidation[int] = None  # type: ignore
2161
    """Maximum number of tokens to be processed in a single iteration.
2162

2163
2164
    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."""
2165

2166
    max_num_seqs: SkipValidation[int] = None  # type: ignore
2167
    """Maximum number of sequences to be processed in a single iteration.
2168

2169
2170
    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."""
2171

2172
    max_model_len: SkipValidation[int] = None  # type: ignore
2173
2174
2175
    """Maximum length of a sequence (including prompt and generated text). This
    is primarily set in `ModelConfig` and that value should be manually
    duplicated here."""
2176

2177
    max_num_partial_prefills: int = 1
2178
2179
    """For chunked prefill, the maximum number of sequences that can be
    partially prefilled concurrently."""
2180
2181

    max_long_partial_prefills: int = 1
2182
2183
2184
2185
    """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."""
2186
2187

    long_prefill_token_threshold: int = 0
2188
2189
    """For chunked prefill, a request is considered long if the prompt is
    longer than this number of tokens."""
2190

2191
    num_lookahead_slots: int = 0
2192
2193
2194
2195
2196
2197
2198
    """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."""
2199

2200
2201
2202
2203
    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
2204
    pattern: [1, 2, 4] + [i for i in range(8, cuda_graph_sizes + 1, 8)]
2205
    3. more than one value (e.g. 1 2 128) is provided, then the capture list
2206
    will follow the provided list."""
2207

2208
    delay_factor: float = 0.0
2209
2210
    """Apply a delay (of delay factor multiplied by previous
    prompt latency) before scheduling next prompt."""
2211

2212
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2213
2214
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""
2215
2216

    is_multimodal_model: bool = False
2217
2218
2219
2220
2221
    """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.
2222

2223
2224
2225
2226
2227
2228
2229
2230
2231
    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."""
2232

2233
    preemption_mode: Optional[PreemptionMode] = None
2234
2235
2236
2237
2238
2239
    """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."""
2240
2241

    num_scheduler_steps: int = 1
2242
    """Maximum number of forward steps per scheduler call."""
2243

2244
2245
    multi_step_stream_outputs: bool = True
    """If False, then multi-step will stream outputs at the end of all steps"""
2246
2247

    send_delta_data: bool = False
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
    """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)."""
2259
2260

    chunked_prefill_enabled: bool = field(init=False)
2261
    """True if chunked prefill is enabled."""
2262

2263
    disable_chunked_mm_input: bool = False
2264
2265
2266
2267
2268
2269
    """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."""
2270

2271
2272
    # scheduler class or path. "vllm.core.scheduler.Scheduler" (default)
    # or "mod.custom_class".
2273
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
2274
2275
2276
    """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"."""
2277

2278
2279
2280
2281
2282
2283
    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.
    """

2284
2285
2286
2287
2288
2289
2290
    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.
    """

2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
    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.
2305
        factors: list[Any] = []
2306
2307
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2308
2309
        return hash_str

2310
    def __post_init__(self) -> None:
2311
2312
2313
2314
2315
2316
        if self.max_model_len is None:
            self.max_model_len = 8192

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

2317
2318
2319
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
2320
2321
2322
2323
                    # 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.
2324
                    self.max_num_batched_tokens = max(
2325
                        self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2326
                else:
2327
                    self.max_num_batched_tokens = (
2328
                        DEFAULT_MAX_NUM_BATCHED_TOKENS)
2329
            else:
2330
                # If max_model_len is too short, use
2331
                # DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
2332
                # for higher throughput.
2333
                self.max_num_batched_tokens = max(
2334
                    self.max_model_len, DEFAULT_MAX_NUM_BATCHED_TOKENS)
2335

2336
2337
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
2338
2339
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2340
                    POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
2341
                )
2342
            if self.is_multimodal_model:
2343
                # The value needs to be at least the number of multimodal tokens
2344
2345
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2346
                    MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
2347
2348
                )

2349
2350
2351
2352
2353
2354
2355
            # 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)

2356
2357
2358
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

2359
        if self.enable_chunked_prefill:
2360
2361
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
2362
                self.max_num_batched_tokens)
2363

2364
        self.chunked_prefill_enabled = self.enable_chunked_prefill
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
        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)

2377
2378
2379
2380
2381
2382
2383
        # 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)]

2384
2385
2386
2387
        if self.async_scheduling:
            self.scheduler_cls = (
                "vllm.v1.core.sched.async_scheduler.AsyncScheduler")

2388
2389
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2390
2391
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
2392
2393
2394
2395
2396
2397
2398
            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.")
2399

2400
2401
2402
2403
2404
        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}).")
2405

2406
2407
        if self.max_num_batched_tokens > self.max_num_seqs * self.max_model_len:
            logger.warning(
2408
                "max_num_batched_tokens (%d) exceeds max_num_seqs "
2409
2410
2411
2412
                "* max_model_len (%d). This may lead to unexpected behavior.",
                self.max_num_batched_tokens,
                self.max_num_seqs * self.max_model_len)

2413
2414
2415
2416
2417
2418
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

2419
2420
2421
2422
2423
2424
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
        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}).")

2448
2449
        return self

2450
2451
2452
2453
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

2454

2455
Device = Literal["auto", "cuda", "neuron", "cpu", "tpu", "xpu"]
2456
2457
2458


@config
2459
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
2460
class DeviceConfig:
2461
2462
    """Configuration for the device to use for vLLM execution."""

2463
    device: SkipValidation[Optional[Union[Device, torch.device]]] = "auto"
2464
    """Device type for vLLM execution.
2465
2466
2467
    This parameter is deprecated and will be
    removed in a future release.
    It will now be set automatically based
2468
    on the current platform."""
2469
2470
2471
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
2472

2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
    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.
2488
        factors: list[Any] = []
2489
2490
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2491
2492
        return hash_str

2493
2494
    def __post_init__(self):
        if self.device == "auto":
2495
            # Automated device type detection
2496
            from vllm.platforms import current_platform
2497
            self.device_type = current_platform.device_type
2498
            if not self.device_type:
2499
2500
2501
2502
                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.")
2503
2504
        else:
            # Device type is assigned explicitly
2505
2506
2507
2508
            if isinstance(self.device, str):
                self.device_type = self.device
            elif isinstance(self.device, torch.device):
                self.device_type = self.device.type
2509
2510

        # Some device types require processing inputs on CPU
2511
        if self.device_type in ["neuron"]:
2512
            self.device = torch.device("cpu")
2513
2514
        elif self.device_type in ["tpu"]:
            self.device = None
2515
2516
2517
2518
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

2519

2520
2521
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
                            "mlp_speculator", "draft_model", "deepseek_mtp"]
2522
2523
2524
2525
2526
SpeculativeAcceptanceMethod = Literal["rejection_sampler",
                                      "typical_acceptance_sampler"]


@config
2527
@dataclass
2528
class SpeculativeConfig:
2529
    """Configuration for speculative decoding."""
2530

2531
    # General speculative decoding control
2532
    num_speculative_tokens: SkipValidation[int] = None  # type: ignore
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
    """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."""
    acceptance_method: SpeculativeAcceptanceMethod = "rejection_sampler"
    """The method to use for accepting draft tokens:\n
    - "rejection_sampler" maps to `RejectionSampler`.\n
    - "typical_acceptance_sampler" maps to `TypicalAcceptanceSampler`.

    If using `typical_acceptance_sampler`, the related configuration
    `posterior_threshold` and `posterior_alpha` should be considered."""
2553
    draft_tensor_parallel_size: Optional[int] = None
2554
2555
    """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."""
2556
    disable_logprobs: bool = True
2557
2558
2559
    """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."""
2560

2561
    # Draft model configuration
2562
    quantization: Optional[me_quant.QuantizationMethods] = None
2563
2564
2565
    """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."""
2566
    max_model_len: Optional[int] = None
2567
2568
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
2569
    revision: Optional[str] = None
2570
2571
2572
    """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."""
2573
    code_revision: Optional[str] = None
2574
2575
2576
    """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."""
2577

2578
    # Advanced control
2579
    disable_mqa_scorer: bool = False
2580
2581
    """Disable the MQA scorer and fall back to batch expansion for scoring
    proposals."""
2582
    disable_by_batch_size: Optional[int] = None
2583
2584
2585
2586
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
2587
    prompt_lookup_max: Optional[int] = None
2588
2589
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
2590
    prompt_lookup_min: Optional[int] = None
2591
2592
2593
2594
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

    # Typical acceptance sampler configuration
2595
    posterior_threshold: Optional[float] = None
2596
2597
2598
2599
    """A threshold value that sets a lower bound on the posterior probability
    of a token in the target model for it to be accepted. This threshold is
    used only when we use the `TypicalAcceptanceSampler` for token acceptance.
    """
2600
    posterior_alpha: Optional[float] = None
2601
2602
    """Scaling factor for entropy-based threshold, applied when using
    `TypicalAcceptanceSampler`."""
2603

2604
    speculative_token_tree: Optional[str] = None
2605
    """Specifies the tree structure for speculative token generation.
2606
    """
2607
    # required configuration params passed from engine
2608
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2609
    """The configuration of the target model."""
2610
2611
    target_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2612
    """The parallel configuration for the target model."""
2613
    enable_chunked_prefill: SkipValidation[bool] = None  # type: ignore
2614
2615
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
2616
    disable_log_stats: SkipValidation[bool] = None  # type: ignore
2617
2618
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
2619
2620

    # params generated in the post-init stage
2621
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
2622
    """The configuration of the draft model initialized internal."""
2623
2624
    draft_parallel_config: SkipValidation[
        ParallelConfig] = None  # type: ignore
2625
    """The parallel configuration for the draft model initialized internal."""
2626

2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
    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.
        """
2639
        factors: list[Any] = []
2640
2641
2642
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2643
2644
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2645
2646
        return hash_str

2647
2648
2649
2650
2651
    @classmethod
    def from_dict(cls, dict_value: dict) -> "SpeculativeConfig":
        """Parse the CLI value for the speculative config."""
        return cls(**dict_value)

2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
    @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"]
            })
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672

        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"]
            })
            return hf_config

2673
2674
        return hf_config

2675
    def __post_init__(self):
2676

2677
2678
2679
2680
2681
2682
2683
        # 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.
2684
2685
2686
2687

        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
2688
            if self.target_model_config and \
2689
2690
2691
2692
                (self.target_model_config.hf_text_config.model_type \
                        == "deepseek_v3" or
                    self.target_model_config.hf_text_config.model_type \
                        == "mimo"):
2693
2694
2695
2696
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2697
            else:
2698
2699
2700
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2701
2702
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2703
2704
2705
2706
2707
2708
2709
        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"
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
            # 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
2724
            if self.prompt_lookup_min < 1:
2725
2726
2727
2728
2729
                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")
2730
            if self.prompt_lookup_min > self.prompt_lookup_max:
2731
2732
2733
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2734

2735
2736
2737
            # 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.
2738
2739
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2740
        else:
2741
2742
2743
2744
2745
2746
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
2747
                    runner="draft",
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
                    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,
                )
2769

2770
                # Automatically detect the method
2771
                if self.method in ('eagle', 'eagle3'):
2772
                    pass
2773
2774
                elif "eagle-" in self.draft_model_config.model.lower() or \
                        "eagle3-" in self.draft_model_config.model.lower():
2775
2776
2777
2778
2779
2780
                    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"
Jiayi Yao's avatar
Jiayi Yao committed
2781
2782
2783
2784
2785
2786
2787
2788
2789
                elif (self.draft_model_config.hf_config.model_type ==
                      "deepseek_mtp"):
                    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."
                            )
2790
                else:
2791
2792
2793
                    self.method = "draft_model"

                # Replace hf_config for EAGLE draft_model
2794
                if self.method in ("eagle", "eagle3"):
2795
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2796
                        raise ValueError(
2797
2798
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2799
2800
2801
2802

                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
                    if isinstance(self.draft_model_config.hf_config,
2803
                                  EAGLEConfig):
2804
2805
2806
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2807
                            self.draft_model_config.hf_config,
2808
2809
                            method=self.method,
                            model_type="eagle")
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
                        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
                )
2837

2838
2839
2840
2841
2842
2843
                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,
                    ))
2844

2845
2846
2847
2848
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2849

2850
2851
2852
2853
2854
        if self.acceptance_method == "typical_acceptance_sampler":
            if self.posterior_threshold is None:
                self.posterior_threshold = 0.09
            if self.posterior_alpha is None:
                self.posterior_alpha = 0.3
2855

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
2883
2884
2885
2886
2887
2888
2889
2890
    @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,
        )

2891
    @staticmethod
2892
    def _verify_and_get_draft_tp(
2893
2894
2895
2896
2897
2898
            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.
2899
        """
2900
2901
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2902
        if speculative_draft_tensor_parallel_size is None:
2903
2904
2905
2906
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2907
2908
2909
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2910
2911
2912
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2913
2914
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2915
            raise ValueError(
2916
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2917
                f"other value than 1 or target model tensor_parallel_size")
2918
        return speculative_draft_tensor_parallel_size
2919

2920
2921
2922
2923
2924
2925
2926
2927
2928
    @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.
        """
2929
2930
2931
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2932
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2933
2934
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
            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

2946
2947
    @model_validator(mode='after')
    def _verify_args(self) -> Self:
2948
2949
2950
2951
2952
2953
        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.")

2954
2955
2956
2957
2958
2959
2960
        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)
2961
2962
            # Validate and set draft token acceptance related settings.

2963
2964
        if self.acceptance_method is None:
            raise ValueError("acceptance_method is not set. "
2965
2966
2967
                             "Expected values are rejection_sampler or "
                             "typical_acceptance_sampler.")

2968
2969
        if (self.acceptance_method != 'rejection_sampler'
                and self.acceptance_method != 'typical_acceptance_sampler'):
2970
            raise ValueError(
2971
                "Expected acceptance_method to be either "
2972
                "rejection_sampler or typical_acceptance_sampler. Instead it "
2973
                f"is {self.acceptance_method}")
2974

2975
2976
2977
2978
        if self.acceptance_method == "typical_acceptance_sampler" and (
            (self.posterior_threshold is not None
             and self.posterior_threshold < 0) or
            (self.posterior_alpha is not None and self.posterior_alpha < 0)):
2979
            raise ValueError(
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
                "Expected the posterior_threshold and posterior_alpha of "
                "typical_acceptance_sampler to be > 0. "
                "Instead found posterior_threshold = "
                f"{self.posterior_threshold} and posterior_alpha = "
                f"{self.posterior_alpha}")

        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=}")
2991

2992
2993
2994
2995
2996
2997
        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=}")

2998
2999
        return self

3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
    @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

3010
    def use_eagle(self) -> bool:
Jiayi Yao's avatar
Jiayi Yao committed
3011
        return self.method in ("eagle", "eagle3", "deepseek_mtp")
3012

3013
    def __repr__(self) -> str:
3014
3015
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
3016
        num_spec_tokens = self.num_speculative_tokens
3017
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
3018
3019


3020
3021
3022
3023
LoRADType = Literal["auto", "float16", "bfloat16"]


@config
3024
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3025
class LoRAConfig:
3026
3027
3028
3029
3030
3031
    """Configuration for LoRA."""

    max_lora_rank: int = 16
    """Max LoRA rank."""
    max_loras: int = 1
    """Max number of LoRAs in a single batch."""
3032
    fully_sharded_loras: bool = False
3033
3034
3035
3036
    """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.
    """
3037
    max_cpu_loras: Optional[int] = None
3038
3039
3040
3041
    """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."""
3042
    lora_extra_vocab_size: int = 256
3043
3044
    """Maximum size of extra vocabulary that can be present in a LoRA adapter
    (added to the base model vocabulary)."""
3045
3046
    lora_vocab_padding_size: ClassVar[int] = current_platform\
        .get_lora_vocab_padding_size()
3047
3048
3049
3050
3051
3052
    long_lora_scaling_factors: Optional[tuple[float, ...]] = None
    """Specify multiple scaling factors (which can be different from base model
    scaling factor - see eg. Long LoRA) to allow for multiple LoRA adapters
    trained with those scaling factors to be used at the same time. If not
    specified, only adapters trained with the base model scaling factor are
    allowed."""
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
    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."""
3063
    bias_enabled: bool = False
3064
    """Enable bias for LoRA adapters."""
3065

3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
    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.
        """
3078
        factors: list[Any] = []
3079
3080
3081
3082
3083
        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)
3084
        factors.append(self.lora_vocab_padding_size)
3085
3086
        factors.append(self.long_lora_scaling_factors)
        factors.append(self.bias_enabled)
3087
3088
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3089
3090
        return hash_str

3091
    def __post_init__(self):
3092
        # Setting the maximum rank to 512 should be able to satisfy the vast
3093
        # majority of applications.
3094
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
3095
        possible_lora_extra_vocab_size = (256, 512)
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
        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
3111
                f"max_loras ({self.max_loras})")
3112

3113
    def verify_with_cache_config(self, cache_config: CacheConfig):
3114
3115
3116
        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.")
3117

3118
3119
3120
3121
3122
3123
    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)

3124
3125
3126
3127
3128
    def verify_lora_support(self):
        if self.long_lora_scaling_factors is not None and envs.VLLM_USE_V1:
            raise ValueError(
                "V1 LoRA does not support long LoRA, please use V0.")

3129

3130
@config
3131
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
3132
class PromptAdapterConfig:
3133
3134
    """Configuration for PromptAdapters."""

3135
3136
3137
3138
    max_prompt_adapters: int = 1
    """Max number of PromptAdapters in a batch."""
    max_prompt_adapter_token: int = 0
    """Max number of PromptAdapters tokens."""
3139
    max_cpu_prompt_adapters: Optional[int] = None
3140
3141
3142
3143
3144
    """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.
    """
3145

3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
    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.
3160
        factors: list[Any] = []
3161
3162
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3163
3164
        return hash_str

3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
    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):
3176
        if self.prompt_adapter_dtype == "auto":
3177
3178
3179
3180
3181
3182
            self.prompt_adapter_dtype = model_config.dtype
        elif isinstance(self.prompt_adapter_dtype, str):
            self.prompt_adapter_dtype = getattr(torch,
                                                self.prompt_adapter_dtype)


3183
@config
3184
@dataclass
3185
class MultiModalConfig:
3186
3187
    """Controls the behavior of multimodal models."""

3188
3189
    limit_per_prompt: dict[str, int] = \
        cast(dict[str, int], get_field(ModelConfig, "limit_mm_per_prompt"))
3190
    """
3191
    The maximum number of input items allowed per prompt for each modality.
3192
    Defaults to 1 (V0) or 999 (V1) for each modality.
3193
3194

    For example, to allow up to 16 images and 2 videos per prompt:
3195
    `{"images": 16, "videos": 2}`
3196
3197
    """

3198
3199
3200
3201
3202
    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
    """Additional args passed to process media inputs, keyed by modalities. 
    For example, to set num_frames for video, set 
    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """

3203
3204
3205
    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
3206
    `transformers.AutoProcessor.from_pretrained`.
3207
3208
3209
3210

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

    For example, for Phi-3-Vision:
3211
    `{"num_crops": 4}`.
3212
3213
3214
3215
    """

    disable_mm_preprocessor_cache: bool = False
    """
3216
    If `True`, disable caching of the processed multi-modal inputs.
3217
3218
    """

3219
3220
3221
3222
3223
    interleave_mm_strings: bool = False
    """
    Enable fully interleaved support for multimodal prompts.
    """

3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
    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.
3238
        factors: list[Any] = []
3239
3240
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3241
3242
        return hash_str

3243
3244
3245
3246
3247
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
3248
3249
3250
3251
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
3252

3253
    # TODO: Add configs to init vision tower or not.
3254

3255

3256
@config
3257
3258
@dataclass
class PoolerConfig:
3259
    """Controls the behavior of output pooling in pooling models."""
3260
3261

    pooling_type: Optional[str] = None
3262
    """
3263
    The pooling method of the pooling model. This should be a key in
3264
    [`vllm.model_executor.layers.pooler.PoolingType`][].
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
    """

    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
    """
3281
    If set, only the score corresponding to the ``step_tag_id`` in the
3282
3283
3284
3285
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

3286
    returned_token_ids: Optional[list[int]] = None
3287
    """
3288
3289
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
3290
3291
3292
    ``math-shepherd-mistral-7b-prm`` model.
    """

3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
    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.
3307
        factors: list[Any] = []
3308
3309
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3310
3311
        return hash_str

3312

3313
3314
3315
3316
3317
3318
3319
3320
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

3321
3322
3323
3324
3325
3326
3327
# 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.",
}
3328

3329

3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
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,
3348
    config: PretrainedConfig,
3349
3350
3351
    *,
    revision: Optional[str],
):
3352
3353
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
3354
    config_dtype = getattr(config, "torch_dtype", None)
3355

3356
    # Fallbacks for multi-modal models if the root config
3357
    # does not define torch_dtype
3358
3359
    if config_dtype is None:
        config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
3360
3361
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)
3362
3363
    if config_dtype is None and hasattr(config, "encoder_config"):
        config_dtype = getattr(config.encoder_config, "torch_dtype", None)
3364

3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
    # 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)

3380
3381
3382
    if config_dtype is None:
        config_dtype = torch.float32

3383
    return config_dtype
3384

Shinichi Hemmi's avatar
Shinichi Hemmi committed
3385

3386
3387
3388
3389
3390
3391
3392
def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform
3393

3394
3395
3396
3397
    supported_dtypes = [
        dtype for dtype in current_platform.supported_dtypes
        if _is_valid_dtype(model_type, dtype)
    ]
3398

3399
3400
3401
3402
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
    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,
            )
3452
        else:
3453
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
3454
                raise ValueError(f"Unknown dtype: {dtype!r}")
3455
3456
3457
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
3458
    else:
3459
        raise ValueError(f"Unknown dtype: {dtype}")
3460

3461
3462
    _check_valid_dtype(model_type, torch_dtype)

3463
3464
3465
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
3466
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
3467
3468
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
3469
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
3470
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
3471
            # Casting between float16 and bfloat16 is allowed with a warning.
3472
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
3473
3474

    return torch_dtype
3475
3476
3477
3478


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
3479
    tokenizer_config: Optional[dict],
3480
    max_model_len: Optional[int],
3481
    disable_sliding_window: bool,
3482
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
3483
    spec_target_max_model_len: Optional[int] = None,
3484
    encoder_config: Optional[Any] = None,
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
) -> 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",
3495
3496
        # ChatGLM2
        "seq_length",
3497
3498
        # Command-R
        "model_max_length",
3499
3500
        # Whisper
        "max_target_positions",
3501
3502
3503
3504
3505
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
3506
    # Choose the smallest "max_length" from the possible keys
3507
    max_len_key = None
3508
    for key in possible_keys:
3509
3510
3511
3512
3513
        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
3514
3515
3516
3517
    # 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
3518
3519
3520
3521

    # 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:
3522
3523

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
3524
        max_len_key = "sliding_window" \
3525
3526
3527
            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)
3528

3529
3530
3531
3532
3533
3534
3535
    # 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)

3536
3537
    # If none of the keys were found in the config, use a default and
    # log a warning.
3538
    if derived_max_model_len == float("inf"):
3539
3540
3541
3542
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

3543
3544
3545
3546
3547
        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

3548
3549
3550
3551
        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: "
3552
            "%s. Assuming the model's maximum length is %d.", possible_keys,
3553
            default_max_len)
3554
        derived_max_model_len = default_max_len
3555

3556
    rope_scaling = getattr(hf_config, "rope_scaling", None)
3557
3558
3559
    # 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:
3560
3561
3562
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572

        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.")

3573
3574
3575
3576
            # 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)

3577
3578
3579
3580
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
3581

3582
3583
3584
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

3585
3586
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
3587
    if max_model_len is None:
3588
        max_model_len = int(derived_max_model_len)
3589
3590
3591
3592
3593
3594
3595
3596
        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)
3597
    elif max_model_len > derived_max_model_len:
3598
3599
3600
3601
3602
        # 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:
3603
3604
3605
3606
3607
3608
3609
            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.")
3610
        else:
3611
            msg = (
3612
                f"User-specified max_model_len ({max_model_len}) is greater "
3613
3614
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
3615
                f"{model_max_length} in model's config.json). This may lead "
3616
3617
3618
3619
3620
3621
3622
3623
3624
                "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")
3625
    return int(max_model_len)
3626
3627


3628
def get_min_sliding_window(
3629
        sliding_window: Union[int, list[Optional[int]]]) -> int:
3630
3631
3632
3633
3634
3635
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


3636
def get_served_model_name(model: str,
3637
                          served_model_name: Optional[Union[str, list[str]]]):
3638
    """
3639
3640
3641
3642
    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
3643
3644
3645
3646
3647
3648
3649
3650
3651
    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


3652
GuidedDecodingBackendV0 = Literal["auto", "outlines", "lm-format-enforcer",
3653
                                  "xgrammar", "guidance"]
3654
3655

GuidedDecodingBackendV1 = Literal["auto", "xgrammar", "guidance", "outlines"]
3656
3657
GuidedDecodingBackend = Literal[GuidedDecodingBackendV0,
                                GuidedDecodingBackendV1]
3658
3659
3660


@config
3661
3662
@dataclass
class DecodingConfig:
3663
    """Dataclass which contains the decoding strategy of the engine."""
3664

3665
    backend: GuidedDecodingBackend = "auto" if envs.VLLM_USE_V1 else "xgrammar"
3666
3667
3668
3669
    """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."""
3670

3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
    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`."""

3683
    reasoning_backend: str = ""
3684
    """Select the reasoning parser depending on the model that you're using.
3685
    This is used to parse the reasoning content into OpenAI API format."""
3686

3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
    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.
3701
        factors: list[Any] = []
3702
3703
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3704
3705
        return hash_str

3706
    def __post_init__(self):
3707
        if envs.VLLM_USE_V1:
3708
            valid_guided_backends = get_args(GuidedDecodingBackendV1)
3709
        else:
3710
            valid_guided_backends = get_args(GuidedDecodingBackendV0)
3711
3712
        if self.backend not in valid_guided_backends:
            raise ValueError(f"Invalid backend '{self.backend}',"
3713
                             f" must be one of {valid_guided_backends}")
3714
3715
3716
3717
3718
3719
3720
3721
        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.")

3722

3723
3724
3725
3726
DetailedTraceModules = Literal["model", "worker", "all"]


@config
3727
3728
@dataclass
class ObservabilityConfig:
3729
    """Configuration for observability - metrics and tracing."""
3730

3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
    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)
3746

3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
    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))
3772

3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
    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.
3787
        factors: list[Any] = []
3788
3789
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3790
3791
        return hash_str

3792
    def __post_init__(self):
3793
3794
3795
3796
3797
        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()

3798
        from vllm.tracing import is_otel_available, otel_import_error_traceback
3799
3800
3801
3802
3803
        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}")
3804

3805
3806
3807
3808
3809
3810
    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(","))

3811

3812
3813
3814
3815
3816
3817
3818
3819
KVProducer = Literal["kv_producer", "kv_both"]
KVConsumer = Literal["kv_consumer", "kv_both"]
KVRole = Literal[KVProducer, KVConsumer]


@config
@dataclass
class KVTransferConfig:
3820
3821
3822
    """Configuration for distributed KV cache transfer."""

    kv_connector: Optional[str] = None
3823
3824
    """The KV connector for vLLM to transmit KV caches between vLLM instances.
    """
3825

3826
    engine_id: Optional[str] = None
Robert Shaw's avatar
Robert Shaw committed
3827
3828
    """The engine id for KV transfers."""

3829
    kv_buffer_device: Optional[str] = "cuda"
3830
3831
    """The device used by kv connector to buffer the KV cache.
    Currently only support 'cuda'."""
3832
3833

    kv_buffer_size: float = 1e9
3834
3835
    """The buffer size for TorchDistributedConnector. Measured in number of
    bytes. Recommended value: 1e9 (about 1GB)."""
3836

3837
3838
    kv_role: Optional[KVRole] = None
    """Whether this vLLM instance produces, consumes KV cache, or both. Choices
Robert Shaw's avatar
Robert Shaw committed
3839
    are 'kv_producer', 'kv_consumer', and 'kv_both'."""
3840
3841

    kv_rank: Optional[int] = None
3842
3843
3844
    """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."""
3845
3846

    kv_parallel_size: int = 1
3847
3848
    """The number of parallel instances for KV cache transfer. For
    PyNcclConnector, this should be 2."""
3849
3850

    kv_ip: str = "127.0.0.1"
3851
    """The KV connector ip, used to build distributed connection."""
3852
3853

    kv_port: int = 14579
3854
    """The KV connector port, used to build distributed connection."""
3855

3856
3857
    kv_connector_extra_config: dict[str, Any] = field(default_factory=dict)
    """any extra config that the connector may need."""
3858

3859
3860
3861
3862
    kv_connector_module_path: Optional[str] = None
    """The Python module path to dynamically load the KV connector from.
    Only supported in V1."""

3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
    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.
3877
        factors: list[Any] = []
3878
3879
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3880
3881
        return hash_str

3882
    def __post_init__(self) -> None:
3883
3884
3885
        if self.engine_id is None:
            self.engine_id = str(uuid.uuid4())

3886
3887
3888
        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)}")
3889
3890
3891

        if self.kv_connector is not None and self.kv_role is None:
            raise ValueError("Please specify kv_disagg_role when kv_connector "
3892
                             f"is set, supported roles are {get_args(KVRole)}")
3893
3894
3895
3896

    @property
    def is_kv_transfer_instance(self) -> bool:
        return self.kv_connector is not None and \
3897
            self.kv_role in get_args(KVRole)
3898
3899
3900
3901

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

    @property
    def is_kv_consumer(self) -> bool:
        return self.kv_connector is not None and \
3907
            self.kv_role in get_args(KVConsumer)
3908

3909
3910
3911
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

3912

3913
3914
3915
@config
@dataclass
class KVEventsConfig:
3916
3917
3918
3919
3920
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
    """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.
    """


3955
3956
3957
3958
3959
3960
3961
3962
class CompilationLevel:
    # constants for the levels of the compilation process
    NO_COMPILATION = 0
    DYNAMO_AS_IS = 1
    DYNAMO_ONCE = 2
    PIECEWISE = 3


3963
3964
3965
3966
3967
3968
3969
3970
3971
@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."""

3972
    enable_fusion: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
3973
3974
3975
    """Whether to enable the custom fusion (RMSNorm/SiluMul+quant) pass."""
    enable_attn_fusion: bool = False
    """Whether to enable the custom attention+quant fusion pass."""
3976
    enable_noop: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
3977
3978
3979
    """Whether to enable the custom no-op elimination pass."""
    enable_sequence_parallelism: bool = False
    """Whether to enable sequence parallelism."""
3980
3981
    enable_async_tp: bool = False
    """Whether to enable async TP."""
3982
3983
3984
3985
    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."""
3986

3987
3988
    # TODO(luka) better pass enabling system.

3989
3990
3991
3992
    def uuid(self):
        """
        Produces a hash unique to the pass configuration.
        Any new fields that affect compilation should be added to the hash.
3993
        Any future fields that don't affect compilation should be excluded.
3994
        """
3995
        return InductorPass.hash_dict(asdict(self))
3996
3997

    def __post_init__(self) -> None:
3998
3999
4000
4001
4002
4003
4004
4005
4006
        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")
4007
4008
4009
4010
4011
4012
4013


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

4014
    - Top-level Compilation control:
4015
4016
4017
4018
4019
4020
        - [`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]
4021
    - CudaGraph capture:
4022
4023
4024
4025
4026
4027
4028
4029
        - [`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]
4030
    - Inductor compilation:
4031
4032
4033
4034
4035
        - [`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]
4036
        - custom inductor passes
4037

4038
4039
4040
4041
4042
4043
4044
4045
4046
    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.
4047
4048
    """
    # Top-level Compilation control
4049
    level: int = 0
4050
4051
4052
4053
4054
4055
    """The level of compilation:

    - 0: no compilation.
    - 1: dynamo as is.
    - 2: dynamo once.
    - 3: piecewise compilation."""
4056
    debug_dump_path: str = ""
4057
    """The path to dump the debug information."""
4058
    cache_dir: str = ""
4059
4060
4061
    """The directory to store the compiled graph, to accelerate Inductor
    compilation. By default, it will use model-related information to generate
    a cache directory."""
4062
    backend: str = ""
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
    """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
4085
4086
    disabled when running with Inductor: level>=PIECEWISE and use_inductor=True.
    Inductor generates (fused) Triton kernels for disabled custom ops."""
4087
4088
4089
4090
4091
    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
4092
    use_inductor: bool = True
4093
4094
    """Whether to use inductor compilation:

4095
4096
4097
4098
4099
4100
4101
    - 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.
        
    This setting is ignored if level<PIECEWISE."""
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
    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
4117
    use_cudagraph: bool = field(default_factory=lambda: envs.VLLM_USE_V1)
4118
4119
4120
4121
4122
    """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.
4123
4124
    In the vLLM V1 Engine, this flag only applies for
    CompilationLevel.PIECEWISE (aka -O3).
4125
4126
4127
4128
    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."""
4129
    cudagraph_num_of_warmups: int = 0
4130
4131
4132
4133
    """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."""
4134
    cudagraph_capture_sizes: Optional[list[int]] = None
4135
4136
4137
    """Sizes to capture cudagraph.
    - None (default): capture sizes are inferred from vllm config.
    - list[int]: capture sizes are specified as given."""
4138
    cudagraph_copy_inputs: bool = False
4139
4140
4141
4142
4143
    """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."""
4144
    full_cuda_graph: bool = False
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
    """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."""
4164

4165
    # keep track of enabled and disabled custom ops
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
    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."""
4182

4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
    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.
        """
4195
        factors: list[Any] = []
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
        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()

4206
4207
    def __repr__(self) -> str:
        exclude = {
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
            "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,
            },
4218
        }
4219
4220
4221
4222
        # 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(
4223
4224
4225
                self,
                exclude=exclude,  # type: ignore[arg-type]
                exclude_unset=True).decode())
4226
4227
4228

    __str__ = __repr__

4229
4230
    @classmethod
    def from_cli(cls, cli_value: str) -> "CompilationConfig":
4231
4232
4233
        """Parse the CLI value for the compilation config.
        -O1, -O2, -O3, etc. is handled in FlexibleArgumentParser.
        """
4234
        return TypeAdapter(CompilationConfig).validate_json(cli_value)
4235

4236
    def __post_init__(self) -> None:
4237
4238
4239
4240
        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
4241
4242
4243
4244
4245
4246
4247
4248
        # 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

4249
        if is_torch_equal_or_newer("2.6"):
Michael Goin's avatar
Michael Goin committed
4250
4251
4252
4253
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

4254
4255
4256
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
4257
4258
4259
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
4260
4261
4262
4263
4264
4265
4266
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
4267
4268
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
4269

4270
4271
        if isinstance(self.pass_config, dict):
            self.pass_config = PassConfig(**self.pass_config)
4272

4273
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
        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
4291

4292
        from vllm.compilation.backends import VllmBackend
4293
        return VllmBackend(vllm_config)
4294

4295
    def init_with_cudagraph_sizes(self,
4296
                                  cudagraph_capture_sizes: list[int]) -> None:
4297
        """To complete the initialization of config,
4298
4299
        we need to know the cudagraph sizes."""

4300
        if self.cudagraph_capture_sizes is None:
4301
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
4302
        else:
4303
            # de-duplicate the sizes provided by the config
4304
4305
4306
4307
4308
4309
            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
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324

        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
4325

4326
        # sort to make sure cudagraph capture sizes are in descending order
4327
4328
4329
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
4330

4331
4332
4333
4334
        # 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)
        ]
4335
4336
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
4337
4338
4339
4340
4341
4342
4343
            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
4344

4345
4346
    def set_splitting_ops_for_v1(self):
        # NOTE: this function needs to be called
4347
4348
4349
4350
4351
        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}")

4352
        if not self.splitting_ops:
4353
            self.splitting_ops = [] if self.full_cuda_graph else [
4354
4355
4356
4357
                "vllm.unified_attention",
                "vllm.unified_attention_with_output",
            ]

4358

4359
@config
4360
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
4361
4362
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
4363
4364
4365
    simplifies passing around the distinct configurations in the codebase.
    """

4366
4367
4368
    # TODO: use default_factory once default constructing ModelConfig doesn't
    # try to download a model
    model_config: ModelConfig = None  # type: ignore
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
    """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."""
4380
    lora_config: Optional[LoRAConfig] = None
4381
4382
4383
    """LoRA configuration."""
    speculative_config: Optional[SpeculativeConfig] = None
    """Speculative decoding configuration."""
4384
    decoding_config: DecodingConfig = field(default_factory=DecodingConfig)
4385
    """Decoding configuration."""
4386
    observability_config: Optional[ObservabilityConfig] = None
4387
    """Observability configuration."""
4388
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
4389
    """Prompt adapter configuration."""
4390
    quant_config: Optional[QuantizationConfig] = None
4391
4392
4393
    """Quantization configuration."""
    compilation_config: CompilationConfig = field(
        default_factory=CompilationConfig)
4394
    """`torch.compile` and cudagraph capture configuration for the model.
4395

4396
4397
4398
4399
    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}'`).
    Currently, -O <n> and -O=<n> are supported as well but this will likely be 
    removed in favor of clearer -O<n> syntax in the future.
4400
4401
4402

    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
4403
    production, also default in V1.
4404
4405
4406
4407
4408
4409

    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."""
4410
    kv_events_config: Optional[KVEventsConfig] = None
4411
    """The configurations for event publishing."""
4412
    # some opaque config, only used to provide additional information
4413
4414
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
4415
4416
4417
4418
    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."""
4419
    instance_id: str = ""
4420
    """The ID of the vLLM instance."""
4421

4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
    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.
        """
4434
        factors: list[Any] = []
4435
4436

        # summarize vllm config
4437
        vllm_factors: list[Any] = []
4438
4439
        from vllm import __version__
        vllm_factors.append(__version__)
4440
        vllm_factors.append(envs.VLLM_USE_V1)
4441
4442
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
4443
4444
        else:
            vllm_factors.append("None")
4445
4446
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
4447
4448
        else:
            vllm_factors.append("None")
4449
4450
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
4451
4452
        else:
            vllm_factors.append("None")
4453
4454
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
4455
4456
        else:
            vllm_factors.append("None")
4457
4458
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
4459
4460
        else:
            vllm_factors.append("None")
4461
4462
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
4463
4464
        else:
            vllm_factors.append("None")
4465
4466
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
4467
4468
4469
4470
4471
            # 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))
4472
4473
        else:
            vllm_factors.append("None")
4474
4475
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
4476
4477
        else:
            vllm_factors.append("None")
4478
4479
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
4480
4481
        else:
            vllm_factors.append("None")
4482
4483
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
4484
4485
        else:
            vllm_factors.append("None")
4486
4487
        if self.prompt_adapter_config:
            vllm_factors.append(self.prompt_adapter_config.compute_hash())
4488
4489
        else:
            vllm_factors.append("None")
4490
4491
4492
4493
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
4494
4495
        else:
            vllm_factors.append("None")
4496
4497
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
4498
4499
4500
        else:
            vllm_factors.append("None")
        if self.additional_config:
4501
4502
4503
4504
4505
4506
4507
4508
            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)
4509
4510
        else:
            vllm_factors.append("None")
4511
4512
        factors.append(vllm_factors)

4513
4514
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
4515
4516
        return hash_str

4517
4518
4519
4520
4521
4522
    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]
4523

4524
4525
4526
4527
4528
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
4529
        from vllm.platforms import current_platform
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
        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
4552

4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
    @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)

4564
4565
4566
4567
4568
4569
4570
4571
4572
    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

4573
4574
4575
4576
4577
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

4578
4579
4580
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
4581
4582
4583

        self.try_verify_and_update_config()

4584
4585
4586
4587
4588
        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)
4589
4590
            self.model_config.verify_dual_chunk_attention_config(
                self.load_config)
4591

4592
        self.cache_config.verify_with_parallel_config(self.parallel_config)
4593

4594
        if self.lora_config is not None:
4595
            self.lora_config.verify_with_cache_config(self.cache_config)
4596
            self.lora_config.verify_with_model_config(self.model_config)
4597
            self.lora_config.verify_lora_support()
4598
        if self.prompt_adapter_config is not None:
4599
4600
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
4601

4602
        if self.quant_config is None and self.model_config is not None:
4603
4604
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
4605

4606
        from vllm.platforms import current_platform
4607
        if self.model_config is not None and \
4608
4609
4610
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
4611
            logger.warning_once(
4612
4613
4614
4615
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

4616
4617
4618
4619
4620
        # 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
4621
4622
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
4623
4624
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
4625
4626
            # By default, V1 uses piecewise CUDA graphs. If full_cuda_graph
            # is set to True, full CUDA graphs will be used.
4627
            self.compilation_config.cudagraph_num_of_warmups = 1
4628
            self.compilation_config.level = CompilationLevel.PIECEWISE
4629
            self.compilation_config.set_splitting_ops_for_v1()
4630

4631
        self._set_cudagraph_sizes()
4632

4633
        if self.cache_config.cpu_offload_gb > 0 and \
4634
4635
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
4636
            logger.warning(
4637
                "CPU offload is not supported with `torch.compile` in v0 yet."
4638
4639
4640
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4641
4642
4643
4644
4645
4646
        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`.")
4647
4648
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4649
4650
        if self.compilation_config.full_cuda_graph and \
            not self.model_config.disable_cascade_attn:
4651
4652
            logger.info("full_cuda_graph is not supported with "
                        "cascade attention. Disabling cascade attention.")
4653
            self.model_config.disable_cascade_attn = True
4654

4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
        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

4676
        if (self.kv_events_config is not None
4677
4678
4679
4680
4681
                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.")
4682
4683
        if (self.kv_events_config is not None
                and self.kv_events_config.publisher != "null"
4684
4685
4686
4687
4688
                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.")
4689
4690
        current_platform.check_and_update_config(self)

4691
4692
4693
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

4694
4695
4696
4697
4698
4699
4700
        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.
4701
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4702
4703
            if self.kv_transfer_config is not None:
                # Hybrid KV cache manager is not compatible with KV transfer.
4704
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4705
4706
            if self.kv_events_config is not None:
                # Hybrid KV cache manager is not compatible with KV events.
4707
                self.scheduler_config.disable_hybrid_kv_cache_manager = True
4708

4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
    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
        ]

4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
    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.

4745
4746
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
4747
4748

        During runtime, if batchsize is larger than
4749
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4750
4751
        no cudagraph will be used.
        If the batch size is no larger than
4752
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
        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)]
4765
4766
4767
4768
4769
                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)

4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
                # 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:
4791
4792
4793
4794
4795
4796
4797
4798
                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
4799
                    raise TypeError(f"Invalid value for {cuda_graph_sizes=}.")
4800
4801
4802
4803
                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)
4804
4805
4806
4807
4808
                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
                ]
4809
4810
4811
4812

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

4813
    def recalculate_max_model_len(self, max_model_len: int):
4814
        # Can only be called in try_verify_and_update_config
4815
        model_config = self.model_config
4816
        max_model_len = model_config.get_and_verify_max_len(max_model_len)
4817
4818
        self.model_config.max_model_len = max_model_len
        self.scheduler_config.max_model_len = max_model_len
4819
4820
4821
4822
4823
4824

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

4825
4826
        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP, HybridAttentionMambaModelConfig)
4827
4828
4829
        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_config(self)
4830

4831
4832
4833
        if self.model_config.is_hybrid:
            HybridAttentionMambaModelConfig.verify_and_update_config(self)

4834
4835
4836
4837
4838
4839
        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)

4840
    def __str__(self):
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
        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}, "
4871
4872
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
4873
4874
4875


_current_vllm_config: Optional[VllmConfig] = None
4876
_current_prefix: Optional[str] = None
4877
4878
4879


@contextmanager
4880
4881
4882
def set_current_vllm_config(vllm_config: VllmConfig,
                            check_compile=False,
                            prefix: Optional[str] = None):
4883
    """
4884
    Temporarily set the current vLLM config.
4885
    Used during model initialization.
4886
    We save the current vLLM config in a global variable,
4887
    so that all modules can access it, e.g. custom ops
4888
    can access the vLLM config to determine how to dispatch.
4889
    """
4890
    global _current_vllm_config, _current_prefix
4891
    old_vllm_config = _current_vllm_config
4892
    old_prefix = _current_prefix
4893
4894
4895
4896
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
4897
        _current_prefix = prefix
4898
        yield
4899
4900
4901
    except Exception:
        raise
    else:
4902
4903
4904
4905
        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)
4906
4907
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
4908
4909
4910
4911
4912
4913
4914
4915
4916
            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"
4917
                " if you want it to be supported.",
4918
                vllm_config.model_config.model)
4919
    finally:
4920
        _current_vllm_config = old_vllm_config
4921
        _current_prefix = old_prefix
4922
4923
4924
4925
4926
4927
4928


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.
4929
        logger.warning("Current vLLM config is not set.")
4930
4931
4932
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
4933
4934


4935
4936
4937
4938
4939
4940
4941
4942
4943
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


4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
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:
4955
        result (bool): `True` if a match is found, `False` otherwise.
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
    """
    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}")
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981


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)
    }
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011


@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:
5012
        return self.min_energy_split_window_size is not None
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030


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)