config.py 189 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import ast
4
import copy
5
import enum
6
import hashlib
7
import inspect
8
import json
9
import re
10
import sys
11
import textwrap
12
import warnings
13
from collections import Counter
14
from contextlib import contextmanager
15
16
from dataclasses import (MISSING, dataclass, field, fields, is_dataclass,
                         replace)
17
from functools import cached_property
18
from importlib.util import find_spec
19
from pathlib import Path
20
21
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Literal, Optional,
                    Protocol, TypeVar, Union, cast, get_args, get_origin)
22
23

import torch
24
from pydantic import BaseModel, Field, PrivateAttr
25
from torch.distributed import ProcessGroup, ReduceOp
26
from transformers import PretrainedConfig
27
from typing_extensions import deprecated
28

29
import vllm.envs as envs
30
from vllm import version
31
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
32
from vllm.logger import init_logger
33
from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
34
                                                     QuantizationMethods,
35
                                                     get_quantization_config)
36
from vllm.model_executor.models import ModelRegistry
37
from vllm.platforms import CpuArchEnum, current_platform
38
from vllm.tracing import is_otel_available, otel_import_error_traceback
39
40
41
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
42
43
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
    try_get_generation_config, uses_mrope)
44
from vllm.transformers_utils.s3_utils import S3Model
45
from vllm.transformers_utils.utils import is_s3, maybe_model_redirect
46
from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
47
48
                        get_cpu_memory, get_open_port, is_torch_equal_or_newer,
                        random_uuid, resolve_obj_by_qualname)
49

50
if TYPE_CHECKING:
51
    from _typeshed import DataclassInstance
52
53
    from ray.util.placement_group import PlacementGroup

54
    from vllm.executor.executor_base import ExecutorBase
55
56
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
57
    from vllm.model_executor.model_loader.loader import BaseModelLoader
58

59
    ConfigType = type[DataclassInstance]
60
61
else:
    QuantizationConfig = None
62
    ConfigType = type
63

64
65
logger = init_logger(__name__)

66
67
ConfigT = TypeVar("ConfigT", bound=ConfigType)

68
69
70
# This value is chosen to have a balance between ITL and TTFT. Note it is
# not optimized for throughput.
_DEFAULT_MAX_NUM_BATCHED_TOKENS = 2048
71
_POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
72
_MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
73

74
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
75
                     "score", "reward", "transcription"]
76

77
_ResolvedTask = Literal["generate", "embed", "classify", "score", "reward",
78
                        "draft", "transcription"]
79

80
RunnerType = Literal["generate", "pooling", "draft", "transcription"]
81

82
_RUNNER_TASKS: dict[RunnerType, list[_ResolvedTask]] = {
83
84
85
    "generate": ["generate"],
    "pooling": ["embed", "classify", "score", "reward"],
    "draft": ["draft"],
86
    "transcription": ["transcription"],
87
88
}

89
_TASK_RUNNER: dict[_ResolvedTask, RunnerType] = {
90
    task: runner
91
92
    for runner, tasks in _RUNNER_TASKS.items()
    for task in tasks
93
}
94

95
HfOverrides = Union[dict[str, Any], Callable[[PretrainedConfig],
96
97
                                             PretrainedConfig]]

98

99
100
101
102
103
104
class SupportsHash(Protocol):

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


105
106
class SupportsMetricsInfo(Protocol):

107
    def metrics_info(self) -> dict[str, str]:
108
109
110
        ...


111
112
113
114
115
116
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"


117
118
119
120
121
122
123
124
125
126
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
127

128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
        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


169
def config(cls: ConfigT) -> ConfigT:
170
171
172
173
174
175
176
177
178
179
180
181
    """
    A decorator that ensures all fields in a dataclass have default values
    and that each field has a docstring.
    """
    if not is_dataclass(cls):
        raise TypeError("The decorated class must be a dataclass.")
    attr_docs = get_attr_docs(cls)
    for f in fields(cls):
        if f.init and f.default is MISSING and f.default_factory is MISSING:
            raise ValueError(
                f"Field '{f.name}' in {cls.__name__} must have a default value."
            )
182

183
184
185
        if f.name not in attr_docs:
            raise ValueError(
                f"Field '{f.name}' in {cls.__name__} must have a docstring.")
186
187
188
189
190
191
192
193
194

        if get_origin(f.type) is Union:
            args = get_args(f.type)
            literal_args = [arg for arg in args if get_origin(arg) is Literal]
            if len(literal_args) > 1:
                raise ValueError(
                    f"Field '{f.name}' in {cls.__name__} must use a single "
                    "Literal type. Please use 'Literal[Literal1, Literal2]' "
                    "instead of 'Union[Literal1, Literal2]'.")
195
196
197
    return cls


198
def get_field(cls: ConfigType, name: str) -> Field:
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
    """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__}.")
    named_field: Field = cls_fields.get(name)
    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.")


215
216
217
218
219
220
TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]


@config
@dataclass
221
class ModelConfig:
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
    """Configuration for the model."""

    model: str = "facebook/opt-125m"
    """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."""
    task: Literal[TaskOption, Literal["draft"]] = "auto"
    """The task to use the model for. Each vLLM instance only supports one
    task, even if the same model can be used for multiple tasks. When the model
    only supports one task, "auto" can be used to select it; otherwise, you
    must specify explicitly which task to use."""
    tokenizer: str = None  # type: ignore
    """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
    """Random seed for reproducibility."""
    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)
    """RoPE scaling configuration in JSON format. For example,
    `{"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."""
    max_model_len: int = None  # type: ignore
    """Model context length (prompt and output). If unspecified, will be
    automatically derived from the model config.
283

284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
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
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
    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
    """Specify the the maximum length for spec decoding draft models."""
    quantization: Optional[QuantizationMethods] = None
    """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."""
    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."""
    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
    config. If a callable, it is called to update the HuggingFace config. When
    specified via CLI, the argument must be a valid JSON string."""
    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}`.
    When specified via CLI, the argument must be a valid JSON string."""
    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
    arguments. e.g. `{"cast_logits_dtype": "bloat16"}`. When specified via CLI,
    the argument must be a valid JSON string."""
    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}`.
    When specified via CLI, the argument must be a valid JSON string."""
    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
    `--generation-config vllm`, only the override parameters are used.
    When specified via CLI, the argument must be a valid JSON string."""
    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."""
399

400
401
402
403
404
405
406
407
408
409
410
411
    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.
        """
412
        factors: list[Any] = []
413
414
415
416
417
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
418
419
420
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
421
        factors.append(self.trust_remote_code)
422
423
424
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
425
426
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
427
428
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
429
430
        str_factors = str(factors)
        assert_hashable(str_factors)
431
432
        return hashlib.sha256(str(factors).encode()).hexdigest()

433
434
435
436
437
438
439
440
441
442
443
444
445
    def __post_init__(self) -> None:
        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):
446
            hf_overrides_kw = {}
447
            hf_overrides_fn = self.hf_overrides
448
        else:
449
            hf_overrides_kw = self.hf_overrides
450
            hf_overrides_fn = None
451

452
453
        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
454
            hf_overrides_kw.update(hf_override)
455
            hf_overrides_str = json.dumps(hf_overrides_kw)
456
457
458
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
459
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
460
461
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
462
            hf_overrides_kw.update(hf_override)
463
            hf_overrides_str = json.dumps(hf_overrides_kw)
464
465
466
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
467
468
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

469
        self.maybe_pull_model_tokenizer_for_s3(self.model, self.tokenizer)
470

471
472
473
474
        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 "
475
476
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
477
478
                "for instructions on how to install it.")

479
480
        from vllm.platforms import current_platform

481
482
483
484
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
485

486
487
488
        if isinstance(self.config_format, str):
            self.config_format = ConfigFormat(self.config_format)

489
        hf_config = get_config(self.hf_config_path or self.model,
490
491
                               self.trust_remote_code, self.revision,
                               self.code_revision, self.config_format)
492
493
494
495
496
497
498
499

        if hf_overrides_kw:
            logger.info("Overriding HF config with %s", hf_overrides_kw)
            hf_config.update(hf_overrides_kw)
        if hf_overrides_fn:
            logger.info("Overriding HF config with %s", hf_overrides_fn)
            hf_config = hf_overrides_fn(hf_config)

500
501
        self.hf_config = hf_config

502
        self.hf_text_config = get_hf_text_config(self.hf_config)
503
504
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
505
        self.encoder_config = self._get_encoder_config()
506
        self.hf_image_processor_config = get_hf_image_processor_config(
507
508
            self.model, hf_token=self.hf_token, revision=self.revision)
        self.dtype = _get_and_verify_dtype(self.hf_config, self.dtype)
509

510
        interleaved_attn_models = ["gemma2", "gemma3_text", "cohere2"]
511
512
513
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
        has_interleaved_attention = (sliding_window is not None) and (
            isinstance(sliding_window, list) or
514
            (self.hf_text_config.model_type in interleaved_attn_models))
515
516

        if (not self.disable_sliding_window and has_interleaved_attention):
517
518
            if (backend :=
                    envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"):
519
520
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
521

522
                logger.warning_once(
523
524
525
526
527
                    "%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,
                )
528
529
530
531
532
533
534
535
536
537
                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
                delattr(self.hf_text_config, "sliding_window")
                sliding_window = None
Woosuk Kwon's avatar
Woosuk Kwon committed
538

539
540
        self.max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
541
            max_model_len=self.max_model_len,
542
            disable_sliding_window=self.disable_sliding_window,
543
            sliding_window_len=self.get_hf_config_sliding_window(),
544
            spec_target_max_model_len=self.spec_target_max_model_len,
545
            encoder_config=self.encoder_config)
546
547
548
        self.served_model_name = get_served_model_name(self.model,
                                                       self.served_model_name)
        self.multimodal_config = self._init_multimodal_config()
549
550
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
551

552
        self.is_attention_free = self._init_attention_free()
553
        self.is_hybrid = self._init_is_hybrid()
554
        self.has_noops = self._init_has_noops()
555
556
        self.has_inner_state = self._init_has_inner_state()

557
558
559
        if (not current_platform.is_neuron() and self.override_neuron_config):
            raise ValueError(
                "`override_neuron_config` is only supported on Neuron.")
560

561
        supported_tasks, task = self._resolve_task(self.task)
562
        self.supported_tasks = supported_tasks
563
        self.task = task
564
565
566
567
        if self.task in ("draft", "generate"):
            self.truncation_side = "left"
        else:
            self.truncation_side = "right"
568

569
        self.pooler_config = self._init_pooler_config()
570

571
        self._verify_quantization()
572
        self._verify_cuda_graph()
573
        self._verify_bnb_config()
574

575
576
577
578
579
580
581
582
    @property
    def registry(self):
        return ModelRegistry

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

583
584
585
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
        """
586
        Pull the model config or tokenizer to a temporary
587
588
589
590
591
592
593
594
595
        directory in case of S3.

        Args:
            model: The model name or path.
            tokenizer: The tokenizer name or path.

        """
        if is_s3(model) or is_s3(tokenizer):
            if is_s3(model):
596
                s3_model = S3Model()
597
598
                s3_model.pull_files(
                    model, allow_pattern=["*.model", "*.py", "*.json"])
599
                self.model_weights = self.model
600
                self.model = s3_model.dir
601
602

            if is_s3(tokenizer):
603
604
                s3_tokenizer = S3Model()
                s3_tokenizer.pull_files(
605
                    model, ignore_pattern=["*.pt", "*.safetensors", "*.bin"])
606
                self.tokenizer = s3_tokenizer.dir
607

608
    def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
609
        if self.registry.is_multimodal_model(self.architectures):
610
            return MultiModalConfig(
611
612
613
614
                limit_per_prompt=self.limit_mm_per_prompt,
                mm_processor_kwargs=self.mm_processor_kwargs,
                disable_mm_preprocessor_cache=self.
                disable_mm_preprocessor_cache)
615

616
        if self.limit_mm_per_prompt:
617
618
            raise ValueError("`limit_mm_per_prompt` is only supported for "
                             "multimodal models.")
619
        if self.mm_processor_kwargs:
620
621
            raise ValueError("`mm_processor_kwargs` is only supported for "
                             "multimodal models.")
622
        if self.disable_mm_preprocessor_cache:
623
624
            raise ValueError("`disable_mm_preprocessor_cache` is only "
                             "supported for multimodal models.")
625
626

        return None
627

628
629
630
631
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

632
    def _init_pooler_config(self) -> Optional["PoolerConfig"]:
633

634
        if self.runner_type == "pooling":
635
636
637
638
639
            if isinstance(self.override_pooler_config, dict):
                self.override_pooler_config = PoolerConfig(
                    **self.override_pooler_config)

            pooler_config = self.override_pooler_config or PoolerConfig()
640
641
642
643
644

            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():
645
646
                    if getattr(pooler_config, k) is None:
                        setattr(pooler_config, k, v)
647

648
            if self.is_matryoshka:
649
650
651
                if pooler_config.normalize is None:
                    pooler_config.normalize = True
                elif not pooler_config.normalize:
652
653
654
655
656
                    raise ValueError(
                        "`normalize` must be enabled (set to True) "
                        "for models that are compatible with "
                        "Matryoshka Representation.")

657
            return pooler_config
658

659
660
        return None

661
    def _init_attention_free(self) -> bool:
662
        return self.registry.is_attention_free_model(self.architectures)
663

664
    def _init_is_hybrid(self) -> bool:
665
        return self.registry.is_hybrid_model(self.architectures)
666

667
668
669
670
    def _init_has_noops(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return self.registry.is_noops_model(architectures)

671
    def _init_has_inner_state(self) -> bool:
672
        return self.registry.model_has_inner_state(self.architectures)
673

674
    def _verify_tokenizer_mode(self) -> None:
675
676
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
677
678
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
679
                f"one of {get_args(TokenizerMode)}.")
680
        self.tokenizer_mode = tokenizer_mode
681

682
683
    def _get_preferred_task(
        self,
684
685
        architectures: list[str],
        supported_tasks: set[_ResolvedTask],
686
687
688
689
    ) -> Optional[_ResolvedTask]:
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
690
        if self.registry.is_cross_encoder_model(architectures):
691
            return "score"
692
        if self.registry.is_transcription_model(architectures):
693
            return "transcription"
694

695
        suffix_to_preferred_task: list[tuple[str, _ResolvedTask]] = [
696
697
698
699
700
701
702
703
704
            # Other models follow this pattern
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ForSequenceClassification", "classify"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]
705
        _, arch = self.registry.inspect_model_cls(architectures)
706
707
708
709
710
711
712

        for suffix, pref_task in suffix_to_preferred_task:
            if arch.endswith(suffix) and pref_task in supported_tasks:
                return pref_task

        return None

713
714
    def _resolve_task(
        self,
715
        task_option: Literal[TaskOption, Literal["draft"]],
716
    ) -> tuple[set[_ResolvedTask], _ResolvedTask]:
717
718
719
        if task_option == "draft":
            return {"draft"}, "draft"

720
721
        registry = self.registry
        architectures = self.architectures
722

723
        runner_support: dict[RunnerType, bool] = {
724
725
            # NOTE: Listed from highest to lowest priority,
            # in case the model supports multiple of them
726
727
728
            "transcription": registry.is_transcription_model(architectures),
            "generate": registry.is_text_generation_model(architectures),
            "pooling": registry.is_pooling_model(architectures),
729
        }
730
        supported_runner_types_lst: list[RunnerType] = [
731
732
733
734
735
            runner_type
            for runner_type, is_supported in runner_support.items()
            if is_supported
        ]

736
        supported_tasks_lst: list[_ResolvedTask] = [
737
738
            task for runner_type in supported_runner_types_lst
            for task in _RUNNER_TASKS[runner_type]
739
740
741
742
743
        ]
        supported_tasks = set(supported_tasks_lst)

        if task_option == "auto":
            selected_task = next(iter(supported_tasks_lst))
744

745
746
747
748
749
            if len(supported_tasks_lst) > 1:
                preferred_task = self._get_preferred_task(
                    architectures, supported_tasks)
                if preferred_task is not None:
                    selected_task = preferred_task
750

751
752
753
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
754
        else:
755
756
757
758
759
760
761
762
763
764
765
766
767
768
            # Aliases
            if task_option == "embedding":
                preferred_task = self._get_preferred_task(
                    architectures, supported_tasks)
                if preferred_task != "embed":
                    msg = ("The 'embedding' task will be restricted to "
                           "embedding models in a future release. Please "
                           "pass `--task classify`, `--task score`, or "
                           "`--task reward` explicitly for other pooling "
                           "models.")
                    warnings.warn(msg, DeprecationWarning, stacklevel=2)

                task_option = preferred_task or "embed"

769
770
771
772
773
774
775
            if task_option not in supported_tasks:
                msg = (
                    f"This model does not support the '{task_option}' task. "
                    f"Supported tasks: {supported_tasks}")
                raise ValueError(msg)

            selected_task = task_option
776

777
        return supported_tasks, selected_task
778

779
780
781
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
782
            # compressed-tensors uses a "compression_config" key
783
            quant_cfg = getattr(self.hf_config, "compression_config", None)
784
785
        return quant_cfg

786
    def _verify_quantization(self) -> None:
787
        supported_quantization = QUANTIZATION_METHODS
788
        optimized_quantization_methods = [
789
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
790
791
            "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
            "quark", "nvfp4", "bitblas", "gptq_bitblas"
792
        ]
793
        if self.quantization is not None:
794
795
            self.quantization = cast(QuantizationMethods,
                                     self.quantization.lower())
796
797

        # Parse quantization method from the HF model config, if available.
798
799
        quant_cfg = self._parse_quant_hf_config()

800
801
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
802
803
804
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
            quant_cfg["quant_method"] = quant_method
805

806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
            # 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

827
            # Detect which checkpoint is it
828
            for name in quantization_methods:
829
                method = get_quantization_config(name)
830
831
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
832
833
834
835
836
837
838
839
840
841
842
                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.
                    if (name in get_args(QuantizationMethods)
                            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.")
843
844
845
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
846

847
            # Verify quantization configurations.
848
            if self.quantization is None:
849
850
                self.quantization = quant_method
            elif self.quantization != quant_method:
851
852
                raise ValueError(
                    "Quantization method specified in the model config "
853
                    f"({quant_method}) does not match the quantization "
854
855
856
857
858
859
860
861
                    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}.")
862
            from vllm.platforms import current_platform
863
            current_platform.verify_quantization(self.quantization)
864
            if self.quantization not in optimized_quantization_methods:
865
                logger.warning(
866
                    "%s quantization is not fully "
867
                    "optimized yet. The speed can be slower than "
868
                    "non-quantized models.", self.quantization)
869

870
    def _verify_cuda_graph(self) -> None:
871
872
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
873
874
875
876
877
878
879
        ROCM_UNSUPPORTED_MODELS = ['mllama']
        if (self.hf_config.model_type in ROCM_UNSUPPORTED_MODELS
                and not self.enforce_eager and current_platform.is_rocm()):
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
                "to the eager mode.", self.hf_config.model_type)
            self.enforce_eager = True
880

881
882
    def _verify_bnb_config(self) -> None:
        """
883
        The current version of bitsandbytes (0.45.3) with 8-bit models does not
884
        yet support CUDA graph.
885
        # TODO Remove this when bitsandbytes supports.
886
887
888
889
890
891
892
893
894
895
896
897
898
899
        """
        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(
900
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
901
                "fallback to the eager mode.")
902

903
904
            self.enforce_eager = True

905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
    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.")

922
923
924
925
926
927
928
929
930
931
    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

932
        # Reminder: Please update docs/source/features/compatibility_matrix.md
933
        # If the feature combo become valid
934
        from vllm.platforms import current_platform
935
        if not current_platform.is_async_output_supported(self.enforce_eager):
936
937
938
939
940
941
942
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

943
        # Async postprocessor is not necessary for pooling models
944
        # since there is no token generation
945
        if self.runner_type == "pooling":
946
947
            self.use_async_output_proc = False

948
        # Reminder: Please update docs/source/features/compatibility_matrix.md
949
        # If the feature combo become valid
950
951
952
        if speculative_config:
            self.use_async_output_proc = False

953
954
955
956
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
957
958
959
960
961
962

        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.")

963
964
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
965
966
967
968
969
970
971
        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}).")

972
        if parallel_config.enable_expert_parallel:
973
974
            self._verify_with_expert_parallelism()

975
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
976
        if pipeline_parallel_size > 1:
977
            if not self.registry.is_pp_supported_model(self.architectures):
978
979
980
981
982
983
                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
984

985
    def get_hf_config_sliding_window(
986
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
987
        """Get the sliding window size, or None if disabled."""
988
989
990
991

        # 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.
992
993
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
994
            return None
995
        return getattr(self.hf_text_config, "sliding_window", None)
996

997
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
998
999
1000
1001
1002
1003
1004
1005
        """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()

1006
    def get_vocab_size(self) -> int:
1007
        return self.hf_text_config.vocab_size
1008

1009
    def get_hidden_size(self) -> int:
1010
        return self.hf_text_config.hidden_size
1011

1012
1013
    @property
    def is_deepseek_mla(self) -> bool:
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
        if not hasattr(self.hf_text_config, "model_type"):
            return False
        elif self.hf_text_config.model_type in \
            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp'):
            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
1026

1027
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1028
        # TODO remove hard code
1029
        if self.is_deepseek_mla:
1030
1031
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1032
            if self.use_mla:
1033
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1034
1035
1036
1037
1038
            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
1039

1040
1041
1042
1043
1044
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1045
1046
1047
        if self.is_attention_free:
            return 0

1048
1049
        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
1050
            return self.hf_text_config.head_dim
1051

1052
        # FIXME(woosuk): This may not be true for all models.
1053
1054
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1055

1056
1057
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1058
        # For GPTBigCode & Falcon:
1059
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1060
1061
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1062
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1063
        new_decoder_arch_falcon = (
1064
            self.hf_config.model_type in falcon_model_types
1065
            and getattr(self.hf_config, "new_decoder_architecture", False))
1066
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1067
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1068
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1069
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1070
            return 1
1071

1072
        # For DBRX and MPT
1073
1074
1075
1076
1077
        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":
1078
1079
1080
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1081
1082
1083
1084
1085
1086
1087
1088
        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")

1089
1090
1091
        if self.is_attention_free:
            return 0

1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
        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:
1102
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1103
1104
1105
1106
1107
            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.
1108
        return self.hf_text_config.num_attention_heads
1109
1110
1111

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1112
1113
1114
1115
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1116
1117
1118
1119
1120
1121
1122
        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)
1123

1124
1125
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1126
1127
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1128

1129
    def get_layers_start_end_indices(
1130
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1131
        from vllm.distributed.utils import get_pp_indices
1132
1133
1134
1135
1136
1137
        if self.hf_text_config.model_type == "deepseek_mtp":
            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)
1138
1139
1140
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1141
1142
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1143
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1144

1145
1146
1147
    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
1148

1149
1150
1151
1152
1153
1154
1155
1156
    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
1157
1158
1159
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
        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
1170
1171
1172
1173
        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])
1174
        else:
1175
            # Hybrid model Jamba
1176
1177
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
            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
1203

1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
    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

1216
    def try_get_generation_config(self) -> dict[str, Any]:
1217
        if self.generation_config in ("auto", "vllm"):
1218
            config = try_get_generation_config(
1219
                self.hf_config_path or self.model,
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
                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()

1234
    def get_diff_sampling_param(self) -> dict[str, Any]:
1235
        """
1236
        This method returns a dictionary containing the parameters
1237
1238
        that differ from the default sampling parameters. If
        `generation_config` is `"vllm"`, an empty dictionary is returned.
1239
1240

        Returns:
1241
            dict[str, Any]: A dictionary with the differing sampling
1242
            parameters, if `generation_config` is `"vllm"` an empty dictionary.
1243
        """
1244
        if self.generation_config == "vllm":
1245
1246
1247
1248
1249
1250
1251
            config = {}
        else:
            config = self.try_get_generation_config()

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

1252
1253
1254
1255
1256
1257
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1258
            "max_new_tokens",
1259
1260
1261
1262
1263
1264
        ]
        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
            }
1265
1266
1267
1268
1269
            # 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")
1270
1271
        else:
            diff_sampling_param = {}
1272
1273
1274
1275
1276
1277
1278

        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`.")
1279
1280
        return diff_sampling_param

1281
    @property
1282
    def is_encoder_decoder(self) -> bool:
1283
        """Extract the HF encoder/decoder model flag."""
1284
1285
1286
1287
1288
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1289

1290
1291
1292
1293
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1294
1295
    @property
    def is_cross_encoder(self) -> bool:
1296
        return self.registry.is_cross_encoder_model(self.architectures)
1297

1298
1299
    @property
    def use_mla(self) -> bool:
1300
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1301

1302
    @property
1303
    def supported_runner_types(self) -> set[RunnerType]:
1304
1305
1306
1307
        return {_TASK_RUNNER[task] for task in self.supported_tasks}

    @property
    def runner_type(self) -> RunnerType:
1308
        return _TASK_RUNNER[cast(_ResolvedTask, self.task)]
1309

1310
1311
1312
1313
1314
    @property
    def is_v1_compatible(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_v1_compatible(architectures)

1315
1316
1317
1318
1319
    @property
    def is_matryoshka(self) -> bool:
        return (hasattr(self.hf_config, "matryoshka_dimensions")
                or getattr(self.hf_config, "is_matryoshka", False))

1320
1321
1322
1323
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)

1324

1325
BlockSize = Literal[1, 8, 16, 32, 64, 128]
1326
1327
1328
1329
1330
1331
CacheDType = Literal["auto", "fp8", "fp8_e4m3", "fp8_e5m2"]
PrefixCachingHashAlgo = Literal["builtin", "sha256"]


@config
@dataclass
1332
class CacheConfig:
1333
    """Configuration for the KV cache."""
1334

1335
    block_size: BlockSize = None  # type: ignore
1336
1337
1338
    """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.
1339
1340
1341
1342

    This config has no static default. If left unspecified by the user, it will
    be set in `Platform.check_and_update_configs()` based on the current
    platform."""
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
    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
    fp8 (=fp8_e4m3)."""
    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
    - "sha256" is collision resistant but with certain overheads."""
    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."""

    # 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."""
1392

1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
    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.
        """
1405
        factors: list[Any] = []
1406
1407
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
1408
1409
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1410
1411
        return hash_str

1412
1413
1414
    def __post_init__(self) -> None:
        self.swap_space_bytes = self.swap_space * GiB_bytes

1415
        self._verify_args()
1416
        self._verify_cache_dtype()
1417
        self._verify_prefix_caching()
1418

1419
    def metrics_info(self):
1420
1421
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1422
1423
        return {key: str(value) for key, value in self.__dict__.items()}

1424
    def _verify_args(self) -> None:
1425
1426
1427
1428
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1429
1430
1431
1432
1433
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1434
1435
1436
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1437
        elif self.cache_dtype in get_args(CacheDType):
1438
            logger.info(
1439
1440
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1441
1442
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
1443
1444
1445
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1446
1447
1448
1449
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

1450
        if self.sliding_window is not None and not envs.VLLM_USE_V1:
1451
1452
1453
1454
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

1455
1456
        if (self.enable_prefix_caching and self.prefix_caching_hash_algo
                not in get_args(PrefixCachingHashAlgo)):
1457
1458
            raise ValueError(
                "Unknown prefix caching hash algorithm: "
1459
1460
                f"{self.prefix_caching_hash_algo}. Must be one of "
                f"{get_args(PrefixCachingHashAlgo)}.")
1461

1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
    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

1472
1473
1474
        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.")
1475
1476
1477
        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:
1478
            logger.warning("Possibly too large swap space. %s", msg)
1479

1480

1481
@config
1482
1483
@dataclass
class TokenizerPoolConfig:
1484
    """This config is deprecated and will be removed in a future release.
1485

1486
1487
1488
    Passing these parameters will have no effect. Please remove them from your
    configurations.
    """
1489

1490
1491
1492
1493
1494
1495
1496
1497
    pool_size: int = 0
    """This parameter is deprecated and will be removed in a future release.
    Passing this parameter will have no effect. Please remove it from your
    configurations."""
    pool_type: str = "ray"
    """This parameter is deprecated and will be removed in a future release.
    Passing this parameter will have no effect. Please remove it from your
    configurations."""
1498
    extra_config: dict = field(default_factory=dict)
1499
1500
1501
    """This parameter is deprecated and will be removed in a future release.
    Passing this parameter will have no effect. Please remove it from your
    configurations."""
1502

1503
1504
1505
1506
1507
    def __post_init__(self) -> None:
        logger.warning_once(
            "TokenizerPoolConfig is deprecated and will be removed in a "
            "future release. Passing this parameter will have no effect. "
            "Please remove it from your configurations.")
1508
1509


1510
1511
1512
1513
1514
1515
1516
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
1517
    SHARDED_STATE = "sharded_state"
1518
    GGUF = "gguf"
1519
    BITSANDBYTES = "bitsandbytes"
1520
    MISTRAL = "mistral"
1521
    RUNAI_STREAMER = "runai_streamer"
1522
    RUNAI_STREAMER_SHARDED = "runai_streamer_sharded"
1523
    FASTSAFETENSORS = "fastsafetensors"
1524
1525


1526
@config
1527
1528
@dataclass
class LoadConfig:
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
    """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."""
1554
    download_dir: Optional[str] = None
1555
1556
    """Directory to download and load the weights, default to the default
    cache directory of Hugging Face."""
1557
    model_loader_extra_config: dict = field(default_factory=dict)
1558
1559
1560
    """Extra config for model loader. This will be passed to the model loader
    corresponding to the chosen load_format. This should be a JSON string that
    will be parsed into a dictionary."""
1561
    ignore_patterns: Optional[Union[list[str], str]] = None
1562
1563
    """The list of patterns to ignore when loading the model. Default to
    "original/**/*" to avoid repeated loading of llama's checkpoints."""
1564
    use_tqdm_on_load: bool = True
1565
1566
    """Whether to enable tqdm for showing progress bar when loading model
    weights."""
1567

1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
    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.
1582
        factors: list[Any] = []
1583
1584
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1585
1586
        return hash_str

1587
    def __post_init__(self):
1588
1589
1590
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1591

1592
1593
1594
1595
1596
1597
1598
        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/**/*"]

1599

1600
1601
1602
DistributedExecutorBackend = Literal["ray", "mp", "uni", "external_launcher"]


1603
@config
1604
@dataclass
1605
class ParallelConfig:
1606
    """Configuration for the distributed execution."""
1607

1608
1609
1610
1611
1612
1613
1614
1615
1616
    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."""
    data_parallel_rank: int = 0
    """Rank of the data parallel group."""
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
    _data_parallel_rank_local: Optional[int] = field(default=None, init=False)
    """Private field to store the local rank of the data parallel group."""

    @property
    def data_parallel_rank_local(self) -> int:
        """Local rank of the data parallel group, defaults to global rank."""
        if self._data_parallel_rank_local is None:
            return self.data_parallel_rank
        return self._data_parallel_rank_local

    @data_parallel_rank_local.setter
    def data_parallel_rank_local(self, value: int) -> None:
        """Set the local rank of the data parallel group."""
        self._data_parallel_rank_local = value

1632
    data_parallel_master_ip: str = "127.0.0.1"
1633
1634
1635
1636
1637
    """IP of the data parallel master."""
    data_parallel_master_port: int = 29500
    """Port of the data parallel master."""
    enable_expert_parallel: bool = False
    """Use expert parallelism instead of tensor parallelism for MoE layers."""
1638

1639
    max_parallel_loading_workers: Optional[int] = None
1640
    """Maximum number of parallel loading workers when loading model
1641
1642
    sequentially in multiple batches. To avoid RAM OOM when using tensor
    parallel and large models."""
1643
1644

    disable_custom_all_reduce: bool = False
1645
    """Disable the custom all-reduce kernel and fall back to NCCL."""
1646
1647

    tokenizer_pool_config: Optional[TokenizerPoolConfig] = None
1648
1649
    """This parameter is deprecated and will be removed in a future release.
    Please remove it from your configs"""
1650
1651

    ray_workers_use_nsight: bool = False
1652
    """Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler."""
1653
1654

    placement_group: Optional["PlacementGroup"] = None
1655
    """ray distributed model workers placement group."""
1656

1657
    distributed_executor_backend: Optional[Union[DistributedExecutorBackend,
1658
                                                 type["ExecutorBase"]]] = None
1659
1660
1661
1662
1663
1664
1665
    """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
    and hpu only support Ray for distributed inference."""
1666
1667

    worker_cls: str = "auto"
1668
1669
    """The full name of the worker class to use. If "auto", the worker class
    will be determined based on the platform."""
1670
    sd_worker_cls: str = "auto"
1671
    """The full name of the worker class to use for speculative decofing.
1672
    If "auto", the worker class will be determined based on the platform."""
1673
    worker_extension_cls: str = ""
1674
1675
1676
1677
    """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."""
1678
1679

    world_size: int = field(init=False)
1680
    """world_size is TPxPP, it affects the number of workers we create."""
1681
    world_size_across_dp: int = field(init=False)
1682
1683
    """world_size_across_dp is TPxPPxDP, it is the size of the world
    including data parallelism."""
1684
1685

    rank: int = 0
1686
    """Global rank in distributed setup."""
1687

1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
    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":
        from vllm.distributed.utils import (
            stateless_init_torch_distributed_process_group)

        # use gloo since the engine process might not have cuda device
        dp_group = 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")

        return dp_group

    @staticmethod
    def has_unfinished_dp(dp_group: "ProcessGroup",
youkaichao's avatar
youkaichao committed
1717
                          has_unfinished: bool) -> bool:
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
        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

1729
1730
1731
1732
1733
1734
1735
1736
    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.
        """
1737
        factors: list[Any] = []
1738
1739
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
1740
        factors.append(self.enable_expert_parallel)
1741
1742
        return hashlib.sha256(str(factors).encode()).hexdigest()

1743
1744
1745
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
1746

1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
        if self.data_parallel_size > 1:
            # Data parallel was specified in the engine args.
            self.data_parallel_master_port = get_open_port()
            # TODO multi-node
        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

1759
        self.world_size_across_dp = self.world_size * self.data_parallel_size
1760

1761
1762
1763
1764
1765
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

1766
        ray_only_devices: list[str] = []
1767
        from vllm.platforms import current_platform
1768
1769
        if (current_platform.device_type in ray_only_devices
                and self.world_size > 1):
1770
1771
1772
1773
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
1774
1775
                    f"{current_platform.device_type.upper()} backend only "
                    "supports Ray for distributed inference.")
1776

1777
        if self.distributed_executor_backend is None and self.world_size > 1:
1778
1779
1780
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1781
            from vllm.executor import ray_utils
1782
            backend: DistributedExecutorBackend = "mp"
1783
            ray_found = ray_utils.ray_is_available()
1784
1785
1786
1787
1788
            if current_platform.is_neuron():
                # neuron uses single process to control multiple devices
                backend = "uni"
            elif (current_platform.is_cuda()
                  and cuda_device_count_stateless() < self.world_size):
1789
1790
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1791
1792
1793
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1794
1795
                backend = "ray"
            elif ray_found:
1796
                if self.placement_group:
1797
                    backend = "ray"
1798
1799
1800
1801
1802
1803
                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"
1804
1805
1806
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1807

1808
1809
1810
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

1811
1812
        self._verify_args()

1813
1814
1815
1816
1817
1818
    @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)

1819
    def _verify_args(self) -> None:
1820
1821
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
1822
        from vllm.platforms import current_platform
1823
        if self.distributed_executor_backend not in (
1824
1825
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
1826
1827
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
1828
            raise ValueError(
1829
1830
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
1831
1832
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
1833
        if self.use_ray:
1834
1835
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1836
1837

        if not current_platform.use_custom_allreduce():
1838
1839
1840
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
1841
                "supported on current platform.")
1842
        if self.ray_workers_use_nsight and not self.use_ray:
1843
1844
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1845

1846
1847
1848
        assert isinstance(self.worker_extension_cls, str), (
            "worker_extension_cls must be a string (qualified class name).")

1849

1850
PreemptionMode = Literal["swap", "recompute"]
1851
1852
1853
1854
SchedulerPolicy = Literal["fcfs", "priority"]


@config
1855
@dataclass
1856
class SchedulerConfig:
1857
    """Scheduler configuration."""
1858

1859
1860
    runner_type: RunnerType = "generate"
    """The runner type to launch for the model."""
1861

1862
1863
    max_num_batched_tokens: int = None  # type: ignore
    """Maximum number of tokens to be processed in a single iteration.
1864

1865
1866
    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."""
1867

1868
1869
1870
1871
1872
1873
1874
    cuda_graph_sizes: list[int] = field(default_factory=lambda: [512])
    """Cuda graph capture sizes, default is 512.
    1. if one value is provided, then the capture list would follow the pattern:
        [1, 2, 4] + [i for i in range(8, cuda_graph_sizes + 1, 8)]
    2. more than one value (e.g. 1 2 128) is provided,
        then the capture list will follow the provided list."""

1875
1876
    max_num_seqs: int = None  # type: ignore
    """Maximum number of sequences to be processed in a single iteration.
1877

1878
1879
    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."""
1880

1881
1882
1883
1884
    max_model_len: int = None  # type: ignore
    """Maximum length of a sequence (including prompt and generated text). This
    is primarily set in `ModelConfig` and that value should be manually
    duplicated here."""
1885

1886
    max_num_partial_prefills: int = 1
1887
1888
    """For chunked prefill, the maximum number of sequences that can be
    partially prefilled concurrently."""
1889
1890

    max_long_partial_prefills: int = 1
1891
1892
1893
1894
    """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."""
1895
1896

    long_prefill_token_threshold: int = 0
1897
1898
    """For chunked prefill, a request is considered long if the prompt is
    longer than this number of tokens."""
1899

1900
    num_lookahead_slots: int = 0
1901
1902
1903
1904
1905
1906
1907
    """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."""
1908
1909

    delay_factor: float = 0.0
1910
1911
    """Apply a delay (of delay factor multiplied by previous
    prompt latency) before scheduling next prompt."""
1912

1913
1914
1915
    enable_chunked_prefill: bool = None  # type: ignore
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""
1916
1917

    is_multimodal_model: bool = False
1918
1919
1920
1921
1922
    """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.
1923

1924
1925
1926
1927
1928
1929
1930
1931
1932
    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."""
1933

1934
    preemption_mode: Optional[PreemptionMode] = None
1935
1936
1937
1938
1939
1940
    """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."""
1941
1942

    num_scheduler_steps: int = 1
1943
    """Maximum number of forward steps per scheduler call."""
1944

1945
1946
    multi_step_stream_outputs: bool = True
    """If False, then multi-step will stream outputs at the end of all steps"""
1947
1948

    send_delta_data: bool = False
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
    """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)."""
1960
1961

    chunked_prefill_enabled: bool = field(init=False)
1962
    """True if chunked prefill is enabled."""
1963

1964
    disable_chunked_mm_input: bool = False
1965
1966
1967
1968
1969
1970
    """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."""
1971

1972
1973
    # scheduler class or path. "vllm.core.scheduler.Scheduler" (default)
    # or "mod.custom_class".
1974
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
1975
1976
1977
    """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"."""
1978

1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
    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.
1993
        factors: list[Any] = []
1994
1995
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1996
1997
        return hash_str

1998
    def __post_init__(self) -> None:
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
        if self.max_model_len is None:
            self.max_model_len = 8192
            logger.warning(
                "max_model_len was is not set. Defaulting to arbitrary value "
                "of %d.", self.max_model_len)

        if self.max_num_seqs is None:
            self.max_num_seqs = 128
            logger.warning(
                "max_num_seqs was is not set. Defaulting to arbitrary value "
                "of %d.", self.max_num_seqs)

2011
2012
2013
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
2014
2015
2016
2017
                    # 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.
2018
2019
                    self.max_num_batched_tokens = max(
                        self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
2020
                else:
2021
2022
                    self.max_num_batched_tokens = (
                        _DEFAULT_MAX_NUM_BATCHED_TOKENS)
2023
            else:
2024
2025
                # If max_model_len is too short, use
                # _DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
2026
                # for higher throughput.
2027
2028
                self.max_num_batched_tokens = max(
                    self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
2029

2030
2031
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
2032
2033
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2034
                    _POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
2035
                )
2036
            if self.is_multimodal_model:
2037
                # The value needs to be at least the number of multimodal tokens
2038
2039
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2040
2041
2042
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

2043
2044
2045
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

2046
        if self.enable_chunked_prefill:
2047
2048
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
2049
                self.max_num_batched_tokens)
2050

2051
        self.chunked_prefill_enabled = self.enable_chunked_prefill
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
        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)

2064
2065
2066
        self._verify_args()

    def _verify_args(self) -> None:
2067
2068
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
2069
2070
2071
2072
2073
2074
2075
            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.")
2076

2077
2078
2079
2080
2081
        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}).")
2082

2083
2084
2085
2086
2087
2088
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

2089
2090
2091
2092
2093
2094
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
        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}).")

2118
2119
2120
2121
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

2122

2123
2124
2125
2126
2127
Device = Literal["auto", "cuda", "neuron", "cpu", "tpu", "xpu", "hpu"]


@config
@dataclass
2128
class DeviceConfig:
2129
2130
2131
2132
2133
2134
2135
    """Configuration for the device to use for vLLM execution."""

    device: Union[Device, torch.device] = "auto"
    """Device type for vLLM execution."""
    device_type: str = field(init=False)
    """Device type from the current platform. This is set in
    `__post_init__`."""
2136

2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
    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.
2152
        factors: list[Any] = []
2153
2154
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2155
2156
        return hash_str

2157
2158
    def __post_init__(self):
        if self.device == "auto":
2159
            # Automated device type detection
2160
            from vllm.platforms import current_platform
2161
            self.device_type = current_platform.device_type
2162
            if not self.device_type:
2163
2164
2165
2166
                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.")
2167
2168
        else:
            # Device type is assigned explicitly
2169
            self.device_type = self.device
2170
2171

        # Some device types require processing inputs on CPU
2172
        if self.device_type in ["neuron"]:
2173
            self.device = torch.device("cpu")
2174
2175
        elif self.device_type in ["tpu"]:
            self.device = None
2176
2177
2178
2179
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

2180

2181
2182
2183
2184
2185
2186
2187
SpeculativeMethod = Literal["ngram", "eagle", "medusa", "mlp_speculator",
                            "draft_model"]
SpeculativeAcceptanceMethod = Literal["rejection_sampler",
                                      "typical_acceptance_sampler"]


@config
2188
@dataclass
2189
class SpeculativeConfig:
2190
    """Configuration for speculative decoding."""
2191

2192
    # General speculative decoding control
2193
2194
    num_speculative_tokens: int = field(default=None,
                                        init=True)  # type: ignore
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
    """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."""
2215
    draft_tensor_parallel_size: Optional[int] = None
2216
2217
    """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."""
2218
    disable_logprobs: bool = True
2219
2220
2221
    """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."""
2222

2223
    # Draft model configuration
2224
    quantization: Optional[QuantizationMethods] = None
2225
2226
2227
    """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."""
2228
    max_model_len: Optional[int] = None
2229
2230
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
2231
    revision: Optional[str] = None
2232
2233
2234
    """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."""
2235
    code_revision: Optional[str] = None
2236
2237
2238
    """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."""
2239

2240
    # Advanced control
2241
    disable_mqa_scorer: bool = False
2242
2243
    """Disable the MQA scorer and fall back to batch expansion for scoring
    proposals."""
2244
    disable_by_batch_size: Optional[int] = None
2245
2246
2247
2248
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
2249
    prompt_lookup_max: Optional[int] = None
2250
2251
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
2252
    prompt_lookup_min: Optional[int] = None
2253
2254
2255
2256
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

    # Typical acceptance sampler configuration
2257
    posterior_threshold: Optional[float] = None
2258
2259
2260
2261
    """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.
    """
2262
    posterior_alpha: Optional[float] = None
2263
2264
    """Scaling factor for entropy-based threshold, applied when using
    `TypicalAcceptanceSampler`."""
2265
2266
2267
2268

    # required configuration params passed from engine
    target_model_config: ModelConfig = field(default=None,
                                             init=True)  # type: ignore
2269
    """The configuration of the target model."""
2270
2271
    target_parallel_config: ParallelConfig = field(default=None,
                                                   init=True)  # type: ignore
2272
    """The parallel configuration for the target model."""
2273
2274
    enable_chunked_prefill: bool = field(default=None,
                                         init=True)  # type: ignore
2275
2276
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
2277
    disable_log_stats: bool = field(default=None, init=True)  # type: ignore
2278
2279
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
2280
2281
2282
2283

    # params generated in the post-init stage
    draft_model_config: ModelConfig = field(default=None,
                                            init=True)  # type: ignore
2284
    """The configuration of the draft model initialized internal."""
2285
2286
    draft_parallel_config: ParallelConfig = field(default=None,
                                                  init=True)  # type: ignore
2287
    """The parallel configuration for the draft model initialized internal."""
2288

2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
    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.
        """
2301
        factors: list[Any] = []
2302
2303
2304
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2305
2306
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2307
2308
        return hash_str

2309
2310
2311
2312
2313
    @classmethod
    def from_dict(cls, dict_value: dict) -> "SpeculativeConfig":
        """Parse the CLI value for the speculative config."""
        return cls(**dict_value)

2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
    @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"]
            })
        return hf_config

2326
    def __post_init__(self):
2327

2328
2329
2330
2331
2332
2333
2334
        # 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.
2335
2336
2337
2338

        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
2339
2340
            if self.target_model_config and \
                self.target_model_config.hf_text_config.model_type \
2341
                        == "deepseek_v3":
2342
2343
2344
2345
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2346
            else:
2347
2348
2349
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2350
2351
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2352
2353
2354
2355
2356
2357
2358
        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"
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
            # 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
2373
            if self.prompt_lookup_min < 1:
2374
2375
2376
2377
2378
                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")
2379
            if self.prompt_lookup_min > self.prompt_lookup_max:
2380
2381
2382
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2383

2384
2385
2386
            # 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.
2387
2388
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2389
        else:
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
                    task="draft",
                    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,
                )
2418

2419
                # Automatically detect the method
2420
                if self.method in ('eagle', 'eagle3'):
2421
                    pass
2422
2423
                elif "eagle-" in self.draft_model_config.model.lower() or \
                        "eagle3-" in self.draft_model_config.model.lower():
2424
2425
2426
2427
2428
2429
                    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"
2430
                else:
2431
2432
2433
                    self.method = "draft_model"

                # Replace hf_config for EAGLE draft_model
2434
                if self.method in ("eagle", "eagle3"):
2435
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2436
                        raise ValueError(
2437
2438
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2439
2440
2441
2442
2443
2444
2445
2446

                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
                    if isinstance(self.draft_model_config.hf_config,
                                  EAGLEConfig):
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2447
2448
                            self.draft_model_config.hf_config,
                            method=self.method)
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
                        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
                )
2476

2477
2478
2479
2480
2481
2482
                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,
                    ))
2483

2484
2485
2486
2487
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2488

2489
2490
2491
2492
2493
        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
2494

2495
        self._verify_args()
2496

2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
    @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,
        )

2532
    @staticmethod
2533
    def _verify_and_get_draft_tp(
2534
2535
2536
2537
2538
2539
            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.
2540
        """
2541
2542
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2543
        if speculative_draft_tensor_parallel_size is None:
2544
2545
2546
2547
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2548
2549
2550
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2551
2552
2553
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2554
2555
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2556
            raise ValueError(
2557
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2558
                f"other value than 1 or target model tensor_parallel_size")
2559
        return speculative_draft_tensor_parallel_size
2560

2561
2562
2563
2564
2565
2566
2567
2568
2569
    @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.
        """
2570
2571
2572
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2573
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2574
2575
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
            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

    def _verify_args(self) -> None:
2588
2589
2590
2591
2592
2593
        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.")

2594
2595
2596
2597
2598
2599
2600
        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)
2601
2602
            # Validate and set draft token acceptance related settings.

2603
2604
        if self.acceptance_method is None:
            raise ValueError("acceptance_method is not set. "
2605
2606
2607
                             "Expected values are rejection_sampler or "
                             "typical_acceptance_sampler.")

2608
2609
        if (self.acceptance_method != 'rejection_sampler'
                and self.acceptance_method != 'typical_acceptance_sampler'):
2610
            raise ValueError(
2611
                "Expected acceptance_method to be either "
2612
                "rejection_sampler or typical_acceptance_sampler. Instead it "
2613
                f"is {self.acceptance_method}")
2614

2615
2616
2617
2618
        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)):
2619
            raise ValueError(
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
                "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=}")
2631

2632
2633
2634
2635
2636
2637
        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=}")

2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
    @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

2648
2649
2650
    def use_eagle(self) -> bool:
        return self.method in ("eagle", "eagle3")

2651
    def __repr__(self) -> str:
2652
2653
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
2654
        num_spec_tokens = self.num_speculative_tokens
2655
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
2656
2657


2658
2659
2660
2661
LoRADType = Literal["auto", "float16", "bfloat16"]


@config
2662
2663
@dataclass
class LoRAConfig:
2664
2665
2666
2667
2668
2669
    """Configuration for LoRA."""

    max_lora_rank: int = 16
    """Max LoRA rank."""
    max_loras: int = 1
    """Max number of LoRAs in a single batch."""
2670
    fully_sharded_loras: bool = False
2671
2672
2673
2674
    """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.
    """
2675
    max_cpu_loras: Optional[int] = None
2676
2677
2678
2679
    """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."""
2680
    lora_extra_vocab_size: int = 256
2681
2682
    """Maximum size of extra vocabulary that can be present in a LoRA adapter
    (added to the base model vocabulary)."""
2683
2684
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
2685
2686
2687
2688
2689
2690
    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."""
2691
    bias_enabled: bool = False
2692
    """Enable bias for LoRA adapters."""
2693

2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
    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.
        """
2706
        factors: list[Any] = []
2707
2708
2709
2710
2711
2712
2713
        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)
        factors.append(self.long_lora_scaling_factors)
        factors.append(self.bias_enabled)
2714
2715
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2716
2717
        return hash_str

2718
    def __post_init__(self):
2719
        # Setting the maximum rank to 512 should be able to satisfy the vast
2720
        # majority of applications.
2721
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
2722
        possible_lora_extra_vocab_size = (256, 512)
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
        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
2738
                f"max_loras ({self.max_loras})")
2739

2740
    def verify_with_cache_config(self, cache_config: CacheConfig):
2741
2742
2743
        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.")
2744

2745
2746
2747
2748
2749
2750
    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)

2751
2752
2753
2754
2755
    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.")

2756

2757
@config
2758
2759
@dataclass
class PromptAdapterConfig:
2760
2761
    """Configuration for PromptAdapters."""

2762
2763
2764
2765
    max_prompt_adapters: int = 1
    """Max number of PromptAdapters in a batch."""
    max_prompt_adapter_token: int = 0
    """Max number of PromptAdapters tokens."""
2766
    max_cpu_prompt_adapters: Optional[int] = None
2767
2768
2769
2770
2771
    """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.
    """
2772

2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
    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.
2787
        factors: list[Any] = []
2788
2789
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2790
2791
        return hash_str

2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
    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):
2803
        if self.prompt_adapter_dtype == "auto":
2804
2805
2806
2807
2808
2809
            self.prompt_adapter_dtype = model_config.dtype
        elif isinstance(self.prompt_adapter_dtype, str):
            self.prompt_adapter_dtype = getattr(torch,
                                                self.prompt_adapter_dtype)


2810
@config
2811
@dataclass
2812
class MultiModalConfig:
2813
2814
    """Controls the behavior of multimodal models."""

2815
2816
    limit_per_prompt: dict[str, int] = get_field(ModelConfig,
                                                 "limit_mm_per_prompt")
2817
    """
2818
    The maximum number of input items allowed per prompt for each modality.
2819
2820
    This should be a JSON string that will be parsed into a dictionary.
    Defaults to 1 (V0) or 999 (V1) for each modality.
2821
2822

    For example, to allow up to 16 images and 2 videos per prompt:
2823
    `{"images": 16, "videos": 2}`
2824
2825
2826
2827
2828
    """

    mm_processor_kwargs: Optional[dict[str, object]] = None
    """
    Overrides for the multi-modal processor obtained from
2829
    `transformers.AutoProcessor.from_pretrained`.
2830
2831
2832
2833

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

    For example, for Phi-3-Vision:
2834
    `{"num_crops": 4}`.
2835
2836
2837
2838
    """

    disable_mm_preprocessor_cache: bool = False
    """
2839
    If `True`, disable caching of the processed multi-modal inputs.
2840
2841
    """

2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
    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.
2856
        factors: list[Any] = []
2857
2858
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2859
2860
        return hash_str

2861
2862
2863
2864
2865
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
2866
2867
2868
2869
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
2870

2871
    # TODO: Add configs to init vision tower or not.
2872

2873

2874
@config
2875
2876
@dataclass
class PoolerConfig:
2877
    """Controls the behavior of output pooling in pooling models."""
2878
2879

    pooling_type: Optional[str] = None
2880
    """
2881
    The pooling method of the pooling model. This should be a key in
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
    :class:`vllm.model_executor.layers.pooler.PoolingType`.
    """

    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
    """
2899
    If set, only the score corresponding to the ``step_tag_id`` in the
2900
2901
2902
2903
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

2904
    returned_token_ids: Optional[list[int]] = None
2905
    """
2906
2907
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2908
2909
2910
    ``math-shepherd-mistral-7b-prm`` model.
    """

2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
    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.
2925
        factors: list[Any] = []
2926
2927
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2928
2929
        return hash_str

2930

2931
2932
2933
2934
2935
2936
2937
2938
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2939
_ROCM_NOT_SUPPORTED_DTYPE: list[str] = []  #
2940

2941
2942
2943

def _get_and_verify_dtype(
    config: PretrainedConfig,
2944
    dtype: Union[str, torch.dtype],
2945
2946
2947
) -> torch.dtype:
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
2948
    config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
2949

2950
    # Fallback for multi-modal models if the root config
2951
2952
2953
2954
    # does not define torch_dtype
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)

2955
2956
2957
    if config_dtype is None:
        config_dtype = torch.float32

2958
2959
2960
2961
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
2962
2963
                # Following common practice, we use float16 for float32 models
                torch_dtype = torch.float16
2964
2965
            else:
                torch_dtype = config_dtype
2966

Shinichi Hemmi's avatar
Shinichi Hemmi committed
2967
2968
2969
2970
2971
2972
2973
            if config.model_type == "plamo2":
                logger.info(
                    "For PLaMo2, we cast models to bfloat16 instead of using "
                    "float16 by default. This is because float16 does not work."
                )
                torch_dtype = torch.bfloat16

2974
            from vllm.platforms import current_platform
2975
2976
            if (current_platform.is_cpu()
                    and current_platform.get_cpu_architecture()
2977
                    == CpuArchEnum.POWERPC
2978
2979
2980
2981
2982
2983
2984
2985
                    and (config_dtype == torch.float16
                         or config_dtype == torch.float32)):
                logger.info(
                    "For POWERPC, we cast models to bfloat16 instead of "
                    "using float16 by default. Float16 is not currently "
                    "supported for POWERPC.")
                torch_dtype = torch.bfloat16

2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
            # TODO: change this condition to check if the platform support bf16
            # instead of checking the OS. For instance M2 shall supports bf16
            # already. But we need to modify `cpu_extension.cmake` to activate
            # the feature in the build.
            if (current_platform.is_cpu() and sys.platform.startswith("darwin")
                    and current_platform.get_cpu_architecture()
                    == CpuArchEnum.ARM and config_dtype == torch.bfloat16):
                logger.info("For macOS with Apple Silicon, currently bfloat16 "
                            "is not supported. Setting dtype to float16.")
                torch_dtype = torch.float16

2997
2998
            if current_platform.is_hpu() and config_dtype == torch.float16:
                logger.info(
2999
                    "For HPU, we cast models to bfloat16 instead of "
3000
3001
3002
                    "using float16 by default. Please specify `dtype` if you "
                    "want to use float16.")
                torch_dtype = torch.bfloat16
Shinichi Hemmi's avatar
Shinichi Hemmi committed
3003
3004
3005
3006
3007
        elif dtype == "float16" and config.model_type == "plamo2":
            logger.warning(
                "For PLaMo2, using float16 is unstable and might cause "
                "unexpected behavior. Please use bfloat16 or float32 instead.")
            torch_dtype = torch.float16
3008
        else:
3009
3010
3011
3012
3013
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
                raise ValueError(f"Unknown dtype: {dtype}")
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
3014
    else:
3015
        raise ValueError(f"Unknown dtype: {dtype}")
3016
3017
3018
3019
3020

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
3021
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
3022
3023
3024
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
3025
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
3026
3027
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
3028
            # Casting between float16 and bfloat16 is allowed with a warning.
3029
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
3030
3031

    return torch_dtype
3032
3033
3034
3035
3036


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
3037
    disable_sliding_window: bool,
3038
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
3039
    spec_target_max_model_len: Optional[int] = None,
3040
    encoder_config: Optional[Any] = None,
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
) -> 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",
3051
3052
        # ChatGLM2
        "seq_length",
3053
3054
        # Command-R
        "model_max_length",
3055
3056
        # Whisper
        "max_target_positions",
3057
3058
3059
3060
3061
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
3062
    # Choose the smallest "max_length" from the possible keys.
3063
    max_len_key = None
3064
    for key in possible_keys:
3065
3066
3067
3068
3069
        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
3070
3071
3072
3073
    # 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
3074
3075
3076
3077

    # 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:
3078
3079

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
3080
        max_len_key = "sliding_window" \
3081
3082
3083
            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)
3084
3085
3086

    # If none of the keys were found in the config, use a default and
    # log a warning.
3087
    if derived_max_model_len == float("inf"):
3088
3089
3090
3091
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

3092
3093
3094
3095
3096
        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

3097
3098
3099
3100
        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: "
3101
            "%s. Assuming the model's maximum length is %d.", possible_keys,
3102
            default_max_len)
3103
        derived_max_model_len = default_max_len
3104

3105
    rope_scaling = getattr(hf_config, "rope_scaling", None)
3106
3107
3108
    # 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:
3109
3110
3111
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121

        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.")

3122
3123
3124
3125
            # 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)

3126
3127
3128
3129
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
3130

3131
3132
3133
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

3134
3135
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
3136
    if max_model_len is None:
3137
        max_model_len = int(derived_max_model_len)
3138
    elif max_model_len > derived_max_model_len:
3139
3140
3141
3142
3143
        # 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:
3144
3145
3146
3147
3148
3149
3150
            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.")
3151
        else:
3152
            msg = (
3153
                f"User-specified max_model_len ({max_model_len}) is greater "
3154
3155
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
3156
                f"{model_max_length} in model's config.json). This may lead "
3157
3158
3159
3160
3161
3162
3163
3164
3165
                "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")
3166
    return int(max_model_len)
3167
3168


3169
def get_min_sliding_window(
3170
        sliding_window: Union[int, list[Optional[int]]]) -> int:
3171
3172
3173
3174
3175
3176
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


3177
def get_served_model_name(model: str,
3178
                          served_model_name: Optional[Union[str, list[str]]]):
3179
    """
3180
3181
3182
3183
    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
3184
3185
3186
3187
3188
3189
3190
3191
3192
    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


3193
GuidedDecodingBackendV0 = Literal["auto", "outlines", "lm-format-enforcer",
3194
                                  "xgrammar", "guidance"]
3195
GuidedDecodingBackendV1 = Literal["auto", "xgrammar", "guidance"]
3196
3197
GuidedDecodingBackend = Literal[GuidedDecodingBackendV0,
                                GuidedDecodingBackendV1]
3198
3199
3200


@config
3201
3202
@dataclass
class DecodingConfig:
3203
    """Dataclass which contains the decoding strategy of the engine."""
3204

3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
    @property
    @deprecated(
        "`guided_decoding_backend` is deprecated and has been renamed to "
        "`backend`. This will be removed in v0.10.0. Please use the "
        "`backend` argument instead.")
    def guided_decoding_backend(self) -> GuidedDecodingBackend:
        return self.backend

    @guided_decoding_backend.setter
    def guided_decoding_backend(self, value: GuidedDecodingBackend):
        self.backend = value

    backend: GuidedDecodingBackend = "auto" if envs.VLLM_USE_V1 else "xgrammar"
3218
3219
3220
3221
    """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."""
3222

3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
    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`."""

3235
    reasoning_backend: str = ""
3236
    """Select the reasoning parser depending on the model that you're using.
3237
    This is used to parse the reasoning content into OpenAI API format."""
3238

3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
    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.
3253
        factors: list[Any] = []
3254
3255
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3256
3257
        return hash_str

3258
    def __post_init__(self):
3259
3260
3261
        if ":" in self.backend:
            self._extract_backend_options()

3262
        if envs.VLLM_USE_V1:
3263
            valid_guided_backends = get_args(GuidedDecodingBackendV1)
3264
        else:
3265
            valid_guided_backends = get_args(GuidedDecodingBackendV0)
3266
3267
        if self.backend not in valid_guided_backends:
            raise ValueError(f"Invalid backend '{self.backend}',"
3268
                             f" must be one of {valid_guided_backends}")
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
        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.")

    @deprecated(
        "Passing guided decoding backend options inside backend in the format "
        "'backend:...' is deprecated. This will be removed in v0.10.0. Please "
        "use the dedicated arguments '--disable-fallback', "
        "'--disable-any-whitespace' and '--disable-additional-properties' "
        "instead.")
    def _extract_backend_options(self):
        """Extract backend options from the backend string."""
        backend, options = self.backend.split(":")
        self.backend = cast(GuidedDecodingBackend, backend)
        options_set = set(options.strip().split(","))
        if "no-fallback" in options_set:
            self.disable_fallback = True
        if "disable-any-whitespace" in options_set:
            self.disable_any_whitespace = True
        if "no-additional-properties" in options_set:
            self.disable_additional_properties = True
3294
3295


3296
3297
3298
3299
DetailedTraceModules = Literal["model", "worker", "all"]


@config
3300
3301
@dataclass
class ObservabilityConfig:
3302
    """Configuration for observability - metrics and tracing."""
3303

3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
    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)
3319

3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
    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))
3345

3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
    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.
3360
        factors: list[Any] = []
3361
3362
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3363
3364
        return hash_str

3365
    def __post_init__(self):
3366
3367
3368
3369
3370
        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()

3371
3372
3373
3374
3375
        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}")
3376

3377
3378
3379
3380
3381
3382
    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(","))

3383

3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
class KVTransferConfig(BaseModel):
    """Configuration for distributed KV cache transfer."""

    # The KV connector for vLLM to transmit KV caches between vLLM instances.
    kv_connector: Optional[str] = None

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

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

    # Whether this vLLM instance produces, consumes KV cache, or both. Choices
    # are 'kv_producer', 'kv_consumer', and 'both'.
    kv_role: Optional[str] = None

    # 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.
    kv_rank: Optional[int] = None

    # The number of parallel instances for KV cache transfer. For
    # PyNcclConnector, this should be 2.
    kv_parallel_size: int = 1

    # The KV connector ip, used to build distributed connection
    kv_ip: str = "127.0.0.1"

    # The KV connector port, used to build distributed connection
    kv_port: int = 14579

3417
3418
3419
    # any extra config that the connector may need
    kv_connector_extra_config: dict[str, Any] = {}

3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
    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.
3434
        factors: list[Any] = []
3435
3436
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3437
3438
        return hash_str

3439
3440
    @classmethod
    def from_cli(cls, cli_value: str) -> "KVTransferConfig":
youkaichao's avatar
youkaichao committed
3441
        """Parse the CLI value for the kv cache transfer config."""
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
        return KVTransferConfig.model_validate_json(cli_value)

    def model_post_init(self, __context: Any) -> None:

        if self.kv_role is not None and self.kv_role not in [
                "kv_producer", "kv_consumer", "kv_both"
        ]:
            raise ValueError(
                f"Unsupported kv_role: {self.kv_role}. "
                f"Supported roles are `kv_producer`, `kv_consumer`, "
                f"and `kv_both`")

        if self.kv_connector is not None and self.kv_role is None:
            raise ValueError("Please specify kv_disagg_role when kv_connector "
                             "is set, supported roles are `kv_producer`, "
                             "`kv_consumer`, and `kv_both`")

    @property
    def is_kv_transfer_instance(self) -> bool:
        return self.kv_connector is not None and \
            self.kv_role in ["kv_producer", "kv_consumer", "kv_both"]

    @property
    def is_kv_producer(self) -> bool:
        return self.kv_connector is not None and \
            self.kv_role in ["kv_producer", "kv_both"]

    @property
    def is_kv_consumer(self) -> bool:
        return self.kv_connector is not None and \
            self.kv_role in ["kv_consumer", "kv_both"]

3474
3475
3476
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

3477

3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
class KVEventsConfig(BaseModel):
    """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.
    """

    @classmethod
    def from_cli(cls, cli_value: str) -> "KVEventsConfig":
        """Parse the CLI value for the event publisher config."""
        return KVEventsConfig.model_validate_json(cli_value)


3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
class CompilationLevel:
    # constants for the levels of the compilation process
    NO_COMPILATION = 0
    DYNAMO_AS_IS = 1
    DYNAMO_ONCE = 2
    PIECEWISE = 3


class CompilationConfig(BaseModel):
    """
    Configuration for compilation.
    It has three parts:
    - Top-level Compilation control:
        - level: the level of compilation.
            - 0: no compilation.
            - 1: dynamo as is.
            - 2: dynamo once.
            - 3: piecewise compilation.
3541
        - debug_dump_path: the path to dump the debug information.
3542
3543
3544
        - cache_dir: the directory to store the compiled graph, to
            accelerate Inductor compilation. By default, it will use
            model-related information to generate a cache directory.
3545
3546
3547
3548
3549
3550
3551
        - backend: 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).
3552
3553
3554
3555
3556
3557
3558
3559
3560
        - custom_ops: 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 disabled when running with Inductor (compile_level >= Inductor).
3561
        - splitting_ops: a list of ops to split the full graph into subgraphs, used in piecewise compilation.
3562
3563
3564
3565
    - CudaGraph capture:
        - use_cudagraph: whether to use cudagraph inside compilation.
            - False: cudagraph inside compilation is not used.
            - True: cudagraph inside compilation is used. It requires
3566
3567
3568
3569
                that all input buffers have fixed addresses, and all
                splitting ops write their outputs to input buffers.
            Note that this is orthogonal to the cudagraph capture logic
            outside of compilation.
3570
3571
3572
            TODO: move outside cudagraph logic into compilation.
            torch.compile will handle cudagraph capture logic in the future.
        - cudagraph_capture_sizes: sizes to capture cudagraph.
3573
            - None (default): capture sizes are inferred from vllm config.
3574
            - list[int]: capture sizes are specified as given.
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
        - cudagraph_num_of_warmups: 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.
        - cudagraph_copy_inputs: 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.
    - Inductor compilation:
        - use_inductor: whether to use inductor compilation.
            - False: inductor compilation is not used. graph runs in eager.
            - True: inductor compilation is used. one graph for symbolic shape
3588
3589
3590
3591
3592
                is compiled. In addition, compile for compile_sizes,
                using configurations in inductor_compile_config.
        - compile_sizes: sizes to compile for inductor. In addition
            to integers, it also supports "cudagraph_capture_sizes" to
            specify the sizes for cudagraph capture.
3593
3594
3595
3596
3597
3598
3599
        - inductor_compile_config: additional configurations for inductor.
            - None: use default configurations.
        - inductor_passes: 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})`
3600
        - custom inductor passes: see PassConfig for more details
3601

3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
    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.
    """ # noqa
    level: int = 0
3613
    debug_dump_path: str = ""
3614
    cache_dir: str = ""
3615
    backend: str = ""
3616
3617
    custom_ops: list[str] = Field(default_factory=list)
    splitting_ops: list[str] = Field(default=None)  # type: ignore
3618
3619

    use_inductor: bool = True
3620
3621
3622
    compile_sizes: Optional[list[Union[int, str]]] = Field(default=None)
    inductor_compile_config: dict = Field(default_factory=dict)
    inductor_passes: dict[str, str] = Field(default_factory=dict)
3623
3624
3625

    use_cudagraph: bool = False
    cudagraph_num_of_warmups: int = 0
3626
    cudagraph_capture_sizes: Optional[list[int]] = None
3627
3628
    cudagraph_copy_inputs: bool = False

3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
    class PassConfig(BaseModel):
        """
        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.
        - dump_graph_stages: list of stages for which we want to dump the graph.
            Each pass defines its own stages (before, after, maybe in-between).
        - dump_graph_dir: directory to dump the graphs. Default is .
        - enable_fusion: whether to enable the custom fusion pass.
3639
3640
        - enable_noop: whether to enable the custom no-op elimination pass.
            TODO(luka) better pass enabling system.
3641
        - enable_sequence_parallelism: whether to enable sequence parallelism.
3642
        """
3643
        dump_graph_stages: list[str] = Field(default_factory=list)
3644
3645
        dump_graph_dir: Path = Field(default=Path("."))
        enable_fusion: bool = True
3646
        enable_noop: bool = True
3647
        enable_sequence_parallelism: bool = False
3648
3649
3650
3651
3652
3653
3654
3655

        def uuid(self):
            """
            Produces a hash unique to the pass configuration.
            Any new fields that affect compilation should be added to the hash.
            Do not include dump_graph_* in the hash - they don't affect
            compilation.
            """
3656
3657
            dict_ = self.model_dump(include={"enable_fusion", "enable_noop", \
                "enable_sequence_parallelism"})
3658
            return InductorPass.hash_dict(dict_)
3659
3660

        def model_post_init(self, __context: Any) -> None:
3661
            if not self.enable_noop and self.enable_fusion:
3662
                logger.warning_once(
3663
                    "Fusion enabled but reshape elimination disabled. "
3664
3665
3666
                    "RMSNorm + quant (fp8) fusion might not work")

    pass_config: PassConfig = Field(default_factory=PassConfig)
3667
3668

    # not configurable, computed after init
3669
    max_capture_size: int = PrivateAttr
3670
    local_cache_dir: str = PrivateAttr  # local cache dir for each rank
3671
    # optimization:
3672
    # Intuitively, bs_to_padded_graph_size should be dict[int, int].
3673
    # since we know all keys are in a range [0, max_capture_size],
3674
3675
    # we can optimize it to list[int] for better lookup performance.
    bs_to_padded_graph_size: list[int] = PrivateAttr
3676

3677
3678
3679
    # keep track of enabled and disabled custom ops
    enabled_custom_ops: Counter[str] = PrivateAttr
    disabled_custom_ops: Counter[str] = PrivateAttr
3680
    traced_files: set[str] = PrivateAttr
3681
    compilation_time: float = PrivateAttr
3682

3683
    # Per-model forward context
3684
3685
    # Map from layer name to layer objects that need to be accessed outside
    # model code, e.g., Attention, FusedMOE when dp_size>1.
3686
    static_forward_context: dict[str, Any] = PrivateAttr
3687

3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
    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.
        """
3700
        factors: list[Any] = []
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
        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()

3711
3712
3713
3714
3715
3716
3717
3718
    def __repr__(self) -> str:
        exclude = {
            "static_forward_context",
            "enabled_custom_ops",
            "disabled_custom_ops",
            "compilation_time",
            "bs_to_padded_graph_size",
            "pass_config",
3719
            "traced_files",
3720
3721
3722
3723
3724
        }
        return self.model_dump_json(exclude=exclude, exclude_unset=True)

    __str__ = __repr__

3725
3726
3727
3728
3729
    @classmethod
    def from_cli(cls, cli_value: str) -> "CompilationConfig":
        """Parse the CLI value for the compilation config."""
        if cli_value in ["0", "1", "2", "3"]:
            return cls(level=int(cli_value))
3730
3731
3732
        # do not use `eval`, it is dangerous and can execute arbitrary code
        dict_value = ast.literal_eval(cli_value)
        return CompilationConfig.model_validate(dict_value)
3733

3734
3735
3736
3737
3738
3739
    def model_post_init(self, __context: Any) -> None:

        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
3740
3741
3742
3743
3744
3745
3746
3747
        # 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

3748
        if is_torch_equal_or_newer("2.6"):
Michael Goin's avatar
Michael Goin committed
3749
3750
3751
3752
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

3753
        if self.splitting_ops is None:
3754
            self.splitting_ops = []
3755

3756
3757
3758
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
3759
3760
3761
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
3762
3763
3764
3765
3766
3767
3768
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
3769
3770
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
3771

3772
3773
        self.enabled_custom_ops = Counter()
        self.disabled_custom_ops = Counter()
3774
        self.traced_files = set()
3775
        self.static_forward_context = {}
3776
        self.compilation_time = 0.0
3777

3778
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
        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
3796

3797
        from vllm.compilation.backends import VllmBackend
3798
        return VllmBackend(vllm_config)
3799

3800
    def init_with_cudagraph_sizes(self,
3801
                                  cudagraph_capture_sizes: list[int]) -> None:
3802
        """To complete the initialization of config,
3803
3804
        we need to know the cudagraph sizes."""

3805
        if self.cudagraph_capture_sizes is None:
3806
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
3807
        else:
3808
3809
3810
            # de-duplicate the sizes provided by the config
            self.cudagraph_capture_sizes = list(
                set(self.cudagraph_capture_sizes))
3811
3812
            logger.info(("cudagraph sizes specified by model runner"
                         " %s is overridden by config %s"),
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
                        cudagraph_capture_sizes, self.cudagraph_capture_sizes)

        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
3829

3830
        # sort to make sure cudagraph capture sizes are in descending order
3831
3832
3833
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
3834

3835
3836
3837
3838
        # 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)
        ]
3839
3840
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
3841
3842
3843
3844
3845
3846
3847
            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
3848

3849
3850
3851
3852
3853
3854
3855
3856
3857
    def set_splitting_ops_for_v1(self):
        # If default, override splitting ops for piecewise cudagraph on V1.
        # NOTE: this function needs to be called
        if not self.splitting_ops:
            self.splitting_ops = [
                "vllm.unified_attention",
                "vllm.unified_attention_with_output",
            ]

3858

3859
3860
3861
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
3862
3863
3864
    simplifies passing around the distinct configurations in the codebase.
    """

3865
3866
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
3867
3868
3869
3870
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig,
                                            init=True)
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig,
                                              init=True)
3871
3872
3873
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
3874
    lora_config: Optional[LoRAConfig] = None
3875
3876
    speculative_config: SpeculativeConfig = field(default=None,
                                                  init=True)  # type: ignore
3877
3878
3879
    decoding_config: Optional[DecodingConfig] = None
    observability_config: Optional[ObservabilityConfig] = None
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
3880
    quant_config: Optional[QuantizationConfig] = None
3881
3882
    compilation_config: CompilationConfig = field(default=None,
                                                  init=True)  # type: ignore
3883
3884
    kv_transfer_config: KVTransferConfig = field(default=None,
                                                 init=True)  # type: ignore
3885
    kv_events_config: Optional[KVEventsConfig] = None
3886
    # some opaque config, only used to provide additional information
3887
3888
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
3889
3890
    additional_config: SupportsHash = field(default=None,
                                            init=True)  # type: ignore
3891
    instance_id: str = ""
3892

3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
    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.
        """
3905
        factors: list[Any] = []
3906
3907

        # summarize vllm config
3908
        vllm_factors: list[Any] = []
3909
3910
        from vllm import __version__
        vllm_factors.append(__version__)
3911
        vllm_factors.append(envs.VLLM_USE_V1)
3912
3913
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3914
3915
        else:
            vllm_factors.append("None")
3916
3917
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3918
3919
        else:
            vllm_factors.append("None")
3920
3921
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3922
3923
        else:
            vllm_factors.append("None")
3924
3925
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3926
3927
        else:
            vllm_factors.append("None")
3928
3929
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3930
3931
        else:
            vllm_factors.append("None")
3932
3933
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3934
3935
        else:
            vllm_factors.append("None")
3936
3937
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3938
3939
3940
3941
3942
            # 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))
3943
3944
        else:
            vllm_factors.append("None")
3945
3946
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3947
3948
        else:
            vllm_factors.append("None")
3949
3950
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3951
3952
        else:
            vllm_factors.append("None")
3953
3954
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3955
3956
        else:
            vllm_factors.append("None")
3957
3958
        if self.prompt_adapter_config:
            vllm_factors.append(self.prompt_adapter_config.compute_hash())
3959
3960
        else:
            vllm_factors.append("None")
3961
3962
3963
3964
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3965
3966
        else:
            vllm_factors.append("None")
3967
3968
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3969
3970
3971
3972
3973
3974
        else:
            vllm_factors.append("None")
        if self.additional_config:
            vllm_factors.append(self.additional_config.compute_hash())
        else:
            vllm_factors.append("None")
3975
3976
        factors.append(vllm_factors)

3977
3978
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
3979
3980
        return hash_str

3981
3982
3983
3984
3985
3986
    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]
3987

3988
3989
3990
3991
3992
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3993
        from vllm.platforms import current_platform
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
        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
4016

4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
    @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)

4028
4029
4030
4031
4032
4033
4034
4035
4036
    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

4037
4038
4039
4040
4041
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

4042
4043
4044
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
4045
4046
4047
4048
4049
4050
4051
4052
        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)

        if self.cache_config is not None:
            self.cache_config.verify_with_parallel_config(self.parallel_config)
4053
4054

        if self.lora_config:
4055
            self.lora_config.verify_with_cache_config(self.cache_config)
4056
            self.lora_config.verify_with_model_config(self.model_config)
4057
            self.lora_config.verify_lora_support()
4058
4059
4060
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
4061
4062
4063
4064
4065

        if self.quant_config is None and \
            self.model_config is not None and self.load_config is not None:
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
4066

4067
        from vllm.platforms import current_platform
4068
4069
4070
4071
4072
        if self.scheduler_config is not None and \
            self.model_config is not None and \
            self.scheduler_config.chunked_prefill_enabled and \
            self.model_config.dtype == torch.float32 and \
            current_platform.get_device_capability() == (7, 5):
4073
            logger.warning_once(
4074
4075
4076
4077
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

4078
        if self.compilation_config is None:
4079
            self.compilation_config = CompilationConfig()
4080
4081
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
4082
4083
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
4084
4085
4086
4087
            # NOTE(woosuk): Currently, we use inductor because the piecewise
            # CUDA graphs do not work properly with the custom CUDA kernels.
            # FIXME(woosuk): Disable inductor to reduce the compilation time
            # and avoid any potential issues with the inductor.
4088
            # FIXME(rob): Add function to set all of these.
4089
4090
            if not self.compilation_config.custom_ops:
                self.compilation_config.custom_ops = ["none"]
4091
4092
            self.compilation_config.use_cudagraph = True
            self.compilation_config.use_inductor = True
4093
            self.compilation_config.cudagraph_num_of_warmups = 1
4094
            self.compilation_config.pass_config.enable_fusion = False
4095
            self.compilation_config.pass_config.enable_noop = False
4096
            self.compilation_config.level = CompilationLevel.PIECEWISE
4097
            self.compilation_config.set_splitting_ops_for_v1()
4098

4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
        if self.parallel_config is not None and \
            self.parallel_config.tensor_parallel_size > 1 and \
            self.parallel_config.pipeline_parallel_size > 1 and \
            self.compilation_config is not None and \
                self.compilation_config.pass_config is not None and \
            self.compilation_config.pass_config.enable_sequence_parallelism:
            logger.warning_once(
                "Sequence parallelism is not supported with pipeline "
                "parallelism. Disabling sequence parallelism.")
            self.compilation_config.pass_config.\
                enable_sequence_parallelism = False

4111
        self._set_cudagraph_sizes()
4112

4113
4114
        if self.cache_config is not None and \
            self.cache_config.cpu_offload_gb > 0 and \
4115
4116
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
4117
            logger.warning(
4118
                "CPU offload is not supported with `torch.compile` in v0 yet."
4119
4120
4121
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4122
4123
4124
4125
4126
4127
        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`.")
4128
4129
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

4130

4131
        if self.model_config and self.model_config.use_mla and \
4132
            not (current_platform.is_cuda() or current_platform.is_rocm()):
4133
            logger.info(
4134
                "MLA is enabled on a non-GPU platform; forcing chunked "
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
                "prefill and prefix caching to be disabled.")
            self.scheduler_config.enable_chunked_prefill = False
            self.scheduler_config.chunked_prefill_enabled = False
            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

4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
        if (self.kv_events_config
                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.")
        if (self.kv_events_config and self.kv_events_config.publisher != "null"
                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.")
4157
4158
        current_platform.check_and_update_config(self)

4159
4160
4161
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
    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
        ]

4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
    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.

4198
4199
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
4200
4201

        During runtime, if batchsize is larger than
4202
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4203
4204
        no cudagraph will be used.
        If the batch size is no larger than
4205
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
        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 = []
            max_batchsize_to_capture = 0
            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)]
4219
4220
4221
4222
4223
                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)

4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
                # 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:
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
                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:
                    raise TypeError(
                        f"Invalid value for {cuda_graph_sizes=}.")
4255
4256
4257
4258
                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)
4259
4260
4261
4262
4263
                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
                ]
4264
4265
4266
4267

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

4268
    def __str__(self):
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
        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}, "
4299
4300
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
4301
4302
4303
4304
4305
4306


_current_vllm_config: Optional[VllmConfig] = None


@contextmanager
4307
def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False):
4308
    """
4309
    Temporarily set the current vLLM config.
4310
    Used during model initialization.
4311
    We save the current vLLM config in a global variable,
4312
    so that all modules can access it, e.g. custom ops
4313
    can access the vLLM config to determine how to dispatch.
4314
4315
4316
4317
4318
4319
4320
4321
    """
    global _current_vllm_config
    old_vllm_config = _current_vllm_config
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
        yield
4322
4323
4324
    except Exception:
        raise
    else:
4325
4326
4327
4328
        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)
4329
4330
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
4331
4332
4333
4334
4335
4336
4337
4338
4339
            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"
4340
                " if you want it to be supported.",
4341
                vllm_config.model_config.model)
4342
    finally:
4343
4344
4345
4346
4347
4348
4349
4350
        _current_vllm_config = old_vllm_config


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.
4351
        logger.warning("Current vLLM config is not set.")
4352
4353
4354
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381


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:
        bool: True if a match is found, False otherwise
    """
    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}")
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394


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)
    }