config.py 181 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 importlib.util import find_spec
18
from pathlib import Path
19
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Final, Literal,
20
                    Optional, Protocol, TypeVar, Union, get_args, get_origin)
21
22

import torch
23
from pydantic import BaseModel, Field, PrivateAttr
24
from torch.distributed import ProcessGroup, ReduceOp
25
from transformers import PretrainedConfig
26

27
import vllm.envs as envs
28
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
29
from vllm.logger import init_logger
30
from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
31
                                                     QuantizationMethods,
32
                                                     get_quantization_config)
33
from vllm.model_executor.models import ModelRegistry
34
from vllm.platforms import CpuArchEnum, current_platform
35
from vllm.sampling_params import GuidedDecodingParams
36
from vllm.tracing import is_otel_available, otel_import_error_traceback
37
38
39
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
40
41
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
    try_get_generation_config, uses_mrope)
42
from vllm.transformers_utils.s3_utils import S3Model
43
from vllm.transformers_utils.utils import is_s3, maybe_model_redirect
44
from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
45
46
                        get_cpu_memory, get_open_port, is_torch_equal_or_newer,
                        random_uuid, resolve_obj_by_qualname)
47

48
if TYPE_CHECKING:
49
    from _typeshed import DataclassInstance
50
51
    from ray.util.placement_group import PlacementGroup

52
    from vllm.executor.executor_base import ExecutorBase
53
54
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
55
    from vllm.model_executor.model_loader.loader import BaseModelLoader
56

57
    ConfigType = type[DataclassInstance]
58
59
else:
    QuantizationConfig = None
60
    ConfigType = type
61

62
63
logger = init_logger(__name__)

64
65
ConfigT = TypeVar("ConfigT", bound=ConfigType)

66
67
68
# 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
69
_POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
70
_MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
71

72
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
73
                     "score", "reward", "transcription"]
74

75
_ResolvedTask = Literal["generate", "embed", "classify", "score", "reward",
76
                        "draft", "transcription"]
77

78
RunnerType = Literal["generate", "pooling", "draft", "transcription"]
79

80
_RUNNER_TASKS: dict[RunnerType, list[_ResolvedTask]] = {
81
82
83
    "generate": ["generate"],
    "pooling": ["embed", "classify", "score", "reward"],
    "draft": ["draft"],
84
    "transcription": ["transcription"],
85
86
}

87
_TASK_RUNNER: dict[_ResolvedTask, RunnerType] = {
88
    task: runner
89
90
    for runner, tasks in _RUNNER_TASKS.items()
    for task in tasks
91
}
92

93
HfOverrides = Union[dict[str, Any], Callable[[PretrainedConfig],
94
95
                                             PretrainedConfig]]

96

97
98
99
100
101
102
class SupportsHash(Protocol):

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


103
104
class SupportsMetricsInfo(Protocol):

105
    def metrics_info(self) -> dict[str, str]:
106
107
108
        ...


109
110
111
112
113
114
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"


115
116
117
118
119
120
121
122
123
124
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
125

126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
        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


167
def config(cls: ConfigT) -> ConfigT:
168
169
170
171
172
173
174
175
176
177
178
179
    """
    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."
            )
180

181
182
183
        if f.name not in attr_docs:
            raise ValueError(
                f"Field '{f.name}' in {cls.__name__} must have a docstring.")
184
185
186
187
188
189
190
191
192

        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]'.")
193
194
195
    return cls


196
def get_field(cls: ConfigType, name: str) -> Field:
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
    """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.")


213
class ModelConfig:
214
215
216
217
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
218
            It is also used as the content for `model_name` tag in metrics
219
220
221
222
223
            output when `served_model_name` is not specified.
        task: 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.
224
        tokenizer: Name or path of the huggingface tokenizer to use.
225
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
226
227
228
            available, "slow" will always use the slow tokenizer,
            "mistral" will always use the tokenizer from `mistral_common`, and
            "custom" will use --tokenizer to select the preregistered tokenizer.
229
230
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
231
232
233
234
        allowed_local_media_path: 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.
235
236
237
238
        dtype: Data type for model weights and activations. The "auto" option
            will use FP16 precision for FP32 and FP16 models, and BF16 precision
            for BF16 models.
        seed: Random seed for reproducibility.
Jasmond L's avatar
Jasmond L committed
239
240
241
        revision: 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.
242
        code_revision: The specific revision to use for the model code on
243
            Hugging Face Hub. It can be a branch name, a tag name, or a
244
            commit id. If unspecified, will use the default version.
245
246
247
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id. If unspecified, will use
            the default version.
248
249
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
250
251
        spec_target_max_model_len: Specify the the maximum length for spec
            decoding draft models.
252
253
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
254
255
256
        enforce_eager: Whether to enforce eager execution. 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.
257
            If None, the user did not specify, so default to False.
258
259
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
260
261
262
            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.
263
        max_logprobs: Maximum number of log probabilities. Defaults to 20.
264
265
266
267
        disable_sliding_window: Whether to disable sliding window. If True,
            we will disable the sliding window functionality of the model.
            If the model does not support sliding window, this argument is
            ignored.
268
269
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
270
        served_model_name: The model name used in metrics tag `model_name`,
271
272
            matches the model name exposed via the APIs. If multiple model
            names provided, the first name will be used. If not specified,
273
            the model name will be the same as `model`.
274
        limit_mm_per_prompt: Maximum number of data items per modality
275
            per prompt. Only applicable for multimodal models.
276
277
278
279
        mm_processor_kwargs: Overrides for the multi-modal processor obtained
            from `AutoProcessor.from_pretrained`.
        disable_mm_preprocessor_cache: If True, disable caching of the
            processed multi-modal inputs.
280
281
        use_async_output_proc: Whether to use async output processor.
            Defaults to True.
282
283
        config_format: The config format which shall be loaded.
            Defaults to 'auto' which defaults to 'hf'.
284
        hf_token: The token to use as HTTP bearer authorization for remote files
285
            . If `True`, will use the token generated when running
286
            `huggingface-cli login` (stored in `~/.huggingface`).
287
288
289
        hf_overrides: If a dictionary, contains arguments to be forwarded to the
            HuggingFace config. If a callable, it is called to update the
            HuggingFace config.
290
291
292
293
        override_neuron_config: 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.
294
        override_pooler_config: Initialize non default pooling config or
295
            override default pooling config for the pooling model.
296
297
        logits_processor_pattern: Optional regex pattern specifying valid
            logits processor qualified names that can be passed with the
298
            `logits_processors` extra completion argument. Defaults to None,
299
            which allows no processors.
300
        generation_config: Configuration parameter file for generation.
301
302
303
304
305
306
        model_impl: Which implementation of the model to use:
            "auto" will try to use the vLLM implementation if it exists and
                fall back to the Transformers implementation if no vLLM
                implementation is available.
            "vllm" will use the vLLM model implementation.
            "transformers" will use the Transformers model implementation.
307
308
        override_generation_config: Override the generation config with the
            given config.
309
    """
310

311
312
313
314
315
316
317
318
319
320
321
322
    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.
        """
323
        factors: list[Any] = []
324
325
326
327
328
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
329
330
331
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
332
        factors.append(self.trust_remote_code)
333
334
335
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
336
337
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
338
339
        # hf_config can control how the model looks!
        factors.append(self.hf_config.to_json_string())
340
341
        str_factors = str(factors)
        assert_hashable(str_factors)
342
343
        return hashlib.sha256(str(factors).encode()).hexdigest()

344
345
346
347
348
349
350
351
352
    def __init__(
        self,
        model: str,
        task: Union[TaskOption, Literal["draft"]],
        tokenizer: str,
        tokenizer_mode: str,
        trust_remote_code: bool,
        dtype: Union[str, torch.dtype],
        seed: int,
353
        hf_config_path: Optional[str] = None,
354
355
356
        allowed_local_media_path: str = "",
        revision: Optional[str] = None,
        code_revision: Optional[str] = None,
357
        rope_scaling: Optional[dict[str, Any]] = None,
358
359
360
361
362
363
364
365
366
        rope_theta: Optional[float] = None,
        tokenizer_revision: Optional[str] = None,
        max_model_len: Optional[int] = None,
        spec_target_max_model_len: Optional[int] = None,
        quantization: Optional[str] = None,
        enforce_eager: Optional[bool] = None,
        max_seq_len_to_capture: Optional[int] = None,
        max_logprobs: int = 20,
        disable_sliding_window: bool = False,
367
        disable_cascade_attn: bool = False,
368
        skip_tokenizer_init: bool = False,
369
        served_model_name: Optional[Union[str, list[str]]] = None,
370
        limit_mm_per_prompt: Optional[dict[str, int]] = None,
371
372
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
        disable_mm_preprocessor_cache: bool = False,
373
374
        use_async_output_proc: bool = True,
        config_format: ConfigFormat = ConfigFormat.AUTO,
375
        hf_token: Optional[Union[bool, str]] = None,
376
        hf_overrides: Optional[HfOverrides] = None,
377
        override_neuron_config: Optional[dict[str, Any]] = None,
378
379
        override_pooler_config: Optional["PoolerConfig"] = None,
        logits_processor_pattern: Optional[str] = None,
380
        generation_config: str = "auto",
381
        enable_sleep_mode: bool = False,
382
        override_generation_config: Optional[dict[str, Any]] = None,
383
        model_impl: Union[str, ModelImpl] = ModelImpl.AUTO,
384
    ) -> None:
385
386
387
        self.model = maybe_model_redirect(model)
        self.tokenizer = maybe_model_redirect(tokenizer)

388
        self.hf_config_path = hf_config_path
389
390
391
        if isinstance(hf_config_path, str):
            self.hf_config_path = maybe_model_redirect(hf_config_path)

392
        self.tokenizer_mode = tokenizer_mode
393
        self.trust_remote_code = trust_remote_code
394
        self.allowed_local_media_path = allowed_local_media_path
395
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
396
        self.revision = revision
397
        self.code_revision = code_revision
398
399
        self.rope_scaling = rope_scaling
        self.rope_theta = rope_theta
400
        self.model_impl = model_impl
401
402
403

        if hf_overrides is None:
            hf_overrides = {}
404
405
406
407
408
409

        if callable(hf_overrides):
            hf_overrides_kw = {}
            hf_overrides_fn = hf_overrides
        else:
            hf_overrides_kw = hf_overrides
410
            hf_overrides_fn = None
411

412
        if rope_scaling is not None:
413
            hf_override: dict[str, Any] = {"rope_scaling": rope_scaling}
414
            hf_overrides_kw.update(hf_override)
415
416
417
418
            hf_overrides_str = json.dumps(hf_overrides)
            msg = (
                "`--rope-scaling` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
419
420
421
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
        if rope_theta is not None:
            hf_override = {"rope_theta": rope_theta}
422
            hf_overrides_kw.update(hf_override)
423
424
425
426
            hf_overrides_str = json.dumps(hf_overrides)
            msg = (
                "`--rope-theta` will be removed in a future release. "
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`")
427
428
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

429
430
        self.maybe_pull_model_tokenizer_for_s3(model, tokenizer)

431
432
433
434
        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 "
435
436
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
437
438
                "for instructions on how to install it.")

439
440
441
442
443
        # The tokenizer version is consistent with the model version by default.
        if tokenizer_revision is None:
            self.tokenizer_revision = revision
        else:
            self.tokenizer_revision = tokenizer_revision
444
        self.quantization = quantization
445
        self.enforce_eager = enforce_eager
446
        self.max_seq_len_to_capture = max_seq_len_to_capture
447
        self.max_logprobs = max_logprobs
448
        self.disable_sliding_window = disable_sliding_window
449
        self.disable_cascade_attn = disable_cascade_attn
450
        self.skip_tokenizer_init = skip_tokenizer_init
451
452
453
454
        self.enable_sleep_mode = enable_sleep_mode

        from vllm.platforms import current_platform

455
456
457
458
        if (self.enable_sleep_mode
                and not current_platform.is_sleep_mode_available()):
            raise ValueError(
                "Sleep mode is not supported on current platform.")
459

460
461
462
        hf_config = get_config(self.hf_config_path or self.model,
                               trust_remote_code, revision, code_revision,
                               config_format)
463
464
465
466
467
468
469
470

        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)

471
472
        self.hf_config = hf_config

473
        self.hf_text_config = get_hf_text_config(self.hf_config)
474
475
        self.attention_chunk_size = getattr(self.hf_text_config,
                                            "attention_chunk_size", None)
476
        self.encoder_config = self._get_encoder_config()
477
        self.hf_image_processor_config = get_hf_image_processor_config(
478
            self.model, hf_token=hf_token, revision=revision)
479
        self.dtype = _get_and_verify_dtype(self.hf_config, dtype)
480
        self.use_async_output_proc = use_async_output_proc
Woosuk Kwon's avatar
Woosuk Kwon committed
481

482
483
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
484
485
            self.enforce_eager = False

486
        interleaved_attn_models = ["gemma2", "gemma3_text", "cohere2"]
487
488
489
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
        has_interleaved_attention = (sliding_window is not None) and (
            isinstance(sliding_window, list) or
490
            (self.hf_text_config.model_type in interleaved_attn_models))
491
492

        if (not self.disable_sliding_window and has_interleaved_attention):
493
494
            if (backend :=
                    envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"):
495
496
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
497

498
                logger.warning_once(
499
500
                    f"{self.hf_text_config.model_type} has interleaved "
                    "attention, which is currently not supported by the "
501
                    f"{backend} backend. Disabling sliding window and capping "
502
503
504
505
506
507
508
509
510
511
512
513
                    "the max length to the sliding window size "
                    f"({sliding_window_len_min}).")
                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
514

515
516
517
518
        self.max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
519
            sliding_window_len=self.get_hf_config_sliding_window(),
520
521
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
522
523
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
524
        self.multimodal_config = self._init_multimodal_config(
525
526
527
528
            limit_mm_per_prompt=limit_mm_per_prompt,
            mm_processor_kwargs=mm_processor_kwargs,
            disable_mm_preprocessor_cache=disable_mm_preprocessor_cache,
        )
529
530
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
531

532
        self.is_attention_free = self._init_attention_free()
533
        self.is_hybrid = self._init_is_hybrid()
534
        self.has_noops = self._init_has_noops()
535
536
        self.has_inner_state = self._init_has_inner_state()

537
538
539
540
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
541

542
        supported_tasks, task = self._resolve_task(task)
543
544
        self.supported_tasks = supported_tasks
        self.task: Final = task
545
546
547
548
        if self.task in ("draft", "generate"):
            self.truncation_side = "left"
        else:
            self.truncation_side = "right"
549

550
        self.pooler_config = self._init_pooler_config(override_pooler_config)
551
        self.logits_processor_pattern = logits_processor_pattern
552

553
        self.generation_config = generation_config
554
        self.override_generation_config = override_generation_config or {}
555

556
        self._verify_quantization()
557
        self._verify_cuda_graph()
558
        self._verify_bnb_config()
559

560
561
562
563
564
565
566
567
    @property
    def registry(self):
        return ModelRegistry

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

568
569
570
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
        """
571
        Pull the model config or tokenizer to a temporary
572
573
574
575
576
577
578
579
580
        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):
581
                s3_model = S3Model()
582
583
                s3_model.pull_files(
                    model, allow_pattern=["*.model", "*.py", "*.json"])
584
                self.model_weights = self.model
585
                self.model = s3_model.dir
586
587

            if is_s3(tokenizer):
588
589
                s3_tokenizer = S3Model()
                s3_tokenizer.pull_files(
590
                    model, ignore_pattern=["*.pt", "*.safetensors", "*.bin"])
591
                self.tokenizer = s3_tokenizer.dir
592

593
    def _init_multimodal_config(
594
595
596
597
        self,
        limit_mm_per_prompt: Optional[dict[str, int]],
        mm_processor_kwargs: Optional[dict[str, Any]],
        disable_mm_preprocessor_cache: bool,
598
    ) -> Optional["MultiModalConfig"]:
599
        if self.registry.is_multimodal_model(self.architectures):
600
601
602
603
604
            return MultiModalConfig(
                limit_per_prompt=limit_mm_per_prompt or {},
                mm_processor_kwargs=mm_processor_kwargs or {},
                disable_mm_preprocessor_cache=disable_mm_preprocessor_cache,
            )
605
606
607
608

        if limit_mm_per_prompt:
            raise ValueError("`limit_mm_per_prompt` is only supported for "
                             "multimodal models.")
609
610
611
612
613
614
        if mm_processor_kwargs:
            raise ValueError("`mm_processor_kwargs` is only supported for "
                             "multimodal models.")
        if disable_mm_preprocessor_cache:
            raise ValueError("`disable_mm_preprocessor_cache` is only "
                             "supported for multimodal models.")
615
616

        return None
617

618
619
620
621
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

622
623
    def _init_pooler_config(
        self,
624
        override_pooler_config: Optional["PoolerConfig"],
625
    ) -> Optional["PoolerConfig"]:
626

627
        if self.runner_type == "pooling":
628
629
630
631
632
633
634
635
636
            user_config = override_pooler_config or PoolerConfig()

            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():
                    if getattr(user_config, k) is None:
                        setattr(user_config, k, v)

637
638
639
640
641
642
643
644
645
            if self.is_matryoshka:
                if user_config.normalize is None:
                    user_config.normalize = True
                elif not user_config.normalize:
                    raise ValueError(
                        "`normalize` must be enabled (set to True) "
                        "for models that are compatible with "
                        "Matryoshka Representation.")

646
647
            return user_config

648
649
        return None

650
    def _init_attention_free(self) -> bool:
651
        return self.registry.is_attention_free_model(self.architectures)
652

653
    def _init_is_hybrid(self) -> bool:
654
        return self.registry.is_hybrid_model(self.architectures)
655

656
657
658
659
    def _init_has_noops(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return self.registry.is_noops_model(architectures)

660
    def _init_has_inner_state(self) -> bool:
661
        return self.registry.model_has_inner_state(self.architectures)
662

663
664
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
665
        if tokenizer_mode not in ["auto", "slow", "mistral", "custom"]:
666
667
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
668
                "either 'auto', 'slow', 'mistral' or 'custom'.")
669
        self.tokenizer_mode = tokenizer_mode
670

671
672
    def _get_preferred_task(
        self,
673
674
        architectures: list[str],
        supported_tasks: set[_ResolvedTask],
675
676
677
678
    ) -> Optional[_ResolvedTask]:
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
679
        if self.registry.is_cross_encoder_model(architectures):
680
            return "score"
681
        if self.registry.is_transcription_model(architectures):
682
            return "transcription"
683

684
        suffix_to_preferred_task: list[tuple[str, _ResolvedTask]] = [
685
686
687
688
689
690
691
692
693
            # Other models follow this pattern
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ForSequenceClassification", "classify"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]
694
        _, arch = self.registry.inspect_model_cls(architectures)
695
696
697
698
699
700
701

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

        return None

702
703
    def _resolve_task(
        self,
704
        task_option: Union[TaskOption, Literal["draft"]],
705
    ) -> tuple[set[_ResolvedTask], _ResolvedTask]:
706
707
708
        if task_option == "draft":
            return {"draft"}, "draft"

709
710
        registry = self.registry
        architectures = self.architectures
711

712
        runner_support: dict[RunnerType, bool] = {
713
714
            # NOTE: Listed from highest to lowest priority,
            # in case the model supports multiple of them
715
716
717
            "transcription": registry.is_transcription_model(architectures),
            "generate": registry.is_text_generation_model(architectures),
            "pooling": registry.is_pooling_model(architectures),
718
        }
719
        supported_runner_types_lst: list[RunnerType] = [
720
721
722
723
724
            runner_type
            for runner_type, is_supported in runner_support.items()
            if is_supported
        ]

725
        supported_tasks_lst: list[_ResolvedTask] = [
726
727
            task for runner_type in supported_runner_types_lst
            for task in _RUNNER_TASKS[runner_type]
728
729
730
731
732
        ]
        supported_tasks = set(supported_tasks_lst)

        if task_option == "auto":
            selected_task = next(iter(supported_tasks_lst))
733

734
735
736
737
738
            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
739

740
741
742
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
743
        else:
744
745
746
747
748
749
750
751
752
753
754
755
756
757
            # 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"

758
759
760
761
762
763
764
            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
765

766
        return supported_tasks, selected_task
767

768
769
770
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
771
            # compressed-tensors uses a "compression_config" key
772
            quant_cfg = getattr(self.hf_config, "compression_config", None)
773
774
        return quant_cfg

775
    def _verify_quantization(self) -> None:
776
        supported_quantization = QUANTIZATION_METHODS
777
        optimized_quantization_methods = [
778
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
779
780
            "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
            "quark", "nvfp4", "bitblas", "gptq_bitblas"
781
        ]
782
783
784
785
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
786
787
        quant_cfg = self._parse_quant_hf_config()

788
789
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
790
791
792
            quant_method = quant_method.replace("compressed_tensors",
                                                "compressed-tensors")
            quant_cfg["quant_method"] = quant_method
793

794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
            # 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

815
            # Detect which checkpoint is it
816
            for name in quantization_methods:
817
                method = get_quantization_config(name)
818
819
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
820
821
822
823
824
825
826
827
828
829
830
                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.")
831
832
833
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
834

835
            # Verify quantization configurations.
836
            if self.quantization is None:
837
838
                self.quantization = quant_method
            elif self.quantization != quant_method:
839
840
                raise ValueError(
                    "Quantization method specified in the model config "
841
                    f"({quant_method}) does not match the quantization "
842
843
844
845
846
847
848
849
                    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}.")
850
            from vllm.platforms import current_platform
851
            current_platform.verify_quantization(self.quantization)
852
            if self.quantization not in optimized_quantization_methods:
853
                logger.warning(
854
                    "%s quantization is not fully "
855
                    "optimized yet. The speed can be slower than "
856
                    "non-quantized models.", self.quantization)
857

858
    def _verify_cuda_graph(self) -> None:
859
860
861
862
        if self.max_seq_len_to_capture is None:
            self.max_seq_len_to_capture = self.max_model_len
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
863
864
865
866
867
868
869
        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
870

871
872
    def _verify_bnb_config(self) -> None:
        """
873
        The current version of bitsandbytes (0.45.3) with 8-bit models does not
874
        yet support CUDA graph.
875
        # TODO Remove this when bitsandbytes supports.
876
877
878
879
880
881
882
883
884
885
886
887
888
889
        """
        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(
890
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
891
                "fallback to the eager mode.")
892

893
894
            self.enforce_eager = True

895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
    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.")

912
913
914
915
916
917
918
919
920
921
    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

922
        # Reminder: Please update docs/source/features/compatibility_matrix.md
923
        # If the feature combo become valid
924
        from vllm.platforms import current_platform
925
        if not current_platform.is_async_output_supported(self.enforce_eager):
926
927
928
929
930
931
932
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

933
        # Async postprocessor is not necessary for pooling models
934
        # since there is no token generation
935
        if self.runner_type == "pooling":
936
937
            self.use_async_output_proc = False

938
        # Reminder: Please update docs/source/features/compatibility_matrix.md
939
        # If the feature combo become valid
940
941
942
        if speculative_config:
            self.use_async_output_proc = False

943
944
945
946
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
947
948
949
950
951
952

        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.")

953
954
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
955
956
957
958
959
960
961
        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}).")

962
        if parallel_config.enable_expert_parallel:
963
964
            self._verify_with_expert_parallelism()

965
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
966
        if pipeline_parallel_size > 1:
967
            if not self.registry.is_pp_supported_model(self.architectures):
968
969
970
971
972
973
                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
974

975
    def get_hf_config_sliding_window(
976
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
977
        """Get the sliding window size, or None if disabled."""
978
979
980
981

        # 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.
982
983
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
984
            return None
985
        return getattr(self.hf_text_config, "sliding_window", None)
986

987
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
988
989
990
991
992
993
994
995
        """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()

996
    def get_vocab_size(self) -> int:
997
        return self.hf_text_config.vocab_size
998

999
    def get_hidden_size(self) -> int:
1000
        return self.hf_text_config.hidden_size
1001

1002
1003
    @property
    def is_deepseek_mla(self) -> bool:
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
        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
1016

1017
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
1018
        # TODO remove hard code
1019
        if self.is_deepseek_mla:
1020
1021
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
1022
            if self.use_mla:
1023
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
1024
1025
1026
1027
1028
            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
1029

1030
1031
1032
1033
1034
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

1035
1036
1037
        if self.is_attention_free:
            return 0

1038
1039
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
1040
        # FIXME(woosuk): This may not be true for all models.
1041
1042
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
1043

1044
1045
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
1046
        # For GPTBigCode & Falcon:
1047
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
1048
1049
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
1050
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
1051
        new_decoder_arch_falcon = (
1052
            self.hf_config.model_type in falcon_model_types
1053
            and getattr(self.hf_config, "new_decoder_architecture", False))
1054
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
1055
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
1056
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
1057
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
1058
            return 1
1059

1060
        # For DBRX and MPT
1061
1062
1063
1064
1065
        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":
1066
1067
1068
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

1069
1070
1071
1072
1073
1074
1075
1076
        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")

1077
1078
1079
        if self.is_attention_free:
            return 0

1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
        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:
1090
            num_kv_heads = getattr(self.hf_text_config, attr, None)
1091
1092
1093
1094
1095
            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.
1096
        return self.hf_text_config.num_attention_heads
1097
1098
1099

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
1100
1101
1102
1103
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

1104
1105
1106
1107
1108
1109
1110
        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)
1111

1112
1113
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
1114
1115
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
1116

1117
    def get_layers_start_end_indices(
1118
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
1119
        from vllm.distributed.utils import get_pp_indices
1120
1121
1122
1123
1124
1125
        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)
1126
1127
1128
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
1129
1130
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
1131
        return start, end
Mor Zusman's avatar
Mor Zusman committed
1132

1133
1134
1135
    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
1136

1137
1138
1139
1140
1141
1142
1143
1144
    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
1145
1146
1147
        is_transformer = not self.is_hybrid and \
                            not self.has_noops and \
                            not self.is_attention_free
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
        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
1158
1159
1160
1161
        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])
1162
        else:
1163
            # Hybrid model Jamba
1164
1165
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
            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
1191

1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
    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

1204
    def try_get_generation_config(self) -> dict[str, Any]:
1205
        if self.generation_config in ("auto", "vllm"):
1206
            config = try_get_generation_config(
1207
                self.hf_config_path or self.model,
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
                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()

1222
    def get_diff_sampling_param(self) -> dict[str, Any]:
1223
        """
1224
        This method returns a dictionary containing the parameters
1225
1226
        that differ from the default sampling parameters. If
        `generation_config` is `"vllm"`, an empty dictionary is returned.
1227
1228

        Returns:
1229
            dict[str, Any]: A dictionary with the differing sampling
1230
            parameters, if `generation_config` is `"vllm"` an empty dictionary.
1231
        """
1232
        if self.generation_config == "vllm":
1233
1234
1235
1236
1237
1238
1239
            config = {}
        else:
            config = self.try_get_generation_config()

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

1240
1241
1242
1243
1244
1245
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1246
            "max_new_tokens",
1247
1248
1249
1250
1251
1252
        ]
        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
            }
1253
1254
1255
1256
1257
            # 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")
1258
1259
        else:
            diff_sampling_param = {}
1260
1261
1262
1263
1264
1265
1266

        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`.")
1267
1268
        return diff_sampling_param

1269
    @property
1270
    def is_encoder_decoder(self) -> bool:
1271
        """Extract the HF encoder/decoder model flag."""
1272
1273
1274
1275
1276
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1277

1278
1279
1280
1281
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1282
1283
    @property
    def is_cross_encoder(self) -> bool:
1284
        return self.registry.is_cross_encoder_model(self.architectures)
1285

1286
1287
    @property
    def use_mla(self) -> bool:
1288
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1289

1290
    @property
1291
    def supported_runner_types(self) -> set[RunnerType]:
1292
1293
1294
1295
1296
1297
        return {_TASK_RUNNER[task] for task in self.supported_tasks}

    @property
    def runner_type(self) -> RunnerType:
        return _TASK_RUNNER[self.task]

1298
1299
1300
1301
1302
    @property
    def is_v1_compatible(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_v1_compatible(architectures)

1303
1304
1305
1306
1307
    @property
    def is_matryoshka(self) -> bool:
        return (hasattr(self.hf_config, "matryoshka_dimensions")
                or getattr(self.hf_config, "is_matryoshka", False))

1308
1309
1310
1311
    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)

1312

1313
BlockSize = Literal[1, 8, 16, 32, 64, 128]
1314
1315
1316
1317
1318
1319
CacheDType = Literal["auto", "fp8", "fp8_e4m3", "fp8_e5m2"]
PrefixCachingHashAlgo = Literal["builtin", "sha256"]


@config
@dataclass
1320
class CacheConfig:
1321
    """Configuration for the KV cache."""
1322

1323
    block_size: BlockSize = None  # type: ignore
1324
1325
1326
    """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.
1327
1328
1329
1330

    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."""
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
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
    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."""
1380

1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
    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.
        """
1393
        factors: list[Any] = []
1394
1395
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
1396
1397
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1398
1399
        return hash_str

1400
1401
1402
    def __post_init__(self) -> None:
        self.swap_space_bytes = self.swap_space * GiB_bytes

1403
        self._verify_args()
1404
        self._verify_cache_dtype()
1405
        self._verify_prefix_caching()
1406

1407
    def metrics_info(self):
1408
1409
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1410
1411
        return {key: str(value) for key, value in self.__dict__.items()}

1412
    def _verify_args(self) -> None:
1413
1414
1415
1416
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1417
1418
1419
1420
1421
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1422
1423
1424
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1425
        elif self.cache_dtype in get_args(CacheDType):
1426
            logger.info(
1427
1428
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1429
1430
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
1431
1432
1433
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1434
1435
1436
1437
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

1438
        if self.sliding_window is not None and not envs.VLLM_USE_V1:
1439
1440
1441
1442
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

1443
1444
        if (self.enable_prefix_caching and self.prefix_caching_hash_algo
                not in get_args(PrefixCachingHashAlgo)):
1445
1446
            raise ValueError(
                "Unknown prefix caching hash algorithm: "
1447
1448
                f"{self.prefix_caching_hash_algo}. Must be one of "
                f"{get_args(PrefixCachingHashAlgo)}.")
1449

1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
    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

1460
1461
1462
        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.")
1463
1464
1465
        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:
1466
            logger.warning("Possibly too large swap space. %s", msg)
1467

1468

1469
@config
1470
1471
@dataclass
class TokenizerPoolConfig:
1472
    """This config is deprecated and will be removed in a future release.
1473

1474
1475
1476
    Passing these parameters will have no effect. Please remove them from your
    configurations.
    """
1477

1478
1479
1480
1481
1482
1483
1484
1485
    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."""
1486
    extra_config: dict = field(default_factory=dict)
1487
1488
1489
    """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."""
1490

1491
1492
1493
1494
1495
    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.")
1496
1497


1498
1499
1500
1501
1502
1503
1504
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
1505
    SHARDED_STATE = "sharded_state"
1506
    GGUF = "gguf"
1507
    BITSANDBYTES = "bitsandbytes"
1508
    MISTRAL = "mistral"
1509
    RUNAI_STREAMER = "runai_streamer"
1510
    RUNAI_STREAMER_SHARDED = "runai_streamer_sharded"
1511
    FASTSAFETENSORS = "fastsafetensors"
1512
1513


1514
@config
1515
1516
@dataclass
class LoadConfig:
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
    """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."""
1542
    download_dir: Optional[str] = None
1543
1544
    """Directory to download and load the weights, default to the default
    cache directory of Hugging Face."""
1545
    model_loader_extra_config: dict = field(default_factory=dict)
1546
1547
1548
    """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."""
1549
    ignore_patterns: Optional[Union[list[str], str]] = None
1550
1551
    """The list of patterns to ignore when loading the model. Default to
    "original/**/*" to avoid repeated loading of llama's checkpoints."""
1552
    use_tqdm_on_load: bool = True
1553
1554
    """Whether to enable tqdm for showing progress bar when loading model
    weights."""
1555

1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
    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.
1570
        factors: list[Any] = []
1571
1572
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1573
1574
        return hash_str

1575
    def __post_init__(self):
1576
1577
1578
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1579

1580
1581
1582
1583
1584
1585
1586
        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/**/*"]

1587

1588
1589
1590
DistributedExecutorBackend = Literal["ray", "mp", "uni", "external_launcher"]


1591
@config
1592
@dataclass
1593
class ParallelConfig:
1594
    """Configuration for the distributed execution."""
1595

1596
1597
1598
1599
1600
1601
1602
1603
1604
    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."""
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
    _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

1620
    data_parallel_master_ip: str = "127.0.0.1"
1621
1622
1623
1624
1625
    """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."""
1626

1627
    max_parallel_loading_workers: Optional[int] = None
1628
1629
1630
    """Maximum number of parallal loading workers when loading model
    sequentially in multiple batches. To avoid RAM OOM when using tensor
    parallel and large models."""
1631
1632

    disable_custom_all_reduce: bool = False
1633
    """Disable the custom all-reduce kernel and fall back to NCCL."""
1634
1635

    tokenizer_pool_config: Optional[TokenizerPoolConfig] = None
1636
1637
    """This parameter is deprecated and will be removed in a future release.
    Please remove it from your configs"""
1638
1639

    ray_workers_use_nsight: bool = False
1640
    """Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler."""
1641
1642

    placement_group: Optional["PlacementGroup"] = None
1643
    """ray distributed model workers placement group."""
1644

1645
    distributed_executor_backend: Optional[Union[DistributedExecutorBackend,
1646
                                                 type["ExecutorBase"]]] = None
1647
1648
1649
1650
1651
1652
1653
    """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."""
1654
1655

    worker_cls: str = "auto"
1656
1657
    """The full name of the worker class to use. If "auto", the worker class
    will be determined based on the platform."""
1658
    sd_worker_cls: str = "auto"
1659
    """The full name of the worker class to use for speculative decofing.
1660
    If "auto", the worker class will be determined based on the platform."""
1661
    worker_extension_cls: str = ""
1662
1663
1664
1665
    """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."""
1666
1667

    world_size: int = field(init=False)
1668
    """world_size is TPxPP, it affects the number of workers we create."""
1669
    world_size_across_dp: int = field(init=False)
1670
1671
    """world_size_across_dp is TPxPPxDP, it is the size of the world
    including data parallelism."""
1672
1673

    rank: int = 0
1674
    """Global rank in distributed setup."""
1675

1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
    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
1705
                          has_unfinished: bool) -> bool:
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
        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

1717
1718
1719
1720
1721
1722
1723
1724
    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.
        """
1725
        factors: list[Any] = []
1726
1727
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
1728
        factors.append(self.enable_expert_parallel)
1729
1730
        return hashlib.sha256(str(factors).encode()).hexdigest()

1731
1732
1733
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
1734

1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
        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

1747
        self.world_size_across_dp = self.world_size * self.data_parallel_size
1748

1749
1750
1751
1752
1753
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

1754
        ray_only_devices: list[str] = []
1755
        from vllm.platforms import current_platform
1756
1757
        if (current_platform.device_type in ray_only_devices
                and self.world_size > 1):
1758
1759
1760
1761
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
1762
1763
                    f"{current_platform.device_type.upper()} backend only "
                    "supports Ray for distributed inference.")
1764

1765
        if self.distributed_executor_backend is None and self.world_size > 1:
1766
1767
1768
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1769
            from vllm.executor import ray_utils
1770
            backend: DistributedExecutorBackend = "mp"
1771
            ray_found = ray_utils.ray_is_available()
1772
1773
1774
1775
1776
            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):
1777
1778
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1779
1780
1781
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1782
1783
                backend = "ray"
            elif ray_found:
1784
                if self.placement_group:
1785
                    backend = "ray"
1786
1787
1788
1789
1790
1791
                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"
1792
1793
1794
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1795

1796
1797
1798
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

1799
1800
        self._verify_args()

1801
1802
1803
1804
1805
1806
    @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)

1807
    def _verify_args(self) -> None:
1808
1809
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
1810
        from vllm.platforms import current_platform
1811
        if self.distributed_executor_backend not in (
1812
1813
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
1814
1815
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
1816
            raise ValueError(
1817
1818
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
1819
1820
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
1821
        if self.use_ray:
1822
1823
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1824
1825

        if not current_platform.use_custom_allreduce():
1826
1827
1828
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
1829
                "supported on current platform.")
1830
        if self.ray_workers_use_nsight and not self.use_ray:
1831
1832
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1833

1834
1835
1836
        assert isinstance(self.worker_extension_cls, str), (
            "worker_extension_cls must be a string (qualified class name).")

1837

1838
PreemptionMode = Literal["swap", "recompute"]
1839
1840
1841
1842
SchedulerPolicy = Literal["fcfs", "priority"]


@config
1843
@dataclass
1844
class SchedulerConfig:
1845
    """Scheduler configuration."""
1846

1847
1848
    runner_type: RunnerType = "generate"
    """The runner type to launch for the model."""
1849

1850
1851
    max_num_batched_tokens: int = None  # type: ignore
    """Maximum number of tokens to be processed in a single iteration.
1852

1853
1854
    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."""
1855

1856
1857
    max_num_seqs: int = None  # type: ignore
    """Maximum number of sequences to be processed in a single iteration.
1858

1859
1860
    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."""
1861

1862
1863
1864
1865
    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."""
1866

1867
    max_num_partial_prefills: int = 1
1868
1869
    """For chunked prefill, the maximum number of sequences that can be
    partially prefilled concurrently."""
1870
1871

    max_long_partial_prefills: int = 1
1872
1873
1874
1875
    """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."""
1876
1877

    long_prefill_token_threshold: int = 0
1878
1879
    """For chunked prefill, a request is considered long if the prompt is
    longer than this number of tokens."""
1880

1881
    num_lookahead_slots: int = 0
1882
1883
1884
1885
1886
1887
1888
    """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."""
1889
1890

    delay_factor: float = 0.0
1891
1892
    """Apply a delay (of delay factor multiplied by previous
    prompt latency) before scheduling next prompt."""
1893

1894
1895
1896
    enable_chunked_prefill: bool = None  # type: ignore
    """If True, prefill requests can be chunked based
    on the remaining max_num_batched_tokens."""
1897
1898

    is_multimodal_model: bool = False
1899
1900
1901
1902
1903
    """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.
1904

1905
1906
1907
1908
1909
1910
1911
1912
1913
    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."""
1914

1915
    preemption_mode: Optional[PreemptionMode] = None
1916
1917
1918
1919
1920
1921
    """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."""
1922
1923

    num_scheduler_steps: int = 1
1924
    """Maximum number of forward steps per scheduler call."""
1925

1926
1927
    multi_step_stream_outputs: bool = True
    """If False, then multi-step will stream outputs at the end of all steps"""
1928
1929

    send_delta_data: bool = False
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
    """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)."""
1941
1942

    chunked_prefill_enabled: bool = field(init=False)
1943
    """True if chunked prefill is enabled."""
1944

1945
    disable_chunked_mm_input: bool = False
1946
1947
1948
1949
1950
1951
    """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."""
1952

1953
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
1954
1955
1956
    """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"."""
1957

1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
    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.
1972
        factors: list[Any] = []
1973
1974
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
1975
1976
        return hash_str

1977
    def __post_init__(self) -> None:
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
        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)

1990
1991
1992
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
1993
1994
1995
1996
                    # 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.
1997
1998
                    self.max_num_batched_tokens = max(
                        self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
1999
                else:
2000
2001
                    self.max_num_batched_tokens = (
                        _DEFAULT_MAX_NUM_BATCHED_TOKENS)
2002
            else:
2003
2004
                # If max_model_len is too short, use
                # _DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
2005
                # for higher throughput.
2006
2007
                self.max_num_batched_tokens = max(
                    self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
2008

2009
2010
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
2011
2012
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2013
                    _POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
2014
                )
2015
            if self.is_multimodal_model:
2016
                # The value needs to be at least the number of multimodal tokens
2017
2018
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
2019
2020
2021
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

2022
2023
2024
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

2025
        if self.enable_chunked_prefill:
2026
2027
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
2028
                self.max_num_batched_tokens)
2029

2030
        self.chunked_prefill_enabled = self.enable_chunked_prefill
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
        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)

2043
2044
2045
        self._verify_args()

    def _verify_args(self) -> None:
2046
2047
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
2048
2049
2050
2051
2052
2053
2054
            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.")
2055

2056
2057
2058
2059
2060
        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}).")
2061

2062
2063
2064
2065
2066
2067
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

2068
2069
2070
2071
2072
2073
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
        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}).")

2097
2098
2099
2100
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

2101

2102
2103
2104
2105
2106
Device = Literal["auto", "cuda", "neuron", "cpu", "tpu", "xpu", "hpu"]


@config
@dataclass
2107
class DeviceConfig:
2108
2109
2110
2111
2112
2113
2114
    """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__`."""
2115

2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
    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.
2131
        factors: list[Any] = []
2132
2133
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2134
2135
        return hash_str

2136
2137
    def __post_init__(self):
        if self.device == "auto":
2138
            # Automated device type detection
2139
            from vllm.platforms import current_platform
2140
            self.device_type = current_platform.device_type
2141
            if not self.device_type:
2142
2143
2144
2145
                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.")
2146
2147
        else:
            # Device type is assigned explicitly
2148
            self.device_type = self.device
2149
2150

        # Some device types require processing inputs on CPU
2151
        if self.device_type in ["neuron"]:
2152
            self.device = torch.device("cpu")
2153
2154
        elif self.device_type in ["tpu"]:
            self.device = None
2155
2156
2157
2158
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

2159

2160
2161
2162
2163
2164
2165
2166
SpeculativeMethod = Literal["ngram", "eagle", "medusa", "mlp_speculator",
                            "draft_model"]
SpeculativeAcceptanceMethod = Literal["rejection_sampler",
                                      "typical_acceptance_sampler"]


@config
2167
@dataclass
2168
class SpeculativeConfig:
2169
    """Configuration for speculative decoding."""
2170

2171
    # General speculative decoding control
2172
2173
    num_speculative_tokens: int = field(default=None,
                                        init=True)  # type: ignore
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
    """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."""
2194
    draft_tensor_parallel_size: Optional[int] = None
2195
2196
    """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."""
2197
    disable_logprobs: bool = True
2198
2199
2200
    """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."""
2201

2202
    # Draft model configuration
2203
    quantization: Optional[str] = None
2204
2205
2206
    """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."""
2207
    max_model_len: Optional[int] = None
2208
2209
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
2210
    revision: Optional[str] = None
2211
2212
2213
    """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."""
2214
    code_revision: Optional[str] = None
2215
2216
2217
    """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."""
2218

2219
    # Advanced control
2220
    disable_mqa_scorer: bool = False
2221
2222
    """Disable the MQA scorer and fall back to batch expansion for scoring
    proposals."""
2223
    disable_by_batch_size: Optional[int] = None
2224
2225
2226
2227
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""

    # Ngram proposer configuration
2228
    prompt_lookup_max: Optional[int] = None
2229
2230
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
2231
    prompt_lookup_min: Optional[int] = None
2232
2233
2234
2235
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

    # Typical acceptance sampler configuration
2236
    posterior_threshold: Optional[float] = None
2237
2238
2239
2240
    """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.
    """
2241
    posterior_alpha: Optional[float] = None
2242
2243
    """Scaling factor for entropy-based threshold, applied when using
    `TypicalAcceptanceSampler`."""
2244
2245
2246
2247

    # required configuration params passed from engine
    target_model_config: ModelConfig = field(default=None,
                                             init=True)  # type: ignore
2248
    """The configuration of the target model."""
2249
2250
    target_parallel_config: ParallelConfig = field(default=None,
                                                   init=True)  # type: ignore
2251
    """The parallel configuration for the target model."""
2252
2253
    enable_chunked_prefill: bool = field(default=None,
                                         init=True)  # type: ignore
2254
2255
    """Whether vLLM is configured to use chunked prefill or not. Used for
    raising an error since it's not yet compatible with speculative decode."""
2256
    disable_log_stats: bool = field(default=None, init=True)  # type: ignore
2257
2258
    """Whether to disable the periodic printing of stage times in speculative
    decoding."""
2259
2260
2261
2262

    # params generated in the post-init stage
    draft_model_config: ModelConfig = field(default=None,
                                            init=True)  # type: ignore
2263
    """The configuration of the draft model initialized internal."""
2264
2265
    draft_parallel_config: ParallelConfig = field(default=None,
                                                  init=True)  # type: ignore
2266
    """The parallel configuration for the draft model initialized internal."""
2267

2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
    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.
        """
2280
        factors: list[Any] = []
2281
2282
2283
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
2284
2285
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2286
2287
        return hash_str

2288
2289
2290
2291
2292
    @classmethod
    def from_dict(cls, dict_value: dict) -> "SpeculativeConfig":
        """Parse the CLI value for the speculative config."""
        return cls(**dict_value)

2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
    @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

2305
    def __post_init__(self):
2306

2307
2308
2309
2310
2311
2312
2313
        # 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.
2314
2315
2316
2317

        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
2318
2319
            if self.target_model_config and \
                self.target_model_config.hf_text_config.model_type \
2320
                        == "deepseek_v3":
2321
2322
2323
2324
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2325
            else:
2326
2327
2328
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

2329
2330
        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
2331
2332
2333
2334
2335
2336
2337
        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"
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
            # 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
2352
            if self.prompt_lookup_min < 1:
2353
2354
2355
2356
2357
                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")
2358
            if self.prompt_lookup_min > self.prompt_lookup_max:
2359
2360
2361
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2362

2363
2364
2365
            # 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.
2366
2367
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2368
        else:
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
            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,
                    max_model_len=None,
                    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,
                )
2398

2399
                # Automatically detect the method
2400
                if self.method in ('eagle', 'eagle3'):
2401
                    pass
2402
2403
                elif "eagle-" in self.draft_model_config.model.lower() or \
                        "eagle3-" in self.draft_model_config.model.lower():
2404
2405
2406
2407
2408
2409
                    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"
2410
                else:
2411
2412
2413
                    self.method = "draft_model"

                # Replace hf_config for EAGLE draft_model
2414
                if self.method in ("eagle", "eagle3"):
2415
                    if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
2416
                        raise ValueError(
2417
2418
                            "Chunked prefill and EAGLE are not compatible "
                            "when using V0.")
2419
2420
2421
2422
2423
2424
2425
2426

                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
                    if isinstance(self.draft_model_config.hf_config,
                                  EAGLEConfig):
                        pass
                    else:
                        eagle_config = EAGLEConfig(
2427
2428
                            self.draft_model_config.hf_config,
                            method=self.method)
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
                        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
                )
2456

2457
2458
2459
2460
2461
2462
                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,
                    ))
2463

2464
2465
2466
2467
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2468

2469
2470
2471
2472
2473
        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
2474

2475
        self._verify_args()
2476

2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
    @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,
        )

2512
    @staticmethod
2513
    def _verify_and_get_draft_tp(
2514
2515
2516
2517
2518
2519
            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.
2520
        """
2521
2522
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2523
        if speculative_draft_tensor_parallel_size is None:
2524
2525
2526
2527
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2528
2529
2530
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2531
2532
2533
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2534
2535
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2536
            raise ValueError(
2537
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2538
                f"other value than 1 or target model tensor_parallel_size")
2539
        return speculative_draft_tensor_parallel_size
2540

2541
2542
2543
2544
2545
2546
2547
2548
2549
    @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.
        """
2550
2551
2552
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2553
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2554
2555
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
            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:
2568
2569
2570
2571
2572
2573
        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.")

2574
2575
2576
2577
2578
2579
2580
        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)
2581
2582
            # Validate and set draft token acceptance related settings.

2583
2584
        if self.acceptance_method is None:
            raise ValueError("acceptance_method is not set. "
2585
2586
2587
                             "Expected values are rejection_sampler or "
                             "typical_acceptance_sampler.")

2588
2589
        if (self.acceptance_method != 'rejection_sampler'
                and self.acceptance_method != 'typical_acceptance_sampler'):
2590
            raise ValueError(
2591
                "Expected acceptance_method to be either "
2592
                "rejection_sampler or typical_acceptance_sampler. Instead it "
2593
                f"is {self.acceptance_method}")
2594

2595
2596
2597
2598
        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)):
2599
            raise ValueError(
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
                "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=}")
2611

2612
2613
2614
2615
2616
2617
        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=}")

2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
    @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

2628
2629
2630
    def use_eagle(self) -> bool:
        return self.method in ("eagle", "eagle3")

2631
    def __repr__(self) -> str:
2632
2633
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
2634
        num_spec_tokens = self.num_speculative_tokens
2635
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"
2636
2637


2638
2639
2640
2641
LoRADType = Literal["auto", "float16", "bfloat16"]


@config
2642
2643
@dataclass
class LoRAConfig:
2644
2645
2646
2647
2648
2649
    """Configuration for LoRA."""

    max_lora_rank: int = 16
    """Max LoRA rank."""
    max_loras: int = 1
    """Max number of LoRAs in a single batch."""
2650
    fully_sharded_loras: bool = False
2651
2652
2653
2654
    """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.
    """
2655
    max_cpu_loras: Optional[int] = None
2656
2657
2658
2659
    """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."""
2660
    lora_extra_vocab_size: int = 256
2661
2662
    """Maximum size of extra vocabulary that can be present in a LoRA adapter
    (added to the base model vocabulary)."""
2663
2664
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
2665
2666
2667
2668
2669
2670
    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."""
2671
    bias_enabled: bool = False
2672
    """Enable bias for LoRA adapters."""
2673

2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
    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.
        """
2686
        factors: list[Any] = []
2687
2688
2689
2690
2691
2692
2693
        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)
2694
2695
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2696
2697
        return hash_str

2698
    def __post_init__(self):
2699
        # Setting the maximum rank to 512 should be able to satisfy the vast
2700
        # majority of applications.
2701
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
2702
        possible_lora_extra_vocab_size = (256, 512)
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
        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
2718
                f"max_loras ({self.max_loras})")
2719

2720
    def verify_with_cache_config(self, cache_config: CacheConfig):
2721
2722
2723
        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.")
2724

2725
2726
2727
2728
2729
2730
    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)

2731
2732
2733
2734
2735
    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.")

2736

2737
@config
2738
2739
@dataclass
class PromptAdapterConfig:
2740
2741
    """Configuration for PromptAdapters."""

2742
2743
2744
2745
    max_prompt_adapters: int = 1
    """Max number of PromptAdapters in a batch."""
    max_prompt_adapter_token: int = 0
    """Max number of PromptAdapters tokens."""
2746
    max_cpu_prompt_adapters: Optional[int] = None
2747
2748
2749
2750
2751
    """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.
    """
2752

2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
    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.
2767
        factors: list[Any] = []
2768
2769
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2770
2771
        return hash_str

2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
    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):
2783
        if self.prompt_adapter_dtype == "auto":
2784
2785
2786
2787
2788
2789
            self.prompt_adapter_dtype = model_config.dtype
        elif isinstance(self.prompt_adapter_dtype, str):
            self.prompt_adapter_dtype = getattr(torch,
                                                self.prompt_adapter_dtype)


2790
@config
2791
@dataclass
2792
class MultiModalConfig:
2793
2794
    """Controls the behavior of multimodal models."""

2795
    limit_per_prompt: dict[str, int] = field(default_factory=dict)
2796
    """
2797
    The maximum number of input items allowed per prompt for each modality.
2798
2799
    This should be a JSON string that will be parsed into a dictionary.
    Defaults to 1 (V0) or 999 (V1) for each modality.
2800
2801

    For example, to allow up to 16 images and 2 videos per prompt:
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
    :code:`{"images": 16, "videos": 2}`
    """

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

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

    For example, for Phi-3-Vision:
    :code:`{"num_crops": 4}`.
    """

    disable_mm_preprocessor_cache: bool = False
    """
    If :code:`True`, disable caching of the processed multi-modal inputs.
2819
2820
    """

2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
    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.
2835
        factors: list[Any] = []
2836
2837
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2838
2839
        return hash_str

2840
2841
2842
2843
2844
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.
        """
2845
2846
2847
2848
        return self.limit_per_prompt.get(
            modality,
            999 if envs.VLLM_USE_V1 else 1,
        )
2849

2850
    # TODO: Add configs to init vision tower or not.
2851

2852

2853
@config
2854
2855
@dataclass
class PoolerConfig:
2856
    """Controls the behavior of output pooling in pooling models."""
2857
2858

    pooling_type: Optional[str] = None
2859
    """
2860
    The pooling method of the pooling model. This should be a key in
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
    :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
    """
2878
    If set, only the score corresponding to the ``step_tag_id`` in the
2879
2880
2881
2882
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

2883
    returned_token_ids: Optional[list[int]] = None
2884
    """
2885
2886
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2887
2888
2889
    ``math-shepherd-mistral-7b-prm`` model.
    """

2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
    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.
2904
        factors: list[Any] = []
2905
2906
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
2907
2908
        return hash_str

2909
2910
2911
    @staticmethod
    def from_json(json_str: str) -> "PoolerConfig":
        return PoolerConfig(**json.loads(json_str))
2912
2913


2914
2915
2916
2917
2918
2919
2920
2921
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2922
_ROCM_NOT_SUPPORTED_DTYPE: list[str] = []  #
2923

2924
2925
2926

def _get_and_verify_dtype(
    config: PretrainedConfig,
2927
    dtype: Union[str, torch.dtype],
2928
2929
2930
) -> torch.dtype:
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
2931
    config_dtype = getattr(config.get_text_config(), "torch_dtype", None)
2932

2933
    # Fallback for multi-modal models if the root config
2934
2935
2936
2937
    # does not define torch_dtype
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)

2938
2939
2940
    if config_dtype is None:
        config_dtype = torch.float32

2941
2942
2943
2944
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
2945
2946
                # Following common practice, we use float16 for float32 models
                torch_dtype = torch.float16
2947
2948
            else:
                torch_dtype = config_dtype
2949

Shinichi Hemmi's avatar
Shinichi Hemmi committed
2950
2951
2952
2953
2954
2955
2956
            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

2957
            from vllm.platforms import current_platform
2958
2959
            if (current_platform.is_cpu()
                    and current_platform.get_cpu_architecture()
2960
                    == CpuArchEnum.POWERPC
2961
2962
2963
2964
2965
2966
2967
2968
                    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

2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
            # 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

2980
2981
            if current_platform.is_hpu() and config_dtype == torch.float16:
                logger.info(
2982
                    "For HPU, we cast models to bfloat16 instead of "
2983
2984
2985
                    "using float16 by default. Please specify `dtype` if you "
                    "want to use float16.")
                torch_dtype = torch.bfloat16
Shinichi Hemmi's avatar
Shinichi Hemmi committed
2986
2987
2988
2989
2990
        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
2991
        else:
2992
2993
2994
2995
2996
            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
2997
    else:
2998
        raise ValueError(f"Unknown dtype: {dtype}")
2999
3000
3001
3002
3003

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
3004
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
3005
3006
3007
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
3008
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
3009
3010
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
3011
            # Casting between float16 and bfloat16 is allowed with a warning.
3012
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
3013
3014

    return torch_dtype
3015
3016
3017
3018
3019


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
3020
    disable_sliding_window: bool,
3021
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
3022
    spec_target_max_model_len: Optional[int] = None,
3023
    encoder_config: Optional[Any] = None,
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
) -> 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",
3034
3035
        # ChatGLM2
        "seq_length",
3036
3037
        # Command-R
        "model_max_length",
3038
3039
        # Whisper
        "max_target_positions",
3040
3041
3042
3043
3044
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
3045
    # Choose the smallest "max_length" from the possible keys.
3046
    max_len_key = None
3047
    for key in possible_keys:
3048
3049
3050
3051
3052
        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
3053
3054
3055
3056
    # 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
3057
3058
3059
3060

    # 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:
3061
3062

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
3063
        max_len_key = "sliding_window" \
3064
3065
3066
            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)
3067
3068
3069

    # If none of the keys were found in the config, use a default and
    # log a warning.
3070
    if derived_max_model_len == float("inf"):
3071
3072
3073
3074
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

3075
3076
3077
3078
3079
        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

3080
3081
3082
3083
        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: "
3084
            "%s. Assuming the model's maximum length is %d.", possible_keys,
3085
            default_max_len)
3086
        derived_max_model_len = default_max_len
3087

3088
    rope_scaling = getattr(hf_config, "rope_scaling", None)
3089
3090
3091
    # 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:
3092
3093
3094
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104

        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.")

3105
3106
3107
3108
            # 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)

3109
3110
3111
3112
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
3113

3114
3115
3116
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

3117
3118
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
3119
    if max_model_len is None:
3120
        max_model_len = int(derived_max_model_len)
3121
    elif max_model_len > derived_max_model_len:
3122
3123
3124
3125
3126
        # 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:
3127
3128
3129
3130
3131
3132
3133
            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.")
3134
        else:
3135
            msg = (
3136
                f"User-specified max_model_len ({max_model_len}) is greater "
3137
3138
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
3139
                f"{model_max_length} in model's config.json). This may lead "
3140
3141
3142
3143
3144
3145
3146
3147
3148
                "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")
3149
    return int(max_model_len)
3150
3151


3152
def get_min_sliding_window(
3153
        sliding_window: Union[int, list[Optional[int]]]) -> int:
3154
3155
3156
3157
3158
3159
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


3160
def get_served_model_name(model: str,
3161
                          served_model_name: Optional[Union[str, list[str]]]):
3162
    """
3163
3164
3165
3166
    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
3167
3168
3169
3170
3171
3172
3173
3174
3175
    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


3176
GuidedDecodingBackendV0 = Literal["auto", "outlines", "lm-format-enforcer",
3177
                                  "xgrammar", "guidance"]
3178
GuidedDecodingBackendV1 = Literal["auto", "xgrammar", "guidance"]
3179
3180
GuidedDecodingBackend = Literal[GuidedDecodingBackendV0,
                                GuidedDecodingBackendV1]
3181
3182
3183


@config
3184
3185
@dataclass
class DecodingConfig:
3186
    """Dataclass which contains the decoding strategy of the engine."""
3187

3188
3189
    guided_decoding_backend: GuidedDecodingBackend = \
        "auto" if envs.VLLM_USE_V1 else "xgrammar"
3190
3191
3192
3193
    """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."""
3194

3195
    reasoning_backend: Optional[str] = None
3196
3197
3198
    """Select the reasoning parser depending on the model that you're using.
    This is used to parse the reasoning content into OpenAI API format.
    Required for `--enable-reasoning`."""
3199

3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
    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.
3214
        factors: list[Any] = []
3215
3216
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3217
3218
        return hash_str

3219
    def __post_init__(self):
3220
3221
        backend = GuidedDecodingParams(
            backend=self.guided_decoding_backend).backend_name
3222
        if envs.VLLM_USE_V1:
3223
            valid_guided_backends = get_args(GuidedDecodingBackendV1)
3224
        else:
3225
            valid_guided_backends = get_args(GuidedDecodingBackendV0)
3226
        if backend not in valid_guided_backends:
3227
            raise ValueError(f"Invalid guided_decoding_backend '{backend}',"
3228
                             f" must be one of {valid_guided_backends}")
3229
3230


3231
3232
@dataclass
class ObservabilityConfig:
3233
3234
3235
    """Configuration for observability - metrics and tracing."""
    show_hidden_metrics: bool = False

3236
3237
    otlp_traces_endpoint: Optional[str] = None

3238
3239
3240
3241
3242
3243
3244
3245
    # Collecting detailed timing information for each request can be expensive.

    # If set, collects the model forward time for the request.
    collect_model_forward_time: bool = False

    # If set, collects the model execute time for the request.
    collect_model_execute_time: bool = False

3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
    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.
3260
        factors: list[Any] = []
3261
3262
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3263
3264
        return hash_str

3265
    def __post_init__(self):
3266
3267
3268
3269
3270
        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}")
3271
3272


3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
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

3306
3307
3308
    # any extra config that the connector may need
    kv_connector_extra_config: dict[str, Any] = {}

3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
    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.
3323
        factors: list[Any] = []
3324
3325
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()
3326
3327
        return hash_str

3328
3329
    @classmethod
    def from_cli(cls, cli_value: str) -> "KVTransferConfig":
youkaichao's avatar
youkaichao committed
3330
        """Parse the CLI value for the kv cache transfer config."""
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
        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"]

3363
3364
3365
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

3366

3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
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.
3385
        - debug_dump_path: the path to dump the debug information.
3386
3387
3388
        - 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.
3389
3390
3391
3392
3393
3394
3395
        - 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).
3396
3397
3398
3399
3400
3401
3402
3403
3404
        - 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).
3405
        - splitting_ops: a list of ops to split the full graph into subgraphs, used in piecewise compilation.
3406
3407
3408
3409
    - 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
3410
3411
3412
3413
                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.
3414
3415
3416
            TODO: move outside cudagraph logic into compilation.
            torch.compile will handle cudagraph capture logic in the future.
        - cudagraph_capture_sizes: sizes to capture cudagraph.
3417
            - None (default): capture sizes are inferred from vllm config.
3418
            - list[int]: capture sizes are specified as given.
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
        - 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
3432
3433
3434
3435
3436
                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.
3437
3438
3439
3440
3441
3442
3443
        - 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})`
3444
        - custom inductor passes: see PassConfig for more details
3445

3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
    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
3457
    debug_dump_path: str = ""
3458
    cache_dir: str = ""
3459
    backend: str = ""
3460
3461
    custom_ops: list[str] = Field(default_factory=list)
    splitting_ops: list[str] = Field(default=None)  # type: ignore
3462
3463

    use_inductor: bool = True
3464
3465
3466
    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)
3467
3468
3469

    use_cudagraph: bool = False
    cudagraph_num_of_warmups: int = 0
3470
    cudagraph_capture_sizes: Optional[list[int]] = None
3471
3472
    cudagraph_copy_inputs: bool = False

3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
    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.
3483
3484
        - enable_noop: whether to enable the custom no-op elimination pass.
            TODO(luka) better pass enabling system.
3485
        - enable_sequence_parallelism: whether to enable sequence parallelism.
3486
        """
3487
        dump_graph_stages: list[str] = Field(default_factory=list)
3488
3489
        dump_graph_dir: Path = Field(default=Path("."))
        enable_fusion: bool = True
3490
        enable_noop: bool = True
3491
        enable_sequence_parallelism: bool = False
3492
3493
3494
3495
3496
3497
3498
3499

        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.
            """
3500
3501
            dict_ = self.model_dump(include={"enable_fusion", "enable_noop", \
                "enable_sequence_parallelism"})
3502
            return InductorPass.hash_dict(dict_)
3503
3504

        def model_post_init(self, __context: Any) -> None:
3505
            if not self.enable_noop and self.enable_fusion:
3506
                logger.warning_once(
3507
                    "Fusion enabled but reshape elimination disabled. "
3508
3509
3510
                    "RMSNorm + quant (fp8) fusion might not work")

    pass_config: PassConfig = Field(default_factory=PassConfig)
3511
3512

    # not configurable, computed after init
3513
    max_capture_size: int = PrivateAttr
3514
    local_cache_dir: str = PrivateAttr  # local cache dir for each rank
3515
    # optimization:
3516
    # Intuitively, bs_to_padded_graph_size should be dict[int, int].
3517
    # since we know all keys are in a range [0, max_capture_size],
3518
3519
    # we can optimize it to list[int] for better lookup performance.
    bs_to_padded_graph_size: list[int] = PrivateAttr
3520

3521
3522
3523
    # keep track of enabled and disabled custom ops
    enabled_custom_ops: Counter[str] = PrivateAttr
    disabled_custom_ops: Counter[str] = PrivateAttr
3524
    traced_files: set[str] = PrivateAttr
3525
    compilation_time: float = PrivateAttr
3526

3527
    # Per-model forward context
3528
3529
    # Map from layer name to layer objects that need to be accessed outside
    # model code, e.g., Attention, FusedMOE when dp_size>1.
3530
    static_forward_context: dict[str, Any] = PrivateAttr
3531

3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
    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.
        """
3544
        factors: list[Any] = []
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
        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()

3555
3556
3557
3558
3559
3560
3561
3562
    def __repr__(self) -> str:
        exclude = {
            "static_forward_context",
            "enabled_custom_ops",
            "disabled_custom_ops",
            "compilation_time",
            "bs_to_padded_graph_size",
            "pass_config",
3563
            "traced_files",
3564
3565
3566
3567
3568
        }
        return self.model_dump_json(exclude=exclude, exclude_unset=True)

    __str__ = __repr__

3569
3570
3571
3572
3573
    @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))
3574
3575
3576
        # 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)
3577

3578
3579
3580
3581
3582
3583
    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
3584
3585
3586
3587
3588
3589
3590
3591
        # 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

3592
        if is_torch_equal_or_newer("2.6"):
Michael Goin's avatar
Michael Goin committed
3593
3594
3595
3596
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

3597
        if self.splitting_ops is None:
3598
            self.splitting_ops = []
3599

3600
3601
3602
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
3603
3604
3605
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
3606
3607
3608
3609
3610
3611
3612
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
3613
3614
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
3615

3616
3617
        self.enabled_custom_ops = Counter()
        self.disabled_custom_ops = Counter()
3618
        self.traced_files = set()
3619
        self.static_forward_context = {}
3620
        self.compilation_time = 0.0
3621

3622
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
        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
3640

3641
        from vllm.compilation.backends import VllmBackend
3642
        return VllmBackend(vllm_config)
3643

3644
    def init_with_cudagraph_sizes(self,
3645
                                  cudagraph_capture_sizes: list[int]) -> None:
3646
        """To complete the initialization of config,
3647
3648
        we need to know the cudagraph sizes."""

3649
        if self.cudagraph_capture_sizes is None:
3650
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
3651
        else:
3652
3653
3654
            # de-duplicate the sizes provided by the config
            self.cudagraph_capture_sizes = list(
                set(self.cudagraph_capture_sizes))
3655
3656
            logger.info(("cudagraph sizes specified by model runner"
                         " %s is overridden by config %s"),
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
                        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
3673

3674
        # sort to make sure cudagraph capture sizes are in descending order
3675
3676
3677
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
3678

3679
3680
3681
3682
        # 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)
        ]
3683
3684
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
3685
3686
3687
3688
3689
3690
3691
            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
3692

3693
3694
3695
3696
3697
3698
3699
3700
3701
    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",
            ]

3702

3703
3704
3705
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
3706
3707
3708
    simplifies passing around the distinct configurations in the codebase.
    """

3709
3710
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
3711
3712
3713
3714
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig,
                                            init=True)
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig,
                                              init=True)
3715
3716
3717
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
3718
    lora_config: Optional[LoRAConfig] = None
3719
3720
    speculative_config: SpeculativeConfig = field(default=None,
                                                  init=True)  # type: ignore
3721
3722
3723
    decoding_config: Optional[DecodingConfig] = None
    observability_config: Optional[ObservabilityConfig] = None
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
3724
    quant_config: Optional[QuantizationConfig] = None
3725
3726
    compilation_config: CompilationConfig = field(default=None,
                                                  init=True)  # type: ignore
3727
3728
    kv_transfer_config: KVTransferConfig = field(default=None,
                                                 init=True)  # type: ignore
3729
    # some opaque config, only used to provide additional information
3730
3731
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
3732
3733
    additional_config: SupportsHash = field(default=None,
                                            init=True)  # type: ignore
3734
    instance_id: str = ""
3735

3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
    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.
        """
3748
        factors: list[Any] = []
3749
3750

        # summarize vllm config
3751
        vllm_factors: list[Any] = []
3752
3753
        from vllm import __version__
        vllm_factors.append(__version__)
3754
        vllm_factors.append(envs.VLLM_USE_V1)
3755
3756
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3757
3758
        else:
            vllm_factors.append("None")
3759
3760
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3761
3762
        else:
            vllm_factors.append("None")
3763
3764
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3765
3766
        else:
            vllm_factors.append("None")
3767
3768
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3769
3770
        else:
            vllm_factors.append("None")
3771
3772
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3773
3774
        else:
            vllm_factors.append("None")
3775
3776
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3777
3778
        else:
            vllm_factors.append("None")
3779
3780
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3781
3782
3783
3784
3785
            # 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))
3786
3787
        else:
            vllm_factors.append("None")
3788
3789
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3790
3791
        else:
            vllm_factors.append("None")
3792
3793
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3794
3795
        else:
            vllm_factors.append("None")
3796
3797
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3798
3799
        else:
            vllm_factors.append("None")
3800
3801
        if self.prompt_adapter_config:
            vllm_factors.append(self.prompt_adapter_config.compute_hash())
3802
3803
        else:
            vllm_factors.append("None")
3804
3805
3806
3807
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3808
3809
        else:
            vllm_factors.append("None")
3810
3811
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3812
3813
3814
3815
3816
3817
        else:
            vllm_factors.append("None")
        if self.additional_config:
            vllm_factors.append(self.additional_config.compute_hash())
        else:
            vllm_factors.append("None")
3818
3819
        factors.append(vllm_factors)

3820
3821
        hash_str = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]
3822
3823
        return hash_str

3824
3825
3826
3827
3828
3829
    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]
3830

3831
3832
3833
3834
3835
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3836
        from vllm.platforms import current_platform
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
        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
3859

3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
    @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)

3871
3872
3873
3874
3875
3876
3877
3878
3879
    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

3880
3881
3882
3883
3884
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

3885
3886
3887
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
3888
3889
3890
3891
3892
3893
3894
3895
        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)
3896
3897

        if self.lora_config:
3898
            self.lora_config.verify_with_cache_config(self.cache_config)
3899
            self.lora_config.verify_with_model_config(self.model_config)
3900
            self.lora_config.verify_lora_support()
3901
3902
3903
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
3904
3905
3906
3907
3908

        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)
3909

3910
        from vllm.platforms import current_platform
3911
3912
3913
3914
3915
        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):
3916
            logger.warning_once(
3917
3918
3919
3920
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

3921
        if self.compilation_config is None:
3922
            self.compilation_config = CompilationConfig()
3923
3924
        if self.compilation_config.pass_config.enable_sequence_parallelism:
            self.compilation_config.custom_ops.append("+rms_norm")
3925
3926
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
3927
3928
3929
3930
            # 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.
3931
            # FIXME(rob): Add function to set all of these.
3932
3933
            if not self.compilation_config.custom_ops:
                self.compilation_config.custom_ops = ["none"]
3934
3935
            self.compilation_config.use_cudagraph = True
            self.compilation_config.use_inductor = True
3936
            self.compilation_config.cudagraph_num_of_warmups = 1
3937
            self.compilation_config.pass_config.enable_fusion = False
3938
            self.compilation_config.pass_config.enable_noop = False
3939
            self.compilation_config.level = CompilationLevel.PIECEWISE
3940
            self.compilation_config.set_splitting_ops_for_v1()
3941

3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
        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

3954
        self._set_cudagraph_sizes()
3955

3956
3957
        if self.cache_config is not None and \
            self.cache_config.cpu_offload_gb > 0 and \
3958
3959
            self.compilation_config.level != CompilationLevel.NO_COMPILATION \
                and not envs.VLLM_USE_V1:
3960
            logger.warning(
3961
                "CPU offload is not supported with `torch.compile` in v0 yet."
3962
3963
3964
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3965
3966
3967
3968
3969
3970
        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`.")
3971
3972
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3973

3974
        if self.model_config and self.model_config.use_mla and \
3975
            not (current_platform.is_cuda() or current_platform.is_rocm()):
3976
            logger.info(
3977
                "MLA is enabled on a non-GPU platform; forcing chunked "
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
                "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

3988
3989
        current_platform.check_and_update_config(self)

3990
3991
3992
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
    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
        ]

4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
    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.

4029
4030
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
4031
4032

        During runtime, if batchsize is larger than
4033
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4034
4035
        no cudagraph will be used.
        If the batch size is no larger than
4036
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
        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)]
4050
4051
4052
4053
4054
                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)

4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
                # 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:
                batch_size_capture_list = [1, 2, 4
                                           ] + [i for i in range(8, 513, 8)]
4078
4079
4080
4081
4082
                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)

4083
4084
4085
4086
4087
                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
                ]
4088
4089
4090
4091

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

4092
    def __str__(self):
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
        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}, "
4123
4124
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
4125
4126
4127
4128
4129
4130


_current_vllm_config: Optional[VllmConfig] = None


@contextmanager
4131
def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False):
4132
    """
4133
    Temporarily set the current vLLM config.
4134
    Used during model initialization.
4135
    We save the current vLLM config in a global variable,
4136
    so that all modules can access it, e.g. custom ops
4137
    can access the vLLM config to determine how to dispatch.
4138
4139
4140
4141
4142
4143
4144
4145
    """
    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
4146
4147
4148
    except Exception:
        raise
    else:
4149
4150
4151
4152
        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)
4153
4154
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
4155
4156
4157
4158
4159
4160
4161
4162
4163
            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"
4164
                " if you want it to be supported.",
4165
                vllm_config.model_config.model)
4166
    finally:
4167
4168
4169
4170
4171
4172
4173
4174
        _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.
4175
        logger.warning("Current vLLM config is not set.")
4176
4177
4178
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205


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}")
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218


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)
    }