config.py 158 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 json
8
import sys
9
import warnings
10
11
from collections import Counter
from collections.abc import Mapping
12
from contextlib import contextmanager
13
from dataclasses import dataclass, field, replace
14
from importlib.util import find_spec
15
from pathlib import Path
16
17
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Final, Literal,
                    Optional, Protocol, Union)
18
19

import torch
20
from pydantic import BaseModel, Field, PrivateAttr
21
from torch.distributed import ProcessGroup, ReduceOp
22
from transformers import PretrainedConfig
23

24
import vllm.envs as envs
25
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
26
from vllm.logger import init_logger
27
28
from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
                                                     get_quantization_config)
29
from vllm.model_executor.models import ModelRegistry
30
from vllm.platforms import CpuArchEnum
31
from vllm.sampling_params import GuidedDecodingParams
32
from vllm.tracing import is_otel_available, otel_import_error_traceback
33
34
35
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
36
37
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
    try_get_generation_config, uses_mrope)
38
from vllm.transformers_utils.s3_utils import S3Model
39
from vllm.transformers_utils.utils import is_s3
40
from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
41
                        get_cpu_memory, random_uuid, resolve_obj_by_qualname)
42

43
44
45
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

46
    from vllm.executor.executor_base import ExecutorBase
47
48
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
49
    from vllm.model_executor.model_loader.loader import BaseModelLoader
50
51
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
52
53
else:
    QuantizationConfig = None
54

Michael Goin's avatar
Michael Goin committed
55
56
from packaging.version import Version

57
58
logger = init_logger(__name__)

59
60
61
# 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
62
_POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
63
_MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
64

65
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
66
                     "score", "reward", "transcription"]
67

68
_ResolvedTask = Literal["generate", "embed", "classify", "score", "reward",
69
                        "draft", "transcription"]
70

71
RunnerType = Literal["generate", "pooling", "draft", "transcription"]
72

73
_RUNNER_TASKS: dict[RunnerType, list[_ResolvedTask]] = {
74
75
76
    "generate": ["generate"],
    "pooling": ["embed", "classify", "score", "reward"],
    "draft": ["draft"],
77
    "transcription": ["transcription"],
78
79
}

80
_TASK_RUNNER: dict[_ResolvedTask, RunnerType] = {
81
    task: runner
82
83
    for runner, tasks in _RUNNER_TASKS.items()
    for task in tasks
84
}
85

86
HfOverrides = Union[dict[str, Any], Callable[[PretrainedConfig],
87
88
                                             PretrainedConfig]]

89

90
91
92
93
94
95
class SupportsHash(Protocol):

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


96
97
class SupportsMetricsInfo(Protocol):

98
    def metrics_info(self) -> dict[str, str]:
99
100
101
        ...


102
103
104
105
106
107
class ModelImpl(str, enum.Enum):
    AUTO = "auto"
    VLLM = "vllm"
    TRANSFORMERS = "transformers"


108
class ModelConfig:
109
110
111
112
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
113
            It is also used as the content for `model_name` tag in metrics
114
115
116
117
118
            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.
119
        tokenizer: Name or path of the huggingface tokenizer to use.
120
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
121
122
123
            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.
124
125
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
126
127
128
129
        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.
130
131
132
133
        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
134
135
136
        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.
137
        code_revision: The specific revision to use for the model code on
138
            Hugging Face Hub. It can be a branch name, a tag name, or a
139
            commit id. If unspecified, will use the default version.
140
141
142
        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.
143
144
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
145
146
        spec_target_max_model_len: Specify the the maximum length for spec
            decoding draft models.
147
148
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
149
150
151
        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.
152
            If None, the user did not specify, so default to False.
153
154
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
155
156
157
            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.
158
        max_logprobs: Maximum number of log probabilities. Defaults to 20.
159
160
161
162
        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.
163
164
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
165
        served_model_name: The model name used in metrics tag `model_name`,
166
167
            matches the model name exposed via the APIs. If multiple model
            names provided, the first name will be used. If not specified,
168
            the model name will be the same as `model`.
169
        limit_mm_per_prompt: Maximum number of data items per modality
170
            per prompt. Only applicable for multimodal models.
171
172
        use_async_output_proc: Whether to use async output processor.
            Defaults to True.
173
174
        config_format: The config format which shall be loaded.
            Defaults to 'auto' which defaults to 'hf'.
175
176
177
        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.
178
179
        mm_processor_kwargs: Arguments to be forwarded to the model's processor
            for multi-modal data, e.g., image processor.
180
181
        disable_mm_preprocessor_cache: If true, then disables caching of the
            multi-modal preprocessor/mapper. (not recommended)
182
183
184
185
        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.
186
        override_pooler_config: Initialize non default pooling config or
187
            override default pooling config for the pooling model.
188
189
        logits_processor_pattern: Optional regex pattern specifying valid
            logits processor qualified names that can be passed with the
190
            `logits_processors` extra completion argument. Defaults to None,
191
            which allows no processors.
192
        generation_config: Configuration parameter file for generation.
193
194
195
196
197
198
        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.
199
200
        override_generation_config: Override the generation config with the
            given config.
201
    """
202

203
204
205
206
207
208
209
210
211
212
213
214
    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.
        """
215
        factors: list[Any] = []
216
217
218
219
220
221
222
223
224
225
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
        factors.append(self.trust_remote_code)
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
        return hashlib.sha256(str(factors).encode()).hexdigest()

226
227
228
229
230
231
232
233
234
    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,
235
        hf_config_path: Optional[str] = None,
236
237
238
        allowed_local_media_path: str = "",
        revision: Optional[str] = None,
        code_revision: Optional[str] = None,
239
        rope_scaling: Optional[dict[str, Any]] = None,
240
241
242
243
244
245
246
247
248
249
        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,
        skip_tokenizer_init: bool = False,
250
        served_model_name: Optional[Union[str, list[str]]] = None,
251
252
253
254
        limit_mm_per_prompt: Optional[Mapping[str, int]] = None,
        use_async_output_proc: bool = True,
        config_format: ConfigFormat = ConfigFormat.AUTO,
        hf_overrides: Optional[HfOverrides] = None,
255
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
256
        disable_mm_preprocessor_cache: bool = False,
257
        override_neuron_config: Optional[dict[str, Any]] = None,
258
259
        override_pooler_config: Optional["PoolerConfig"] = None,
        logits_processor_pattern: Optional[str] = None,
260
        generation_config: str = "auto",
261
        enable_sleep_mode: bool = False,
262
        override_generation_config: Optional[dict[str, Any]] = None,
263
        model_impl: Union[str, ModelImpl] = ModelImpl.AUTO,
264
    ) -> None:
265
        self.model = model
266
        self.hf_config_path = hf_config_path
267
        self.tokenizer = tokenizer
268
        self.tokenizer_mode = tokenizer_mode
269
        self.trust_remote_code = trust_remote_code
270
        self.allowed_local_media_path = allowed_local_media_path
271
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
272
        self.revision = revision
273
        self.code_revision = code_revision
274
275
        self.rope_scaling = rope_scaling
        self.rope_theta = rope_theta
276
        self.model_impl = model_impl
277
278
279

        if hf_overrides is None:
            hf_overrides = {}
280
281
282
283
284
285

        if callable(hf_overrides):
            hf_overrides_kw = {}
            hf_overrides_fn = hf_overrides
        else:
            hf_overrides_kw = hf_overrides
286
            hf_overrides_fn = None
287

288
        if rope_scaling is not None:
289
            hf_override: dict[str, Any] = {"rope_scaling": rope_scaling}
290
            hf_overrides_kw.update(hf_override)
291
292
293
294
            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}'`")
295
296
297
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
        if rope_theta is not None:
            hf_override = {"rope_theta": rope_theta}
298
            hf_overrides_kw.update(hf_override)
299
300
301
302
            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}'`")
303
304
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

305
306
        self.maybe_pull_model_tokenizer_for_s3(model, tokenizer)

307
308
309
310
311
312
313
314
        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 "
                "module was not found."
                "See https://github.com/vllm-project/vllm/blob/main/Dockerfile"
                "for instructions on how to install it.")

315
316
317
318
319
        # 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
320
        self.quantization = quantization
321
        self.enforce_eager = enforce_eager
322
        self.max_seq_len_to_capture = max_seq_len_to_capture
323
        self.max_logprobs = max_logprobs
324
        self.disable_sliding_window = disable_sliding_window
325
        self.skip_tokenizer_init = skip_tokenizer_init
326
327
328
329
330
331
        self.enable_sleep_mode = enable_sleep_mode

        from vllm.platforms import current_platform

        if self.enable_sleep_mode and not current_platform.is_cuda():
            raise ValueError("Sleep mode is only supported on CUDA devices.")
332

333
334
335
        hf_config = get_config(self.hf_config_path or self.model,
                               trust_remote_code, revision, code_revision,
                               config_format)
336
337
338
339
340
341
342
343

        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)

344
345
        self.hf_config = hf_config

346
        self.hf_text_config = get_hf_text_config(self.hf_config)
347
        self.encoder_config = self._get_encoder_config()
348
349
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, revision)
350
        self.dtype = _get_and_verify_dtype(self.hf_config, dtype)
351
        self.use_async_output_proc = use_async_output_proc
352
        self.mm_processor_kwargs = mm_processor_kwargs
353
        self.disable_mm_preprocessor_cache = disable_mm_preprocessor_cache
Woosuk Kwon's avatar
Woosuk Kwon committed
354

355
356
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
357
358
            self.enforce_eager = False

359
        interleaved_attn_models = ["gemma2", "gemma3_text", "cohere2"]
360
361
362
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
        has_interleaved_attention = (sliding_window is not None) and (
            isinstance(sliding_window, list) or
363
            (self.hf_text_config.model_type in interleaved_attn_models))
364
365

        if (not self.disable_sliding_window and has_interleaved_attention):
366
367
            if (backend :=
                    envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"):
368
369
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
370

371
                logger.warning_once(
372
373
                    f"{self.hf_text_config.model_type} has interleaved "
                    "attention, which is currently not supported by the "
374
                    f"{backend} backend. Disabling sliding window and capping "
375
376
377
378
379
380
381
382
383
384
385
386
                    "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
387

388
389
390
391
        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,
392
            sliding_window_len=self.get_hf_config_sliding_window(),
393
394
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
395
396
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
397
398
        self.multimodal_config = self._init_multimodal_config(
            limit_mm_per_prompt)
399
400
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
401

402
        self.is_attention_free = self._init_attention_free()
403
        self.is_hybrid = self._init_is_hybrid()
404
405
        self.has_inner_state = self._init_has_inner_state()

406
407
408
409
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
410

411
        supported_tasks, task = self._resolve_task(task)
412
413
        self.supported_tasks = supported_tasks
        self.task: Final = task
414
415
416
417
        if self.task in ("draft", "generate"):
            self.truncation_side = "left"
        else:
            self.truncation_side = "right"
418

419
        self.pooler_config = self._init_pooler_config(override_pooler_config)
420
        self.logits_processor_pattern = logits_processor_pattern
421

422
        self.generation_config = generation_config
423
        self.override_generation_config = override_generation_config or {}
424

425
        self._verify_quantization()
426
        self._verify_cuda_graph()
427
        self._verify_bnb_config()
428

429
430
431
432
433
434
435
436
    @property
    def registry(self):
        return ModelRegistry

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

437
438
439
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
        """
440
        Pull the model config or tokenizer to a temporary
441
442
443
444
445
446
447
448
449
        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):
450
                s3_model = S3Model()
451
452
                s3_model.pull_files(
                    model, allow_pattern=["*.model", "*.py", "*.json"])
453
                self.model_weights = self.model
454
                self.model = s3_model.dir
455
456

            if is_s3(tokenizer):
457
458
                s3_tokenizer = S3Model()
                s3_tokenizer.pull_files(
459
                    model, ignore_pattern=["*.pt", "*.safetensors", "*.bin"])
460
                self.tokenizer = s3_tokenizer.dir
461

462
463
464
    def _init_multimodal_config(
        self, limit_mm_per_prompt: Optional[Mapping[str, int]]
    ) -> Optional["MultiModalConfig"]:
465
        if self.registry.is_multimodal_model(self.architectures):
466
            return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
467
468
469
470
471
472

        if limit_mm_per_prompt:
            raise ValueError("`limit_mm_per_prompt` is only supported for "
                             "multimodal models.")

        return None
473

474
475
476
477
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

478
479
    def _init_pooler_config(
        self,
480
        override_pooler_config: Optional["PoolerConfig"],
481
    ) -> Optional["PoolerConfig"]:
482

483
        if self.runner_type == "pooling":
484
485
486
487
488
489
490
491
492
493
494
            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)

            return user_config

495
496
        return None

497
    def _init_attention_free(self) -> bool:
498
        return self.registry.is_attention_free_model(self.architectures)
499

500
    def _init_is_hybrid(self) -> bool:
501
        return self.registry.is_hybrid_model(self.architectures)
502

503
    def _init_has_inner_state(self) -> bool:
504
        return self.registry.model_has_inner_state(self.architectures)
505

506
507
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
508
        if tokenizer_mode not in ["auto", "slow", "mistral", "custom"]:
509
510
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
511
                "either 'auto', 'slow', 'mistral' or 'custom'.")
512
        self.tokenizer_mode = tokenizer_mode
513

514
515
    def _get_preferred_task(
        self,
516
517
        architectures: list[str],
        supported_tasks: set[_ResolvedTask],
518
519
520
521
    ) -> Optional[_ResolvedTask]:
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
522
        if self.registry.is_cross_encoder_model(architectures):
523
            return "score"
524
        if self.registry.is_transcription_model(architectures):
525
            return "transcription"
526

527
        suffix_to_preferred_task: list[tuple[str, _ResolvedTask]] = [
528
529
530
531
532
533
534
535
536
            # Other models follow this pattern
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ForSequenceClassification", "classify"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]
537
        _, arch = self.registry.inspect_model_cls(architectures)
538
539
540
541
542
543
544

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

        return None

545
546
    def _resolve_task(
        self,
547
        task_option: Union[TaskOption, Literal["draft"]],
548
    ) -> tuple[set[_ResolvedTask], _ResolvedTask]:
549
550
551
        if task_option == "draft":
            return {"draft"}, "draft"

552
553
        registry = self.registry
        architectures = self.architectures
554

555
        runner_support: dict[RunnerType, bool] = {
556
557
            # NOTE: Listed from highest to lowest priority,
            # in case the model supports multiple of them
558
559
560
            "transcription": registry.is_transcription_model(architectures),
            "generate": registry.is_text_generation_model(architectures),
            "pooling": registry.is_pooling_model(architectures),
561
        }
562
        supported_runner_types_lst: list[RunnerType] = [
563
564
565
566
567
            runner_type
            for runner_type, is_supported in runner_support.items()
            if is_supported
        ]

568
        supported_tasks_lst: list[_ResolvedTask] = [
569
570
            task for runner_type in supported_runner_types_lst
            for task in _RUNNER_TASKS[runner_type]
571
572
573
574
575
        ]
        supported_tasks = set(supported_tasks_lst)

        if task_option == "auto":
            selected_task = next(iter(supported_tasks_lst))
576

577
578
579
580
581
            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
582

583
584
585
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
586
        else:
587
588
589
590
591
592
593
594
595
596
597
598
599
600
            # 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"

601
602
603
604
605
606
607
            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
608

609
        return supported_tasks, selected_task
610

611
612
613
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
614
            # compressed-tensors uses a "compression_config" key
615
            quant_cfg = getattr(self.hf_config, "compression_config", None)
616
617
        return quant_cfg

618
    def _verify_quantization(self) -> None:
619
        supported_quantization = QUANTIZATION_METHODS
620
        optimized_quantization_methods = [
621
622
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
            "awq_marlin", "fbgemm_fp8", "compressed_tensors",
623
            "compressed-tensors", "experts_int8", "quark", "nvfp4"
624
        ]
625
626
627
628
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
629
630
        quant_cfg = self._parse_quant_hf_config()

631
632
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
633
634

            # Detect which checkpoint is it
635
636
            for name in QUANTIZATION_METHODS:
                method = get_quantization_config(name)
637
638
639
640
641
642
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
643

644
            # Verify quantization configurations.
645
            if self.quantization is None:
646
647
                self.quantization = quant_method
            elif self.quantization != quant_method:
648
649
                raise ValueError(
                    "Quantization method specified in the model config "
650
                    f"({quant_method}) does not match the quantization "
651
652
653
654
655
656
657
658
                    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}.")
659
            from vllm.platforms import current_platform
660
            current_platform.verify_quantization(self.quantization)
661
            if self.quantization not in optimized_quantization_methods:
662
                logger.warning(
663
                    "%s quantization is not fully "
664
                    "optimized yet. The speed can be slower than "
665
                    "non-quantized models.", self.quantization)
666

667
    def _verify_cuda_graph(self) -> None:
668
669
670
671
        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)
672

673
        MODEL_NOT_SUPPORT_CUDA_GRAPH = ['mllama']
674
        if (self.hf_config.model_type in MODEL_NOT_SUPPORT_CUDA_GRAPH
Simon Mo's avatar
Simon Mo committed
675
                and not self.enforce_eager):
676
677
678
            logger.warning(
                "CUDA graph is not supported for %s yet, fallback to the eager "
                "mode.", self.hf_config.model_type)
Simon Mo's avatar
Simon Mo committed
679
680
            self.enforce_eager = True

681
682
    def _verify_bnb_config(self) -> None:
        """
683
        The current version of bitsandbytes (0.44.0) with 8-bit models does not
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
        yet support CUDA graph.
        """
        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(
                "CUDA graph is not supported on BitAndBytes 8bit yet, "
                "fallback to the eager mode.")
            self.enforce_eager = True

703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
    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.")

720
721
722
723
724
725
726
727
728
729
    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

730
        # Reminder: Please update docs/source/features/compatibility_matrix.md
731
        # If the feature combo become valid
732
        from vllm.platforms import current_platform
733
        if not current_platform.is_async_output_supported(self.enforce_eager):
734
735
736
737
738
739
740
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

741
        # Async postprocessor is not necessary for pooling models
742
        # since there is no token generation
743
        if self.runner_type == "pooling":
744
745
            self.use_async_output_proc = False

746
        # Reminder: Please update docs/source/features/compatibility_matrix.md
747
        # If the feature combo become valid
748
749
750
        if speculative_config:
            self.use_async_output_proc = False

751
752
753
754
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
755
756
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
757
758
759
760
761
762
763
        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}).")

764
        if parallel_config.enable_expert_parallel:
765
766
            self._verify_with_expert_parallelism()

767
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
768
        if pipeline_parallel_size > 1:
769
            if not self.registry.is_pp_supported_model(self.architectures):
770
771
772
773
774
775
                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
776

777
    def get_hf_config_sliding_window(
778
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
779
        """Get the sliding window size, or None if disabled."""
780
781
782
783

        # 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.
784
785
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
786
            return None
787
        return getattr(self.hf_text_config, "sliding_window", None)
788

789
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
790
791
792
793
794
795
796
797
        """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()

798
    def get_vocab_size(self) -> int:
799
        return self.hf_text_config.vocab_size
800

801
    def get_hidden_size(self) -> int:
802
        return self.hf_text_config.hidden_size
803

804
805
    @property
    def is_deepseek_mla(self) -> bool:
806
807
        return (hasattr(self.hf_text_config, "model_type")) \
                and (self.hf_text_config.model_type in \
808
                    ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp'))\
809
                and (self.hf_text_config.kv_lora_rank is not None)
810

811
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
812
        # TODO remove hard code
813
        if self.is_deepseek_mla:
814
815
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
816
            if self.use_mla:
817
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
818
819
820
821
822
            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
823

824
825
826
827
828
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

829
830
831
        if self.is_attention_free:
            return 0

832
833
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
834
        # FIXME(woosuk): This may not be true for all models.
835
836
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
837

838
839
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
840
        # For GPTBigCode & Falcon:
841
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
842
843
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
844
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
845
        new_decoder_arch_falcon = (
846
            self.hf_config.model_type in falcon_model_types
847
            and getattr(self.hf_config, "new_decoder_architecture", False))
848
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
849
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
850
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
851
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
852
            return 1
853

854
        # For DBRX and MPT
855
856
857
858
859
        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":
860
861
862
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

863
864
865
        if self.is_attention_free:
            return 0

866
867
868
869
870
871
872
873
874
875
        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:
876
            num_kv_heads = getattr(self.hf_text_config, attr, None)
877
878
879
880
881
            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.
882
        return self.hf_text_config.num_attention_heads
883
884
885

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
886
887
888
889
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

890
891
892
893
894
895
896
        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)
897

898
899
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
900
901
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
902

903
    def get_layers_start_end_indices(
904
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
905
        from vllm.distributed.utils import get_pp_indices
906
907
908
909
910
911
        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)
912
913
914
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
915
916
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
917
        return start, end
Mor Zusman's avatar
Mor Zusman committed
918

919
920
921
    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
922

923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
    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
        is_transformer = not self.is_hybrid and not self.is_attention_free
        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
        else:
            # Hybrid model
            layers_block_type_value = getattr(self.hf_config,
                                              "layers_block_type", None)
            if layers_block_type_value is None:
947
948
                raise ValueError("The model is an hybrid without a "
                                 "layers_block_type in the hf_config, "
949
950
951
                                 "cannot determine the num of "
                                 f"{block_type.value} layers")

952
953
954
955
956
957
958
959
960
            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)

961
962
            return sum(t == block_type.value
                       for t in layers_block_type_value[start:end])
Mor Zusman's avatar
Mor Zusman committed
963

964
965
966
967
968
969
970
971
972
973
974
975
    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

976
    def try_get_generation_config(self) -> dict[str, Any]:
977
        if self.generation_config in ("auto", "vllm"):
978
            config = try_get_generation_config(
979
                self.hf_config_path or self.model,
980
981
982
983
984
985
986
987
988
989
990
991
992
993
                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()

994
    def get_diff_sampling_param(self) -> dict[str, Any]:
995
        """
996
        This method returns a dictionary containing the parameters
997
998
        that differ from the default sampling parameters. If
        `generation_config` is `"vllm"`, an empty dictionary is returned.
999
1000

        Returns:
1001
            dict[str, Any]: A dictionary with the differing sampling
1002
            parameters, if `generation_config` is `"vllm"` an empty dictionary.
1003
        """
1004
        if self.generation_config == "vllm":
1005
1006
1007
1008
1009
1010
1011
            config = {}
        else:
            config = self.try_get_generation_config()

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

1012
1013
1014
1015
1016
1017
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1018
            "max_new_tokens",
1019
1020
1021
1022
1023
1024
        ]
        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
            }
1025
1026
1027
1028
1029
            # 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")
1030
1031
1032
1033
        else:
            diff_sampling_param = {}
        return diff_sampling_param

1034
    @property
1035
    def is_encoder_decoder(self) -> bool:
1036
        """Extract the HF encoder/decoder model flag."""
1037
1038
1039
1040
1041
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1042

1043
1044
1045
1046
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1047
1048
    @property
    def is_cross_encoder(self) -> bool:
1049
        return self.registry.is_cross_encoder_model(self.architectures)
1050

1051
1052
    @property
    def use_mla(self) -> bool:
1053
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1054

1055
    @property
1056
    def supported_runner_types(self) -> set[RunnerType]:
1057
1058
1059
1060
1061
1062
        return {_TASK_RUNNER[task] for task in self.supported_tasks}

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

1063
1064
1065
1066
1067
    @property
    def is_v1_compatible(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_v1_compatible(architectures)

1068
1069

class CacheConfig:
1070
1071
1072
1073
1074
    """Configuration for the KV cache.

    Args:
        block_size: Size of a cache block in number of tokens.
        gpu_memory_utilization: Fraction of GPU memory to use for the
Woosuk Kwon's avatar
Woosuk Kwon committed
1075
            vLLM execution.
1076
        swap_space: Size of the CPU swap space per GPU (in GiB).
1077
        cache_dtype: Data type for kv cache storage.
1078
        is_attention_free: Whether the model is attention-free.
1079
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
1080
            profiled num_gpu_blocks if specified. Does nothing if None.
1081
1082
1083
1084
        sliding_window: Sliding window size for the KV cache. Can not work with
            prefix caching enabled.
        enable_prefix_caching: Whether to enable prefix caching.
        cpu_offload_gb: Size of the CPU offload buffer in GiB.
1085
    """
1086

1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
    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.
        """
1099
        factors: list[Any] = []
1100
1101
1102
1103
1104
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1105
1106
1107
1108
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
1109
        swap_space: float,
1110
        cache_dtype: str,
1111
        is_attention_free: bool = False,
1112
        num_gpu_blocks_override: Optional[int] = None,
1113
        sliding_window: Optional[int] = None,
1114
        enable_prefix_caching: bool = False,
1115
        cpu_offload_gb: float = 0,
1116
        calculate_kv_scales: Optional[bool] = None,
1117
1118
1119
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
1120
        self.swap_space_bytes = swap_space * GiB_bytes
1121
        self.num_gpu_blocks_override = num_gpu_blocks_override
1122
        self.cache_dtype = cache_dtype
1123
        self.is_attention_free = is_attention_free
1124
        self.sliding_window = sliding_window
1125
        self.enable_prefix_caching = enable_prefix_caching
1126
        self.cpu_offload_gb = cpu_offload_gb
1127
        self.calculate_kv_scales = calculate_kv_scales
1128
        self._verify_args()
1129
        self._verify_cache_dtype()
1130
        self._verify_prefix_caching()
1131
1132

        # Will be set after profiling.
1133
1134
        self.num_gpu_blocks: Optional[int] = None
        self.num_cpu_blocks: Optional[int] = None
1135

1136
1137
1138
1139
        # Set calculate_kv_scales to False if the value is unset.
        if self.calculate_kv_scales is None:
            self.calculate_kv_scales = False

1140
    def metrics_info(self):
1141
1142
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1143
1144
        return {key: str(value) for key, value in self.__dict__.items()}

1145
    def _verify_args(self) -> None:
1146
1147
1148
1149
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1150
1151
1152
1153
1154
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1155
1156
1157
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1158
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
1159
1160
1161
1162
            if envs.VLLM_USE_V1:
                raise NotImplementedError(
                    "V1 does not yet support fp8 KV cache. "
                    "Set VLLM_USE_V1=0 to enable fp8 kv cache.")
1163
            logger.info(
1164
1165
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1166
1167
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
1168
1169
1170
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1171
1172
1173
1174
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

1175
        if self.sliding_window is not None and not envs.VLLM_USE_V1:
1176
1177
1178
1179
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
    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

1190
1191
1192
        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.")
1193
1194
1195
        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:
1196
            logger.warning("Possibly too large swap space. %s", msg)
1197

1198

1199
1200
1201
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
1202

1203
1204
1205
1206
1207
1208
1209
1210
    Args:
        pool_size: Number of tokenizer workers in the pool.
        pool_type: Type of the pool.
        extra_config: Additional config for the pool.
            The way the config will be used depends on the
            pool type.
    """
    pool_size: int
1211
    pool_type: Union[str, type["BaseTokenizerGroup"]]
1212
1213
    extra_config: dict

1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
    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.
1228
        factors: list[Any] = []
1229
1230
1231
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1232
    def __post_init__(self):
1233
1234
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
1235
1236
1237
1238
1239
1240
            raise ValueError(f"Unknown pool type: {self.pool_type}")
        if not isinstance(self.extra_config, dict):
            raise ValueError("extra_config must be a dictionary.")

    @classmethod
    def create_config(
1241
        cls, tokenizer_pool_size: int,
1242
        tokenizer_pool_type: Union[str, type["BaseTokenizerGroup"]],
1243
1244
1245
        tokenizer_pool_extra_config: Optional[Union[str, dict]]
    ) -> Optional["TokenizerPoolConfig"]:
        """Create a TokenizerPoolConfig from the given parameters.
1246

1247
        If tokenizer_pool_size is 0, return None.
1248

1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
        Args:
            tokenizer_pool_size: Number of tokenizer workers in the pool.
            tokenizer_pool_type: Type of the pool.
            tokenizer_pool_extra_config: Additional config for the pool.
                The way the config will be used depends on the
                pool type. This can be a JSON string (will be parsed).
        """
        if tokenizer_pool_size:
            if isinstance(tokenizer_pool_extra_config, str):
                tokenizer_pool_extra_config_parsed = json.loads(
                    tokenizer_pool_extra_config)
            else:
                tokenizer_pool_extra_config_parsed = (
                    tokenizer_pool_extra_config or {})
            tokenizer_pool_config = cls(tokenizer_pool_size,
                                        tokenizer_pool_type,
                                        tokenizer_pool_extra_config_parsed)
        else:
            tokenizer_pool_config = None
        return tokenizer_pool_config


1271
1272
1273
1274
1275
1276
1277
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
1278
    SHARDED_STATE = "sharded_state"
1279
    GGUF = "gguf"
1280
    BITSANDBYTES = "bitsandbytes"
1281
    MISTRAL = "mistral"
1282
    RUNAI_STREAMER = "runai_streamer"
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301


@dataclass
class LoadConfig:
    """
        download_dir: Directory to download and load the weights, default to the
            default cache directory of huggingface.
        load_format: The format of the model weights to load:
            "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.
            "pt" will load the weights in the pytorch bin format.
            "safetensors" will load the weights in the safetensors format.
            "npcache" will load the weights in pytorch format and store
                a numpy cache to speed up the loading.
            "dummy" will initialize the weights with random values, which is
                mainly for profiling.
            "tensorizer" will use CoreWeave's tensorizer library for
                fast weight loading.
1302
            "bitsandbytes" will load nf4 type weights.
1303
        model_loader_extra_config: The extra config for the model loader.
1304
        ignore_patterns: The list of patterns to ignore when loading the model.
1305
            Default to "original/**/*" to avoid repeated loading of llama's
1306
            checkpoints.
1307
1308
        use_tqdm_on_load: Whether to enable tqdm for showing progress bar during
            loading. Default to True
1309
1310
1311
1312
1313
1314
    """

    load_format: Union[str, LoadFormat, "BaseModelLoader"] = LoadFormat.AUTO
    download_dir: Optional[str] = None
    model_loader_extra_config: Optional[Union[str, dict]] = field(
        default_factory=dict)
1315
    ignore_patterns: Optional[Union[list[str], str]] = None
1316
    use_tqdm_on_load: bool = True
1317

1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
    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.
1332
        factors: list[Any] = []
1333
1334
1335
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1336
1337
1338
1339
1340
    def __post_init__(self):
        model_loader_extra_config = self.model_loader_extra_config or {}
        if isinstance(model_loader_extra_config, str):
            self.model_loader_extra_config = json.loads(
                model_loader_extra_config)
1341
1342
1343
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1344

1345
1346
1347
1348
1349
1350
1351
        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/**/*"]

1352

1353
@dataclass
1354
class ParallelConfig:
1355
    """Configuration for the distributed execution."""
1356

1357
1358
    pipeline_parallel_size: int = 1  # Number of pipeline parallel groups.
    tensor_parallel_size: int = 1  # Number of tensor parallel groups.
1359
1360
1361
1362
1363
    data_parallel_size: int = 1  # Number of data parallel groups.
    data_parallel_rank: int = 0  # Rank of the data parallel group.
    # IP of the data parallel master.
    data_parallel_master_ip: str = "127.0.0.1"
    data_parallel_master_port: int = 29500  # Port of the data parallel master.
1364
    enable_expert_parallel: bool = False  # Use EP instead of TP for MoE layers.
1365

1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
    # Maximum number of multiple batches
    # when load model sequentially. To avoid RAM OOM when using tensor
    # parallel and large models.
    max_parallel_loading_workers: Optional[int] = None

    # Disable the custom all-reduce kernel and fall back to NCCL.
    disable_custom_all_reduce: bool = False

    # Config for the tokenizer pool. If None, will use synchronous tokenization.
    tokenizer_pool_config: Optional[TokenizerPoolConfig] = None

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

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

    # 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.
    distributed_executor_backend: Optional[Union[str,
1391
                                                 type["ExecutorBase"]]] = None
1392
1393
1394
1395

    # the full name of the worker class to use. If "auto", the worker class
    # will be determined based on the platform.
    worker_cls: str = "auto"
1396
    sd_worker_cls: str = "auto"
1397
    worker_extension_cls: str = ""
1398

1399
    # world_size is TPxPP, it affects the number of workers we create.
1400
    world_size: int = field(init=False)
1401
1402
1403
    # world_size_across_dp is TPxPPxDP, it is the size of the world
    # including data parallelism.
    world_size_across_dp: int = field(init=False)
1404
1405
1406

    rank: int = 0

1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
    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
1436
                          has_unfinished: bool) -> bool:
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
        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

1448
1449
1450
1451
1452
1453
1454
1455
    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.
        """
1456
        factors: list[Any] = []
1457
1458
1459
1460
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
        return hashlib.sha256(str(factors).encode()).hexdigest()

1461
1462
1463
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
1464
1465
1466
1467
1468
1469

        self.data_parallel_size = envs.VLLM_DP_SIZE
        self.data_parallel_rank = envs.VLLM_DP_RANK
        self.data_parallel_master_ip = envs.VLLM_DP_MASTER_IP
        self.data_parallel_master_port = envs.VLLM_DP_MASTER_PORT
        self.world_size_across_dp = self.world_size * self.data_parallel_size
1470

1471
1472
1473
1474
1475
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

1476
        ray_only_devices: list[str] = []
1477
        from vllm.platforms import current_platform
1478
1479
        if (current_platform.device_type in ray_only_devices
                and self.world_size > 1):
1480
1481
1482
1483
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
1484
1485
                    f"{current_platform.device_type.upper()} backend only "
                    "supports Ray for distributed inference.")
1486

1487
        if self.distributed_executor_backend is None and self.world_size > 1:
1488
1489
1490
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1491
            from vllm.executor import ray_utils
1492
            backend = "mp"
1493
            ray_found = ray_utils.ray_is_available()
1494
1495
1496
1497
1498
            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):
1499
1500
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1501
1502
1503
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1504
1505
                backend = "ray"
            elif ray_found:
1506
                if self.placement_group:
1507
                    backend = "ray"
1508
1509
1510
1511
1512
1513
                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"
1514
1515
1516
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1517

1518
1519
1520
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

1521
1522
        self._verify_args()

1523
1524
1525
1526
1527
1528
    @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)

1529
    def _verify_args(self) -> None:
1530
1531
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
1532
        from vllm.platforms import current_platform
1533
        if self.distributed_executor_backend not in (
1534
1535
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
1536
1537
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
1538
            raise ValueError(
1539
1540
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
1541
1542
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
1543
        if self.use_ray:
1544
1545
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1546
        if current_platform.is_rocm():
1547
1548
1549
1550
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
1551
        if self.ray_workers_use_nsight and not self.use_ray:
1552
1553
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1554

1555
1556
1557
        assert isinstance(self.worker_extension_cls, str), (
            "worker_extension_cls must be a string (qualified class name).")

1558

1559
@dataclass
1560
class SchedulerConfig:
1561
    """Scheduler configuration."""
1562

1563
    runner_type: str = "generate"  # The runner type to launch for the model.
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573

    # Maximum number of tokens to be processed in a single iteration.
    max_num_batched_tokens: int = field(default=None)  # type: ignore

    # Maximum number of sequences to be processed in a single iteration.
    max_num_seqs: int = 128

    # Maximum length of a sequence (including prompt and generated text).
    max_model_len: int = 8192

1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
    # Maximum number of sequences that can be partially prefilled concurrently
    max_num_partial_prefills: int = 1

    # Maximum number of "very long prompt" sequences that can be prefilled
    # concurrently (long is defined by long_prefill_threshold)
    max_long_partial_prefills: int = 1

    # calculate context length that determines which sequences are
    # considered "long"
    long_prefill_token_threshold: int = 0

1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
    # 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.
    num_lookahead_slots: int = 0

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

    # If True, prefill requests can be chunked based
    # on the remaining max_num_batched_tokens.
    enable_chunked_prefill: bool = False

    is_multimodal_model: bool = False
1600

1601
1602
1603
1604
1605
1606
    # NOTE: The following multimodal encoder budget will be initialized to
    # max_num_batched_tokens and overridden in case max multimodal embedding
    # size is larger.
    # TODO (ywang96): Make these configurable.
    # Multimodal encoder compute budget, only used in V1
    max_num_encoder_input_tokens: int = field(default=None)  # type: ignore
1607
1608

    # Multimodal encoder cache size, only used in V1
1609
    encoder_cache_size: int = field(default=None)  # type: ignore
1610

1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
    # 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.
    preemption_mode: Optional[str] = None

    num_scheduler_steps: int = 1

    multi_step_stream_outputs: bool = False

    # 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
    send_delta_data: bool = False

    # The scheduling policy to use. "fcfs" (default) or "priority".
    policy: str = "fcfs"

    chunked_prefill_enabled: bool = field(init=False)

1634
1635
    # scheduler class or path. "vllm.core.scheduler.Scheduler" (default)
    # or "mod.custom_class".
1636
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
1637

1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
    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.
1652
        factors: list[Any] = []
1653
1654
1655
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1656
1657
1658
1659
    def __post_init__(self) -> None:
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
1660
1661
1662
1663
                    # 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.
1664
1665
                    self.max_num_batched_tokens = max(
                        self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
1666
                else:
1667
1668
                    self.max_num_batched_tokens = (
                        _DEFAULT_MAX_NUM_BATCHED_TOKENS)
1669
            else:
1670
1671
                # If max_model_len is too short, use
                # _DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
1672
                # for higher throughput.
1673
1674
                self.max_num_batched_tokens = max(
                    self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
1675

1676
1677
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
1678
1679
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1680
                    _POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
1681
                )
1682
            if self.is_multimodal_model:
1683
                # The value needs to be at least the number of multimodal tokens
1684
1685
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1686
1687
1688
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

1689
1690
1691
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

1692
        if self.enable_chunked_prefill:
1693
1694
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
1695
                self.max_num_batched_tokens)
1696

1697
        self.chunked_prefill_enabled = self.enable_chunked_prefill
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
        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)

1710
1711
1712
        self._verify_args()

    def _verify_args(self) -> None:
1713
1714
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
1715
1716
1717
1718
1719
1720
1721
            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.")
1722

1723
1724
1725
1726
1727
        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}).")
1728

1729
1730
1731
1732
1733
1734
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

1735
1736
1737
1738
1739
1740
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
        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}).")

1764
1765
1766
1767
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

1768

1769
class DeviceConfig:
1770
    device: Optional[torch.device]
1771
    device_type: str
1772

1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
    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.
1788
        factors: list[Any] = []
1789
1790
1791
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1792
1793
1794
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
1795
            from vllm.platforms import current_platform
1796
            self.device_type = current_platform.device_type
1797
            if not self.device_type:
1798
                raise RuntimeError("Failed to infer device type")
1799
1800
1801
1802
1803
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
1804
        if self.device_type in ["neuron", "openvino"]:
1805
            self.device = torch.device("cpu")
1806
1807
        elif self.device_type in ["tpu"]:
            self.device = None
1808
1809
1810
1811
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1812

1813
1814
1815
1816
1817
1818
1819
class SpeculativeConfig:
    """Configuration for speculative decoding.

    The configuration is currently specialized to draft-model speculative
    decoding with top-1 proposals.
    """

1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
    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.
        # spec decode does not use `torch.compile` yet.
1834
        factors: list[Any] = []
1835
1836
1837
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
    @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

1850
1851
1852
1853
1854
1855
    @staticmethod
    def maybe_create_spec_config(
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        target_dtype: str,
        speculative_model: Optional[str],
1856
        speculative_model_quantization: Optional[str],
1857
        speculative_draft_tensor_parallel_size: Optional[int],
1858
        num_speculative_tokens: Optional[int],
1859
        speculative_disable_mqa_scorer: Optional[bool],
1860
1861
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
1862
        disable_log_stats: bool,
1863
        speculative_disable_by_batch_size: Optional[int],
1864
1865
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1866
1867
1868
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
1869
        disable_logprobs: Optional[bool],
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
    ) -> Optional["SpeculativeConfig"]:
        """Create a SpeculativeConfig if possible, else return None.

        This function attempts to create a SpeculativeConfig object based on the
        provided parameters. If the necessary conditions are met, it returns an
        instance of SpeculativeConfig. Otherwise, it returns None.

        Args:
            target_model_config (ModelConfig): The configuration of the target
                model.
            target_parallel_config (ParallelConfig): The parallel configuration
                for the target model.
            target_dtype (str): The data type used for the target model.
            speculative_model (Optional[str]): The name of the speculative
                model, if provided.
1885
1886
1887
            speculative_model_quantization (Optional[str]): Quantization method
                that was used to quantize the speculative model weights. If
                None, we assume the model weights are not quantized.
1888
1889
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
1890
            num_speculative_tokens (Optional[int]): The number of speculative
1891
1892
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
1893
1894
1895
            speculative_disable_mqa_scorer (Optional[bool]): Disable the MQA
                scorer for the speculative model and fall back to batch
                expansion for scoring.
1896
1897
1898
1899
1900
1901
            speculative_max_model_len (Optional[int]): The maximum model len of
                the speculative model. Used when testing the ability to skip
                speculation for some sequences.
            enable_chunked_prefill (bool): Whether vLLM is configured to use
                chunked prefill or not. Used for raising an error since its not
                yet compatible with spec decode.
1902
1903
1904
            speculative_disable_by_batch_size (Optional[int]): Disable
                speculative decoding for new incoming requests when the number
                of enqueue requests  is larger than this value, if provided.
1905
1906
1907
1908
            ngram_prompt_lookup_max (Optional[int]): Max size of ngram token
                window, if provided.
            ngram_prompt_lookup_min (Optional[int]): Min size of ngram token
                window, if provided.
1909
1910
1911
1912
1913
1914
1915
1916
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
1917
                accepted. This threshold is used only when we use the
1918
1919
1920
1921
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1922
1923
1924
1925
1926
            disable_logprobs (Optional[bool]): 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.
                If not specified, it defaults to True.
1927

1928
1929
1930
1931
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """
1932
1933
        if speculative_model is None:
            if num_speculative_tokens is not None:
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
                if target_model_config.hf_text_config.model_type \
                        == "deepseek_v3":
                    # use the draft model from the same model:
                    speculative_model = target_model_config.model
                else:
                    raise ValueError(
                        "num_speculative_tokens was provided without "
                        "speculative_model.")
            else:
                return None
1944

1945
1946
1947
1948
1949
        if (speculative_disable_by_batch_size is not None
                and speculative_disable_by_batch_size < 2):
            raise ValueError("Expect the batch size threshold of disabling "
                             "speculative decoding is > 1, but got "
                             f"{speculative_disable_by_batch_size=}")
1950
1951
        if (enable_chunked_prefill and speculative_model == "eagle"):
            raise ValueError("Chunked prefill and EAGLE are not compatible.")
1952
1953
        # TODO: The user should be able to specify revision/max model len
        # for the draft model. It is not currently supported.
1954
1955
        draft_revision = None
        draft_code_revision = None
1956
        draft_quantization = speculative_model_quantization
1957

1958
1959
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1960
1961
1962
1963
1964
1965
1966
1967
                ngram_prompt_lookup_min = 1
            if ngram_prompt_lookup_max is None or ngram_prompt_lookup_max < 1:
                raise ValueError(f"{ngram_prompt_lookup_max=} must be > 0")
            if ngram_prompt_lookup_min < 1:
                raise ValueError(f"{ngram_prompt_lookup_min=} must be > 0")
            if ngram_prompt_lookup_min > ngram_prompt_lookup_max:
                raise ValueError(f"{ngram_prompt_lookup_min=} cannot be "
                                 f"larger than {ngram_prompt_lookup_max=}")
1968

1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
            # 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.
            draft_model_config = target_model_config
            draft_parallel_config = target_parallel_config
        else:
            ngram_prompt_lookup_max = 0
            ngram_prompt_lookup_min = 0
            draft_model_config = ModelConfig(
                model=speculative_model,
1979
                task="draft",
1980
1981
1982
                tokenizer=target_model_config.tokenizer,
                tokenizer_mode=target_model_config.tokenizer_mode,
                trust_remote_code=target_model_config.trust_remote_code,
1983
1984
                allowed_local_media_path=target_model_config.
                allowed_local_media_path,
1985
1986
1987
1988
1989
1990
                dtype=target_model_config.dtype,
                seed=target_model_config.seed,
                revision=draft_revision,
                code_revision=draft_code_revision,
                tokenizer_revision=target_model_config.tokenizer_revision,
                max_model_len=None,
1991
                spec_target_max_model_len=target_model_config.max_model_len,
1992
1993
                quantization=draft_quantization,
                enforce_eager=target_model_config.enforce_eager,
1994
1995
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1996
                max_logprobs=target_model_config.max_logprobs,
1997
                hf_overrides=SpeculativeConfig.hf_config_override,
1998
1999
            )

2000
            draft_hf_config = draft_model_config.hf_config
2001

2002
2003
2004
2005
2006
2007
2008
2009
2010
            # Detect EAGLE prefix to replace hf_config for EAGLE draft_model
            if "eagle-" in draft_model_config.model.lower():
                from vllm.transformers_utils.configs.eagle import EAGLEConfig
                if isinstance(draft_model_config.hf_config, EAGLEConfig):
                    pass
                else:
                    eagle_config = EAGLEConfig(draft_model_config.hf_config)
                    draft_model_config.hf_config = eagle_config

2011
2012
2013
2014
            if (num_speculative_tokens is not None
                    and hasattr(draft_hf_config, "num_lookahead_tokens")):
                draft_hf_config.num_lookahead_tokens = num_speculative_tokens
            n_predict = getattr(draft_hf_config, "n_predict", None)
2015
2016
2017
2018
            if n_predict is not None:
                if num_speculative_tokens is None:
                    # Default to max value defined in draft model config.
                    num_speculative_tokens = n_predict
2019
2020
2021
                elif num_speculative_tokens > n_predict and \
                        num_speculative_tokens % n_predict != 0:
                    # Ensure divisibility for MTP module reuse.
2022
                    raise ValueError(
2023
2024
                        f"{num_speculative_tokens=} must be divisible by "
                        f"{n_predict=}")
2025

2026
2027
2028
2029
2030
2031
2032
            speculative_draft_tensor_parallel_size = \
                SpeculativeConfig._verify_and_get_draft_model_tensor_parallel_size(
                    target_parallel_config,
                    speculative_draft_tensor_parallel_size,
                    draft_hf_config
            )

2033
2034
2035
2036
2037
2038
2039
2040
2041
            draft_model_config.max_model_len = (
                SpeculativeConfig._maybe_override_draft_max_model_len(
                    speculative_max_model_len,
                    draft_model_config.max_model_len,
                    target_model_config.max_model_len,
                ))

            draft_parallel_config = (
                SpeculativeConfig.create_draft_parallel_config(
2042
                    target_parallel_config,
2043
                    speculative_draft_tensor_parallel_size, draft_hf_config))
2044

2045
2046
2047
2048
2049
2050
        if 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.")

2051
2052
2053
2054
        if typical_acceptance_sampler_posterior_threshold is None:
            typical_acceptance_sampler_posterior_threshold = 0.09
        if typical_acceptance_sampler_posterior_alpha is None:
            typical_acceptance_sampler_posterior_alpha = 0.3
2055
2056
        if disable_logprobs is None:
            disable_logprobs = True
2057

2058
2059
2060
2061
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
2062
            speculative_disable_mqa_scorer,
2063
            speculative_disable_by_batch_size,
2064
2065
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
2066
2067
2068
2069
2070
            draft_token_acceptance_method=draft_token_acceptance_method,
            typical_acceptance_sampler_posterior_threshold=\
                typical_acceptance_sampler_posterior_threshold,
            typical_acceptance_sampler_posterior_alpha=\
                typical_acceptance_sampler_posterior_alpha,
2071
2072
            disable_logprobs=disable_logprobs,
            disable_log_stats=disable_log_stats,
2073
2074
        )

2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
    @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,
        )

2110
    @staticmethod
2111
2112
2113
2114
2115
2116
2117
    def _verify_and_get_draft_model_tensor_parallel_size(
            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.
2118
        """
2119
2120
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2121
        if speculative_draft_tensor_parallel_size is None:
2122
2123
2124
2125
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2126
2127
2128
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2129
2130
2131
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2132
2133
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2134
            raise ValueError(
2135
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2136
                f"other value than 1 or target model tensor_parallel_size")
2137
        return speculative_draft_tensor_parallel_size
2138

2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
    @staticmethod
    def create_draft_parallel_config(
        target_parallel_config: ParallelConfig,
        speculative_draft_tensor_parallel_size: int,
        draft_hf_config: PretrainedConfig,
    ) -> 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.
        """
2149
2150
2151
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2152
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2153
2154
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
            max_parallel_loading_workers=target_parallel_config.
            max_parallel_loading_workers,
            disable_custom_all_reduce=target_parallel_config.
            disable_custom_all_reduce,
            tokenizer_pool_config=target_parallel_config.tokenizer_pool_config,
            ray_workers_use_nsight=target_parallel_config.
            ray_workers_use_nsight,
            placement_group=target_parallel_config.placement_group,
        )

        return draft_parallel_config

    def __init__(
        self,
        draft_model_config: ModelConfig,
        draft_parallel_config: ParallelConfig,
        num_speculative_tokens: int,
2172
        speculative_disable_mqa_scorer: Optional[bool],
2173
2174
2175
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
2176
2177
2178
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
2179
        disable_logprobs: bool,
2180
        disable_log_stats: bool,
2181
2182
2183
2184
2185
2186
2187
2188
    ):
        """Create a SpeculativeConfig object.

        Args:
            draft_model_config: ModelConfig for the draft model.
            draft_parallel_config: ParallelConfig for the draft model.
            num_speculative_tokens: The number of tokens to sample from the
                draft model before scoring with the target model.
2189
2190
2191
2192
2193
            speculative_disable_by_batch_size: Disable speculative
                decoding for new incoming requests when the number of
                enqueue requests is larger than this value.
            ngram_prompt_lookup_max: Max size of ngram token window.
            ngram_prompt_lookup_min: Min size of ngram token window.
2194
2195
2196
2197
2198
2199
2200
2201
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
2202
                accepted. This threshold is used only when we use the
2203
2204
2205
2206
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
2207
            disable_logprobs: If set to True, token log probabilities will not
2208
                be returned even if requested by sampling parameters. This
2209
2210
2211
2212
                reduces latency by skipping logprob calculation in proposal
                sampling, target sampling, and after accepted tokens are
                determined. If set to False, log probabilities will be
                returned.
2213
2214
            disable_log_stats: Whether to disable periodic printing of stage
                times in speculative decoding.
2215
2216
2217
2218
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
2219
        self.speculative_disable_mqa_scorer = speculative_disable_mqa_scorer
2220
2221
2222
2223
        self.speculative_disable_by_batch_size = \
            speculative_disable_by_batch_size
        self.ngram_prompt_lookup_max = ngram_prompt_lookup_max or 0
        self.ngram_prompt_lookup_min = ngram_prompt_lookup_min or 0
2224
2225
2226
2227
2228
        self.draft_token_acceptance_method = draft_token_acceptance_method
        self.typical_acceptance_sampler_posterior_threshold = \
            typical_acceptance_sampler_posterior_threshold
        self.typical_acceptance_sampler_posterior_alpha = \
            typical_acceptance_sampler_posterior_alpha
2229
        self.disable_logprobs = disable_logprobs
2230
        self.disable_log_stats = disable_log_stats
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241

        self._verify_args()

    def _verify_args(self) -> None:
        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)
2242
2243
2244
2245
2246
2247
2248
2249
            # Validate and set draft token acceptance related settings.

        if (self.draft_token_acceptance_method is None):
            raise ValueError("draft_token_acceptance_method is not set. "
                             "Expected values are rejection_sampler or "
                             "typical_acceptance_sampler.")

        if (self.draft_token_acceptance_method != 'rejection_sampler'
2250
2251
                and self.draft_token_acceptance_method
                != 'typical_acceptance_sampler'):
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
            raise ValueError(
                "Expected draft_token_acceptance_method to be either "
                "rejection_sampler or typical_acceptance_sampler. Instead it "
                f"is {self.draft_token_acceptance_method}")

        if (self.typical_acceptance_sampler_posterior_threshold < 0
                or self.typical_acceptance_sampler_posterior_alpha < 0):
            raise ValueError(
                "Expected typical_acceptance_sampler_posterior_threshold "
                "and typical_acceptance_sampler_posterior_alpha to be > 0. "
                "Instead found "
                f"typical_acceptance_sampler_posterior_threshold = "
                f"{self.typical_acceptance_sampler_posterior_threshold} and "
                f"typical_acceptance_sampler_posterior_alpha = "
                f"{self.typical_acceptance_sampler_posterior_alpha}")
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278

    @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

    def __repr__(self) -> str:
2279
2280
2281
2282
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
2283
2284
2285
2286
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


2287
2288
2289
2290
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
2291
    fully_sharded_loras: bool = False
2292
    max_cpu_loras: Optional[int] = None
2293
    lora_dtype: Optional[Union[torch.dtype, str]] = None
2294
2295
2296
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
2297
    long_lora_scaling_factors: Optional[tuple[float]] = None
2298
    bias_enabled: bool = False
2299

2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
    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.
        """
2312
        factors: list[Any] = []
2313
2314
2315
2316
2317
2318
2319
        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)
2320
2321
2322
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2323
    def __post_init__(self):
2324
        # Setting the maximum rank to 512 should be able to satisfy the vast
2325
        # majority of applications.
2326
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
2327
        possible_lora_extra_vocab_size = (256, 512)
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
        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
2343
                f"max_loras ({self.max_loras})")
2344

2345
2346
2347
2348
2349
    def verify_with_cache_config(self, cache_config: CacheConfig):
        # TODO LoRA supports CPU offload.
        if cache_config.cpu_offload_gb > 0:
            raise ValueError("CPU offload is not supported with LoRA yet.")

2350
2351
2352
2353
2354
    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)
2355
2356
2357
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
2358
            # TODO support marlin
2359
2360
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
2361
2362

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
2363
        # Reminder: Please update docs/source/features/compatibility_matrix.md
2364
        # If the feature combo become valid
2365
        if scheduler_config.chunked_prefill_enabled:
2366
2367
            logger.warning("LoRA with chunked prefill is still experimental "
                           "and may be unstable.")
2368
2369


2370
2371
2372
2373
2374
2375
2376
@dataclass
class PromptAdapterConfig:
    max_prompt_adapters: int
    max_prompt_adapter_token: int
    max_cpu_prompt_adapters: Optional[int] = None
    prompt_adapter_dtype: Optional[torch.dtype] = None

2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
    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.
2391
        factors: list[Any] = []
2392
2393
2394
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
    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):
        if self.prompt_adapter_dtype in (None, "auto"):
            self.prompt_adapter_dtype = model_config.dtype
        elif isinstance(self.prompt_adapter_dtype, str):
            self.prompt_adapter_dtype = getattr(torch,
                                                self.prompt_adapter_dtype)


2413
@dataclass
2414
class MultiModalConfig:
2415
2416
    """Controls the behavior of multimodal models."""

2417
    limit_per_prompt: Mapping[str, int] = field(default_factory=dict)
2418
    """
2419
    The maximum number of input items allowed per prompt for each modality.
2420
2421
    """

2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
    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.
2436
        factors: list[Any] = []
2437
2438
2439
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2440
2441
2442
2443
2444
2445
2446
2447
2448
    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality.

        If not set by the user, this defaults to `1`.
        """
        return self.limit_per_prompt.get(modality, 1)

2449
    # TODO: Add configs to init vision tower or not.
2450

2451

2452
2453
@dataclass
class PoolerConfig:
2454
    """Controls the behavior of output pooling in pooling models."""
2455
2456

    pooling_type: Optional[str] = None
2457
    """
2458
    The pooling method of the pooling model. This should be a key in
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
    :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
    """
2476
    If set, only the score corresponding to the ``step_tag_id`` in the
2477
2478
2479
2480
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

2481
    returned_token_ids: Optional[list[int]] = None
2482
    """
2483
2484
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2485
2486
2487
    ``math-shepherd-mistral-7b-prm`` model.
    """

2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
    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.
2502
        factors: list[Any] = []
2503
2504
2505
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2506
2507
2508
    @staticmethod
    def from_json(json_str: str) -> "PoolerConfig":
        return PoolerConfig(**json.loads(json_str))
2509
2510


2511
2512
2513
2514
2515
2516
2517
2518
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2519
_ROCM_NOT_SUPPORTED_DTYPE: list[str] = []  #
2520

2521
2522
2523

def _get_and_verify_dtype(
    config: PretrainedConfig,
2524
    dtype: Union[str, torch.dtype],
2525
2526
2527
2528
) -> torch.dtype:
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
    config_dtype = getattr(config, "torch_dtype", None)
2529
2530
2531
2532
2533
2534
2535
2536

    # Fallbacks for multi-modal models if the root config
    # does not define torch_dtype
    if config_dtype is None and hasattr(config, "text_config"):
        config_dtype = getattr(config.text_config, "torch_dtype", None)
    if config_dtype is None and hasattr(config, "vision_config"):
        config_dtype = getattr(config.vision_config, "torch_dtype", None)

2537
2538
2539
    if config_dtype is None:
        config_dtype = torch.float32

2540
2541
2542
2543
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
2544
2545
                # Following common practice, we use float16 for float32 models
                torch_dtype = torch.float16
2546
2547
            else:
                torch_dtype = config_dtype
2548

2549
            from vllm.platforms import current_platform
2550
2551
            if (current_platform.is_cpu()
                    and current_platform.get_cpu_architecture()
2552
                    == CpuArchEnum.POWERPC
2553
2554
2555
2556
2557
2558
2559
2560
                    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

2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
            # 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

2572
2573
            if current_platform.is_hpu() and config_dtype == torch.float16:
                logger.info(
2574
                    "For HPU, we cast models to bfloat16 instead of "
2575
2576
2577
                    "using float16 by default. Please specify `dtype` if you "
                    "want to use float16.")
                torch_dtype = torch.bfloat16
2578
        else:
2579
2580
2581
2582
2583
            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
2584
    else:
2585
        raise ValueError(f"Unknown dtype: {dtype}")
2586
2587
2588
2589
2590

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
2591
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
2592
2593
2594
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
2595
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
2596
2597
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
2598
            # Casting between float16 and bfloat16 is allowed with a warning.
2599
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
2600
2601

    return torch_dtype
2602
2603
2604
2605
2606


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
2607
    disable_sliding_window: bool,
2608
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
2609
    spec_target_max_model_len: Optional[int] = None,
2610
    encoder_config: Optional[Any] = None,
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
) -> 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",
2621
2622
        # ChatGLM2
        "seq_length",
2623
2624
        # Command-R
        "model_max_length",
2625
2626
        # Whisper
        "max_target_positions",
2627
2628
2629
2630
2631
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
2632
    # Choose the smallest "max_length" from the possible keys.
2633
    max_len_key = None
2634
    for key in possible_keys:
2635
2636
2637
2638
2639
        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)
2640
2641
2642
2643

    # 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:
2644
2645

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
2646
        max_len_key = "sliding_window" \
2647
2648
2649
            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)
2650
2651
2652

    # If none of the keys were found in the config, use a default and
    # log a warning.
2653
    if derived_max_model_len == float("inf"):
2654
2655
2656
2657
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

2658
2659
2660
2661
2662
        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

2663
2664
2665
2666
        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: "
2667
            "%s. Assuming the model's maximum length is %d.", possible_keys,
2668
            default_max_len)
2669
        derived_max_model_len = default_max_len
2670

2671
    rope_scaling = getattr(hf_config, "rope_scaling", None)
2672
2673
2674
    # 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:
2675
2676
2677
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687

        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.")

2688
2689
2690
2691
            # 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)

2692
2693
2694
2695
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
2696

2697
2698
2699
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2700
2701
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2702
    if max_model_len is None:
2703
        max_model_len = int(derived_max_model_len)
2704
    elif max_model_len > derived_max_model_len:
2705
2706
2707
2708
2709
        # 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:
2710
2711
2712
2713
2714
2715
2716
            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.")
2717
        else:
2718
            msg = (
2719
                f"User-specified max_model_len ({max_model_len}) is greater "
2720
2721
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2722
                f"{model_max_length} in model's config.json). This may lead "
2723
2724
2725
2726
2727
2728
2729
2730
2731
                "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")
2732
    return int(max_model_len)
2733
2734


2735
def get_min_sliding_window(
2736
        sliding_window: Union[int, list[Optional[int]]]) -> int:
2737
2738
2739
2740
2741
2742
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


2743
def get_served_model_name(model: str,
2744
                          served_model_name: Optional[Union[str, list[str]]]):
2745
    """
2746
2747
2748
2749
    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
2750
2751
2752
2753
2754
2755
2756
2757
2758
    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


2759
2760
2761
2762
@dataclass
class DecodingConfig:
    """Dataclass which contains the decoding strategy of the engine"""

2763
2764
2765
    # Which guided decoding algo to use.
    # 'outlines' / 'lm-format-enforcer' / 'xgrammar'
    guided_decoding_backend: str = 'xgrammar'
2766

2767
2768
    reasoning_backend: Optional[str] = None

2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
    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.
2783
        factors: list[Any] = []
2784
2785
2786
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2787
    def __post_init__(self):
2788
        valid_guided_backends = ['outlines', 'lm-format-enforcer', 'xgrammar']
2789
2790
2791

        backend = GuidedDecodingParams(
            backend=self.guided_decoding_backend).backend_name
2792
2793
        if backend not in valid_guided_backends:
            raise ValueError(f"Invalid guided_decoding_backend '{backend},"
2794
                             f" must be one of {valid_guided_backends}")
2795
2796


2797
2798
@dataclass
class ObservabilityConfig:
2799
2800
2801
    """Configuration for observability - metrics and tracing."""
    show_hidden_metrics: bool = False

2802
2803
    otlp_traces_endpoint: Optional[str] = None

2804
2805
2806
2807
2808
2809
2810
2811
    # 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

2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
    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.
2826
        factors: list[Any] = []
2827
2828
2829
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2830
    def __post_init__(self):
2831
2832
2833
2834
2835
        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}")
2836
2837


2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
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

2871
2872
2873
    # any extra config that the connector may need
    kv_connector_extra_config: dict[str, Any] = {}

2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
    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.
2888
        factors: list[Any] = []
2889
2890
2891
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2892
2893
    @classmethod
    def from_cli(cls, cli_value: str) -> "KVTransferConfig":
youkaichao's avatar
youkaichao committed
2894
        """Parse the CLI value for the kv cache transfer config."""
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
        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 need_kv_parallel_group(self) -> bool:
        # for those database-based connector, vLLM does not need to create
        # parallel group, and in that case the kv parallel size will be 1.
        return self.kv_connector is not None and self.kv_parallel_size > 1

    @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"]

2933
2934
2935
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

2936

2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
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.
2955
        - debug_dump_path: the path to dump the debug information.
2956
2957
2958
        - 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.
2959
2960
2961
2962
2963
2964
2965
        - 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).
2966
2967
2968
2969
2970
2971
2972
2973
2974
        - 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).
2975
        - splitting_ops: a list of ops to split the full graph into subgraphs, used in piecewise compilation.
2976
2977
2978
2979
    - 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
2980
2981
2982
2983
                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.
2984
2985
2986
            TODO: move outside cudagraph logic into compilation.
            torch.compile will handle cudagraph capture logic in the future.
        - cudagraph_capture_sizes: sizes to capture cudagraph.
2987
            - None (default): capture sizes are inferred from vllm config.
2988
            - list[int]: capture sizes are specified as given.
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
        - 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
3002
3003
3004
3005
3006
                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.
3007
3008
3009
3010
3011
3012
3013
        - 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})`
3014
        - custom inductor passes: see PassConfig for more details
3015

3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
    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
3027
    debug_dump_path: str = ""
3028
    cache_dir: str = ""
3029
    backend: str = ""
3030
3031
    custom_ops: list[str] = Field(default_factory=list)
    splitting_ops: list[str] = Field(default=None)  # type: ignore
3032
3033

    use_inductor: bool = True
3034
3035
3036
    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)
3037
3038
3039

    use_cudagraph: bool = False
    cudagraph_num_of_warmups: int = 0
3040
    cudagraph_capture_sizes: Optional[list[int]] = None
3041
3042
    cudagraph_copy_inputs: bool = False

3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
    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.
3053
3054
        - enable_noop: whether to enable the custom no-op elimination pass.
            TODO(luka) better pass enabling system.
3055
        """
3056
        dump_graph_stages: list[str] = Field(default_factory=list)
3057
3058
        dump_graph_dir: Path = Field(default=Path("."))
        enable_fusion: bool = True
3059
        enable_noop: bool = True
3060
3061
3062
3063
3064
3065
3066
3067

        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.
            """
3068
            dict_ = self.model_dump(include={"enable_fusion", "enable_noop"})
3069
3070
3071
3072
            encoded = json.dumps(dict_, sort_keys=True).encode("utf-8")
            return hashlib.sha256(encoded).digest()

        def model_post_init(self, __context: Any) -> None:
3073
            if not self.enable_noop and self.enable_fusion:
3074
                logger.warning_once(
3075
                    "Fusion enabled but reshape elimination disabled. "
3076
3077
3078
                    "RMSNorm + quant (fp8) fusion might not work")

    pass_config: PassConfig = Field(default_factory=PassConfig)
3079
3080

    # not configurable, computed after init
3081
    max_capture_size: int = PrivateAttr
3082
    local_cache_dir: str = PrivateAttr  # local cache dir for each rank
3083
    # optimization:
3084
    # Intuitively, bs_to_padded_graph_size should be dict[int, int].
3085
    # since we know all keys are in a range [0, max_capture_size],
3086
3087
    # we can optimize it to list[int] for better lookup performance.
    bs_to_padded_graph_size: list[int] = PrivateAttr
3088

3089
3090
3091
    # keep track of enabled and disabled custom ops
    enabled_custom_ops: Counter[str] = PrivateAttr
    disabled_custom_ops: Counter[str] = PrivateAttr
3092
    traced_files: set[str] = PrivateAttr
3093
    compilation_time: float = PrivateAttr
3094

3095
3096
    # Per-model forward context
    # Map from layer name to the attention cls
3097
    static_forward_context: dict[str, Any] = PrivateAttr
3098

3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
    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.
        """
3111
        factors: list[Any] = []
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
        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()

3122
3123
3124
3125
3126
3127
3128
3129
    def __repr__(self) -> str:
        exclude = {
            "static_forward_context",
            "enabled_custom_ops",
            "disabled_custom_ops",
            "compilation_time",
            "bs_to_padded_graph_size",
            "pass_config",
3130
            "traced_files",
3131
3132
3133
3134
3135
        }
        return self.model_dump_json(exclude=exclude, exclude_unset=True)

    __str__ = __repr__

3136
3137
3138
3139
3140
    @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))
3141
3142
3143
        # 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)
3144

3145
3146
3147
3148
3149
3150
    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
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
        # 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

        if Version(torch.__version__) >= Version("2.6"):
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

3164
        if self.splitting_ops is None:
3165
            self.splitting_ops = []
3166

3167
3168
3169
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
3170
3171
3172
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
3173
3174
3175
3176
3177
3178
3179
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
3180
3181
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
3182

3183
3184
        self.enabled_custom_ops = Counter()
        self.disabled_custom_ops = Counter()
3185
        self.traced_files = set()
3186
        self.static_forward_context = {}
3187
        self.compilation_time = 0.0
3188

3189
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
        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
3207

3208
        from vllm.compilation.backends import VllmBackend
3209
        return VllmBackend(vllm_config)
3210

3211
    def init_with_cudagraph_sizes(self,
3212
                                  cudagraph_capture_sizes: list[int]) -> None:
3213
        """To complete the initialization of config,
3214
3215
        we need to know the cudagraph sizes."""

3216
        if self.cudagraph_capture_sizes is None:
3217
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
3218
        else:
3219
3220
3221
            # de-duplicate the sizes provided by the config
            self.cudagraph_capture_sizes = list(
                set(self.cudagraph_capture_sizes))
3222
3223
            logger.info(("cudagraph sizes specified by model runner"
                         " %s is overridden by config %s"),
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
                        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
3240

3241
        # sort to make sure cudagraph capture sizes are in descending order
3242
3243
3244
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
3245

3246
3247
3248
3249
        # 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)
        ]
3250
3251
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
3252
3253
3254
3255
3256
3257
3258
            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
3259

3260
3261
3262
3263
3264
3265
3266
3267
3268
    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",
            ]

3269

3270
3271
3272
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
3273
3274
3275
    simplifies passing around the distinct configurations in the codebase.
    """

3276
3277
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
3278
3279
3280
3281
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig,
                                            init=True)
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig,
                                              init=True)
3282
3283
3284
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
3285
3286
3287
3288
3289
    lora_config: Optional[LoRAConfig] = None
    speculative_config: Optional[SpeculativeConfig] = None
    decoding_config: Optional[DecodingConfig] = None
    observability_config: Optional[ObservabilityConfig] = None
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
3290
    quant_config: Optional[QuantizationConfig] = None
3291
3292
    compilation_config: CompilationConfig = field(default=None,
                                                  init=True)  # type: ignore
3293
3294
    kv_transfer_config: KVTransferConfig = field(default=None,
                                                 init=True)  # type: ignore
3295
    # some opaque config, only used to provide additional information
3296
3297
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
3298
3299
    additional_config: SupportsHash = field(default=None,
                                            init=True)  # type: ignore
3300
    instance_id: str = ""
3301

3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
    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.
        """
3314
        factors: list[Any] = []
3315
3316

        # summarize vllm config
3317
        vllm_factors: list[Any] = []
3318
3319
        from vllm import __version__
        vllm_factors.append(__version__)
3320
        vllm_factors.append(envs.VLLM_USE_V1)
3321
3322
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3323
3324
        else:
            vllm_factors.append("None")
3325
3326
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3327
3328
        else:
            vllm_factors.append("None")
3329
3330
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3331
3332
        else:
            vllm_factors.append("None")
3333
3334
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3335
3336
        else:
            vllm_factors.append("None")
3337
3338
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3339
3340
        else:
            vllm_factors.append("None")
3341
3342
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3343
3344
        else:
            vllm_factors.append("None")
3345
3346
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3347
3348
3349
3350
3351
            # 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))
3352
3353
        else:
            vllm_factors.append("None")
3354
3355
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3356
3357
        else:
            vllm_factors.append("None")
3358
3359
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3360
3361
        else:
            vllm_factors.append("None")
3362
3363
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3364
3365
        else:
            vllm_factors.append("None")
3366
3367
        if self.prompt_adapter_config:
            vllm_factors.append(self.prompt_adapter_config.compute_hash())
3368
3369
        else:
            vllm_factors.append("None")
3370
3371
3372
3373
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3374
3375
        else:
            vllm_factors.append("None")
3376
3377
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3378
3379
3380
3381
3382
3383
        else:
            vllm_factors.append("None")
        if self.additional_config:
            vllm_factors.append(self.additional_config.compute_hash())
        else:
            vllm_factors.append("None")
3384
3385
3386
3387
3388
        factors.append(vllm_factors)

        hash_str = hashlib.md5(str(factors).encode()).hexdigest()[:10]
        return hash_str

3389
3390
3391
3392
3393
3394
    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]
3395

3396
3397
3398
3399
3400
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3401
        from vllm.platforms import current_platform
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
        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
3424

3425
3426
3427
3428
3429
3430
3431
3432
3433
    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

3434
3435
3436
3437
3438
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

3439
3440
3441
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
3442
3443
3444
3445
3446
3447
3448
3449
        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)
3450
3451

        if self.lora_config:
3452
            self.lora_config.verify_with_cache_config(self.cache_config)
3453
3454
3455
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
3456
3457
3458
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
3459
3460
3461
3462
3463

        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)
3464

3465
        from vllm.platforms import current_platform
3466
3467
3468
3469
3470
        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):
3471
            logger.warning_once(
3472
3473
3474
3475
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

3476
        if self.compilation_config is None:
3477
            self.compilation_config = CompilationConfig()
3478
3479
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
3480
3481
3482
3483
            # 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.
3484
            # FIXME(rob): Add function to set all of these.
3485
3486
3487
            self.compilation_config.custom_ops = ["none"]
            self.compilation_config.use_cudagraph = True
            self.compilation_config.use_inductor = True
3488
            self.compilation_config.cudagraph_num_of_warmups = 1
3489
            self.compilation_config.pass_config.enable_fusion = False
3490
            self.compilation_config.pass_config.enable_noop = False
3491
            self.compilation_config.level = CompilationLevel.PIECEWISE
3492
            self.compilation_config.set_splitting_ops_for_v1()
3493

3494
        self._set_cudagraph_sizes()
3495

3496
3497
3498
3499
3500
3501
3502
3503
        if self.cache_config is not None and \
            self.cache_config.cpu_offload_gb > 0 and \
            self.compilation_config.level != CompilationLevel.NO_COMPILATION:
            logger.warning(
                "CPU offload is not supported with `torch.compile` yet."
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3504
3505
3506
3507
3508
3509
        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`.")
3510
3511
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3512

3513
        if self.model_config and self.model_config.use_mla and \
3514
            not (current_platform.is_cuda() or current_platform.is_rocm()):
3515
            logger.info(
3516
                "MLA is enabled on a non-GPU platform; forcing chunked "
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
                "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

3527
3528
        current_platform.check_and_update_config(self)

3529
3530
3531
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
    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.

3548
3549
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
3550
3551

        During runtime, if batchsize is larger than
3552
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3553
3554
        no cudagraph will be used.
        If the batch size is no larger than
3555
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
        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)]
                # 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)]
3592
3593
3594
3595
3596
                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
                ]
3597
3598
3599
3600

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

3601
    def __str__(self):
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
        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}, "
3632
            f"disable_mm_preprocessor_cache={self.model_config.disable_mm_preprocessor_cache!r}, "  # noqa
3633
            f"mm_processor_kwargs={self.model_config.mm_processor_kwargs}, "
3634
3635
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
3636
3637
3638
3639
3640
3641


_current_vllm_config: Optional[VllmConfig] = None


@contextmanager
3642
def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False):
3643
    """
3644
    Temporarily set the current vLLM config.
3645
    Used during model initialization.
3646
    We save the current vLLM config in a global variable,
3647
    so that all modules can access it, e.g. custom ops
3648
    can access the vLLM config to determine how to dispatch.
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
    """
    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
    finally:
        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)
3662
3663
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
3664
3665
3666
3667
3668
3669
3670
3671
3672
            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"
3673
                " if you want it to be supported.",
3674
3675
3676
3677
3678
3679
3680
3681
3682
                vllm_config.model_config.model)
        _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.
3683
        logger.warning("Current vLLM config is not set.")
3684
3685
3686
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config