config.py 159 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 importlib.metadata
8
import json
9
import sys
10
import warnings
11
12
from collections import Counter
from collections.abc import Mapping
13
from contextlib import contextmanager
14
from dataclasses import dataclass, field, replace
15
from importlib.util import find_spec
16
from pathlib import Path
17
18
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Final, Literal,
                    Optional, Protocol, Union)
19
20

import torch
21
from packaging.version import Version
22
from pydantic import BaseModel, Field, PrivateAttr
23
from torch.distributed import ProcessGroup, ReduceOp
24
from transformers import PretrainedConfig
25

26
import vllm.envs as envs
27
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
28
from vllm.logger import init_logger
29
30
from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
                                                     get_quantization_config)
31
from vllm.model_executor.models import ModelRegistry
32
from vllm.platforms import CpuArchEnum
33
from vllm.sampling_params import GuidedDecodingParams
34
from vllm.tracing import is_otel_available, otel_import_error_traceback
35
36
37
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
38
39
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
    try_get_generation_config, uses_mrope)
40
from vllm.transformers_utils.s3_utils import S3Model
41
from vllm.transformers_utils.utils import is_s3
42
from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
43
                        get_cpu_memory, random_uuid, resolve_obj_by_qualname)
44

45
46
47
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

48
    from vllm.executor.executor_base import ExecutorBase
49
50
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
51
    from vllm.model_executor.model_loader.loader import BaseModelLoader
52
53
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
54
55
else:
    QuantizationConfig = None
56

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
        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)
224
225
226
        # rope cos/sin cache depends on the max_position_embeddings
        factors.append(
            getattr(self.hf_config, "max_position_embeddings", "None"))
227
228
        return hashlib.sha256(str(factors).encode()).hexdigest()

229
230
231
232
233
234
235
236
237
    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,
238
        hf_config_path: Optional[str] = None,
239
240
241
        allowed_local_media_path: str = "",
        revision: Optional[str] = None,
        code_revision: Optional[str] = None,
242
        rope_scaling: Optional[dict[str, Any]] = None,
243
244
245
246
247
248
249
250
251
        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,
252
        disable_cascade_attn: bool = False,
253
        skip_tokenizer_init: bool = False,
254
        served_model_name: Optional[Union[str, list[str]]] = None,
255
256
257
258
        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,
259
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
260
        disable_mm_preprocessor_cache: bool = False,
261
        override_neuron_config: Optional[dict[str, Any]] = None,
262
263
        override_pooler_config: Optional["PoolerConfig"] = None,
        logits_processor_pattern: Optional[str] = None,
264
        generation_config: str = "auto",
265
        enable_sleep_mode: bool = False,
266
        override_generation_config: Optional[dict[str, Any]] = None,
267
        model_impl: Union[str, ModelImpl] = ModelImpl.AUTO,
268
    ) -> None:
269
        self.model = model
270
        self.hf_config_path = hf_config_path
271
        self.tokenizer = tokenizer
272
        self.tokenizer_mode = tokenizer_mode
273
        self.trust_remote_code = trust_remote_code
274
        self.allowed_local_media_path = allowed_local_media_path
275
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
276
        self.revision = revision
277
        self.code_revision = code_revision
278
279
        self.rope_scaling = rope_scaling
        self.rope_theta = rope_theta
280
        self.model_impl = model_impl
281
282
283

        if hf_overrides is None:
            hf_overrides = {}
284
285
286
287
288
289

        if callable(hf_overrides):
            hf_overrides_kw = {}
            hf_overrides_fn = hf_overrides
        else:
            hf_overrides_kw = hf_overrides
290
            hf_overrides_fn = None
291

292
        if rope_scaling is not None:
293
            hf_override: dict[str, Any] = {"rope_scaling": rope_scaling}
294
            hf_overrides_kw.update(hf_override)
295
296
297
298
            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}'`")
299
300
301
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
        if rope_theta is not None:
            hf_override = {"rope_theta": rope_theta}
302
            hf_overrides_kw.update(hf_override)
303
304
305
306
            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}'`")
307
308
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

309
310
        self.maybe_pull_model_tokenizer_for_s3(model, tokenizer)

311
312
313
314
315
316
317
318
        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.")

319
320
321
322
323
        # 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
324
        self.quantization = quantization
325
        self.enforce_eager = enforce_eager
326
        self.max_seq_len_to_capture = max_seq_len_to_capture
327
        self.max_logprobs = max_logprobs
328
        self.disable_sliding_window = disable_sliding_window
329
        self.disable_cascade_attn = disable_cascade_attn
330
        self.skip_tokenizer_init = skip_tokenizer_init
331
332
333
334
335
336
        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.")
337

338
339
340
        hf_config = get_config(self.hf_config_path or self.model,
                               trust_remote_code, revision, code_revision,
                               config_format)
341
342
343
344
345
346
347
348

        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)

349
350
        self.hf_config = hf_config

351
        self.hf_text_config = get_hf_text_config(self.hf_config)
352
        self.encoder_config = self._get_encoder_config()
353
354
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, revision)
355
        self.dtype = _get_and_verify_dtype(self.hf_config, dtype)
356
        self.use_async_output_proc = use_async_output_proc
357
        self.mm_processor_kwargs = mm_processor_kwargs
358
        self.disable_mm_preprocessor_cache = disable_mm_preprocessor_cache
Woosuk Kwon's avatar
Woosuk Kwon committed
359

360
361
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
362
363
            self.enforce_eager = False

364
        interleaved_attn_models = ["gemma2", "gemma3_text", "cohere2"]
365
366
367
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
        has_interleaved_attention = (sliding_window is not None) and (
            isinstance(sliding_window, list) or
368
            (self.hf_text_config.model_type in interleaved_attn_models))
369
370

        if (not self.disable_sliding_window and has_interleaved_attention):
371
372
            if (backend :=
                    envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"):
373
374
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
375

376
                logger.warning_once(
377
378
                    f"{self.hf_text_config.model_type} has interleaved "
                    "attention, which is currently not supported by the "
379
                    f"{backend} backend. Disabling sliding window and capping "
380
381
382
383
384
385
386
387
388
389
390
391
                    "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
392

393
394
395
396
        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,
397
            sliding_window_len=self.get_hf_config_sliding_window(),
398
399
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
400
401
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
402
403
        self.multimodal_config = self._init_multimodal_config(
            limit_mm_per_prompt)
404
405
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
406

407
        self.is_attention_free = self._init_attention_free()
408
        self.is_hybrid = self._init_is_hybrid()
409
410
        self.has_inner_state = self._init_has_inner_state()

411
412
413
414
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
415

416
        supported_tasks, task = self._resolve_task(task)
417
418
        self.supported_tasks = supported_tasks
        self.task: Final = task
419
420
421
422
        if self.task in ("draft", "generate"):
            self.truncation_side = "left"
        else:
            self.truncation_side = "right"
423

424
        self.pooler_config = self._init_pooler_config(override_pooler_config)
425
        self.logits_processor_pattern = logits_processor_pattern
426

427
        self.generation_config = generation_config
428
        self.override_generation_config = override_generation_config or {}
429

430
        self._verify_quantization()
431
        self._verify_cuda_graph()
432
        self._verify_bnb_config()
433

434
435
436
437
438
439
440
441
    @property
    def registry(self):
        return ModelRegistry

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

442
443
444
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
        """
445
        Pull the model config or tokenizer to a temporary
446
447
448
449
450
451
452
453
454
        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):
455
                s3_model = S3Model()
456
457
                s3_model.pull_files(
                    model, allow_pattern=["*.model", "*.py", "*.json"])
458
                self.model_weights = self.model
459
                self.model = s3_model.dir
460
461

            if is_s3(tokenizer):
462
463
                s3_tokenizer = S3Model()
                s3_tokenizer.pull_files(
464
                    model, ignore_pattern=["*.pt", "*.safetensors", "*.bin"])
465
                self.tokenizer = s3_tokenizer.dir
466

467
468
469
    def _init_multimodal_config(
        self, limit_mm_per_prompt: Optional[Mapping[str, int]]
    ) -> Optional["MultiModalConfig"]:
470
        if self.registry.is_multimodal_model(self.architectures):
471
            return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
472
473
474
475
476
477

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

        return None
478

479
480
481
482
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

483
484
    def _init_pooler_config(
        self,
485
        override_pooler_config: Optional["PoolerConfig"],
486
    ) -> Optional["PoolerConfig"]:
487

488
        if self.runner_type == "pooling":
489
490
491
492
493
494
495
496
497
498
499
            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

500
501
        return None

502
    def _init_attention_free(self) -> bool:
503
        return self.registry.is_attention_free_model(self.architectures)
504

505
    def _init_is_hybrid(self) -> bool:
506
        return self.registry.is_hybrid_model(self.architectures)
507

508
    def _init_has_inner_state(self) -> bool:
509
        return self.registry.model_has_inner_state(self.architectures)
510

511
512
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
513
        if tokenizer_mode not in ["auto", "slow", "mistral", "custom"]:
514
515
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
516
                "either 'auto', 'slow', 'mistral' or 'custom'.")
517
        self.tokenizer_mode = tokenizer_mode
518

519
520
    def _get_preferred_task(
        self,
521
522
        architectures: list[str],
        supported_tasks: set[_ResolvedTask],
523
524
525
526
    ) -> Optional[_ResolvedTask]:
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
527
        if self.registry.is_cross_encoder_model(architectures):
528
            return "score"
529
        if self.registry.is_transcription_model(architectures):
530
            return "transcription"
531

532
        suffix_to_preferred_task: list[tuple[str, _ResolvedTask]] = [
533
534
535
536
537
538
539
540
541
            # Other models follow this pattern
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ForSequenceClassification", "classify"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]
542
        _, arch = self.registry.inspect_model_cls(architectures)
543
544
545
546
547
548
549

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

        return None

550
551
    def _resolve_task(
        self,
552
        task_option: Union[TaskOption, Literal["draft"]],
553
    ) -> tuple[set[_ResolvedTask], _ResolvedTask]:
554
555
556
        if task_option == "draft":
            return {"draft"}, "draft"

557
558
        registry = self.registry
        architectures = self.architectures
559

560
        runner_support: dict[RunnerType, bool] = {
561
562
            # NOTE: Listed from highest to lowest priority,
            # in case the model supports multiple of them
563
564
565
            "transcription": registry.is_transcription_model(architectures),
            "generate": registry.is_text_generation_model(architectures),
            "pooling": registry.is_pooling_model(architectures),
566
        }
567
        supported_runner_types_lst: list[RunnerType] = [
568
569
570
571
572
            runner_type
            for runner_type, is_supported in runner_support.items()
            if is_supported
        ]

573
        supported_tasks_lst: list[_ResolvedTask] = [
574
575
            task for runner_type in supported_runner_types_lst
            for task in _RUNNER_TASKS[runner_type]
576
577
578
579
580
        ]
        supported_tasks = set(supported_tasks_lst)

        if task_option == "auto":
            selected_task = next(iter(supported_tasks_lst))
581

582
583
584
585
586
            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
587

588
589
590
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
591
        else:
592
593
594
595
596
597
598
599
600
601
602
603
604
605
            # 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"

606
607
608
609
610
611
612
            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
613

614
        return supported_tasks, selected_task
615

616
617
618
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
619
            # compressed-tensors uses a "compression_config" key
620
            quant_cfg = getattr(self.hf_config, "compression_config", None)
621
622
        return quant_cfg

623
    def _verify_quantization(self) -> None:
624
        supported_quantization = QUANTIZATION_METHODS
625
        optimized_quantization_methods = [
626
627
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
            "awq_marlin", "fbgemm_fp8", "compressed_tensors",
628
            "compressed-tensors", "experts_int8", "quark", "nvfp4"
629
        ]
630
631
632
633
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
634
635
        quant_cfg = self._parse_quant_hf_config()

636
637
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
638
639

            # Detect which checkpoint is it
640
641
            for name in QUANTIZATION_METHODS:
                method = get_quantization_config(name)
642
643
644
645
646
647
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
648

649
            # Verify quantization configurations.
650
            if self.quantization is None:
651
652
                self.quantization = quant_method
            elif self.quantization != quant_method:
653
654
                raise ValueError(
                    "Quantization method specified in the model config "
655
                    f"({quant_method}) does not match the quantization "
656
657
658
659
660
661
662
663
                    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}.")
664
            from vllm.platforms import current_platform
665
            current_platform.verify_quantization(self.quantization)
666
            if self.quantization not in optimized_quantization_methods:
667
                logger.warning(
668
                    "%s quantization is not fully "
669
                    "optimized yet. The speed can be slower than "
670
                    "non-quantized models.", self.quantization)
671

672
    def _verify_cuda_graph(self) -> None:
673
674
675
676
        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)
677

678
679
    def _verify_bnb_config(self) -> None:
        """
680
        The current version of bitsandbytes (0.44.0) with 8-bit models does not
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
        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

700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
    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.")

717
718
719
720
721
722
723
724
725
726
    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

727
        # Reminder: Please update docs/source/features/compatibility_matrix.md
728
        # If the feature combo become valid
729
        from vllm.platforms import current_platform
730
        if not current_platform.is_async_output_supported(self.enforce_eager):
731
732
733
734
735
736
737
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            self.use_async_output_proc = False
            return

738
        # Async postprocessor is not necessary for pooling models
739
        # since there is no token generation
740
        if self.runner_type == "pooling":
741
742
            self.use_async_output_proc = False

743
        # Reminder: Please update docs/source/features/compatibility_matrix.md
744
        # If the feature combo become valid
745
746
747
        if speculative_config:
            self.use_async_output_proc = False

748
749
750
751
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
752
753
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
754
755
756
757
758
759
760
        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}).")

761
        if parallel_config.enable_expert_parallel:
762
763
            self._verify_with_expert_parallelism()

764
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
765
        if pipeline_parallel_size > 1:
766
            if not self.registry.is_pp_supported_model(self.architectures):
767
768
769
770
771
772
                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
773

774
    def get_hf_config_sliding_window(
775
            self) -> Union[Optional[int], list[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
776
        """Get the sliding window size, or None if disabled."""
777
778
779
780

        # 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.
781
782
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
783
            return None
784
        return getattr(self.hf_text_config, "sliding_window", None)
785

786
    def get_sliding_window(self) -> Optional[Union[int, list[Optional[int]]]]:
787
788
789
790
791
792
793
794
        """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()

795
    def get_vocab_size(self) -> int:
796
        return self.hf_text_config.vocab_size
797

798
    def get_hidden_size(self) -> int:
799
        return self.hf_text_config.hidden_size
800

801
802
    @property
    def is_deepseek_mla(self) -> bool:
803
804
        return (hasattr(self.hf_text_config, "model_type")) \
                and (self.hf_text_config.model_type in \
805
                    ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp'))\
806
                and (self.hf_text_config.kv_lora_rank is not None)
807

808
    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
809
        # TODO remove hard code
810
        if self.is_deepseek_mla:
811
812
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
813
            if self.use_mla:
814
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
815
816
817
818
819
            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
820

821
822
823
824
825
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      == "zamba2"):
            return self.hf_text_config.attention_head_dim

826
827
828
        if self.is_attention_free:
            return 0

829
830
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
831
        # FIXME(woosuk): This may not be true for all models.
832
833
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
834

835
836
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
837
        # For GPTBigCode & Falcon:
838
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
839
840
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
841
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
842
        new_decoder_arch_falcon = (
843
            self.hf_config.model_type in falcon_model_types
844
            and getattr(self.hf_config, "new_decoder_architecture", False))
845
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
846
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
847
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
848
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
849
            return 1
850

851
        # For DBRX and MPT
852
853
854
855
856
        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":
857
858
859
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

860
861
862
        if self.is_attention_free:
            return 0

863
864
865
866
867
868
869
870
871
872
        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:
873
            num_kv_heads = getattr(self.hf_text_config, attr, None)
874
875
876
877
878
            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.
879
        return self.hf_text_config.num_attention_heads
880
881
882

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
883
884
885
886
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

887
888
889
890
891
892
893
        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)
894

895
896
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
897
898
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
899

900
    def get_layers_start_end_indices(
901
            self, parallel_config: "ParallelConfig") -> tuple[int, int]:
902
        from vllm.distributed.utils import get_pp_indices
903
904
905
906
907
908
        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)
909
910
911
        # the layout order is: DP x PP x TP
        pp_rank = (parallel_config.rank // parallel_config.tensor_parallel_size
                   ) % parallel_config.pipeline_parallel_size
912
913
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
914
        return start, end
Mor Zusman's avatar
Mor Zusman committed
915

916
917
918
    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
919

920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
    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:
944
945
                raise ValueError("The model is an hybrid without a "
                                 "layers_block_type in the hf_config, "
946
947
948
                                 "cannot determine the num of "
                                 f"{block_type.value} layers")

949
950
951
952
953
954
955
956
957
            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)

958
959
            return sum(t == block_type.value
                       for t in layers_block_type_value[start:end])
Mor Zusman's avatar
Mor Zusman committed
960

961
962
963
964
965
966
967
968
969
970
971
972
    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

973
    def try_get_generation_config(self) -> dict[str, Any]:
974
        if self.generation_config in ("auto", "vllm"):
975
            config = try_get_generation_config(
976
                self.hf_config_path or self.model,
977
978
979
980
981
982
983
984
985
986
987
988
989
990
                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()

991
    def get_diff_sampling_param(self) -> dict[str, Any]:
992
        """
993
        This method returns a dictionary containing the parameters
994
995
        that differ from the default sampling parameters. If
        `generation_config` is `"vllm"`, an empty dictionary is returned.
996
997

        Returns:
998
            dict[str, Any]: A dictionary with the differing sampling
999
            parameters, if `generation_config` is `"vllm"` an empty dictionary.
1000
        """
1001
        if self.generation_config == "vllm":
1002
1003
1004
1005
1006
1007
1008
            config = {}
        else:
            config = self.try_get_generation_config()

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

1009
1010
1011
1012
1013
1014
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
1015
            "max_new_tokens",
1016
1017
1018
1019
1020
1021
        ]
        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
            }
1022
1023
1024
1025
1026
            # 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")
1027
1028
        else:
            diff_sampling_param = {}
1029
1030
1031
1032
1033
1034
1035

        if diff_sampling_param:
            logger.warning_once(
                "Default sampling parameters have been overridden by the "
                "model's Hugging Face generation config recommended from the "
                "model creator. If this is not intended, please relaunch "
                "vLLM instance with `--generation-config vllm`.")
1036
1037
        return diff_sampling_param

1038
    @property
1039
    def is_encoder_decoder(self) -> bool:
1040
        """Extract the HF encoder/decoder model flag."""
1041
1042
1043
1044
1045
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
1046

1047
1048
1049
1050
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

1051
1052
    @property
    def is_cross_encoder(self) -> bool:
1053
        return self.registry.is_cross_encoder_model(self.architectures)
1054

1055
1056
    @property
    def use_mla(self) -> bool:
1057
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
1058

1059
    @property
1060
    def supported_runner_types(self) -> set[RunnerType]:
1061
1062
1063
1064
1065
1066
        return {_TASK_RUNNER[task] for task in self.supported_tasks}

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

1067
1068
1069
1070
1071
    @property
    def is_v1_compatible(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_v1_compatible(architectures)

1072
1073

class CacheConfig:
1074
1075
1076
1077
1078
    """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
1079
            vLLM execution.
1080
        swap_space: Size of the CPU swap space per GPU (in GiB).
1081
        cache_dtype: Data type for kv cache storage.
1082
        is_attention_free: Whether the model is attention-free.
1083
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
1084
            profiled num_gpu_blocks if specified. Does nothing if None.
1085
1086
1087
1088
        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.
1089
    """
1090

1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
    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.
        """
1103
        factors: list[Any] = []
1104
1105
1106
1107
1108
        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

1109
1110
1111
1112
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
1113
        swap_space: float,
1114
        cache_dtype: str,
1115
        is_attention_free: bool = False,
1116
        num_gpu_blocks_override: Optional[int] = None,
1117
        sliding_window: Optional[int] = None,
1118
        enable_prefix_caching: bool = False,
1119
        cpu_offload_gb: float = 0,
1120
        calculate_kv_scales: Optional[bool] = None,
1121
1122
1123
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
1124
        self.swap_space_bytes = swap_space * GiB_bytes
1125
        self.num_gpu_blocks_override = num_gpu_blocks_override
1126
        self.cache_dtype = cache_dtype
1127
        self.is_attention_free = is_attention_free
1128
        self.sliding_window = sliding_window
1129
        self.enable_prefix_caching = enable_prefix_caching
1130
        self.cpu_offload_gb = cpu_offload_gb
1131
        self.calculate_kv_scales = calculate_kv_scales
1132
        self._verify_args()
1133
        self._verify_cache_dtype()
1134
        self._verify_prefix_caching()
1135
1136

        # Will be set after profiling.
1137
1138
        self.num_gpu_blocks: Optional[int] = None
        self.num_cpu_blocks: Optional[int] = None
1139

1140
1141
1142
1143
        # Set calculate_kv_scales to False if the value is unset.
        if self.calculate_kv_scales is None:
            self.calculate_kv_scales = False

1144
    def metrics_info(self):
1145
1146
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1147
1148
        return {key: str(value) for key, value in self.__dict__.items()}

1149
    def _verify_args(self) -> None:
1150
1151
1152
1153
        if self.cpu_offload_gb < 0:
            raise ValueError("CPU offload space must be non-negative"
                             f", but got {self.cpu_offload_gb}")

1154
1155
1156
1157
1158
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

1159
1160
1161
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1162
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
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
    FASTSAFETENSORS = "fastsafetensors"
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302


@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.
1303
            "bitsandbytes" will load nf4 type weights.
1304
1305
1306
1307
1308
1309
            "sharded_state" will load weights from pre-sharded checkpoint files,
                supporting efficient loading of tensor-parallel models.
            "gguf" will load weights from GGUF format files.
            "mistral" will load weights from consolidated safetensors files used
                by Mistral models.
            "runai_streamer" will load weights from RunAI streamer format files.
1310
        model_loader_extra_config: The extra config for the model loader.
1311
        ignore_patterns: The list of patterns to ignore when loading the model.
1312
            Default to "original/**/*" to avoid repeated loading of llama's
1313
            checkpoints.
1314
1315
        use_tqdm_on_load: Whether to enable tqdm for showing progress bar during
            loading. Default to True
1316
1317
1318
1319
1320
1321
    """

    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)
1322
    ignore_patterns: Optional[Union[list[str], str]] = None
1323
    use_tqdm_on_load: bool = True
1324

1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
    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.
1339
        factors: list[Any] = []
1340
1341
1342
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1343
1344
1345
1346
1347
    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)
1348
1349
1350
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1351

1352
1353
1354
1355
1356
1357
1358
        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/**/*"]

1359

1360
@dataclass
1361
class ParallelConfig:
1362
    """Configuration for the distributed execution."""
1363

1364
1365
    pipeline_parallel_size: int = 1  # Number of pipeline parallel groups.
    tensor_parallel_size: int = 1  # Number of tensor parallel groups.
1366
1367
1368
1369
1370
    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.
1371
    enable_expert_parallel: bool = False  # Use EP instead of TP for MoE layers.
1372

1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
    # 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,
1398
                                                 type["ExecutorBase"]]] = None
1399
1400
1401
1402

    # 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"
1403
    sd_worker_cls: str = "auto"
1404
    worker_extension_cls: str = ""
1405

1406
    # world_size is TPxPP, it affects the number of workers we create.
1407
    world_size: int = field(init=False)
1408
1409
1410
    # world_size_across_dp is TPxPPxDP, it is the size of the world
    # including data parallelism.
    world_size_across_dp: int = field(init=False)
1411
1412
1413

    rank: int = 0

1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
    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
1443
                          has_unfinished: bool) -> bool:
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
        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

1455
1456
1457
1458
1459
1460
1461
1462
    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.
        """
1463
        factors: list[Any] = []
1464
1465
1466
1467
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
        return hashlib.sha256(str(factors).encode()).hexdigest()

1468
1469
1470
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size
1471
1472
1473
1474
1475
1476

        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
1477

1478
1479
1480
1481
1482
        if self.distributed_executor_backend == "external_launcher":
            import os
            os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
            logger.info("Disabling V1 multiprocessing for external launcher.")

1483
        ray_only_devices: list[str] = []
1484
        from vllm.platforms import current_platform
1485
1486
        if (current_platform.device_type in ray_only_devices
                and self.world_size > 1):
1487
1488
1489
1490
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
1491
1492
                    f"{current_platform.device_type.upper()} backend only "
                    "supports Ray for distributed inference.")
1493

1494
        if self.distributed_executor_backend is None and self.world_size > 1:
1495
1496
1497
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1498
            from vllm.executor import ray_utils
1499
            backend = "mp"
1500
            ray_found = ray_utils.ray_is_available()
1501
1502
1503
1504
1505
            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):
1506
1507
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1508
1509
1510
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1511
1512
                backend = "ray"
            elif ray_found:
1513
                if self.placement_group:
1514
                    backend = "ray"
1515
1516
1517
1518
1519
1520
                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"
1521
1522
1523
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1524

1525
1526
1527
        if self.distributed_executor_backend is None and self.world_size == 1:
            self.distributed_executor_backend = "uni"

1528
1529
        self._verify_args()

1530
1531
1532
1533
1534
1535
    @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)

1536
    def _verify_args(self) -> None:
1537
1538
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
1539
        from vllm.platforms import current_platform
1540
        if self.distributed_executor_backend not in (
1541
1542
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
1543
1544
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
1545
            raise ValueError(
1546
1547
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
1548
1549
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
1550
        if self.use_ray:
1551
1552
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1553
        if current_platform.is_rocm():
1554
1555
1556
1557
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
1558
        if self.ray_workers_use_nsight and not self.use_ray:
1559
1560
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1561

1562
1563
1564
        assert isinstance(self.worker_extension_cls, str), (
            "worker_extension_cls must be a string (qualified class name).")

1565

1566
@dataclass
1567
class SchedulerConfig:
1568
    """Scheduler configuration."""
1569

1570
    runner_type: str = "generate"  # The runner type to launch for the model.
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580

    # 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

1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
    # 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

1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
    # 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
1607

1608
1609
1610
1611
1612
1613
    # 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
1614
1615

    # Multimodal encoder cache size, only used in V1
1616
    encoder_cache_size: int = field(default=None)  # type: ignore
1617

1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
    # 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)

1641
1642
    # scheduler class or path. "vllm.core.scheduler.Scheduler" (default)
    # or "mod.custom_class".
1643
    scheduler_cls: Union[str, type[object]] = "vllm.core.scheduler.Scheduler"
1644

1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
    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.
1659
        factors: list[Any] = []
1660
1661
1662
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1663
1664
1665
1666
    def __post_init__(self) -> None:
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
1667
1668
1669
1670
                    # 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.
1671
1672
                    self.max_num_batched_tokens = max(
                        self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
1673
                else:
1674
1675
                    self.max_num_batched_tokens = (
                        _DEFAULT_MAX_NUM_BATCHED_TOKENS)
1676
            else:
1677
1678
                # If max_model_len is too short, use
                # _DEFAULT_MAX_NUM_BATCHED_TOKENS as the default value
1679
                # for higher throughput.
1680
1681
                self.max_num_batched_tokens = max(
                    self.max_model_len, _DEFAULT_MAX_NUM_BATCHED_TOKENS)
1682

1683
1684
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
1685
1686
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1687
                    _POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
1688
                )
1689
            if self.is_multimodal_model:
1690
                # The value needs to be at least the number of multimodal tokens
1691
1692
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1693
1694
1695
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

1696
1697
1698
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

1699
        if self.enable_chunked_prefill:
1700
1701
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
1702
                self.max_num_batched_tokens)
1703

1704
        self.chunked_prefill_enabled = self.enable_chunked_prefill
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
        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)

1717
1718
1719
        self._verify_args()

    def _verify_args(self) -> None:
1720
1721
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
1722
1723
1724
1725
1726
1727
1728
            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.")
1729

1730
1731
1732
1733
1734
        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}).")
1735

1736
1737
1738
1739
1740
1741
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

1742
1743
1744
1745
1746
1747
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
        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}).")

1771
1772
1773
1774
    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

1775

1776
class DeviceConfig:
1777
    device: Optional[torch.device]
1778
    device_type: str
1779

1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
    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.
1795
        factors: list[Any] = []
1796
1797
1798
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1799
1800
1801
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
1802
            from vllm.platforms import current_platform
1803
            self.device_type = current_platform.device_type
1804
            if not self.device_type:
1805
                raise RuntimeError("Failed to infer device type")
1806
1807
1808
1809
1810
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
1811
        if self.device_type in ["neuron"]:
1812
            self.device = torch.device("cpu")
1813
1814
        elif self.device_type in ["tpu"]:
            self.device = None
1815
1816
1817
1818
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1819

1820
@dataclass
1821
class SpeculativeConfig:
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
    """
    Configuration for speculative decoding.
    Configurable parameters include:
    - General Speculative Decoding Control:
        - num_speculative_tokens (int): The number of speculative
            tokens, if provided. It will default to the number in the draft
            model config if present, otherwise, it is required.
        - model (Optional[str]): The name of the draft model, eagle head,
            or additional weights, if provided.
        - method (Optional[str]): The name of the speculative method to use.
            If users provide and set the `model` param, the speculative method
            type will be detected automatically if possible, if `model` param
            is not provided, the method name must be provided.
            - Possible values:
                - ngram
                    Related additional configuration:
                    - prompt_lookup_max (Optional[int]):
                        Maximum size of ngram token window when using Ngram
                        proposer, required when method is set to ngram.
                    - prompt_lookup_min (Optional[int]):
                        Minimum size of ngram token window when using Ngram
                        proposer, if provided. Defaults to 1.
                - eagle
                - medusa
                - mlp_speculator
                - draft_model
        - 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. If not specified, it
            defaults to 'rejection_sampler'.
            - Possible values:
                - rejection_sampler
                - typical_acceptance_sampler
                    Related additional configuration:
                    - 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 accepted. This threshold is used only
                        when we use the TypicalAcceptanceSampler for token
                        acceptance.
                    - posterior_alpha (Optional[float]):
                        Scaling factor for entropy-based threshold, applied
                        when using TypicalAcceptanceSampler.
        - draft_tensor_parallel_size (Optional[int]): The degree of the tensor
            parallelism for the draft model. Can only be 1 or the same as the
            target model's tensor parallel size.
        - disable_logprobs (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.

    - Draft Model Configuration:
        - quantization (Optional[str]): Quantization method that was used to
            quantize the draft model weights. If None, we assume the
            model weights are not quantized. Note that it only takes effect
            when using the draft model-based speculative method.
        - max_model_len (Optional[int]): The maximum model length of the
            draft model. Used when testing the ability to skip
            speculation for some sequences.
        - revision: The specific model version to use for the draft model. It
            can be a branch name, a tag name, or a commit id. If unspecified,
            will use the default version.
        - code_revision: The specific revision to use for the draft model code
            on Hugging Face Hub. It can be a branch name, a tag name, or a
            commit id. If unspecified, will use the default version.
1888

1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
    - Advanced Control:
        - disable_mqa_scorer (bool): Disable the MQA scorer and fall back to
            batch expansion for scoring proposals. If not specified, it
            defaults to False.
        - disable_by_batch_size (Optional[int]): Disable speculative decoding
            for new incoming requests when the number of enqueued requests is
            larger than this value, if provided.

    Although the parameters above are structured hierarchically, there is no
    need to nest them during configuration.

    Non-configurable internal parameters include:
    - Model Configuration:
        - target_model_config (ModelConfig): The configuration of the target
            model.
        - draft_model_config (ModelConfig): The configuration of the draft
            model initialized internal.
    - Parallelism Configuration:
        - target_parallel_config (ParallelConfig): The parallel configuration
            for the target model.
        - draft_parallel_config (ParallelConfig): The parallel configuration
            for the draft model initialized internal.
    - Execution Control:
        - enable_chunked_prefill (bool): Whether vLLM is configured to use
            chunked prefill or not. Used for raising an error since it's not
            yet compatible with speculative decode.
        - disable_log_stats (bool): Whether to disable the periodic printing of
            stage times in speculative decoding.
1917
    """
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
    # speculative configs from cli args
    num_speculative_tokens: int = field(default=None,
                                        init=True)  # type: ignore
    method: Optional[str] = None
    acceptance_method: str = "rejection_sampler"
    draft_tensor_parallel_size: Optional[int] = None
    disable_logprobs: bool = True

    model: Optional[str] = None
    quantization: Optional[str] = None
    max_model_len: Optional[int] = None
    revision: Optional[str] = None
    code_revision: Optional[str] = None

    disable_mqa_scorer: bool = False
    disable_by_batch_size: Optional[int] = None
    prompt_lookup_max: Optional[int] = None
    prompt_lookup_min: Optional[int] = None
    posterior_threshold: Optional[float] = None
    posterior_alpha: Optional[float] = None

    # required configuration params passed from engine
    target_model_config: ModelConfig = field(default=None,
                                             init=True)  # type: ignore
    target_parallel_config: ParallelConfig = field(default=None,
                                                   init=True)  # type: ignore
    enable_chunked_prefill: bool = field(default=None,
                                         init=True)  # type: ignore
    disable_log_stats: bool = field(default=None, init=True)  # type: ignore

    # params generated in the post-init stage
    draft_model_config: ModelConfig = field(default=None,
                                            init=True)  # type: ignore
    draft_parallel_config: ParallelConfig = field(default=None,
                                                  init=True)  # type: ignore
1953

1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
    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.
1968
        factors: list[Any] = []
1969
1970
1971
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1972
1973
1974
1975
1976
    @classmethod
    def from_dict(cls, dict_value: dict) -> "SpeculativeConfig":
        """Parse the CLI value for the speculative config."""
        return cls(**dict_value)

1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
    @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

1989
    def __post_init__(self):
1990

1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
        # Note: After next release, the method parameter will be used to
        # specify the speculative method, which helps to extend the
        # configuration of non-model-based proposers, and the model parameter
        # will be used when the draft model or head is needed.
        # If users do not specify the method, the speculative method will
        # be detected automatically if possible. If the speculative method can
        # not be detected, it will be considered as the draft-model-based
        # method by default.

        if self.model is None and self.num_speculative_tokens is not None:
            # TODO(Shangming): Refactor mtp configuration logic when supporting
            # mtp acceleration for more models besides deepseek_v3
            if self.target_model_config.hf_text_config.model_type \
2004
                        == "deepseek_v3":
2005
2006
2007
2008
                # use the draft model from the same model:
                self.model = self.target_model_config.model
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
2009
            else:
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative model.")

        # Automatically configure the ngram method during configuration
        # refactoring to ensure a smooth transition.
        if self.method is None and (self.model is not None
                                    and self.model in ("ngram", "[ngram]")):
            self.method = "ngram"

        if self.method in ("ngram", "[ngram]"):
            # Unified to "ngram" internally
            self.method = "ngram"
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
            # Set default values if not provided
            if (self.prompt_lookup_min is None
                    and self.prompt_lookup_max is None):
                # TODO(woosuk): Tune these values. They are arbitrarily chosen.
                self.prompt_lookup_min = 5
                self.prompt_lookup_max = 5
            elif self.prompt_lookup_min is None:
                assert self.prompt_lookup_max is not None
                self.prompt_lookup_min = self.prompt_lookup_max
            elif self.prompt_lookup_max is None:
                assert self.prompt_lookup_min is not None
                self.prompt_lookup_max = self.prompt_lookup_min

            # Validate values
2036
            if self.prompt_lookup_min < 1:
2037
2038
2039
2040
2041
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must be > 0")
            if self.prompt_lookup_max < 1:
                raise ValueError(
                    f"prompt_lookup_max={self.prompt_lookup_max} must be > 0")
2042
            if self.prompt_lookup_min > self.prompt_lookup_max:
2043
2044
2045
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}")
2046

2047
2048
2049
            # 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.
2050
2051
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
2052
        else:
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                self.draft_model_config = ModelConfig(
                    model=self.model,
                    task="draft",
                    tokenizer=self.target_model_config.tokenizer,
                    tokenizer_mode=self.target_model_config.tokenizer_mode,
                    trust_remote_code=self.target_model_config.
                    trust_remote_code,
                    allowed_local_media_path=self.target_model_config.
                    allowed_local_media_path,
                    dtype=self.target_model_config.dtype,
                    seed=self.target_model_config.seed,
                    revision=self.revision,
                    code_revision=self.code_revision,
                    tokenizer_revision=self.target_model_config.
                    tokenizer_revision,
                    max_model_len=None,
                    spec_target_max_model_len=self.target_model_config.
                    max_model_len,
                    quantization=self.quantization,
                    enforce_eager=self.target_model_config.enforce_eager,
                    max_seq_len_to_capture=self.target_model_config.
                    max_seq_len_to_capture,
                    max_logprobs=self.target_model_config.max_logprobs,
                    hf_overrides=SpeculativeConfig.hf_config_override,
                )
2082

2083
2084
2085
2086
2087
2088
2089
2090
                # Automatically detect the method
                if "eagle-" in self.draft_model_config.model.lower():
                    self.method = "eagle"
                elif self.draft_model_config.hf_config.model_type == "medusa":
                    self.method = "medusa"
                elif (self.draft_model_config.hf_config.model_type ==
                      "mlp_speculator"):
                    self.method = "mlp_speculator"
2091
                else:
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
                    self.method = "draft_model"

                # Replace hf_config for EAGLE draft_model
                if self.method == "eagle":
                    if self.enable_chunked_prefill:
                        raise ValueError(
                            "Chunked prefill and EAGLE are not compatible.")

                    from vllm.transformers_utils.configs.eagle import (
                        EAGLEConfig)
                    if isinstance(self.draft_model_config.hf_config,
                                  EAGLEConfig):
                        pass
                    else:
                        eagle_config = EAGLEConfig(
                            self.draft_model_config.hf_config)
                        self.draft_model_config.hf_config = eagle_config

                if (self.num_speculative_tokens is not None
                        and hasattr(self.draft_model_config.hf_config,
                                    "num_lookahead_tokens")):
                    self.draft_model_config.hf_config.num_lookahead_tokens = \
                    self.num_speculative_tokens

                n_predict = getattr(self.draft_model_config.hf_config,
                                    "n_predict", None)
                if n_predict is not None:
                    if self.num_speculative_tokens is None:
                        # Default to max value defined in draft model config.
                        self.num_speculative_tokens = n_predict
                    elif self.num_speculative_tokens > n_predict and \
                            self.num_speculative_tokens % n_predict != 0:
                        # Ensure divisibility for MTP module reuse.
                        raise ValueError(
                            f"num_speculative_tokens:{self.num_speculative_tokens}"
                            f" must be divisible by {n_predict=}")

                self.draft_tensor_parallel_size = \
                    SpeculativeConfig._verify_and_get_draft_tp(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size,
                        self.draft_model_config.hf_config
                )
2135

2136
2137
2138
2139
2140
2141
                self.draft_model_config.max_model_len = (
                    SpeculativeConfig._maybe_override_draft_max_model_len(
                        self.max_model_len,
                        self.draft_model_config.max_model_len,
                        self.target_model_config.max_model_len,
                    ))
2142

2143
2144
2145
2146
                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size))
2147

2148
2149
2150
2151
2152
        if self.acceptance_method == "typical_acceptance_sampler":
            if self.posterior_threshold is None:
                self.posterior_threshold = 0.09
            if self.posterior_alpha is None:
                self.posterior_alpha = 0.3
2153

2154
        self._verify_args()
2155

2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
    @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,
        )

2191
    @staticmethod
2192
    def _verify_and_get_draft_tp(
2193
2194
2195
2196
2197
2198
            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.
2199
        """
2200
2201
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
2202
        if speculative_draft_tensor_parallel_size is None:
2203
2204
2205
2206
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
2207
2208
2209
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
                        draft_hf_config.model_type)
2210
2211
2212
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
2213
2214
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
2215
            raise ValueError(
2216
                f"{speculative_draft_tensor_parallel_size=} cannot be "
2217
                f"other value than 1 or target model tensor_parallel_size")
2218
        return speculative_draft_tensor_parallel_size
2219

2220
2221
2222
2223
2224
2225
2226
2227
2228
    @staticmethod
    def create_draft_parallel_config(
        target_parallel_config: ParallelConfig,
        speculative_draft_tensor_parallel_size: int,
    ) -> ParallelConfig:
        """Create a parallel config for use by the draft worker.

        This is mostly a copy of the target parallel config, except the tp_size.
        """
2229
2230
2231
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
2232
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
2233
2234
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
            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 _verify_args(self) -> None:
2248
2249
2250
2251
2252
2253
        if self.num_speculative_tokens is None:
            raise ValueError(
                "num_speculative_tokens must be provided with "
                "speculative model unless the draft model config contains an "
                "n_predict parameter.")

2254
2255
2256
2257
2258
2259
2260
        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)
2261
2262
            # Validate and set draft token acceptance related settings.

2263
2264
        if self.acceptance_method is None:
            raise ValueError("acceptance_method is not set. "
2265
2266
2267
                             "Expected values are rejection_sampler or "
                             "typical_acceptance_sampler.")

2268
2269
        if (self.acceptance_method != 'rejection_sampler'
                and self.acceptance_method != 'typical_acceptance_sampler'):
2270
            raise ValueError(
2271
                "Expected acceptance_method to be either "
2272
                "rejection_sampler or typical_acceptance_sampler. Instead it "
2273
                f"is {self.acceptance_method}")
2274

2275
2276
2277
2278
        if self.acceptance_method == "typical_acceptance_sampler" and (
            (self.posterior_threshold is not None
             and self.posterior_threshold < 0) or
            (self.posterior_alpha is not None and self.posterior_alpha < 0)):
2279
            raise ValueError(
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
                "Expected the posterior_threshold and posterior_alpha of "
                "typical_acceptance_sampler to be > 0. "
                "Instead found posterior_threshold = "
                f"{self.posterior_threshold} and posterior_alpha = "
                f"{self.posterior_alpha}")

        if (self.disable_by_batch_size is not None
                and self.disable_by_batch_size < 2):
            raise ValueError("Expect the batch size threshold of disabling "
                             "speculative decoding is > 1, but got "
                             f"{self.disable_by_batch_size=}")
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302

    @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:
2303
2304
        if self.prompt_lookup_max is not None and self.prompt_lookup_max > 0:
            draft_model = "ngram"
2305
2306
        else:
            draft_model = self.draft_model_config.model
2307
2308
2309
2310
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


2311
2312
2313
2314
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
2315
    fully_sharded_loras: bool = False
2316
    max_cpu_loras: Optional[int] = None
2317
    lora_dtype: Optional[Union[torch.dtype, str]] = None
2318
2319
2320
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
2321
    long_lora_scaling_factors: Optional[tuple[float]] = None
2322
    bias_enabled: bool = False
2323

2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
    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.
        """
2336
        factors: list[Any] = []
2337
2338
2339
2340
2341
2342
2343
        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)
2344
2345
2346
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2347
    def __post_init__(self):
2348
        # Setting the maximum rank to 512 should be able to satisfy the vast
2349
        # majority of applications.
2350
        possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)
2351
        possible_lora_extra_vocab_size = (256, 512)
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
        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
2367
                f"max_loras ({self.max_loras})")
2368

2369
2370
2371
2372
2373
    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.")

2374
2375
2376
2377
2378
2379
2380
    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)

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
2381
        # Reminder: Please update docs/source/features/compatibility_matrix.md
2382
        # If the feature combo become valid
2383
        if scheduler_config.chunked_prefill_enabled:
2384
2385
            logger.warning("LoRA with chunked prefill is still experimental "
                           "and may be unstable.")
2386
2387


2388
2389
2390
2391
2392
2393
2394
@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

2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
    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.
2409
        factors: list[Any] = []
2410
2411
2412
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
    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)


2431
@dataclass
2432
class MultiModalConfig:
2433
2434
    """Controls the behavior of multimodal models."""

2435
    limit_per_prompt: Mapping[str, int] = field(default_factory=dict)
2436
    """
2437
    The maximum number of input items allowed per prompt for each modality.
2438
2439
    """

2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
    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.
2454
        factors: list[Any] = []
2455
2456
2457
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2458
2459
2460
2461
2462
2463
2464
2465
2466
    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)

2467
    # TODO: Add configs to init vision tower or not.
2468

2469

2470
2471
@dataclass
class PoolerConfig:
2472
    """Controls the behavior of output pooling in pooling models."""
2473
2474

    pooling_type: Optional[str] = None
2475
    """
2476
    The pooling method of the pooling model. This should be a key in
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
    :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
    """
2494
    If set, only the score corresponding to the ``step_tag_id`` in the
2495
2496
2497
2498
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

2499
    returned_token_ids: Optional[list[int]] = None
2500
    """
2501
2502
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2503
2504
2505
    ``math-shepherd-mistral-7b-prm`` model.
    """

2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
    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.
2520
        factors: list[Any] = []
2521
2522
2523
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2524
2525
2526
    @staticmethod
    def from_json(json_str: str) -> "PoolerConfig":
        return PoolerConfig(**json.loads(json_str))
2527
2528


2529
2530
2531
2532
2533
2534
2535
2536
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2537
_ROCM_NOT_SUPPORTED_DTYPE: list[str] = []  #
2538

2539
2540
2541

def _get_and_verify_dtype(
    config: PretrainedConfig,
2542
    dtype: Union[str, torch.dtype],
2543
2544
2545
2546
) -> 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)
2547
2548
2549
2550
2551
2552
2553
2554

    # 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)

2555
2556
2557
    if config_dtype is None:
        config_dtype = torch.float32

2558
2559
2560
2561
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
2562
2563
                # Following common practice, we use float16 for float32 models
                torch_dtype = torch.float16
2564
2565
            else:
                torch_dtype = config_dtype
2566

2567
            from vllm.platforms import current_platform
2568
2569
            if (current_platform.is_cpu()
                    and current_platform.get_cpu_architecture()
2570
                    == CpuArchEnum.POWERPC
2571
2572
2573
2574
2575
2576
2577
2578
                    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

2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
            # 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

2590
2591
            if current_platform.is_hpu() and config_dtype == torch.float16:
                logger.info(
2592
                    "For HPU, we cast models to bfloat16 instead of "
2593
2594
2595
                    "using float16 by default. Please specify `dtype` if you "
                    "want to use float16.")
                torch_dtype = torch.bfloat16
2596
        else:
2597
2598
2599
2600
2601
            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
2602
    else:
2603
        raise ValueError(f"Unknown dtype: {dtype}")
2604
2605
2606
2607
2608

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
2609
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
2610
2611
2612
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
2613
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
2614
2615
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
2616
            # Casting between float16 and bfloat16 is allowed with a warning.
2617
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
2618
2619

    return torch_dtype
2620
2621
2622
2623
2624


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
2625
    disable_sliding_window: bool,
2626
    sliding_window_len: Optional[Union[int, list[Optional[int]]]],
2627
    spec_target_max_model_len: Optional[int] = None,
2628
    encoder_config: Optional[Any] = None,
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
) -> 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",
2639
2640
        # ChatGLM2
        "seq_length",
2641
2642
        # Command-R
        "model_max_length",
2643
2644
        # Whisper
        "max_target_positions",
2645
2646
2647
2648
2649
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
2650
    # Choose the smallest "max_length" from the possible keys.
2651
    max_len_key = None
2652
    for key in possible_keys:
2653
2654
2655
2656
2657
        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)
2658
2659
2660
2661

    # 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:
2662
2663

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
2664
        max_len_key = "sliding_window" \
2665
2666
2667
            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)
2668
2669
2670

    # If none of the keys were found in the config, use a default and
    # log a warning.
2671
    if derived_max_model_len == float("inf"):
2672
2673
2674
2675
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

2676
2677
2678
2679
2680
        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

2681
2682
2683
2684
        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: "
2685
            "%s. Assuming the model's maximum length is %d.", possible_keys,
2686
            default_max_len)
2687
        derived_max_model_len = default_max_len
2688

2689
    rope_scaling = getattr(hf_config, "rope_scaling", None)
2690
2691
2692
    # 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:
2693
2694
2695
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705

        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.")

2706
2707
2708
2709
            # 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)

2710
2711
2712
2713
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
2714

2715
2716
2717
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2718
2719
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2720
    if max_model_len is None:
2721
        max_model_len = int(derived_max_model_len)
2722
    elif max_model_len > derived_max_model_len:
2723
2724
2725
2726
2727
        # 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:
2728
2729
2730
2731
2732
2733
2734
            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.")
2735
        else:
2736
            msg = (
2737
                f"User-specified max_model_len ({max_model_len}) is greater "
2738
2739
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2740
                f"{model_max_length} in model's config.json). This may lead "
2741
2742
2743
2744
2745
2746
2747
2748
2749
                "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")
2750
    return int(max_model_len)
2751
2752


2753
def get_min_sliding_window(
2754
        sliding_window: Union[int, list[Optional[int]]]) -> int:
2755
2756
2757
2758
2759
2760
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


2761
def get_served_model_name(model: str,
2762
                          served_model_name: Optional[Union[str, list[str]]]):
2763
    """
2764
2765
2766
2767
    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
2768
2769
2770
2771
2772
2773
2774
2775
2776
    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


2777
2778
2779
2780
@dataclass
class DecodingConfig:
    """Dataclass which contains the decoding strategy of the engine"""

2781
2782
2783
    # Which guided decoding algo to use.
    # 'outlines' / 'lm-format-enforcer' / 'xgrammar'
    guided_decoding_backend: str = 'xgrammar'
2784

2785
2786
    reasoning_backend: Optional[str] = None

2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
    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.
2801
        factors: list[Any] = []
2802
2803
2804
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2805
    def __post_init__(self):
2806
2807
        v0_valid_guided_backends = [
            'outlines', 'lm-format-enforcer', 'xgrammar'
2808
        ]
2809
        v1_valid_guided_backends = ['xgrammar', 'guidance', 'auto']
2810
2811
2812

        backend = GuidedDecodingParams(
            backend=self.guided_decoding_backend).backend_name
2813
2814
2815
2816
        if envs.VLLM_USE_V1:
            valid_guided_backends = v1_valid_guided_backends
        else:
            valid_guided_backends = v0_valid_guided_backends
2817
        if backend not in valid_guided_backends:
2818
            raise ValueError(f"Invalid guided_decoding_backend '{backend}',"
2819
                             f" must be one of {valid_guided_backends}")
2820
2821


2822
2823
@dataclass
class ObservabilityConfig:
2824
2825
2826
    """Configuration for observability - metrics and tracing."""
    show_hidden_metrics: bool = False

2827
2828
    otlp_traces_endpoint: Optional[str] = None

2829
2830
2831
2832
2833
2834
2835
2836
    # 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

2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
    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.
2851
        factors: list[Any] = []
2852
2853
2854
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2855
    def __post_init__(self):
2856
2857
2858
2859
2860
        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}")
2861
2862


2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
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

2896
2897
2898
    # any extra config that the connector may need
    kv_connector_extra_config: dict[str, Any] = {}

2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
    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.
2913
        factors: list[Any] = []
2914
2915
2916
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2917
2918
    @classmethod
    def from_cli(cls, cli_value: str) -> "KVTransferConfig":
youkaichao's avatar
youkaichao committed
2919
        """Parse the CLI value for the kv cache transfer config."""
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
        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"]

2958
2959
2960
    def get_from_extra_config(self, key, default) -> Any:
        return self.kv_connector_extra_config.get(key, default)

2961

2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
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.
2980
        - debug_dump_path: the path to dump the debug information.
2981
2982
2983
        - 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.
2984
2985
2986
2987
2988
2989
2990
        - 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).
2991
2992
2993
2994
2995
2996
2997
2998
2999
        - 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).
3000
        - splitting_ops: a list of ops to split the full graph into subgraphs, used in piecewise compilation.
3001
3002
3003
3004
    - 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
3005
3006
3007
3008
                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.
3009
3010
3011
            TODO: move outside cudagraph logic into compilation.
            torch.compile will handle cudagraph capture logic in the future.
        - cudagraph_capture_sizes: sizes to capture cudagraph.
3012
            - None (default): capture sizes are inferred from vllm config.
3013
            - list[int]: capture sizes are specified as given.
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
        - 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
3027
3028
3029
3030
3031
                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.
3032
3033
3034
3035
3036
3037
3038
        - 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})`
3039
        - custom inductor passes: see PassConfig for more details
3040

3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
    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
3052
    debug_dump_path: str = ""
3053
    cache_dir: str = ""
3054
    backend: str = ""
3055
3056
    custom_ops: list[str] = Field(default_factory=list)
    splitting_ops: list[str] = Field(default=None)  # type: ignore
3057
3058

    use_inductor: bool = True
3059
3060
3061
    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)
3062
3063
3064

    use_cudagraph: bool = False
    cudagraph_num_of_warmups: int = 0
3065
    cudagraph_capture_sizes: Optional[list[int]] = None
3066
3067
    cudagraph_copy_inputs: bool = False

3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
    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.
3078
3079
        - enable_noop: whether to enable the custom no-op elimination pass.
            TODO(luka) better pass enabling system.
3080
        """
3081
        dump_graph_stages: list[str] = Field(default_factory=list)
3082
3083
        dump_graph_dir: Path = Field(default=Path("."))
        enable_fusion: bool = True
3084
        enable_noop: bool = True
3085
3086
3087
3088
3089
3090
3091
3092

        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.
            """
3093
            dict_ = self.model_dump(include={"enable_fusion", "enable_noop"})
3094
            return InductorPass.hash_dict(dict_)
3095
3096

        def model_post_init(self, __context: Any) -> None:
3097
            if not self.enable_noop and self.enable_fusion:
3098
                logger.warning_once(
3099
                    "Fusion enabled but reshape elimination disabled. "
3100
3101
3102
                    "RMSNorm + quant (fp8) fusion might not work")

    pass_config: PassConfig = Field(default_factory=PassConfig)
3103
3104

    # not configurable, computed after init
3105
    max_capture_size: int = PrivateAttr
3106
    local_cache_dir: str = PrivateAttr  # local cache dir for each rank
3107
    # optimization:
3108
    # Intuitively, bs_to_padded_graph_size should be dict[int, int].
3109
    # since we know all keys are in a range [0, max_capture_size],
3110
3111
    # we can optimize it to list[int] for better lookup performance.
    bs_to_padded_graph_size: list[int] = PrivateAttr
3112

3113
3114
3115
    # keep track of enabled and disabled custom ops
    enabled_custom_ops: Counter[str] = PrivateAttr
    disabled_custom_ops: Counter[str] = PrivateAttr
3116
    traced_files: set[str] = PrivateAttr
3117
    compilation_time: float = PrivateAttr
3118

3119
3120
    # Per-model forward context
    # Map from layer name to the attention cls
3121
    static_forward_context: dict[str, Any] = PrivateAttr
3122

3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
    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.
        """
3135
        factors: list[Any] = []
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
        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()

3146
3147
3148
3149
3150
3151
3152
3153
    def __repr__(self) -> str:
        exclude = {
            "static_forward_context",
            "enabled_custom_ops",
            "disabled_custom_ops",
            "compilation_time",
            "bs_to_padded_graph_size",
            "pass_config",
3154
            "traced_files",
3155
3156
3157
3158
3159
        }
        return self.model_dump_json(exclude=exclude, exclude_unset=True)

    __str__ = __repr__

3160
3161
3162
3163
3164
    @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))
3165
3166
3167
        # 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)
3168

3169
3170
3171
3172
3173
3174
    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
3175
3176
3177
3178
3179
3180
3181
3182
        # 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

3183
        if Version(importlib.metadata.version('torch')) >= Version("2.6"):
Michael Goin's avatar
Michael Goin committed
3184
3185
3186
3187
            KEY = 'enable_auto_functionalized_v2'
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

3188
        if self.splitting_ops is None:
3189
            self.splitting_ops = []
3190

3191
3192
3193
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
3194
3195
3196
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
3197
3198
3199
3200
3201
3202
3203
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
3204
3205
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
3206

3207
3208
        self.enabled_custom_ops = Counter()
        self.disabled_custom_ops = Counter()
3209
        self.traced_files = set()
3210
        self.static_forward_context = {}
3211
        self.compilation_time = 0.0
3212

3213
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
        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
3231

3232
        from vllm.compilation.backends import VllmBackend
3233
        return VllmBackend(vllm_config)
3234

3235
    def init_with_cudagraph_sizes(self,
3236
                                  cudagraph_capture_sizes: list[int]) -> None:
3237
        """To complete the initialization of config,
3238
3239
        we need to know the cudagraph sizes."""

3240
        if self.cudagraph_capture_sizes is None:
3241
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
3242
        else:
3243
3244
3245
            # de-duplicate the sizes provided by the config
            self.cudagraph_capture_sizes = list(
                set(self.cudagraph_capture_sizes))
3246
3247
            logger.info(("cudagraph sizes specified by model runner"
                         " %s is overridden by config %s"),
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
                        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
3264

3265
        # sort to make sure cudagraph capture sizes are in descending order
3266
3267
3268
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
3269

3270
3271
3272
3273
        # 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)
        ]
3274
3275
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
3276
3277
3278
3279
3280
3281
3282
            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
3283

3284
3285
3286
3287
3288
3289
3290
3291
3292
    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",
            ]

3293

3294
3295
3296
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
3297
3298
3299
    simplifies passing around the distinct configurations in the codebase.
    """

3300
3301
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
3302
3303
3304
3305
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig,
                                            init=True)
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig,
                                              init=True)
3306
3307
3308
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
3309
    lora_config: Optional[LoRAConfig] = None
3310
3311
    speculative_config: SpeculativeConfig = field(default=None,
                                                  init=True)  # type: ignore
3312
3313
3314
    decoding_config: Optional[DecodingConfig] = None
    observability_config: Optional[ObservabilityConfig] = None
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
3315
    quant_config: Optional[QuantizationConfig] = None
3316
3317
    compilation_config: CompilationConfig = field(default=None,
                                                  init=True)  # type: ignore
3318
3319
    kv_transfer_config: KVTransferConfig = field(default=None,
                                                 init=True)  # type: ignore
3320
    # some opaque config, only used to provide additional information
3321
3322
    # for the hash computation, mainly used for testing, debugging or out of
    # tree config registration.
3323
3324
    additional_config: SupportsHash = field(default=None,
                                            init=True)  # type: ignore
3325
    instance_id: str = ""
3326

3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
    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.
        """
3339
        factors: list[Any] = []
3340
3341

        # summarize vllm config
3342
        vllm_factors: list[Any] = []
3343
3344
        from vllm import __version__
        vllm_factors.append(__version__)
3345
        vllm_factors.append(envs.VLLM_USE_V1)
3346
3347
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3348
3349
        else:
            vllm_factors.append("None")
3350
3351
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3352
3353
        else:
            vllm_factors.append("None")
3354
3355
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3356
3357
        else:
            vllm_factors.append("None")
3358
3359
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3360
3361
        else:
            vllm_factors.append("None")
3362
3363
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3364
3365
        else:
            vllm_factors.append("None")
3366
3367
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3368
3369
        else:
            vllm_factors.append("None")
3370
3371
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3372
3373
3374
3375
3376
            # 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))
3377
3378
        else:
            vllm_factors.append("None")
3379
3380
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3381
3382
        else:
            vllm_factors.append("None")
3383
3384
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3385
3386
        else:
            vllm_factors.append("None")
3387
3388
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3389
3390
        else:
            vllm_factors.append("None")
3391
3392
        if self.prompt_adapter_config:
            vllm_factors.append(self.prompt_adapter_config.compute_hash())
3393
3394
        else:
            vllm_factors.append("None")
3395
3396
3397
3398
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3399
3400
        else:
            vllm_factors.append("None")
3401
3402
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3403
3404
3405
3406
3407
3408
        else:
            vllm_factors.append("None")
        if self.additional_config:
            vllm_factors.append(self.additional_config.compute_hash())
        else:
            vllm_factors.append("None")
3409
3410
3411
3412
3413
        factors.append(vllm_factors)

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

3414
3415
3416
3417
3418
3419
    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]
3420

3421
3422
3423
3424
3425
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3426
        from vllm.platforms import current_platform
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
        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
3449

3450
3451
3452
3453
3454
3455
3456
3457
3458
    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

3459
3460
3461
3462
3463
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

3464
3465
3466
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
3467
3468
3469
3470
3471
3472
3473
3474
        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)
3475
3476

        if self.lora_config:
3477
            self.lora_config.verify_with_cache_config(self.cache_config)
3478
3479
3480
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
3481
3482
3483
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
3484
3485
3486
3487
3488

        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)
3489

3490
        from vllm.platforms import current_platform
3491
3492
3493
3494
3495
        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):
3496
            logger.warning_once(
3497
3498
3499
3500
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

3501
        if self.compilation_config is None:
3502
            self.compilation_config = CompilationConfig()
3503
3504
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
3505
3506
3507
3508
            # 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.
3509
            # FIXME(rob): Add function to set all of these.
3510
3511
3512
            self.compilation_config.custom_ops = ["none"]
            self.compilation_config.use_cudagraph = True
            self.compilation_config.use_inductor = True
3513
            self.compilation_config.cudagraph_num_of_warmups = 1
3514
            self.compilation_config.pass_config.enable_fusion = False
3515
            self.compilation_config.pass_config.enable_noop = False
3516
            self.compilation_config.level = CompilationLevel.PIECEWISE
3517
            self.compilation_config.set_splitting_ops_for_v1()
3518

3519
        self._set_cudagraph_sizes()
3520

3521
3522
3523
3524
3525
3526
3527
3528
        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

3529
3530
3531
3532
3533
3534
        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`.")
3535
3536
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

3537

3538
        if self.model_config and self.model_config.use_mla and \
3539
            not (current_platform.is_cuda() or current_platform.is_rocm()):
3540
            logger.info(
3541
                "MLA is enabled on a non-GPU platform; forcing chunked "
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
                "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

3552
3553
        current_platform.check_and_update_config(self)

3554
3555
3556
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
    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.

3573
3574
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
3575
3576

        During runtime, if batchsize is larger than
3577
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3578
3579
        no cudagraph will be used.
        If the batch size is no larger than
3580
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
        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)]
3617
3618
3619
3620
3621
                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
                ]
3622
3623
3624
3625

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

3626
    def __str__(self):
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
        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}, "
3657
            f"disable_mm_preprocessor_cache={self.model_config.disable_mm_preprocessor_cache!r}, "  # noqa
3658
            f"mm_processor_kwargs={self.model_config.mm_processor_kwargs}, "
3659
3660
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
3661
3662
3663
3664
3665
3666


_current_vllm_config: Optional[VllmConfig] = None


@contextmanager
3667
def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False):
3668
    """
3669
    Temporarily set the current vLLM config.
3670
    Used during model initialization.
3671
    We save the current vLLM config in a global variable,
3672
    so that all modules can access it, e.g. custom ops
3673
    can access the vLLM config to determine how to dispatch.
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
    """
    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)
3687
3688
        if check_compile and \
            vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
3689
3690
3691
3692
3693
3694
3695
3696
3697
            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"
3698
                " if you want it to be supported.",
3699
3700
3701
3702
3703
3704
3705
3706
3707
                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.
3708
        logger.warning("Current vLLM config is not set.")
3709
3710
3711
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config