config.py 119 KB
Newer Older
1
import ast
2
import copy
3
import enum
4
import hashlib
5
import json
6
import warnings
7
from contextlib import contextmanager
8
from dataclasses import dataclass, field, replace
9
from pathlib import Path
10
11
12
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Counter, Dict,
                    Final, List, Literal, Mapping, Optional, Set, Tuple, Type,
                    Union)
13
14

import torch
15
from pydantic import BaseModel, Field, PrivateAttr
16
from transformers import PretrainedConfig
17

18
import vllm.envs as envs
19
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
20
from vllm.logger import init_logger
21
22
from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
                                                     get_quantization_config)
23
from vllm.model_executor.models import ModelRegistry
24
from vllm.platforms import current_platform
25
from vllm.tracing import is_otel_available, otel_import_error_traceback
26
27
28
29
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
    get_sentence_transformer_tokenizer_config, is_encoder_decoder, uses_mrope)
30
31
from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
                        get_cpu_memory, print_warning_once, random_uuid,
32
                        resolve_obj_by_qualname)
33

34
35
36
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

37
    from vllm.executor.executor_base import ExecutorBase
38
39
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
40
    from vllm.model_executor.model_loader.loader import BaseModelLoader
41
42
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
43
44
else:
    QuantizationConfig = None
45

46
47
logger = init_logger(__name__)

48
_POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
49
_MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
50

51
52
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
                     "score", "reward"]
53

54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
_ResolvedTask = Literal["generate", "embed", "classify", "score", "reward",
                        "draft"]

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

_RUNNER_TASKS: Dict[RunnerType, List[_ResolvedTask]] = {
    "generate": ["generate"],
    "pooling": ["embed", "classify", "score", "reward"],
    "draft": ["draft"],
}

_TASK_RUNNER: Dict[_ResolvedTask, RunnerType] = {
    task: runner
    for runner, tasks in _RUNNER_TASKS.items() for task in tasks
}
69

70
71
72
HfOverrides = Union[Dict[str, Any], Callable[[PretrainedConfig],
                                             PretrainedConfig]]

73
74

class ModelConfig:
75
76
77
78
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
79
            It is also used as the content for `model_name` tag in metrics
80
81
82
83
84
            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.
85
        tokenizer: Name or path of the huggingface tokenizer to use.
86
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
87
88
            available, "slow" will always use the slow tokenizer, and
            "mistral" will always use the tokenizer from `mistral_common`.
89
90
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
91
92
93
94
        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.
95
96
97
98
        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
99
100
101
        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.
102
        code_revision: The specific revision to use for the model code on
103
            Hugging Face Hub. It can be a branch name, a tag name, or a
104
            commit id. If unspecified, will use the default version.
105
106
107
        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.
108
109
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
110
111
        spec_target_max_model_len: Specify the the maximum length for spec
            decoding draft models.
112
113
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
114
115
        quantization_param_path: Path to JSON file containing scaling factors.
            Used to load KV cache scaling factors into the model when KV cache
116
117
            type is FP8_E4M3 on ROCm (AMD GPU). In the future these will also
            be used to load activation and weight scaling factors when the
118
            model dtype is FP8_E4M3 on ROCm.
119
120
121
        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.
122
            If None, the user did not specify, so default to False.
123
124
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
125
126
127
            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.
128
        max_logprobs: Maximum number of log probabilities. Defaults to 20.
129
130
131
132
        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.
133
134
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
135
        served_model_name: The model name used in metrics tag `model_name`,
136
137
            matches the model name exposed via the APIs. If multiple model
            names provided, the first name will be used. If not specified,
138
            the model name will be the same as `model`.
139
        limit_mm_per_prompt: Maximum number of data items per modality
140
            per prompt. Only applicable for multimodal models.
141
142
        use_async_output_proc: Whether to use async output processor.
            Defaults to True.
143
144
        config_format: The config format which shall be loaded.
            Defaults to 'auto' which defaults to 'hf'.
145
146
147
        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.
148
149
        mm_processor_kwargs: Arguments to be forwarded to the model's processor
            for multi-modal data, e.g., image processor.
150
151
152
153
        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.
154
        override_pooler_config: Initialize non default pooling config or
155
            override default pooling config for the pooling model.
156
    """
157

158
159
160
    def __init__(
            self,
            model: str,
161
            task: Union[TaskOption, Literal["draft"]],
162
163
164
165
166
            tokenizer: str,
            tokenizer_mode: str,
            trust_remote_code: bool,
            dtype: Union[str, torch.dtype],
            seed: int,
167
            allowed_local_media_path: str = "",
168
169
            revision: Optional[str] = None,
            code_revision: Optional[str] = None,
170
            rope_scaling: Optional[Dict[str, Any]] = None,
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
            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,
            quantization_param_path: Optional[str] = None,
            enforce_eager: Optional[bool] = None,
            max_seq_len_to_capture: Optional[int] = None,
            max_logprobs: int = 20,
            disable_sliding_window: bool = False,
            skip_tokenizer_init: bool = False,
            served_model_name: Optional[Union[str, List[str]]] = None,
            limit_mm_per_prompt: Optional[Mapping[str, int]] = None,
            use_async_output_proc: bool = True,
            config_format: ConfigFormat = ConfigFormat.AUTO,
186
            hf_overrides: Optional[HfOverrides] = None,
187
            mm_processor_kwargs: Optional[Dict[str, Any]] = None,
188
189
            override_neuron_config: Optional[Dict[str, Any]] = None,
            override_pooler_config: Optional["PoolerConfig"] = None) -> None:
190
        self.model = model
191
        self.tokenizer = tokenizer
192
        self.tokenizer_mode = tokenizer_mode
193
        self.trust_remote_code = trust_remote_code
194
        self.allowed_local_media_path = allowed_local_media_path
195
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
196
        self.revision = revision
197
        self.code_revision = code_revision
198
199
200

        if hf_overrides is None:
            hf_overrides = {}
201
202
203
204
205
206

        if callable(hf_overrides):
            hf_overrides_kw = {}
            hf_overrides_fn = hf_overrides
        else:
            hf_overrides_kw = hf_overrides
207
            hf_overrides_fn = None
208

209
210
        if rope_scaling is not None:
            hf_override: Dict[str, Any] = {"rope_scaling": rope_scaling}
211
            hf_overrides_kw.update(hf_override)
212
213
214
215
216
            msg = ("`--rope-scaling` will be removed in a future release. "
                   f"'Please instead use `--hf-overrides '{hf_override!r}'`")
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
        if rope_theta is not None:
            hf_override = {"rope_theta": rope_theta}
217
            hf_overrides_kw.update(hf_override)
218
219
220
221
            msg = ("`--rope-theta` will be removed in a future release. "
                   f"'Please instead use `--hf-overrides '{hf_override!r}'`")
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

222
223
224
225
226
        # 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
227
        self.quantization = quantization
228
        self.quantization_param_path = quantization_param_path
229
        self.enforce_eager = enforce_eager
230
        self.max_seq_len_to_capture = max_seq_len_to_capture
231
        self.max_logprobs = max_logprobs
232
        self.disable_sliding_window = disable_sliding_window
233
        self.skip_tokenizer_init = skip_tokenizer_init
234
235

        hf_config = get_config(self.model, trust_remote_code, revision,
236
237
238
239
240
241
242
243
244
                               code_revision, config_format)

        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)

245
246
        self.hf_config = hf_config

247
        self.hf_text_config = get_hf_text_config(self.hf_config)
248
        self.encoder_config = self._get_encoder_config()
249
250
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, revision)
251
        self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
252
        self.use_async_output_proc = use_async_output_proc
253
        self.mm_processor_kwargs = mm_processor_kwargs
Woosuk Kwon's avatar
Woosuk Kwon committed
254

255
256
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
257
258
            self.enforce_eager = False

259
260
261
262
263
264
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
        has_interleaved_attention = (sliding_window is not None) and (
            isinstance(sliding_window, list) or
            (self.hf_text_config.model_type in ["gemma2"]))

        if (not self.disable_sliding_window and has_interleaved_attention):
265
266
267
            if envs.VLLM_ATTENTION_BACKEND == "XFORMERS":
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
268

269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
                print_warning_once(
                    f"{self.hf_text_config.model_type} has interleaved "
                    "attention, which is currently not supported by the "
                    "XFORMERS backend. Disabling sliding window and capping "
                    "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
285

286
287
288
289
        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,
290
            sliding_window_len=self.get_hf_config_sliding_window(),
291
292
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
293
294
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
295
296
        self.multimodal_config = self._init_multimodal_config(
            limit_mm_per_prompt)
297
298
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
299

300
        self.is_attention_free = self._init_attention_free()
301
        self.is_hybrid = self._init_is_hybrid()
302
303
        self.has_inner_state = self._init_has_inner_state()

304
305
306
307
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
308
309
310
311

        supported_tasks, task = self._resolve_task(task, self.hf_config)
        self.supported_tasks = supported_tasks
        self.task: Final = task
312

313
        self.pooler_config = self._init_pooler_config(override_pooler_config)
314

315
        self._verify_quantization()
316
        self._verify_cuda_graph()
317
        self._verify_bnb_config()
318

319
320
321
322
    def _init_multimodal_config(
        self, limit_mm_per_prompt: Optional[Mapping[str, int]]
    ) -> Optional["MultiModalConfig"]:
        architectures = getattr(self.hf_config, "architectures", [])
323
        if ModelRegistry.is_multimodal_model(architectures):
324
            return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
325
326
327
328
329
330

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

        return None
331

332
333
334
335
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

336
337
    def _init_pooler_config(
        self,
338
        override_pooler_config: Optional["PoolerConfig"],
339
    ) -> Optional["PoolerConfig"]:
340

341
        if self.runner_type == "pooling":
342
343
344
345
346
347
348
349
350
351
352
            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

353
354
        return None

355
356
357
358
    def _init_attention_free(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_attention_free_model(architectures)

359
360
361
362
    def _init_is_hybrid(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_hybrid_model(architectures)

363
364
365
366
    def _init_has_inner_state(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.model_has_inner_state(architectures)

367
368
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
369
        if tokenizer_mode not in ["auto", "slow", "mistral"]:
370
371
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
372
                "either 'auto', 'slow' or 'mistral'.")
373
        self.tokenizer_mode = tokenizer_mode
374

375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
    def _get_preferred_task(
        self,
        architectures: List[str],
        supported_tasks: Set[_ResolvedTask],
    ) -> Optional[_ResolvedTask]:
        model_id = self.model
        if get_pooling_config(model_id, self.revision):
            return "embed"
        if ModelRegistry.is_cross_encoder_model(architectures):
            return "score"

        suffix_to_preferred_task: List[Tuple[str, _ResolvedTask]] = [
            # Other models follow this pattern
            ("ForCausalLM", "generate"),
            ("ForConditionalGeneration", "generate"),
            ("ForSequenceClassification", "classify"),
            ("ChatModel", "generate"),
            ("LMHeadModel", "generate"),
            ("EmbeddingModel", "embed"),
            ("RewardModel", "reward"),
        ]
        _, arch = ModelRegistry.inspect_model_cls(architectures)

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

        return None

404
405
    def _resolve_task(
        self,
406
        task_option: Union[TaskOption, Literal["draft"]],
407
        hf_config: PretrainedConfig,
408
    ) -> Tuple[Set[_ResolvedTask], _ResolvedTask]:
409
410
411
        if task_option == "draft":
            return {"draft"}, "draft"

412
413
        architectures = getattr(hf_config, "architectures", [])

414
        runner_support: Dict[RunnerType, bool] = {
415
416
417
            # NOTE: Listed from highest to lowest priority,
            # in case the model supports multiple of them
            "generate": ModelRegistry.is_text_generation_model(architectures),
418
            "pooling": ModelRegistry.is_pooling_model(architectures),
419
        }
420
421
422
423
424
425
426
427
428
        supported_runner_types_lst: List[RunnerType] = [
            runner_type
            for runner_type, is_supported in runner_support.items()
            if is_supported
        ]

        supported_tasks_lst: List[_ResolvedTask] = [
            task for runner_type in supported_runner_types_lst
            for task in _RUNNER_TASKS[runner_type]
429
430
431
432
433
        ]
        supported_tasks = set(supported_tasks_lst)

        if task_option == "auto":
            selected_task = next(iter(supported_tasks_lst))
434

435
436
437
438
439
            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
440

441
442
443
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
444
        else:
445
446
447
448
449
450
451
452
453
454
455
456
457
458
            # 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"

459
460
461
462
463
464
465
            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
466

467
        return supported_tasks, selected_task
468

469
470
471
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
472
            # compressed-tensors uses a "compression_config" key
473
            quant_cfg = getattr(self.hf_config, "compression_config", None)
474
475
        return quant_cfg

476
    def _verify_quantization(self) -> None:
477
        supported_quantization = QUANTIZATION_METHODS
478
        optimized_quantization_methods = [
479
480
481
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
            "awq_marlin", "fbgemm_fp8", "compressed_tensors",
            "compressed-tensors", "experts_int8"
482
        ]
483
484
485
486
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
487
488
        quant_cfg = self._parse_quant_hf_config()

489
490
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
491
492

            # Detect which checkpoint is it
493
494
            for name in QUANTIZATION_METHODS:
                method = get_quantization_config(name)
495
496
497
498
499
500
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
501

502
            # Verify quantization configurations.
503
            if self.quantization is None:
504
505
                self.quantization = quant_method
            elif self.quantization != quant_method:
506
507
                raise ValueError(
                    "Quantization method specified in the model config "
508
                    f"({quant_method}) does not match the quantization "
509
510
511
512
513
514
515
516
                    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}.")
517
            current_platform.verify_quantization(self.quantization)
518
            if self.quantization not in optimized_quantization_methods:
519
                logger.warning(
520
                    "%s quantization is not fully "
521
                    "optimized yet. The speed can be slower than "
522
                    "non-quantized models.", self.quantization)
523

524
    def _verify_cuda_graph(self) -> None:
525
526
527
528
        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)
529

530
531
    def _verify_bnb_config(self) -> None:
        """
532
        The current version of bitsandbytes (0.44.0) with 8-bit models does not
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
        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

552
553
554
555
556
557
558
559
560
561
562
563
    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:
            logger.warning("Async output processing can not be enabled "
                           "with pipeline parallel")
            self.use_async_output_proc = False
            return

564
        # Reminder: Please update docs/source/usage/compatibility_matrix.rst
565
        # If the feature combo become valid
566
        if not current_platform.is_async_output_supported(self.enforce_eager):
567
            logger.warning(
568
569
                "Async output processing is not supported on the "
                "current platform type %s.", current_platform.device_type)
570
571
572
573
574
575
576
577
578
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            logger.warning(
                "Async output processing can not be enabled with ray spmd")
            self.use_async_output_proc = False
            return

579
        # Async postprocessor is not necessary for pooling models
580
        # since there is no token generation
581
        if self.runner_type == "pooling":
582
583
            self.use_async_output_proc = False

584
        # Reminder: Please update docs/source/usage/compatibility_matrix.rst
585
        # If the feature combo become valid
586
587
588
589
590
        if speculative_config:
            logger.warning("Async output processing is not supported with"
                           " speculative decoding currently.")
            self.use_async_output_proc = False

591
592
593
594
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
595
596
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
597
598
599
600
601
602
603
604
        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}).")

        pipeline_parallel_size = parallel_config.pipeline_parallel_size
605
606
607
608
609
610
611
612
613
614
615
        if pipeline_parallel_size > 1:
            architectures = getattr(self.hf_config, "architectures", [])
            if not ModelRegistry.is_pp_supported_model(architectures):
                raise NotImplementedError(
                    "Pipeline parallelism is not supported for this model. "
                    "Supported models implement the `SupportsPP` interface.")

            if self.use_async_output_proc:
                logger.warning("Async output processor is not supported with "
                               "pipeline parallelism currently. Disabling it.")
                self.use_async_output_proc = False
616

617
618
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], List[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
619
        """Get the sliding window size, or None if disabled."""
620
621
622
623

        # 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.
624
625
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
626
            return None
627
        return getattr(self.hf_text_config, "sliding_window", None)
628

629
    def get_sliding_window(self) -> Optional[Union[int, List[Optional[int]]]]:
630
631
632
633
634
635
636
637
        """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()

638
    def get_vocab_size(self) -> int:
639
        return self.hf_text_config.vocab_size
640

641
    def get_hidden_size(self) -> int:
642
        return self.hf_text_config.hidden_size
643
644

    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
645
646
647
648
649
650
        # TODO remove hard code
        if hasattr(self.hf_text_config, "model_type"
                   ) and self.hf_text_config.model_type == 'deepseek_v2':
            # FlashAttention supports only head_size 32, 64, 128, 256,
            # we need to pad head_size 192 to 256
            return 256
651
652
653
654

        if self.is_attention_free:
            return 0

655
656
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
657
        # FIXME(woosuk): This may not be true for all models.
658
659
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
660

661
662
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
663
        # For GPTBigCode & Falcon:
664
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
665
666
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
667
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
668
        new_decoder_arch_falcon = (
669
            self.hf_config.model_type in falcon_model_types
670
            and getattr(self.hf_config, "new_decoder_architecture", False))
671
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
672
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
673
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
674
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
675
            return 1
676

677
        # For DBRX and MPT
678
679
680
681
682
        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":
683
684
685
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

686
687
688
        if self.is_attention_free:
            return 0

689
690
691
692
693
694
695
696
697
698
        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:
699
            num_kv_heads = getattr(self.hf_text_config, attr, None)
700
701
702
703
704
            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.
705
        return self.hf_text_config.num_attention_heads
706
707
708
709
710
711
712
713
714
715

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
        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)
716

717
718
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
719
720
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
721

722
723
    def get_layers_start_end_indices(
            self, parallel_config: "ParallelConfig") -> Tuple[int, int]:
724
        from vllm.distributed.utils import get_pp_indices
Mor Zusman's avatar
Mor Zusman committed
725
726
        total_num_hidden_layers = getattr(self.hf_text_config,
                                          "num_hidden_layers", 0)
727
728
729
        pp_rank = parallel_config.rank // parallel_config.tensor_parallel_size
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
730
        return start, end
Mor Zusman's avatar
Mor Zusman committed
731

732
733
734
    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
735

736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
    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:
                raise ValueError("The model is an hybrid without a"
                                 "layers_block_type in the hf_config,"
                                 "cannot determine the num of "
                                 f"{block_type.value} layers")

            return sum(t == block_type.value
                       for t in layers_block_type_value[start:end])
Mor Zusman's avatar
Mor Zusman committed
767

768
769
770
771
772
773
774
775
776
777
778
779
    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

780
    @property
781
    def is_encoder_decoder(self) -> bool:
782
        """Extract the HF encoder/decoder model flag."""
783
784
785
786
787
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
788

789
790
791
792
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

793
794
795
796
797
    @property
    def is_cross_encoder(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_cross_encoder_model(architectures)

798
799
800
801
802
803
804
805
    @property
    def supported_runner_types(self) -> Set[RunnerType]:
        return {_TASK_RUNNER[task] for task in self.supported_tasks}

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

806
807

class CacheConfig:
808
809
810
811
812
    """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
813
            vLLM execution.
814
        swap_space: Size of the CPU swap space per GPU (in GiB).
815
        cache_dtype: Data type for kv cache storage.
816
        is_attention_free: Whether the model is attention-free.
817
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
818
            profiled num_gpu_blocks if specified. Does nothing if None.
819
820
821
822
        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.
823
    """
824

825
826
827
828
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
829
        swap_space: float,
830
        cache_dtype: str,
831
        is_attention_free: bool = False,
832
        num_gpu_blocks_override: Optional[int] = None,
833
        sliding_window: Optional[int] = None,
834
        enable_prefix_caching: bool = False,
835
        cpu_offload_gb: float = 0,
836
837
838
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
839
        self.swap_space_bytes = swap_space * GiB_bytes
840
        self.num_gpu_blocks_override = num_gpu_blocks_override
841
        self.cache_dtype = cache_dtype
842
        self.is_attention_free = is_attention_free
843
        self.sliding_window = sliding_window
844
        self.enable_prefix_caching = enable_prefix_caching
845
        self.cpu_offload_gb = cpu_offload_gb
846

847
        self._verify_args()
848
        self._verify_cache_dtype()
849
        self._verify_prefix_caching()
850
851

        # Will be set after profiling.
852
853
        self.num_gpu_blocks: Optional[int] = None
        self.num_cpu_blocks: Optional[int] = None
854

855
    def metrics_info(self):
856
857
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
858
859
        return {key: str(value) for key, value in self.__dict__.items()}

860
861
862
863
864
865
    def _verify_args(self) -> None:
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

866
867
868
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
869
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
870
            logger.info(
871
872
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
873
874
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
875
876
877
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

878
879
880
881
882
883
884
885
886
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

        if self.sliding_window is not None:
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

887
888
889
890
891
892
893
894
895
896
    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

897
898
899
        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.")
900
901
902
        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:
903
            logger.warning("Possibly too large swap space. %s", msg)
904

905

906
907
908
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
909

910
911
912
913
914
915
916
917
    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
918
    pool_type: Union[str, Type["BaseTokenizerGroup"]]
919
920
921
    extra_config: dict

    def __post_init__(self):
922
923
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
924
925
926
927
928
929
            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(
930
931
        cls, tokenizer_pool_size: int,
        tokenizer_pool_type: Union[str, Type["BaseTokenizerGroup"]],
932
933
934
        tokenizer_pool_extra_config: Optional[Union[str, dict]]
    ) -> Optional["TokenizerPoolConfig"]:
        """Create a TokenizerPoolConfig from the given parameters.
935

936
        If tokenizer_pool_size is 0, return None.
937

938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
        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


960
961
962
963
964
965
966
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
967
    SHARDED_STATE = "sharded_state"
968
    GGUF = "gguf"
969
    BITSANDBYTES = "bitsandbytes"
970
    MISTRAL = "mistral"
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989


@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.
990
            "bitsandbytes" will load nf4 type weights.
991
        model_loader_extra_config: The extra config for the model loader.
992
        ignore_patterns: The list of patterns to ignore when loading the model.
993
            Default to "original/**/*" to avoid repeated loading of llama's
994
            checkpoints.
995
996
997
998
999
1000
    """

    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)
1001
    ignore_patterns: Optional[Union[List[str], str]] = None
1002
1003
1004
1005
1006
1007

    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)
1008
1009
1010
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1011

1012
1013
1014
1015
1016
1017
1018
        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/**/*"]

1019

1020
@dataclass
1021
class ParallelConfig:
1022
    """Configuration for the distributed execution."""
1023

1024
1025
    pipeline_parallel_size: int = 1  # Number of pipeline parallel groups.
    tensor_parallel_size: int = 1  # Number of tensor parallel groups.
1026

1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
    # Deprecated, use distributed_executor_backend instead.
    worker_use_ray: Optional[bool] = None

    # 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,
                                                 Type["ExecutorBase"]]] = None

    # 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"
1060
    sd_worker_cls: str = "auto"
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070

    world_size: int = field(init=False)

    rank: int = 0

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

        if self.worker_use_ray:
1071
1072
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
1073
            elif not self.use_ray:
1074
1075
1076
                raise ValueError(f"worker-use-ray can't be used with "
                                 f"distributed executor backend "
                                 f"'{self.distributed_executor_backend}'.")
1077
1078
1079
        ray_only_devices = ["tpu", "hpu"]
        if (current_platform.device_type in ray_only_devices
                and self.world_size > 1):
1080
1081
1082
1083
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
1084
1085
                    f"{current_platform.device_type.upper()} backend only "
                    "supports Ray for distributed inference.")
1086

1087
        if self.distributed_executor_backend is None and self.world_size > 1:
1088
1089
1090
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1091
            from vllm.executor import ray_utils
1092
            backend = "mp"
1093
            ray_found = ray_utils.ray_is_available()
1094
            if (current_platform.is_cuda()
1095
                    and cuda_device_count_stateless() < self.world_size):
1096
1097
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1098
1099
1100
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1101
1102
                backend = "ray"
            elif ray_found:
1103
                if self.placement_group:
1104
                    backend = "ray"
1105
1106
1107
1108
1109
1110
                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"
1111
1112
1113
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1114

1115
1116
        self._verify_args()

1117
1118
1119
1120
1121
1122
    @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)

1123
    def _verify_args(self) -> None:
1124
1125
1126
1127
1128
1129
1130
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase

        if self.distributed_executor_backend not in (
                "ray", "mp", None) and not (isinstance(
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
1131
            raise ValueError(
1132
1133
1134
1135
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
                "values are 'ray', 'mp' or custom ExecutorBase subclass.")
        if self.use_ray:
1136
1137
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1138
        if current_platform.is_rocm():
1139
1140
1141
1142
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
1143
        if self.ray_workers_use_nsight and not self.use_ray:
1144
1145
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1146

1147

1148
@dataclass
1149
class SchedulerConfig:
1150
    """Scheduler configuration."""
1151

1152
    runner_type: str = "generate"  # The runner type to launch for the model.
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177

    # 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

    # 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
1178

1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
    # 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)

    def __post_init__(self) -> None:
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
1206
1207
1208
1209
                    # 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.
1210
                    self.max_num_batched_tokens = max(self.max_model_len, 2048)
1211
                else:
1212
1213
1214
                    # This value is chosen to have a balance between ITL
                    # and TTFT. Note it is not optimized for throughput.
                    self.max_num_batched_tokens = 2048
1215
1216
1217
            else:
                # If max_model_len is too short, use 2048 as the default value
                # for higher throughput.
1218
                self.max_num_batched_tokens = max(self.max_model_len, 2048)
1219

1220
1221
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
1222
1223
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1224
                    _POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
1225
                )
1226
            if self.is_multimodal_model:
1227
                # The value needs to be at least the number of multimodal tokens
1228
1229
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1230
1231
1232
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

1233
        if self.enable_chunked_prefill:
1234
1235
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
1236
                self.max_num_batched_tokens)
1237

1238
        self.chunked_prefill_enabled = self.enable_chunked_prefill
1239
1240
1241
        self._verify_args()

    def _verify_args(self) -> None:
1242
1243
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
1244
1245
1246
1247
1248
1249
1250
            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.")
1251

1252
1253
1254
1255
1256
        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}).")
1257

1258
1259
1260
1261
1262
1263
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

1274

1275
class DeviceConfig:
1276
    device: Optional[torch.device]
1277
    device_type: str
1278

1279
1280
1281
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
1282
            self.device_type = current_platform.device_type
1283
            if not self.device_type:
1284
                raise RuntimeError("Failed to infer device type")
1285
1286
1287
1288
1289
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
1290
        if self.device_type in ["neuron", "openvino"]:
1291
            self.device = torch.device("cpu")
1292
1293
        elif self.device_type in ["tpu"]:
            self.device = None
1294
1295
1296
1297
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1298

1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
class SpeculativeConfig:
    """Configuration for speculative decoding.

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

    @staticmethod
    def maybe_create_spec_config(
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        target_dtype: str,
        speculative_model: Optional[str],
1312
        speculative_model_quantization: Optional[str],
1313
        speculative_draft_tensor_parallel_size: Optional[int],
1314
        num_speculative_tokens: Optional[int],
1315
        speculative_disable_mqa_scorer: Optional[bool],
1316
1317
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
1318
        disable_log_stats: bool,
1319
        speculative_disable_by_batch_size: Optional[int],
1320
1321
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1322
1323
1324
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
1325
        disable_logprobs: Optional[bool],
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
    ) -> Optional["SpeculativeConfig"]:
        """Create a SpeculativeConfig if possible, else return None.

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

        Args:
            target_model_config (ModelConfig): The configuration of the target
                model.
            target_parallel_config (ParallelConfig): The parallel configuration
                for the target model.
            target_dtype (str): The data type used for the target model.
            speculative_model (Optional[str]): The name of the speculative
                model, if provided.
1341
1342
1343
            speculative_model_quantization (Optional[str]): Quantization method
                that was used to quantize the speculative model weights. If
                None, we assume the model weights are not quantized.
1344
1345
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
1346
            num_speculative_tokens (Optional[int]): The number of speculative
1347
1348
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
1349
1350
1351
            speculative_disable_mqa_scorer (Optional[bool]): Disable the MQA
                scorer for the speculative model and fall back to batch
                expansion for scoring.
1352
1353
1354
1355
1356
1357
            speculative_max_model_len (Optional[int]): The maximum model len of
                the speculative model. Used when testing the ability to skip
                speculation for some sequences.
            enable_chunked_prefill (bool): Whether vLLM is configured to use
                chunked prefill or not. Used for raising an error since its not
                yet compatible with spec decode.
1358
1359
1360
            speculative_disable_by_batch_size (Optional[int]): Disable
                speculative decoding for new incoming requests when the number
                of enqueue requests  is larger than this value, if provided.
1361
1362
1363
1364
            ngram_prompt_lookup_max (Optional[int]): Max size of ngram token
                window, if provided.
            ngram_prompt_lookup_min (Optional[int]): Min size of ngram token
                window, if provided.
1365
1366
1367
1368
1369
1370
1371
1372
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
1373
                accepted. This threshold is used only when we use the
1374
1375
1376
1377
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1378
1379
1380
1381
1382
            disable_logprobs (Optional[bool]): If set to True, token log
                probabilities are not returned during speculative decoding.
                If set to False, token log probabilities are returned
                according to the log probability settings in SamplingParams.
                If not specified, it defaults to True.
1383

1384
1385
1386
1387
1388
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """

1389
1390
1391
1392
        if speculative_model is None:
            if num_speculative_tokens is not None:
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative_model.")
1393
1394
            return None

1395
1396
1397
1398
1399
1400
        if (speculative_disable_by_batch_size is not None
                and speculative_disable_by_batch_size < 2):
            raise ValueError("Expect the batch size threshold of disabling "
                             "speculative decoding is > 1, but got "
                             f"{speculative_disable_by_batch_size=}")

1401
1402
        # TODO: The user should be able to specify revision/max model len
        # for the draft model. It is not currently supported.
1403
1404
        draft_revision = None
        draft_code_revision = None
1405
        draft_quantization = speculative_model_quantization
1406

1407
1408
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1409
1410
1411
1412
1413
1414
1415
1416
                ngram_prompt_lookup_min = 1
            if ngram_prompt_lookup_max is None or ngram_prompt_lookup_max < 1:
                raise ValueError(f"{ngram_prompt_lookup_max=} must be > 0")
            if ngram_prompt_lookup_min < 1:
                raise ValueError(f"{ngram_prompt_lookup_min=} must be > 0")
            if ngram_prompt_lookup_min > ngram_prompt_lookup_max:
                raise ValueError(f"{ngram_prompt_lookup_min=} cannot be "
                                 f"larger than {ngram_prompt_lookup_max=}")
1417

1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
            # TODO: current we still need extract vocab_size from target model
            # config, in future, we may try refactor it out, and set
            # draft related config as None here.
            draft_model_config = target_model_config
            draft_parallel_config = target_parallel_config
        else:
            ngram_prompt_lookup_max = 0
            ngram_prompt_lookup_min = 0
            draft_model_config = ModelConfig(
                model=speculative_model,
1428
                task="draft",
1429
1430
1431
                tokenizer=target_model_config.tokenizer,
                tokenizer_mode=target_model_config.tokenizer_mode,
                trust_remote_code=target_model_config.trust_remote_code,
1432
1433
                allowed_local_media_path=target_model_config.
                allowed_local_media_path,
1434
1435
1436
1437
1438
1439
                dtype=target_model_config.dtype,
                seed=target_model_config.seed,
                revision=draft_revision,
                code_revision=draft_code_revision,
                tokenizer_revision=target_model_config.tokenizer_revision,
                max_model_len=None,
1440
                spec_target_max_model_len=target_model_config.max_model_len,
1441
1442
                quantization=draft_quantization,
                enforce_eager=target_model_config.enforce_eager,
1443
1444
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1445
1446
1447
                max_logprobs=target_model_config.max_logprobs,
            )

1448
            draft_hf_config = draft_model_config.hf_config
1449

1450
1451
1452
1453
1454
            if (num_speculative_tokens is not None
                    and hasattr(draft_hf_config, "num_lookahead_tokens")):
                draft_hf_config.num_lookahead_tokens = num_speculative_tokens

            n_predict = getattr(draft_hf_config, "n_predict", None)
1455
1456
1457
1458
1459
1460
1461
1462
            if n_predict is not None:
                if num_speculative_tokens is None:
                    # Default to max value defined in draft model config.
                    num_speculative_tokens = n_predict
                elif num_speculative_tokens > n_predict:
                    # Verify provided value doesn't exceed the maximum
                    # supported by the draft model.
                    raise ValueError(
1463
1464
1465
                        "This speculative model supports a maximum of "
                        f"num_speculative_tokens={n_predict}, but "
                        f"{num_speculative_tokens=} was provided.")
1466

1467
1468
1469
1470
1471
1472
            if enable_chunked_prefill and draft_hf_config.model_type in (
                    "medusa", "mlp_speculator", "eagle"):
                raise ValueError(
                    "Chunked prefill and hidden-state based draft models are "
                    "not compatible.")

1473
1474
1475
1476
1477
1478
1479
            speculative_draft_tensor_parallel_size = \
                SpeculativeConfig._verify_and_get_draft_model_tensor_parallel_size(
                    target_parallel_config,
                    speculative_draft_tensor_parallel_size,
                    draft_hf_config
            )

1480
1481
1482
1483
1484
1485
1486
1487
1488
            draft_model_config.max_model_len = (
                SpeculativeConfig._maybe_override_draft_max_model_len(
                    speculative_max_model_len,
                    draft_model_config.max_model_len,
                    target_model_config.max_model_len,
                ))

            draft_parallel_config = (
                SpeculativeConfig.create_draft_parallel_config(
1489
                    target_parallel_config,
1490
                    speculative_draft_tensor_parallel_size, draft_hf_config))
1491

1492
1493
1494
1495
1496
1497
        if num_speculative_tokens is None:
            raise ValueError(
                "num_speculative_tokens must be provided with "
                "speculative_model unless the draft model config contains an "
                "n_predict parameter.")

1498
1499
1500
1501
        if typical_acceptance_sampler_posterior_threshold is None:
            typical_acceptance_sampler_posterior_threshold = 0.09
        if typical_acceptance_sampler_posterior_alpha is None:
            typical_acceptance_sampler_posterior_alpha = 0.3
1502
1503
        if disable_logprobs is None:
            disable_logprobs = True
1504

1505
1506
1507
1508
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
1509
            speculative_disable_mqa_scorer,
1510
            speculative_disable_by_batch_size,
1511
1512
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
1513
1514
1515
1516
1517
            draft_token_acceptance_method=draft_token_acceptance_method,
            typical_acceptance_sampler_posterior_threshold=\
                typical_acceptance_sampler_posterior_threshold,
            typical_acceptance_sampler_posterior_alpha=\
                typical_acceptance_sampler_posterior_alpha,
1518
1519
            disable_logprobs=disable_logprobs,
            disable_log_stats=disable_log_stats,
1520
1521
        )

1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
    @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,
        )

1557
    @staticmethod
1558
1559
1560
1561
1562
1563
1564
    def _verify_and_get_draft_model_tensor_parallel_size(
            target_parallel_config: ParallelConfig,
            speculative_draft_tensor_parallel_size: Optional[int],
            draft_hf_config: PretrainedConfig) -> int:
        """
        Verifies and adjusts the tensor parallel size for a draft model
        specified using speculative_draft_tensor_parallel_size.
1565
        """
1566
1567
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
1568
        if speculative_draft_tensor_parallel_size is None:
1569
1570
1571
1572
1573
1574
1575
1576
1577
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
                        "MLPSpeculator cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1")
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
1578
1579
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
1580
            raise ValueError(
1581
                f"{speculative_draft_tensor_parallel_size=} cannot be "
1582
                f"other value than 1 or target model tensor_parallel_size")
1583
        return speculative_draft_tensor_parallel_size
1584

1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
    @staticmethod
    def create_draft_parallel_config(
        target_parallel_config: ParallelConfig,
        speculative_draft_tensor_parallel_size: int,
        draft_hf_config: PretrainedConfig,
    ) -> ParallelConfig:
        """Create a parallel config for use by the draft worker.

        This is mostly a copy of the target parallel config, except the tp_size.
        """
1595
1596
1597
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
1598
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
1599
1600
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
            max_parallel_loading_workers=target_parallel_config.
            max_parallel_loading_workers,
            disable_custom_all_reduce=target_parallel_config.
            disable_custom_all_reduce,
            tokenizer_pool_config=target_parallel_config.tokenizer_pool_config,
            ray_workers_use_nsight=target_parallel_config.
            ray_workers_use_nsight,
            placement_group=target_parallel_config.placement_group,
        )

        return draft_parallel_config

    def __init__(
        self,
        draft_model_config: ModelConfig,
        draft_parallel_config: ParallelConfig,
        num_speculative_tokens: int,
1618
        speculative_disable_mqa_scorer: Optional[bool],
1619
1620
1621
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1622
1623
1624
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
1625
        disable_logprobs: bool,
1626
        disable_log_stats: bool,
1627
1628
1629
1630
1631
1632
1633
1634
    ):
        """Create a SpeculativeConfig object.

        Args:
            draft_model_config: ModelConfig for the draft model.
            draft_parallel_config: ParallelConfig for the draft model.
            num_speculative_tokens: The number of tokens to sample from the
                draft model before scoring with the target model.
1635
1636
1637
1638
1639
            speculative_disable_by_batch_size: Disable speculative
                decoding for new incoming requests when the number of
                enqueue requests is larger than this value.
            ngram_prompt_lookup_max: Max size of ngram token window.
            ngram_prompt_lookup_min: Min size of ngram token window.
1640
1641
1642
1643
1644
1645
1646
1647
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
1648
                accepted. This threshold is used only when we use the
1649
1650
1651
1652
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1653
            disable_logprobs: If set to True, token log probabilities will not
1654
                be returned even if requested by sampling parameters. This
1655
1656
1657
1658
                reduces latency by skipping logprob calculation in proposal
                sampling, target sampling, and after accepted tokens are
                determined. If set to False, log probabilities will be
                returned.
1659
1660
            disable_log_stats: Whether to disable periodic printing of stage
                times in speculative decoding.
1661
1662
1663
1664
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
1665
        self.speculative_disable_mqa_scorer = speculative_disable_mqa_scorer
1666
1667
1668
1669
        self.speculative_disable_by_batch_size = \
            speculative_disable_by_batch_size
        self.ngram_prompt_lookup_max = ngram_prompt_lookup_max or 0
        self.ngram_prompt_lookup_min = ngram_prompt_lookup_min or 0
1670
1671
1672
1673
1674
        self.draft_token_acceptance_method = draft_token_acceptance_method
        self.typical_acceptance_sampler_posterior_threshold = \
            typical_acceptance_sampler_posterior_threshold
        self.typical_acceptance_sampler_posterior_alpha = \
            typical_acceptance_sampler_posterior_alpha
1675
        self.disable_logprobs = disable_logprobs
1676
        self.disable_log_stats = disable_log_stats
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687

        self._verify_args()

    def _verify_args(self) -> None:
        if self.num_speculative_tokens <= 0:
            raise ValueError("Expected num_speculative_tokens to be greater "
                             f"than zero ({self.num_speculative_tokens}).")

        if self.draft_model_config:
            self.draft_model_config.verify_with_parallel_config(
                self.draft_parallel_config)
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
            # Validate and set draft token acceptance related settings.

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

        if (self.draft_token_acceptance_method != 'rejection_sampler'
                and self.draft_token_acceptance_method !=
                'typical_acceptance_sampler'):
            raise ValueError(
                "Expected draft_token_acceptance_method to be either "
                "rejection_sampler or typical_acceptance_sampler. Instead it "
                f"is {self.draft_token_acceptance_method}")

        if (self.typical_acceptance_sampler_posterior_threshold < 0
                or self.typical_acceptance_sampler_posterior_alpha < 0):
            raise ValueError(
                "Expected typical_acceptance_sampler_posterior_threshold "
                "and typical_acceptance_sampler_posterior_alpha to be > 0. "
                "Instead found "
                f"typical_acceptance_sampler_posterior_threshold = "
                f"{self.typical_acceptance_sampler_posterior_threshold} and "
                f"typical_acceptance_sampler_posterior_alpha = "
                f"{self.typical_acceptance_sampler_posterior_alpha}")
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724

    @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:
1725
1726
1727
1728
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
1729
1730
1731
1732
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


1733
1734
1735
1736
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
1737
    fully_sharded_loras: bool = False
1738
    max_cpu_loras: Optional[int] = None
1739
    lora_dtype: Optional[Union[torch.dtype, str]] = None
1740
1741
1742
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
1743
    long_lora_scaling_factors: Optional[Tuple[float]] = None
1744
    bias_enabled: bool = False
1745
1746

    def __post_init__(self):
1747
1748
1749
        # Setting the maximum rank to 256 should be able to satisfy the vast
        # majority of applications.
        possible_max_ranks = (8, 16, 32, 64, 128, 256)
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
        possible_lora_extra_vocab_size = (0, 256, 512)
        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
1766
                f"max_loras ({self.max_loras})")
1767
1768
1769
1770
1771
1772

    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)
1773
1774
1775
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
1776
            # TODO support marlin
1777
1778
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
1779
1780

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
1781
        # Reminder: Please update docs/source/usage/compatibility_matrix.rst
1782
        # If the feature combo become valid
1783
        if scheduler_config.chunked_prefill_enabled:
1784
1785
            logger.warning("LoRA with chunked prefill is still experimental "
                           "and may be unstable.")
1786
1787


1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
@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

    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)


1813
@dataclass
1814
class MultiModalConfig:
1815
1816
    """Controls the behavior of multimodal models."""

1817
    limit_per_prompt: Mapping[str, int] = field(default_factory=dict)
1818
1819
1820
1821
1822
    """
    The maximum number of multi-modal input instances allowed per prompt
    for each :class:`~vllm.multimodal.MultiModalPlugin`.
    """

1823
    # TODO: Add configs to init vision tower or not.
1824

1825

1826
1827
@dataclass
class PoolerConfig:
1828
    """Controls the behavior of output pooling in pooling models."""
1829
1830

    pooling_type: Optional[str] = None
1831
    """
1832
    The pooling method of the pooling model. This should be a key in
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
    :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
    """
1850
    If set, only the score corresponding to the ``step_tag_id`` in the
1851
1852
1853
1854
1855
1856
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

    returned_token_ids: Optional[List[int]] = None
    """
1857
1858
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
1859
1860
1861
1862
1863
1864
    ``math-shepherd-mistral-7b-prm`` model.
    """

    @staticmethod
    def from_json(json_str: str) -> "PoolerConfig":
        return PoolerConfig(**json.loads(json_str))
1865
1866


1867
1868
1869
1870
1871
1872
1873
1874
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

1875
_ROCM_NOT_SUPPORTED_DTYPE: List[str] = []  #
1876

1877
1878
1879

def _get_and_verify_dtype(
    config: PretrainedConfig,
1880
    dtype: Union[str, torch.dtype],
1881
1882
1883
1884
1885
1886
1887
) -> 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)
    if config_dtype is None:
        config_dtype = torch.float32

1888
1889
1890
1891
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
Woosuk Kwon's avatar
Woosuk Kwon committed
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
                if config.model_type == "gemma2":
                    logger.info(
                        "For Gemma 2, we downcast float32 to bfloat16 instead "
                        "of float16 by default. Please specify `dtype` if you "
                        "want to use float16.")
                    torch_dtype = torch.bfloat16
                else:
                    # Following the common practice, we use float16 for float32
                    # models.
                    torch_dtype = torch.float16
1902
1903
            else:
                torch_dtype = config_dtype
1904
1905
1906
1907
1908
1909
1910

            if current_platform.is_hpu() and config_dtype == torch.float16:
                logger.info(
                    "For HPU, we cast models to bfloat16 instead of"
                    "using float16 by default. Please specify `dtype` if you "
                    "want to use float16.")
                torch_dtype = torch.bfloat16
1911
        else:
1912
1913
1914
1915
1916
            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
1917
    else:
1918
        raise ValueError(f"Unknown dtype: {dtype}")
1919
1920
1921
1922
1923

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
1924
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
1925
1926
1927
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
1928
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
1929
1930
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
1931
            # Casting between float16 and bfloat16 is allowed with a warning.
1932
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
1933
1934

    return torch_dtype
1935
1936
1937
1938
1939


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
1940
    disable_sliding_window: bool,
1941
    sliding_window_len: Optional[Union[int, List[Optional[int]]]],
1942
    spec_target_max_model_len: Optional[int] = None,
1943
    encoder_config: Optional[Any] = None,
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
) -> 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",
1954
1955
        # ChatGLM2
        "seq_length",
1956
1957
        # Command-R
        "model_max_length",
1958
1959
1960
1961
1962
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
1963
    # Choose the smallest "max_length" from the possible keys.
1964
    max_len_key = None
1965
    for key in possible_keys:
1966
1967
1968
1969
1970
        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)
1971
1972
1973
1974

    # 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:
1975
1976

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
1977
        max_len_key = "sliding_window" \
1978
1979
1980
            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)
1981
1982
1983

    # If none of the keys were found in the config, use a default and
    # log a warning.
1984
    if derived_max_model_len == float("inf"):
1985
1986
1987
1988
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

1989
1990
1991
1992
1993
        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

1994
1995
1996
1997
        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: "
1998
            "%s. Assuming the model's maximum length is %d.", possible_keys,
1999
            default_max_len)
2000
        derived_max_model_len = default_max_len
2001

2002
    rope_scaling = getattr(hf_config, "rope_scaling", None)
2003
    if rope_scaling is not None:
2004
2005
2006
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016

        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.")

2017
2018
2019
2020
            # 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)

2021
2022
2023
2024
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
2025

2026
2027
2028
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2029
2030
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2031
    if max_model_len is None:
2032
        max_model_len = int(derived_max_model_len)
2033
    elif max_model_len > derived_max_model_len:
2034
2035
2036
2037
2038
        # 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:
2039
2040
2041
2042
2043
2044
2045
            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.")
2046
        else:
2047
            msg = (
2048
                f"User-specified max_model_len ({max_model_len}) is greater "
2049
2050
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2051
                f"{model_max_length} in model's config.json). This may lead "
2052
2053
2054
2055
2056
2057
2058
2059
2060
                "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")
2061
    return int(max_model_len)
2062
2063


2064
2065
2066
2067
2068
2069
2070
2071
def get_min_sliding_window(
        sliding_window: Union[int, List[Optional[int]]]) -> int:
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


2072
2073
2074
def get_served_model_name(model: str,
                          served_model_name: Optional[Union[str, List[str]]]):
    """
2075
2076
2077
2078
    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
2079
2080
2081
2082
2083
2084
2085
2086
2087
    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


2088
2089
2090
2091
@dataclass
class DecodingConfig:
    """Dataclass which contains the decoding strategy of the engine"""

2092
2093
2094
    # Which guided decoding algo to use.
    # 'outlines' / 'lm-format-enforcer' / 'xgrammar'
    guided_decoding_backend: str = 'xgrammar'
2095
2096

    def __post_init__(self):
2097
        valid_guided_backends = ['outlines', 'lm-format-enforcer', 'xgrammar']
2098
2099
2100
2101
2102
2103
        backend = self.guided_decoding_backend
        if backend not in valid_guided_backends:
            raise ValueError(f"Invalid guided_decoding_backend '{backend},"
                             f"must be one of {valid_guided_backends}")


2104
2105
2106
2107
2108
@dataclass
class ObservabilityConfig:
    """Configuration for observability."""
    otlp_traces_endpoint: Optional[str] = None

2109
2110
2111
2112
2113
2114
2115
2116
    # 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

2117
    def __post_init__(self):
2118
2119
2120
2121
2122
        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}")
2123
2124


2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
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

    @classmethod
    def from_cli(cls, cli_value: str) -> "KVTransferConfig":
youkaichao's avatar
youkaichao committed
2160
        """Parse the CLI value for the kv cache transfer config."""
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
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
        return KVTransferConfig.model_validate_json(cli_value)

    def model_post_init(self, __context: Any) -> None:
        if all([
                self.kv_connector is not None,
                self.kv_connector != "PyNcclConnector"
        ]):
            raise ValueError(f"Unsupported kv_connector: {self.kv_connector}. "
                             f"Supported connectors are "
                             f"`PyNcclConnector`.")

        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"]


2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
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.
2225
        - debug_dump_path: the path to dump the debug information.
2226
2227
2228
2229
2230
2231
2232
        - 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).
2233
2234
2235
2236
2237
2238
2239
2240
2241
        - 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).
2242
        - splitting_ops: a list of ops to split the full graph into subgraphs, used in piecewise compilation.
2243
2244
2245
2246
    - 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
2247
2248
2249
2250
                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.
2251
2252
2253
            TODO: move outside cudagraph logic into compilation.
            torch.compile will handle cudagraph capture logic in the future.
        - cudagraph_capture_sizes: sizes to capture cudagraph.
2254
2255
            - None (default): capture sizes are inferred from vllm config.
            - List[int]: capture sizes are specified as given.
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
        - 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
2269
2270
                is compiled. In addition, compile for cudagraph sizes that are
                in candidate_compile_sizes, using configurations
2271
                in inductor_compile_config.
2272
        - candidate_compile_sizes: sizes to compile for inductor.
2273
2274
2275
2276
2277
2278
2279
        - 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})`
2280
        - custom inductor passes: see PassConfig for more details
2281

2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
    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
2293
    debug_dump_path: str = ""
2294
    backend: str = ""
2295
    custom_ops: List[str] = Field(default_factory=list)
2296
    splitting_ops: List[str] = Field(default_factory=lambda: [
2297
        "vllm.unified_attention",
2298
        "vllm.unified_attention_with_output",
2299
    ])
2300
2301

    use_inductor: bool = True
2302
    candidate_compile_sizes: Optional[List[int]] = Field(default=None)
2303
2304
2305
2306
2307
2308
2309
2310
    inductor_compile_config: Dict = Field(default_factory=dict)
    inductor_passes: Dict[str, str] = Field(default_factory=dict)

    use_cudagraph: bool = False
    cudagraph_num_of_warmups: int = 0
    cudagraph_capture_sizes: Optional[List[int]] = None
    cudagraph_copy_inputs: bool = False

2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
    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.
        - enable_reshape: whether to enable the custom reshape elimination pass.
            TODO better pass enabling system.
        """
        dump_graph_stages: List[str] = Field(default_factory=list)
        dump_graph_dir: Path = Field(default=Path("."))
        enable_fusion: bool = True
        enable_reshape: bool = True

        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.
            """
            dict_ = self.model_dump(
                include={"enable_fusion", "enable_reshape"})
            encoded = json.dumps(dict_, sort_keys=True).encode("utf-8")
            return hashlib.sha256(encoded).digest()

        def model_post_init(self, __context: Any) -> None:
            if not self.enable_reshape and self.enable_fusion:
                print_warning_once(
                    "Fusion enabled but reshape elimination disabled."
                    "RMSNorm + quant (fp8) fusion might not work")

    pass_config: PassConfig = Field(default_factory=PassConfig)
2348
2349
2350
2351
2352

    # not configurable, computed after init
    compile_sizes: List[int] = PrivateAttr
    capture_sizes: List[int] = PrivateAttr

2353
2354
2355
    # keep track of enabled and disabled custom ops
    enabled_custom_ops: Counter[str] = PrivateAttr
    disabled_custom_ops: Counter[str] = PrivateAttr
2356
    compilation_time: float = PrivateAttr
2357

2358
2359
2360
2361
2362
    # Per-model forward context
    # Mainly used to store attention cls
    # Map from layer name to the attention cls
    static_forward_context: Dict[str, Any] = PrivateAttr

2363
2364
2365
2366
2367
    @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))
2368
2369
2370
        # 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)
2371

2372
2373
2374
2375
2376
2377
2378
2379
2380
    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'"

        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
2381
2382
2383
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
2384
2385
2386
2387
2388
2389
2390
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
2391
2392
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
2393

2394
2395
        self.enabled_custom_ops = Counter()
        self.disabled_custom_ops = Counter()
2396
        self.static_forward_context = {}
2397
        self.compilation_time = 0.0
2398

2399
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
        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
        from vllm.compilation.backends import VllmBackend
2418
        return VllmBackend(vllm_config)
2419

2420
    def init_with_cudagraph_sizes(self, sizes_to_specialize: List[int]):
2421
        """To complete the initialization of config,
2422
2423
        we need to know the cudagraph sizes."""

2424
2425
2426
2427
2428
2429
2430
        if self.cudagraph_capture_sizes is None:
            self.capture_sizes = sizes_to_specialize
        else:
            self.capture_sizes = self.cudagraph_capture_sizes
            logger.info(("cudagraph sizes specified by model runner"
                         " %s is overridden by config %s"),
                        sizes_to_specialize, self.cudagraph_capture_sizes)
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444

        if self.candidate_compile_sizes is None:
            self.candidate_compile_sizes = []
        self.compile_sizes = [
            x for x in self.candidate_compile_sizes if x in self.capture_sizes
        ]
        ignored_sizes = [
            x for x in self.candidate_compile_sizes
            if x not in self.capture_sizes
        ]
        if ignored_sizes:
            logger.warning(("candidate_compile_sizes %s are ignored "
                            "because they are not cudagraph capture sizes."),
                           ignored_sizes)
2445

2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
        # sort to make sure cudagraph capture sizes are in descending order
        self.capture_sizes.sort(reverse=True)


_BATCH_SIZE_ALIGNMENT = 8
# all the token sizes that **can** be captured by cudagraph.
# they can be arbitrarily large.
# currently it includes: 1, 2, 4, 8, 16, 24, 32, 40, ..., 8192.
# the actual sizes to capture will be determined by the model,
# depending on the model's max_num_seqs.
# NOTE: get_graph_batch_size needs to be updated if this list is changed.
_BATCH_SIZES_TO_CAPTURE = [1, 2, 4] + [
    _BATCH_SIZE_ALIGNMENT * i for i in range(1, 1025)
]

2461

2462
2463
2464
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
2465
2466
2467
    simplifies passing around the distinct configurations in the codebase.
    """

2468
2469
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
2470
2471
2472
2473
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig,
                                            init=True)
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig,
                                              init=True)
2474
2475
2476
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
2477
2478
2479
2480
2481
    lora_config: Optional[LoRAConfig] = None
    speculative_config: Optional[SpeculativeConfig] = None
    decoding_config: Optional[DecodingConfig] = None
    observability_config: Optional[ObservabilityConfig] = None
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
2482
    quant_config: Optional[QuantizationConfig] = None
2483
2484
    compilation_config: CompilationConfig = field(default=None,
                                                  init=True)  # type: ignore
2485
2486
    kv_transfer_config: KVTransferConfig = field(default=None,
                                                 init=True)  # type: ignore
2487
    instance_id: str = ""
2488

2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
    @staticmethod
    def get_graph_batch_size(batch_size: int) -> int:
        """Returns the padded batch size given actual batch size.

        Batch sizes are 1, 2, 4, _BATCH_SIZE_ALIGNMENT,
        2*_BATCH_SIZE_ALIGNMENT, 3*_BATCH_SIZE_ALIGNMENT...
        """
        if batch_size <= 2:
            return batch_size
        elif batch_size <= 4:
            return 4
        else:
            return ((batch_size + _BATCH_SIZE_ALIGNMENT - 1) //
                    _BATCH_SIZE_ALIGNMENT * _BATCH_SIZE_ALIGNMENT)

    @staticmethod
    def get_max_graph_batch_size(max_num_seqs: int) -> int:
        """
        max_num_seqs: Maximum number of sequences in a batch.
        _BATCH_SIZES_TO_CAPTURE: all the sizes that we want to capture.

        pad the max_num_seqs if necessary by calling get_graph_batch_size,
        which will deal with some edge cases like 1, 2, 4.

        if the padded size is in _BATCH_SIZES_TO_CAPTURE, return the padded
        size. if not, it means the padded size is larger than the largest size
        in _BATCH_SIZES_TO_CAPTURE, return the largest size in
        _BATCH_SIZES_TO_CAPTURE.
        """
        padded_size = VllmConfig.get_graph_batch_size(max_num_seqs)
        if padded_size in _BATCH_SIZES_TO_CAPTURE:
            return padded_size
        assert padded_size > _BATCH_SIZES_TO_CAPTURE[-1]
        return _BATCH_SIZES_TO_CAPTURE[-1]

2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
        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
2551

2552
2553
2554
2555
2556
2557
2558
2559
2560
    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

2561
2562
2563
2564
2565
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

2566
2567
2568
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
2569
2570
2571
2572
2573
2574
2575
2576
        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)
2577
2578
2579
2580
2581

        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
2582
2583
2584
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
2585
2586
2587
2588
2589

        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)
2590

2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
        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):
            print_warning_once(
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

2601
        if self.compilation_config is None:
2602
            self.compilation_config = CompilationConfig()
2603
        if envs.VLLM_USE_V1 and not self.model_config.enforce_eager:
2604
2605
2606
2607
2608
2609
2610
            # 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.
            self.compilation_config.custom_ops = ["none"]
            self.compilation_config.use_cudagraph = True
            self.compilation_config.use_inductor = True
2611
            self.compilation_config.cudagraph_num_of_warmups = 1
2612
2613
            self.compilation_config.pass_config.enable_fusion = False
            self.compilation_config.pass_config.enable_reshape = False
2614
            self.compilation_config.level = CompilationLevel.PIECEWISE
2615

2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
        if not envs.VLLM_USE_V1:
            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:
                max_batchsize_to_capture = \
                    self.get_max_graph_batch_size(
                    self.scheduler_config.max_num_seqs)
            batch_size_capture_list = [
                size for size in _BATCH_SIZES_TO_CAPTURE
                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)]

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
        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

        if self.lora_config is not None and self.compilation_config.level !=\
             CompilationLevel.NO_COMPILATION:
            logger.warning("LoRA is not supported with `torch.compile` yet. "
                           "Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

2652
2653
        current_platform.check_and_update_config(self)

2654
2655
2656
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

2657
    def __str__(self):
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
        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"quantization_param_path={self.model_config.quantization_param_path},"
            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}, "
            f"mm_processor_kwargs={self.model_config.mm_processor_kwargs}, "
            f"pooler_config={self.model_config.pooler_config!r},"
            f" compilation_config={self.compilation_config!r}")
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741


_current_vllm_config: Optional[VllmConfig] = None


@contextmanager
def set_current_vllm_config(vllm_config: VllmConfig):
    """
    Temporarily set the current VLLM config.
    Used during model initialization.
    We save the current VLLM config in a global variable,
    so that all modules can access it, e.g. custom ops
    can access the VLLM config to determine how to dispatch.
    """
    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)
        if vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
            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"
                "if you want it to be supported.",
                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.
        logger.warning("Current VLLM config is not set.")
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config