config.py 144 KB
Newer Older
1
import ast
2
import copy
3
import enum
4
import hashlib
5
import json
6
import sys
7
import warnings
8
from contextlib import contextmanager
9
from dataclasses import dataclass, field, replace
10
from pathlib import Path
11
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Counter, Dict,
12
13
                    Final, List, Literal, Mapping, Optional, Protocol, Set,
                    Tuple, Type, Union)
14
15

import torch
16
from pydantic import BaseModel, Field, PrivateAttr
17
from transformers import PretrainedConfig
18

19
import vllm.envs as envs
20
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
21
from vllm.logger import init_logger
22
23
from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
                                                     get_quantization_config)
24
from vllm.model_executor.models import ModelRegistry
25
from vllm.platforms import CpuArchEnum
26
from vllm.tracing import is_otel_available, otel_import_error_traceback
27
28
29
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
30
31
    get_sentence_transformer_tokenizer_config, is_encoder_decoder,
    try_get_generation_config, uses_mrope)
32
from vllm.transformers_utils.s3_utils import S3Model
33
from vllm.transformers_utils.utils import is_s3
34
from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
35
                        get_cpu_memory, random_uuid, resolve_obj_by_qualname)
36

37
38
39
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

40
    from vllm.executor.executor_base import ExecutorBase
41
42
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
43
    from vllm.model_executor.model_loader.loader import BaseModelLoader
44
45
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
46
47
else:
    QuantizationConfig = None
48

49
50
logger = init_logger(__name__)

51
_POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
52
_MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
53

54
55
TaskOption = Literal["auto", "generate", "embedding", "embed", "classify",
                     "score", "reward"]
56

57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
_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
}
72

73
74
75
HfOverrides = Union[Dict[str, Any], Callable[[PretrainedConfig],
                                             PretrainedConfig]]

76

77
78
79
80
81
82
class SupportsHash(Protocol):

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


83
class ModelConfig:
84
85
86
87
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
88
            It is also used as the content for `model_name` tag in metrics
89
90
91
92
93
            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.
94
        tokenizer: Name or path of the huggingface tokenizer to use.
95
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
96
97
            available, "slow" will always use the slow tokenizer, and
            "mistral" will always use the tokenizer from `mistral_common`.
98
99
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
100
101
102
103
        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.
104
105
106
107
        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
108
109
110
        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.
111
        code_revision: The specific revision to use for the model code on
112
            Hugging Face Hub. It can be a branch name, a tag name, or a
113
            commit id. If unspecified, will use the default version.
114
115
116
        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.
117
118
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
119
120
        spec_target_max_model_len: Specify the the maximum length for spec
            decoding draft models.
121
122
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
123
124
125
        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.
126
            If None, the user did not specify, so default to False.
127
128
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
129
130
131
            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.
132
        max_logprobs: Maximum number of log probabilities. Defaults to 20.
133
134
135
136
        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.
137
138
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
139
        served_model_name: The model name used in metrics tag `model_name`,
140
141
            matches the model name exposed via the APIs. If multiple model
            names provided, the first name will be used. If not specified,
142
            the model name will be the same as `model`.
143
        limit_mm_per_prompt: Maximum number of data items per modality
144
            per prompt. Only applicable for multimodal models.
145
146
        use_async_output_proc: Whether to use async output processor.
            Defaults to True.
147
148
        config_format: The config format which shall be loaded.
            Defaults to 'auto' which defaults to 'hf'.
149
150
151
        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.
152
153
        mm_processor_kwargs: Arguments to be forwarded to the model's processor
            for multi-modal data, e.g., image processor.
154
155
        disable_mm_preprocessor_cache: If true, then disables caching of the
            multi-modal preprocessor/mapper. (not recommended)
156
157
158
159
        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.
160
        override_pooler_config: Initialize non default pooling config or
161
            override default pooling config for the pooling model.
162
163
        logits_processor_pattern: Optional regex pattern specifying valid
            logits processor qualified names that can be passed with the
164
            `logits_processors` extra completion argument. Defaults to None,
165
            which allows no processors.
166
        generation_config: Configuration parameter file for generation.
167
    """
168

169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: List[Any] = []
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
        factors.append(self.trust_remote_code)
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
        return hashlib.sha256(str(factors).encode()).hexdigest()

192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
    def __init__(
        self,
        model: str,
        task: Union[TaskOption, Literal["draft"]],
        tokenizer: str,
        tokenizer_mode: str,
        trust_remote_code: bool,
        dtype: Union[str, torch.dtype],
        seed: int,
        allowed_local_media_path: str = "",
        revision: Optional[str] = None,
        code_revision: Optional[str] = None,
        rope_scaling: Optional[Dict[str, Any]] = None,
        rope_theta: Optional[float] = None,
        tokenizer_revision: Optional[str] = None,
        max_model_len: Optional[int] = None,
        spec_target_max_model_len: Optional[int] = None,
        quantization: Optional[str] = None,
        enforce_eager: Optional[bool] = None,
        max_seq_len_to_capture: Optional[int] = None,
        max_logprobs: int = 20,
        disable_sliding_window: bool = False,
        skip_tokenizer_init: bool = False,
        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,
        hf_overrides: Optional[HfOverrides] = None,
        mm_processor_kwargs: Optional[Dict[str, Any]] = None,
        disable_mm_preprocessor_cache: bool = False,
        override_neuron_config: Optional[Dict[str, Any]] = None,
        override_pooler_config: Optional["PoolerConfig"] = None,
        logits_processor_pattern: Optional[str] = None,
        generation_config: Optional[str] = None,
        enable_sleep_mode: bool = False,
    ) -> None:
228
        self.model = model
229
        self.tokenizer = tokenizer
230
        self.tokenizer_mode = tokenizer_mode
231
        self.trust_remote_code = trust_remote_code
232
        self.allowed_local_media_path = allowed_local_media_path
233
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
234
        self.revision = revision
235
        self.code_revision = code_revision
236
237
        self.rope_scaling = rope_scaling
        self.rope_theta = rope_theta
238
239
240

        if hf_overrides is None:
            hf_overrides = {}
241
242
243
244
245
246

        if callable(hf_overrides):
            hf_overrides_kw = {}
            hf_overrides_fn = hf_overrides
        else:
            hf_overrides_kw = hf_overrides
247
            hf_overrides_fn = None
248

249
250
        if rope_scaling is not None:
            hf_override: Dict[str, Any] = {"rope_scaling": rope_scaling}
251
            hf_overrides_kw.update(hf_override)
252
253
254
255
256
            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}
257
            hf_overrides_kw.update(hf_override)
258
259
260
261
            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)

262
263
        self.maybe_pull_model_tokenizer_for_s3(model, tokenizer)

264
265
266
267
268
        # 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
269
        self.quantization = quantization
270
        self.enforce_eager = enforce_eager
271
        self.max_seq_len_to_capture = max_seq_len_to_capture
272
        self.max_logprobs = max_logprobs
273
        self.disable_sliding_window = disable_sliding_window
274
        self.skip_tokenizer_init = skip_tokenizer_init
275
276
277
278
279
280
        self.enable_sleep_mode = enable_sleep_mode

        from vllm.platforms import current_platform

        if self.enable_sleep_mode and not current_platform.is_cuda():
            raise ValueError("Sleep mode is only supported on CUDA devices.")
281
282

        hf_config = get_config(self.model, trust_remote_code, revision,
283
284
285
286
287
288
289
290
291
                               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)

292
293
        self.hf_config = hf_config

294
        self.hf_text_config = get_hf_text_config(self.hf_config)
295
        self.encoder_config = self._get_encoder_config()
296
297
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, revision)
298
        self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
299
        self.use_async_output_proc = use_async_output_proc
300
        self.mm_processor_kwargs = mm_processor_kwargs
301
        self.disable_mm_preprocessor_cache = disable_mm_preprocessor_cache
Woosuk Kwon's avatar
Woosuk Kwon committed
302

303
304
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
305
306
            self.enforce_eager = False

307
308
309
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
        has_interleaved_attention = (sliding_window is not None) and (
            isinstance(sliding_window, list) or
310
            (self.hf_text_config.model_type in ["gemma2", "cohere2"]))
311
312

        if (not self.disable_sliding_window and has_interleaved_attention):
313
314
315
            if envs.VLLM_ATTENTION_BACKEND == "XFORMERS":
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
316

317
                logger.warning_once(
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
                    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
333

334
335
336
337
        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,
338
            sliding_window_len=self.get_hf_config_sliding_window(),
339
340
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
341
342
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
343
344
        self.multimodal_config = self._init_multimodal_config(
            limit_mm_per_prompt)
345
346
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
347

348
        self.is_attention_free = self._init_attention_free()
349
        self.is_hybrid = self._init_is_hybrid()
350
351
        self.has_inner_state = self._init_has_inner_state()

352
353
354
355
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
356
357
358
359

        supported_tasks, task = self._resolve_task(task, self.hf_config)
        self.supported_tasks = supported_tasks
        self.task: Final = task
360
361
362
363
        if self.task in ("draft", "generate"):
            self.truncation_side = "left"
        else:
            self.truncation_side = "right"
364

365
        self.pooler_config = self._init_pooler_config(override_pooler_config)
366
        self.logits_processor_pattern = logits_processor_pattern
367

368
369
        self.generation_config = generation_config

370
        self._verify_quantization()
371
        self._verify_cuda_graph()
372
        self._verify_bnb_config()
373

374
375
376
    def maybe_pull_model_tokenizer_for_s3(self, model: str,
                                          tokenizer: str) -> None:
        """
377
        Pull the model config or tokenizer to a temporary
378
379
380
381
382
383
384
385
386
        directory in case of S3.

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

        """
        if is_s3(model) or is_s3(tokenizer):
            if is_s3(model):
387
388
                s3_model = S3Model()
                s3_model.pull_files(model, allow_pattern=["*config.json"])
389
                self.model_weights = self.model
390
                self.model = s3_model.dir
391
392

            if is_s3(tokenizer):
393
394
                s3_tokenizer = S3Model()
                s3_tokenizer.pull_files(
395
                    model, ignore_pattern=["*.pt", "*.safetensors", "*.bin"])
396
                self.tokenizer = s3_tokenizer.dir
397

398
399
400
401
    def _init_multimodal_config(
        self, limit_mm_per_prompt: Optional[Mapping[str, int]]
    ) -> Optional["MultiModalConfig"]:
        architectures = getattr(self.hf_config, "architectures", [])
402
        if ModelRegistry.is_multimodal_model(architectures):
403
            return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
404
405
406
407
408
409

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

        return None
410

411
412
413
414
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

415
416
    def _init_pooler_config(
        self,
417
        override_pooler_config: Optional["PoolerConfig"],
418
    ) -> Optional["PoolerConfig"]:
419

420
        if self.runner_type == "pooling":
421
422
423
424
425
426
427
428
429
430
431
            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

432
433
        return None

434
435
436
437
    def _init_attention_free(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_attention_free_model(architectures)

438
439
440
441
    def _init_is_hybrid(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_hybrid_model(architectures)

442
443
444
445
    def _init_has_inner_state(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.model_has_inner_state(architectures)

446
447
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
448
        if tokenizer_mode not in ["auto", "slow", "mistral"]:
449
450
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
451
                "either 'auto', 'slow' or 'mistral'.")
452
        self.tokenizer_mode = tokenizer_mode
453

454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
    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

483
484
    def _resolve_task(
        self,
485
        task_option: Union[TaskOption, Literal["draft"]],
486
        hf_config: PretrainedConfig,
487
    ) -> Tuple[Set[_ResolvedTask], _ResolvedTask]:
488
489
490
        if task_option == "draft":
            return {"draft"}, "draft"

491
492
        architectures = getattr(hf_config, "architectures", [])

493
        runner_support: Dict[RunnerType, bool] = {
494
495
496
            # NOTE: Listed from highest to lowest priority,
            # in case the model supports multiple of them
            "generate": ModelRegistry.is_text_generation_model(architectures),
497
            "pooling": ModelRegistry.is_pooling_model(architectures),
498
        }
499
500
501
502
503
504
505
506
507
        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]
508
509
510
511
512
        ]
        supported_tasks = set(supported_tasks_lst)

        if task_option == "auto":
            selected_task = next(iter(supported_tasks_lst))
513

514
515
516
517
518
            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
519

520
521
522
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
523
        else:
524
525
526
527
528
529
530
531
532
533
534
535
536
537
            # 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"

538
539
540
541
542
543
544
            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
545

546
        return supported_tasks, selected_task
547

548
549
550
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
551
            # compressed-tensors uses a "compression_config" key
552
            quant_cfg = getattr(self.hf_config, "compression_config", None)
553
554
        return quant_cfg

555
    def _verify_quantization(self) -> None:
556
        supported_quantization = QUANTIZATION_METHODS
557
        optimized_quantization_methods = [
558
559
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
            "awq_marlin", "fbgemm_fp8", "compressed_tensors",
560
            "compressed-tensors", "experts_int8", "quark"
561
        ]
562
563
564
565
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
566
567
        quant_cfg = self._parse_quant_hf_config()

568
569
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
570
571

            # Detect which checkpoint is it
572
573
            for name in QUANTIZATION_METHODS:
                method = get_quantization_config(name)
574
575
576
577
578
579
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
580

581
            # Verify quantization configurations.
582
            if self.quantization is None:
583
584
                self.quantization = quant_method
            elif self.quantization != quant_method:
585
586
                raise ValueError(
                    "Quantization method specified in the model config "
587
                    f"({quant_method}) does not match the quantization "
588
589
590
591
592
593
594
595
                    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}.")
596
            from vllm.platforms import current_platform
597
            current_platform.verify_quantization(self.quantization)
598
            if self.quantization not in optimized_quantization_methods:
599
                logger.warning(
600
                    "%s quantization is not fully "
601
                    "optimized yet. The speed can be slower than "
602
                    "non-quantized models.", self.quantization)
603

604
    def _verify_cuda_graph(self) -> None:
605
606
607
608
        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)
609

610
        MODEL_NOT_SUPPORT_CUDA_GRAPH = ['mllama']
611
        if (self.hf_config.model_type in MODEL_NOT_SUPPORT_CUDA_GRAPH
Simon Mo's avatar
Simon Mo committed
612
                and not self.enforce_eager):
613
614
615
            logger.warning(
                "CUDA graph is not supported for %s yet, fallback to the eager "
                "mode.", self.hf_config.model_type)
Simon Mo's avatar
Simon Mo committed
616
617
            self.enforce_eager = True

618
619
    def _verify_bnb_config(self) -> None:
        """
620
        The current version of bitsandbytes (0.44.0) with 8-bit models does not
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
        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

640
641
642
643
644
645
646
647
648
649
650
651
    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

652
        # Reminder: Please update docs/source/features/compatibility_matrix.md
653
        # If the feature combo become valid
654
        from vllm.platforms import current_platform
655
        if not current_platform.is_async_output_supported(self.enforce_eager):
656
            logger.warning(
657
658
                "Async output processing is not supported on the "
                "current platform type %s.", current_platform.device_type)
659
660
661
662
663
664
665
666
667
            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

668
        # Async postprocessor is not necessary for pooling models
669
        # since there is no token generation
670
        if self.runner_type == "pooling":
671
672
            self.use_async_output_proc = False

673
        # Reminder: Please update docs/source/features/compatibility_matrix.md
674
        # If the feature combo become valid
675
676
677
678
679
        if speculative_config:
            logger.warning("Async output processing is not supported with"
                           " speculative decoding currently.")
            self.use_async_output_proc = False

680
681
682
683
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
684
685
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
686
687
688
689
690
691
692
693
        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
694
695
696
697
698
699
700
701
702
703
704
        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
705

706
707
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], List[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
708
        """Get the sliding window size, or None if disabled."""
709
710
711
712

        # 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.
713
714
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
715
            return None
716
        return getattr(self.hf_text_config, "sliding_window", None)
717

718
    def get_sliding_window(self) -> Optional[Union[int, List[Optional[int]]]]:
719
720
721
722
723
724
725
726
        """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()

727
    def get_vocab_size(self) -> int:
728
        return self.hf_text_config.vocab_size
729

730
    def get_hidden_size(self) -> int:
731
        return self.hf_text_config.hidden_size
732
733

    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
734
        # TODO remove hard code
Simon Mo's avatar
Simon Mo committed
735
736
737
        if hasattr(self.hf_text_config,
                   "model_type") and (self.hf_text_config.model_type
                                      in ('deepseek_v2', 'deepseek_v3')):
738
739
740
741
742
743
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
                                       0)
            qk_nope_head_dim = getattr(self.hf_text_config, "qk_nope_head_dim",
                                       0)
            if qk_rope_head_dim and qk_nope_head_dim:
                return qk_rope_head_dim + qk_nope_head_dim
744
745
746
747

        if self.is_attention_free:
            return 0

748
749
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
750
        # FIXME(woosuk): This may not be true for all models.
751
752
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
753

754
755
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
756
        # For GPTBigCode & Falcon:
757
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
758
759
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
760
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
761
        new_decoder_arch_falcon = (
762
            self.hf_config.model_type in falcon_model_types
763
            and getattr(self.hf_config, "new_decoder_architecture", False))
764
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
765
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
766
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
767
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
768
            return 1
769

770
        # For DBRX and MPT
771
772
773
774
775
        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":
776
777
778
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

779
780
781
        if self.is_attention_free:
            return 0

782
783
784
785
786
787
788
789
790
791
        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:
792
            num_kv_heads = getattr(self.hf_text_config, attr, None)
793
794
795
796
797
            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.
798
        return self.hf_text_config.num_attention_heads
799
800
801
802
803
804
805
806
807
808

    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)
809

810
811
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
812
813
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
814

815
816
    def get_layers_start_end_indices(
            self, parallel_config: "ParallelConfig") -> Tuple[int, int]:
817
        from vllm.distributed.utils import get_pp_indices
Mor Zusman's avatar
Mor Zusman committed
818
819
        total_num_hidden_layers = getattr(self.hf_text_config,
                                          "num_hidden_layers", 0)
820
821
822
        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)
823
        return start, end
Mor Zusman's avatar
Mor Zusman committed
824

825
826
827
    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
828

829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
    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
860

861
862
863
864
865
866
867
868
869
870
871
872
    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

873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
    def try_get_generation_config(self) -> Dict[str, Any]:
        if self.generation_config is None or self.generation_config == "auto":
            config = try_get_generation_config(
                self.model,
                trust_remote_code=self.trust_remote_code,
                revision=self.revision,
            )
        else:
            config = try_get_generation_config(
                self.generation_config,
                trust_remote_code=self.trust_remote_code,
            )

        if config is None:
            return {}

        return config.to_diff_dict()

    def get_diff_sampling_param(self) -> Dict[str, Any]:
        """
893
894
895
        This method returns a dictionary containing the parameters
        that differ from the default sampling parameters, but only
        if `generation_config` is set. If `generation_config` is not
896
897
898
        set, an empty dictionary is returned.

        Returns:
899
900
            Dict[str, Any]: A dictionary with the differing sampling
            parameters if `generation_config` is set, otherwise an
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
            empty dictionary.
        """
        if self.generation_config is None:
            # When generation_config is not set
            return {}
        config = self.try_get_generation_config()
        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
        ]
        if any(p in config for p in available_params):
            diff_sampling_param = {
                p: config.get(p)
                for p in available_params if config.get(p) is not None
            }
        else:
            diff_sampling_param = {}
        return diff_sampling_param

923
    @property
924
    def is_encoder_decoder(self) -> bool:
925
        """Extract the HF encoder/decoder model flag."""
926
927
928
929
930
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
931

932
933
934
935
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

936
937
938
939
940
    @property
    def is_cross_encoder(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_cross_encoder_model(architectures)

941
942
943
944
945
946
947
948
    @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]

949
950

class CacheConfig:
951
952
953
954
955
    """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
956
            vLLM execution.
957
        swap_space: Size of the CPU swap space per GPU (in GiB).
958
        cache_dtype: Data type for kv cache storage.
959
        is_attention_free: Whether the model is attention-free.
960
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
961
            profiled num_gpu_blocks if specified. Does nothing if None.
962
963
964
965
        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.
966
    """
967

968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: List[Any] = []
        factors.append(self.cache_dtype)
        # `cpu_offload_gb` does not use `torch.compile` yet.
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

986
987
988
989
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
990
        swap_space: float,
991
        cache_dtype: str,
992
        is_attention_free: bool = False,
993
        num_gpu_blocks_override: Optional[int] = None,
994
        sliding_window: Optional[int] = None,
995
        enable_prefix_caching: bool = False,
996
        cpu_offload_gb: float = 0,
997
        calculate_kv_scales: Optional[bool] = None,
998
999
1000
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
1001
        self.swap_space_bytes = swap_space * GiB_bytes
1002
        self.num_gpu_blocks_override = num_gpu_blocks_override
1003
        self.cache_dtype = cache_dtype
1004
        self.is_attention_free = is_attention_free
1005
        self.sliding_window = sliding_window
1006
        self.enable_prefix_caching = enable_prefix_caching
1007
        self.cpu_offload_gb = cpu_offload_gb
1008
        self.calculate_kv_scales = calculate_kv_scales
1009
        self._verify_args()
1010
        self._verify_cache_dtype()
1011
        self._verify_prefix_caching()
1012
1013

        # Will be set after profiling.
1014
1015
        self.num_gpu_blocks: Optional[int] = None
        self.num_cpu_blocks: Optional[int] = None
1016

1017
1018
1019
1020
        # Set calculate_kv_scales to False if the value is unset.
        if self.calculate_kv_scales is None:
            self.calculate_kv_scales = False

1021
    def metrics_info(self):
1022
1023
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
1024
1025
        return {key: str(value) for key, value in self.__dict__.items()}

1026
1027
1028
1029
1030
1031
    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}.")

1032
1033
1034
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
1035
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
1036
            logger.info(
1037
1038
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
1039
1040
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
1041
1042
1043
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

1044
1045
1046
1047
1048
1049
1050
1051
1052
    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.")

1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
    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

1063
1064
1065
        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.")
1066
1067
1068
        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:
1069
            logger.warning("Possibly too large swap space. %s", msg)
1070

1071

1072
1073
1074
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
1075

1076
1077
1078
1079
1080
1081
1082
1083
    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
1084
    pool_type: Union[str, Type["BaseTokenizerGroup"]]
1085
1086
    extra_config: dict

1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1105
    def __post_init__(self):
1106
1107
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
1108
1109
1110
1111
1112
1113
            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(
1114
1115
        cls, tokenizer_pool_size: int,
        tokenizer_pool_type: Union[str, Type["BaseTokenizerGroup"]],
1116
1117
1118
        tokenizer_pool_extra_config: Optional[Union[str, dict]]
    ) -> Optional["TokenizerPoolConfig"]:
        """Create a TokenizerPoolConfig from the given parameters.
1119

1120
        If tokenizer_pool_size is 0, return None.
1121

1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
        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


1144
1145
1146
1147
1148
1149
1150
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
1151
    SHARDED_STATE = "sharded_state"
1152
    GGUF = "gguf"
1153
    BITSANDBYTES = "bitsandbytes"
1154
    MISTRAL = "mistral"
1155
    RUNAI_STREAMER = "runai_streamer"
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174


@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.
1175
            "bitsandbytes" will load nf4 type weights.
1176
        model_loader_extra_config: The extra config for the model loader.
1177
        ignore_patterns: The list of patterns to ignore when loading the model.
1178
            Default to "original/**/*" to avoid repeated loading of llama's
1179
            checkpoints.
1180
1181
1182
1183
1184
1185
    """

    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)
1186
    ignore_patterns: Optional[Union[List[str], str]] = None
1187

1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1206
1207
1208
1209
1210
    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)
1211
1212
1213
        if isinstance(self.load_format, str):
            load_format = self.load_format.lower()
            self.load_format = LoadFormat(load_format)
1214

1215
1216
1217
1218
1219
1220
1221
        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/**/*"]

1222

1223
@dataclass
1224
class ParallelConfig:
1225
    """Configuration for the distributed execution."""
1226

1227
1228
    pipeline_parallel_size: int = 1  # Number of pipeline parallel groups.
    tensor_parallel_size: int = 1  # Number of tensor parallel groups.
1229

1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
    # 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"
1263
    sd_worker_cls: str = "auto"
1264
1265
1266
1267
1268

    world_size: int = field(init=False)

    rank: int = 0

1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
    def compute_hash(self):
        """
        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: List[Any] = []
        factors.append(self.pipeline_parallel_size)
        factors.append(self.tensor_parallel_size)
        return hashlib.sha256(str(factors).encode()).hexdigest()

1282
1283
1284
1285
1286
    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size

        if self.worker_use_ray:
1287
1288
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
1289
            elif not self.use_ray:
1290
1291
1292
                raise ValueError(f"worker-use-ray can't be used with "
                                 f"distributed executor backend "
                                 f"'{self.distributed_executor_backend}'.")
1293
        ray_only_devices = ["tpu"]
1294
        from vllm.platforms import current_platform
1295
1296
        if (current_platform.device_type in ray_only_devices
                and self.world_size > 1):
1297
1298
1299
1300
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
1301
1302
                    f"{current_platform.device_type.upper()} backend only "
                    "supports Ray for distributed inference.")
1303

1304
        if self.distributed_executor_backend is None and self.world_size > 1:
1305
1306
1307
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1308
            from vllm.executor import ray_utils
1309
            backend = "mp"
1310
            ray_found = ray_utils.ray_is_available()
1311
1312
1313
1314
1315
            if current_platform.is_neuron():
                # neuron uses single process to control multiple devices
                backend = "uni"
            elif (current_platform.is_cuda()
                  and cuda_device_count_stateless() < self.world_size):
1316
1317
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1318
1319
1320
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1321
1322
                backend = "ray"
            elif ray_found:
1323
                if self.placement_group:
1324
                    backend = "ray"
1325
1326
1327
1328
1329
1330
                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"
1331
1332
1333
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1334

1335
1336
        self._verify_args()

1337
1338
1339
1340
1341
1342
    @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)

1343
    def _verify_args(self) -> None:
1344
1345
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase
1346
        from vllm.platforms import current_platform
1347
        if self.distributed_executor_backend not in (
1348
1349
                "ray", "mp", "uni",
                "external_launcher", None) and not (isinstance(
1350
1351
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
1352
            raise ValueError(
1353
1354
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
1355
1356
                "values are 'ray', 'mp' 'uni', 'external_launcher' or"
                " custom ExecutorBase subclass.")
1357
        if self.use_ray:
1358
1359
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1360
        if current_platform.is_rocm():
1361
1362
1363
1364
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
1365
        if self.ray_workers_use_nsight and not self.use_ray:
1366
1367
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1368

1369

1370
@dataclass
1371
class SchedulerConfig:
1372
    """Scheduler configuration."""
1373

1374
    runner_type: str = "generate"  # The runner type to launch for the model.
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399

    # 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
1400

1401
1402
1403
1404
1405
1406
    # NOTE: The following multimodal encoder budget will be initialized to
    # max_num_batched_tokens and overridden in case max multimodal embedding
    # size is larger.
    # TODO (ywang96): Make these configurable.
    # Multimodal encoder compute budget, only used in V1
    max_num_encoder_input_tokens: int = field(default=None)  # type: ignore
1407
1408

    # Multimodal encoder cache size, only used in V1
1409
    encoder_cache_size: int = field(default=None)  # type: ignore
1410

1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
    # 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)

1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1452
1453
1454
1455
    def __post_init__(self) -> None:
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
1456
1457
1458
1459
                    # 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.
1460
                    self.max_num_batched_tokens = max(self.max_model_len, 2048)
1461
                else:
1462
1463
1464
                    # 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
1465
1466
1467
            else:
                # If max_model_len is too short, use 2048 as the default value
                # for higher throughput.
1468
                self.max_num_batched_tokens = max(self.max_model_len, 2048)
1469

1470
1471
            if self.runner_type == "pooling":
                # Choose specific value for higher throughput
1472
1473
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1474
                    _POOLING_MODEL_MAX_NUM_BATCHED_TOKENS,
1475
                )
1476
            if self.is_multimodal_model:
1477
                # The value needs to be at least the number of multimodal tokens
1478
1479
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1480
1481
1482
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

1483
1484
1485
        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
        self.encoder_cache_size = self.max_num_batched_tokens

1486
        if self.enable_chunked_prefill:
1487
1488
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
1489
                self.max_num_batched_tokens)
1490

1491
        self.chunked_prefill_enabled = self.enable_chunked_prefill
1492
1493
1494
        self._verify_args()

    def _verify_args(self) -> None:
1495
1496
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
1497
1498
1499
1500
1501
1502
1503
            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.")
1504

1505
1506
1507
1508
1509
        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}).")
1510

1511
1512
1513
1514
1515
1516
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
        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

1527

1528
class DeviceConfig:
1529
    device: Optional[torch.device]
1530
    device_type: str
1531

1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # the device/platform information will be summarized
        # by torch/vllm automatically.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1551
1552
1553
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
1554
            from vllm.platforms import current_platform
1555
            self.device_type = current_platform.device_type
1556
            if not self.device_type:
1557
                raise RuntimeError("Failed to infer device type")
1558
1559
1560
1561
1562
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
1563
        if self.device_type in ["neuron", "openvino"]:
1564
            self.device = torch.device("cpu")
1565
1566
        elif self.device_type in ["tpu"]:
            self.device = None
1567
1568
1569
1570
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1571

1572
1573
1574
1575
1576
1577
1578
class SpeculativeConfig:
    """Configuration for speculative decoding.

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

1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # spec decode does not use `torch.compile` yet.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

1597
1598
1599
1600
1601
1602
    @staticmethod
    def maybe_create_spec_config(
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        target_dtype: str,
        speculative_model: Optional[str],
1603
        speculative_model_quantization: Optional[str],
1604
        speculative_draft_tensor_parallel_size: Optional[int],
1605
        num_speculative_tokens: Optional[int],
1606
        speculative_disable_mqa_scorer: Optional[bool],
1607
1608
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
1609
        disable_log_stats: bool,
1610
        speculative_disable_by_batch_size: Optional[int],
1611
1612
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1613
1614
1615
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
1616
        disable_logprobs: Optional[bool],
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
    ) -> 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.
1632
1633
1634
            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.
1635
1636
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
1637
            num_speculative_tokens (Optional[int]): The number of speculative
1638
1639
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
1640
1641
1642
            speculative_disable_mqa_scorer (Optional[bool]): Disable the MQA
                scorer for the speculative model and fall back to batch
                expansion for scoring.
1643
1644
1645
1646
1647
1648
            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.
1649
1650
1651
            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.
1652
1653
1654
1655
            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.
1656
1657
1658
1659
1660
1661
1662
1663
            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
1664
                accepted. This threshold is used only when we use the
1665
1666
1667
1668
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1669
1670
1671
1672
1673
            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.
1674

1675
1676
1677
1678
1679
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """

1680
1681
1682
1683
        if speculative_model is None:
            if num_speculative_tokens is not None:
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative_model.")
1684
1685
            return None

1686
1687
1688
1689
1690
1691
        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=}")

1692
1693
        # TODO: The user should be able to specify revision/max model len
        # for the draft model. It is not currently supported.
1694
1695
        draft_revision = None
        draft_code_revision = None
1696
        draft_quantization = speculative_model_quantization
1697

1698
1699
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1700
1701
1702
1703
1704
1705
1706
1707
                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=}")
1708

1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
            # 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,
1719
                task="draft",
1720
1721
1722
                tokenizer=target_model_config.tokenizer,
                tokenizer_mode=target_model_config.tokenizer_mode,
                trust_remote_code=target_model_config.trust_remote_code,
1723
1724
                allowed_local_media_path=target_model_config.
                allowed_local_media_path,
1725
1726
1727
1728
1729
1730
                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,
1731
                spec_target_max_model_len=target_model_config.max_model_len,
1732
1733
                quantization=draft_quantization,
                enforce_eager=target_model_config.enforce_eager,
1734
1735
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1736
1737
1738
                max_logprobs=target_model_config.max_logprobs,
            )

1739
            draft_hf_config = draft_model_config.hf_config
1740

1741
1742
1743
1744
1745
            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)
1746
1747
1748
1749
1750
1751
1752
1753
            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(
1754
1755
1756
                        "This speculative model supports a maximum of "
                        f"num_speculative_tokens={n_predict}, but "
                        f"{num_speculative_tokens=} was provided.")
1757

1758
1759
1760
1761
1762
1763
            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.")

1764
1765
1766
1767
1768
1769
1770
            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
            )

1771
1772
1773
1774
1775
1776
1777
1778
1779
            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(
1780
                    target_parallel_config,
1781
                    speculative_draft_tensor_parallel_size, draft_hf_config))
1782

1783
1784
1785
1786
1787
1788
        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.")

1789
1790
1791
1792
        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
1793
1794
        if disable_logprobs is None:
            disable_logprobs = True
1795

1796
1797
1798
1799
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
1800
            speculative_disable_mqa_scorer,
1801
            speculative_disable_by_batch_size,
1802
1803
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
1804
1805
1806
1807
1808
            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,
1809
1810
            disable_logprobs=disable_logprobs,
            disable_log_stats=disable_log_stats,
1811
1812
        )

1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
    @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,
        )

1848
    @staticmethod
1849
1850
1851
1852
1853
1854
1855
    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.
1856
        """
1857
1858
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
1859
        if speculative_draft_tensor_parallel_size is None:
1860
1861
1862
1863
1864
1865
1866
1867
1868
            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
1869
1870
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
1871
            raise ValueError(
1872
                f"{speculative_draft_tensor_parallel_size=} cannot be "
1873
                f"other value than 1 or target model tensor_parallel_size")
1874
        return speculative_draft_tensor_parallel_size
1875

1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
    @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.
        """
1886
1887
1888
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
1889
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
1890
1891
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
            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,
1909
        speculative_disable_mqa_scorer: Optional[bool],
1910
1911
1912
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1913
1914
1915
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
1916
        disable_logprobs: bool,
1917
        disable_log_stats: bool,
1918
1919
1920
1921
1922
1923
1924
1925
    ):
        """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.
1926
1927
1928
1929
1930
            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.
1931
1932
1933
1934
1935
1936
1937
1938
            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
1939
                accepted. This threshold is used only when we use the
1940
1941
1942
1943
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1944
            disable_logprobs: If set to True, token log probabilities will not
1945
                be returned even if requested by sampling parameters. This
1946
1947
1948
1949
                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.
1950
1951
            disable_log_stats: Whether to disable periodic printing of stage
                times in speculative decoding.
1952
1953
1954
1955
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
1956
        self.speculative_disable_mqa_scorer = speculative_disable_mqa_scorer
1957
1958
1959
1960
        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
1961
1962
1963
1964
1965
        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
1966
        self.disable_logprobs = disable_logprobs
1967
        self.disable_log_stats = disable_log_stats
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978

        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)
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
            # 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}")
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015

    @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:
2016
2017
2018
2019
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
2020
2021
2022
2023
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


2024
2025
2026
2027
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
2028
    fully_sharded_loras: bool = False
2029
    max_cpu_loras: Optional[int] = None
2030
    lora_dtype: Optional[Union[torch.dtype, str]] = None
2031
2032
2033
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
2034
    long_lora_scaling_factors: Optional[Tuple[float]] = None
2035
    bias_enabled: bool = False
2036

2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # LoRA is not compatible with `torch.compile` .
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2055
    def __post_init__(self):
2056
2057
2058
        # 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)
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
        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
2075
                f"max_loras ({self.max_loras})")
2076

2077
2078
2079
2080
2081
    def verify_with_cache_config(self, cache_config: CacheConfig):
        # TODO LoRA supports CPU offload.
        if cache_config.cpu_offload_gb > 0:
            raise ValueError("CPU offload is not supported with LoRA yet.")

2082
2083
2084
2085
2086
    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)
2087
2088
2089
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
2090
            # TODO support marlin
2091
2092
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
2093
2094

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
2095
        # Reminder: Please update docs/source/features/compatibility_matrix.md
2096
        # If the feature combo become valid
2097
        if scheduler_config.chunked_prefill_enabled:
2098
2099
            logger.warning("LoRA with chunked prefill is still experimental "
                           "and may be unstable.")
2100
2101


2102
2103
2104
2105
2106
2107
2108
@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

2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
    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)


2145
@dataclass
2146
class MultiModalConfig:
2147
2148
    """Controls the behavior of multimodal models."""

2149
    limit_per_prompt: Mapping[str, int] = field(default_factory=dict)
2150
    """
2151
    The maximum number of input items allowed per prompt for each modality.
2152
2153
    """

2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2172
    # TODO: Add configs to init vision tower or not.
2173

2174

2175
2176
@dataclass
class PoolerConfig:
2177
    """Controls the behavior of output pooling in pooling models."""
2178
2179

    pooling_type: Optional[str] = None
2180
    """
2181
    The pooling method of the pooling model. This should be a key in
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
    :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
    """
2199
    If set, only the score corresponding to the ``step_tag_id`` in the
2200
2201
2202
2203
2204
2205
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

    returned_token_ids: Optional[List[int]] = None
    """
2206
2207
    A list of indices for the vocabulary dimensions to be extracted,
    such as the token IDs of ``good_token`` and ``bad_token`` in the
2208
2209
2210
    ``math-shepherd-mistral-7b-prm`` model.
    """

2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2229
2230
2231
    @staticmethod
    def from_json(json_str: str) -> "PoolerConfig":
        return PoolerConfig(**json.loads(json_str))
2232
2233


2234
2235
2236
2237
2238
2239
2240
2241
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

2242
_ROCM_NOT_SUPPORTED_DTYPE: List[str] = []  #
2243

2244
2245
2246

def _get_and_verify_dtype(
    config: PretrainedConfig,
2247
    dtype: Union[str, torch.dtype],
2248
2249
2250
2251
2252
2253
2254
) -> 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

2255
2256
2257
2258
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
Woosuk Kwon's avatar
Woosuk Kwon committed
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
                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
2269
2270
            else:
                torch_dtype = config_dtype
2271

2272
            from vllm.platforms import current_platform
2273
2274
            if (current_platform.is_cpu()
                    and current_platform.get_cpu_architecture()
2275
                    == CpuArchEnum.POWERPC
2276
2277
2278
2279
2280
2281
2282
2283
                    and (config_dtype == torch.float16
                         or config_dtype == torch.float32)):
                logger.info(
                    "For POWERPC, we cast models to bfloat16 instead of "
                    "using float16 by default. Float16 is not currently "
                    "supported for POWERPC.")
                torch_dtype = torch.bfloat16

2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
            # TODO: change this condition to check if the platform support bf16
            # instead of checking the OS. For instance M2 shall supports bf16
            # already. But we need to modify `cpu_extension.cmake` to activate
            # the feature in the build.
            if (current_platform.is_cpu() and sys.platform.startswith("darwin")
                    and current_platform.get_cpu_architecture()
                    == CpuArchEnum.ARM and config_dtype == torch.bfloat16):
                logger.info("For macOS with Apple Silicon, currently bfloat16 "
                            "is not supported. Setting dtype to float16.")
                torch_dtype = torch.float16

2295
2296
2297
2298
2299
2300
            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
2301
        else:
2302
2303
2304
2305
2306
            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
2307
    else:
2308
        raise ValueError(f"Unknown dtype: {dtype}")
2309
2310
2311
2312
2313

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
2314
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
2315
2316
2317
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
2318
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
2319
2320
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
2321
            # Casting between float16 and bfloat16 is allowed with a warning.
2322
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
2323
2324

    return torch_dtype
2325
2326
2327
2328
2329


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
2330
    disable_sliding_window: bool,
2331
    sliding_window_len: Optional[Union[int, List[Optional[int]]]],
2332
    spec_target_max_model_len: Optional[int] = None,
2333
    encoder_config: Optional[Any] = None,
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
) -> 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",
2344
2345
        # ChatGLM2
        "seq_length",
2346
2347
        # Command-R
        "model_max_length",
2348
2349
        # Whisper
        "max_target_positions",
2350
2351
2352
2353
2354
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
2355
    # Choose the smallest "max_length" from the possible keys.
2356
    max_len_key = None
2357
    for key in possible_keys:
2358
2359
2360
2361
2362
        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)
2363
2364
2365
2366

    # 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:
2367
2368

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
2369
        max_len_key = "sliding_window" \
2370
2371
2372
            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)
2373
2374
2375

    # If none of the keys were found in the config, use a default and
    # log a warning.
2376
    if derived_max_model_len == float("inf"):
2377
2378
2379
2380
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

2381
2382
2383
2384
2385
        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

2386
2387
2388
2389
        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: "
2390
            "%s. Assuming the model's maximum length is %d.", possible_keys,
2391
            default_max_len)
2392
        derived_max_model_len = default_max_len
2393

2394
    rope_scaling = getattr(hf_config, "rope_scaling", None)
2395
    if rope_scaling is not None:
2396
2397
2398
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408

        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.")

2409
2410
2411
2412
            # 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)

2413
2414
2415
2416
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
2417

2418
2419
2420
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2421
2422
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2423
    if max_model_len is None:
2424
        max_model_len = int(derived_max_model_len)
2425
    elif max_model_len > derived_max_model_len:
2426
2427
2428
2429
2430
        # 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:
2431
2432
2433
2434
2435
2436
2437
            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.")
2438
        else:
2439
            msg = (
2440
                f"User-specified max_model_len ({max_model_len}) is greater "
2441
2442
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2443
                f"{model_max_length} in model's config.json). This may lead "
2444
2445
2446
2447
2448
2449
2450
2451
2452
                "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")
2453
    return int(max_model_len)
2454
2455


2456
2457
2458
2459
2460
2461
2462
2463
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


2464
2465
2466
def get_served_model_name(model: str,
                          served_model_name: Optional[Union[str, List[str]]]):
    """
2467
2468
2469
2470
    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
2471
2472
2473
2474
2475
2476
2477
2478
2479
    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


2480
2481
2482
2483
@dataclass
class DecodingConfig:
    """Dataclass which contains the decoding strategy of the engine"""

2484
2485
2486
    # Which guided decoding algo to use.
    # 'outlines' / 'lm-format-enforcer' / 'xgrammar'
    guided_decoding_backend: str = 'xgrammar'
2487

2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2506
    def __post_init__(self):
2507
        valid_guided_backends = ['outlines', 'lm-format-enforcer', 'xgrammar']
2508
2509
2510
2511
2512
2513
        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}")


2514
2515
2516
2517
2518
@dataclass
class ObservabilityConfig:
    """Configuration for observability."""
    otlp_traces_endpoint: Optional[str] = None

2519
2520
2521
2522
2523
2524
2525
2526
    # 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

2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2545
    def __post_init__(self):
2546
2547
2548
2549
2550
        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}")
2551
2552


2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
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

2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        # no factors to consider.
        # this config will not affect the computation graph.
        factors: List[Any] = []
        hash_str = hashlib.md5(str(factors).encode()).hexdigest()
        return hash_str

2604
2605
    @classmethod
    def from_cli(cls, cli_value: str) -> "KVTransferConfig":
youkaichao's avatar
youkaichao committed
2606
        """Parse the CLI value for the kv cache transfer config."""
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
        return KVTransferConfig.model_validate_json(cli_value)

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

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

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

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

    @property
    def need_kv_parallel_group(self) -> bool:
        # for those database-based connector, vLLM does not need to create
        # parallel group, and in that case the kv parallel size will be 1.
        return self.kv_connector is not None and self.kv_parallel_size > 1

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

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


2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
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.
2664
        - debug_dump_path: the path to dump the debug information.
2665
2666
2667
        - cache_dir: the directory to store the compiled graph, to
            accelerate Inductor compilation. By default, it will use
            model-related information to generate a cache directory.
2668
2669
2670
2671
2672
2673
2674
        - 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).
2675
2676
2677
2678
2679
2680
2681
2682
2683
        - 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).
2684
        - splitting_ops: a list of ops to split the full graph into subgraphs, used in piecewise compilation.
2685
2686
2687
2688
    - 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
2689
2690
2691
2692
                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.
2693
2694
2695
            TODO: move outside cudagraph logic into compilation.
            torch.compile will handle cudagraph capture logic in the future.
        - cudagraph_capture_sizes: sizes to capture cudagraph.
2696
2697
            - None (default): capture sizes are inferred from vllm config.
            - List[int]: capture sizes are specified as given.
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
        - 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
2711
2712
2713
2714
2715
                is compiled. In addition, compile for compile_sizes,
                using configurations in inductor_compile_config.
        - compile_sizes: sizes to compile for inductor. In addition
            to integers, it also supports "cudagraph_capture_sizes" to
            specify the sizes for cudagraph capture.
2716
2717
2718
2719
2720
2721
2722
        - 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})`
2723
        - custom inductor passes: see PassConfig for more details
2724

2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
    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
2736
    debug_dump_path: str = ""
2737
    cache_dir: str = ""
2738
    backend: str = ""
2739
    custom_ops: List[str] = Field(default_factory=list)
2740
    splitting_ops: List[str] = Field(default=None)  # type: ignore
2741
2742

    use_inductor: bool = True
2743
    compile_sizes: Optional[List[Union[int, str]]] = Field(default=None)
2744
2745
2746
2747
2748
2749
2750
2751
    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

2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
    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:
2784
                logger.warning_once(
2785
2786
2787
2788
                    "Fusion enabled but reshape elimination disabled."
                    "RMSNorm + quant (fp8) fusion might not work")

    pass_config: PassConfig = Field(default_factory=PassConfig)
2789
2790

    # not configurable, computed after init
2791
    max_capture_size: int = PrivateAttr
2792
    local_cache_dir: str = PrivateAttr  # local cache dir for each rank
2793
2794
2795
2796
2797
    # optimization:
    # Intuitively, bs_to_padded_graph_size should be Dict[int, int].
    # since we know all keys are in a range [0, max_capture_size],
    # we can optimize it to List[int] for better lookup performance.
    bs_to_padded_graph_size: List[int] = PrivateAttr
2798

2799
2800
2801
    # keep track of enabled and disabled custom ops
    enabled_custom_ops: Counter[str] = PrivateAttr
    disabled_custom_ops: Counter[str] = PrivateAttr
2802
    traced_files: Set[str] = PrivateAttr
2803
    compilation_time: float = PrivateAttr
2804

2805
2806
2807
2808
    # Per-model forward context
    # Map from layer name to the attention cls
    static_forward_context: Dict[str, Any] = PrivateAttr

2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: List[Any] = []
        factors.append(self.level)
        factors.append(self.backend)
        factors.append(self.custom_ops)
        factors.append(self.splitting_ops)
        factors.append(self.use_inductor)
        factors.append(self.inductor_compile_config)
        factors.append(self.inductor_passes)
        factors.append(self.pass_config.uuid())
        return hashlib.sha256(str(factors).encode()).hexdigest()

2832
2833
2834
2835
2836
2837
2838
2839
    def __repr__(self) -> str:
        exclude = {
            "static_forward_context",
            "enabled_custom_ops",
            "disabled_custom_ops",
            "compilation_time",
            "bs_to_padded_graph_size",
            "pass_config",
2840
            "traced_files",
2841
2842
2843
2844
2845
        }
        return self.model_dump_json(exclude=exclude, exclude_unset=True)

    __str__ = __repr__

2846
2847
2848
2849
2850
    @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))
2851
2852
2853
        # 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)
2854

2855
2856
2857
2858
2859
2860
    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'"

2861
2862
2863
2864
2865
2866
2867
2868
2869
        if self.splitting_ops is None:
            if envs.VLLM_USE_V1:
                # v1 must split the graph on attention ops
                # for piecewise cudagraph
                self.splitting_ops = [
                    "vllm.unified_attention",
                    "vllm.unified_attention_with_output",
                ]
            else:
2870
2871
                # v0 uses full graph compilation
                self.splitting_ops = []
2872

2873
2874
2875
        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
2876
2877
2878
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
2879
2880
2881
2882
2883
2884
2885
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
2886
2887
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
2888

2889
2890
        self.enabled_custom_ops = Counter()
        self.disabled_custom_ops = Counter()
2891
        self.traced_files = set()
2892
        self.static_forward_context = {}
2893
        self.compilation_time = 0.0
2894

2895
    def init_backend(self, vllm_config: "VllmConfig") -> Union[str, Callable]:
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
        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
2913

2914
        from vllm.compilation.backends import VllmBackend
2915
        return VllmBackend(vllm_config)
2916

2917
2918
    def init_with_cudagraph_sizes(self,
                                  cudagraph_capture_sizes: List[int]) -> None:
2919
        """To complete the initialization of config,
2920
2921
        we need to know the cudagraph sizes."""

2922
        if self.cudagraph_capture_sizes is None:
2923
            self.cudagraph_capture_sizes = cudagraph_capture_sizes
2924
        else:
2925
2926
2927
            # de-duplicate the sizes provided by the config
            self.cudagraph_capture_sizes = list(
                set(self.cudagraph_capture_sizes))
2928
2929
            logger.info(("cudagraph sizes specified by model runner"
                         " %s is overridden by config %s"),
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
                        cudagraph_capture_sizes, self.cudagraph_capture_sizes)

        computed_compile_sizes = []
        if self.compile_sizes is not None:
            # de-duplicate the sizes provided by the config
            self.compile_sizes = list(set(self.compile_sizes))
            for x in self.compile_sizes:
                if isinstance(x, str):
                    assert x == "cudagraph_capture_sizes", \
                    "Unrecognized size type in compile_sizes, " \
                    f"expect 'cudagraph_capture_sizes', got {x}"
                    computed_compile_sizes.extend(self.cudagraph_capture_sizes)
                else:
                    assert isinstance(x, int)
                    computed_compile_sizes.append(x)
        self.compile_sizes = computed_compile_sizes  # type: ignore
2946

2947
        # sort to make sure cudagraph capture sizes are in descending order
2948
2949
2950
        self.cudagraph_capture_sizes.sort(reverse=True)
        self.max_capture_size = self.cudagraph_capture_sizes[
            0] if self.cudagraph_capture_sizes else 0
2951

2952
2953
2954
2955
        # pre-compute the mapping from batch size to padded graph size
        self.bs_to_padded_graph_size = [
            0 for i in range(self.max_capture_size + 1)
        ]
2956
2957
        for end, start in zip(self.cudagraph_capture_sizes,
                              self.cudagraph_capture_sizes[1:] + [0]):
2958
2959
2960
2961
2962
2963
2964
            for bs in range(start, end):
                if bs == start:
                    self.bs_to_padded_graph_size[bs] = start
                else:
                    self.bs_to_padded_graph_size[bs] = end
        self.bs_to_padded_graph_size[
            self.max_capture_size] = self.max_capture_size
2965

2966

2967
2968
2969
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
2970
2971
2972
    simplifies passing around the distinct configurations in the codebase.
    """

2973
2974
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
2975
2976
2977
2978
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig,
                                            init=True)
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig,
                                              init=True)
2979
2980
2981
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
2982
2983
2984
2985
2986
    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
2987
    quant_config: Optional[QuantizationConfig] = None
2988
2989
    compilation_config: CompilationConfig = field(default=None,
                                                  init=True)  # type: ignore
2990
2991
    kv_transfer_config: KVTransferConfig = field(default=None,
                                                 init=True)  # type: ignore
2992
2993
2994
2995
    # some opaque config, only used to provide additional information
    # for the hash computation, mainly used for testing and debugging.
    additional_config: SupportsHash = field(default=None,
                                            init=True)  # type: ignore
2996
    instance_id: str = ""
2997

2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: List[Any] = []
        # summarize system state
        from torch._inductor.codecache import CacheBase
        system_factors = CacheBase.get_system()
        factors.append(system_factors)

        # summarize pytorch state
        from torch._inductor.codecache import torch_key
        torch_factors = torch_key()
        factors.append(torch_factors)

        # summarize vllm config
        vllm_factors: List[Any] = []
        from vllm import __version__
        vllm_factors.append(__version__)
        if self.model_config:
            vllm_factors.append(self.model_config.compute_hash())
3027
3028
        else:
            vllm_factors.append("None")
3029
3030
        if self.cache_config:
            vllm_factors.append(self.cache_config.compute_hash())
3031
3032
        else:
            vllm_factors.append("None")
3033
3034
        if self.parallel_config:
            vllm_factors.append(self.parallel_config.compute_hash())
3035
3036
        else:
            vllm_factors.append("None")
3037
3038
        if self.scheduler_config:
            vllm_factors.append(self.scheduler_config.compute_hash())
3039
3040
        else:
            vllm_factors.append("None")
3041
3042
        if self.device_config:
            vllm_factors.append(self.device_config.compute_hash())
3043
3044
        else:
            vllm_factors.append("None")
3045
3046
        if self.load_config:
            vllm_factors.append(self.load_config.compute_hash())
3047
3048
        else:
            vllm_factors.append("None")
3049
3050
        if self.lora_config:
            vllm_factors.append(self.lora_config.compute_hash())
3051
3052
        else:
            vllm_factors.append("None")
3053
3054
        if self.speculative_config:
            vllm_factors.append(self.speculative_config.compute_hash())
3055
3056
        else:
            vllm_factors.append("None")
3057
3058
        if self.decoding_config:
            vllm_factors.append(self.decoding_config.compute_hash())
3059
3060
        else:
            vllm_factors.append("None")
3061
3062
        if self.observability_config:
            vllm_factors.append(self.observability_config.compute_hash())
3063
3064
        else:
            vllm_factors.append("None")
3065
3066
        if self.prompt_adapter_config:
            vllm_factors.append(self.prompt_adapter_config.compute_hash())
3067
3068
        else:
            vllm_factors.append("None")
3069
3070
3071
3072
        if self.quant_config:
            pass  # should be captured by model_config.quantization
        if self.compilation_config:
            vllm_factors.append(self.compilation_config.compute_hash())
3073
3074
        else:
            vllm_factors.append("None")
3075
3076
        if self.kv_transfer_config:
            vllm_factors.append(self.kv_transfer_config.compute_hash())
3077
3078
3079
3080
3081
3082
        else:
            vllm_factors.append("None")
        if self.additional_config:
            vllm_factors.append(self.additional_config.compute_hash())
        else:
            vllm_factors.append("None")
3083
3084
3085
3086
3087
        factors.append(vllm_factors)

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

3088
3089
3090
3091
3092
3093
    def pad_for_cudagraph(self, batch_size: int) -> int:
        # if batch_size > self.compilation_config.max_capture_size,
        # it should raise an IndexError.
        # the caller should make sure the batch_size is within the range,
        # i.e., batch_size <= self.compilation_config.max_capture_size
        return self.compilation_config.bs_to_padded_graph_size[batch_size]
3094

3095
3096
3097
3098
3099
    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
3100
        from vllm.platforms import current_platform
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
        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
3123

3124
3125
3126
3127
3128
3129
3130
3131
3132
    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

3133
3134
3135
3136
3137
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

3138
3139
3140
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
3141
3142
3143
3144
3145
3146
3147
3148
        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)
3149
3150

        if self.lora_config:
3151
            self.lora_config.verify_with_cache_config(self.cache_config)
3152
3153
3154
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
3155
3156
3157
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
3158
3159
3160
3161
3162

        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)
3163

3164
        from vllm.platforms import current_platform
3165
3166
3167
3168
3169
        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):
3170
            logger.warning_once(
3171
3172
3173
3174
                "Turing devices tensor cores do not support float32 matmul. "
                "To workaround this limitation, vLLM will set 'ieee' input "
                "precision for chunked prefill triton kernels.")

3175
        if self.compilation_config is None:
3176
            self.compilation_config = CompilationConfig()
3177
3178
        if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
3179
3180
3181
3182
3183
3184
3185
            # 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
3186
            self.compilation_config.cudagraph_num_of_warmups = 1
3187
3188
            self.compilation_config.pass_config.enable_fusion = False
            self.compilation_config.pass_config.enable_reshape = False
3189
            self.compilation_config.level = CompilationLevel.PIECEWISE
3190

3191
        self._set_cudagraph_sizes()
3192

3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
        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

3207
3208
        current_platform.check_and_update_config(self)

3209
3210
3211
        if not self.instance_id:
            self.instance_id = random_uuid()[:5]

3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
    def _set_cudagraph_sizes(self):
        """
        cudagraph batchsize padding logic:

        `[1, 2, 4] + [8 * i for i in range(1, 1025)]` is a list of all possible
        batch sizes that cudagraph will capture.

        Depending on the engine's configuration of `max_num_seqs`, the
        candidate batch sizes to capture cudagraph will shrink to the subset
        which just cover the range of `[1, max_num_seqs]`. In the common case,
        `max_num_seqs` is 256, and the cudagraph batch sizes will be
        `[1, 2, 4, 8, 16, 24, 32, 40, ..., 256]`.

        However, if users specify the cudagraph capture sizes through
        compilation config, we will use the specified sizes instead.

3228
3229
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in descending order).
3230
3231

        During runtime, if batchsize is larger than
3232
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3233
3234
        no cudagraph will be used.
        If the batch size is no larger than
3235
        `vllm_config.compilation_config.cudagraph_capture_sizes`,
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
        we can quickly find the padded graph size for a given batch size by
        looking up `vllm_config.compilation_config.bs_to_padded_graph_size`.
        """

        # calculate the default `batch_size_capture_list`
        if not envs.VLLM_USE_V1:
            batch_size_capture_list = []
            max_batchsize_to_capture = 0
            if self.scheduler_config is not None and \
                self.model_config is not None and \
                    not self.model_config.enforce_eager:

                possible_sizes = [1, 2, 4] + [8 * i for i in range(1, 1025)]
                # find the minimum size that is larger than max_num_seqs,
                # which then becomes the max_batchsize_to_capture
                larger_sizes = [
                    x for x in possible_sizes
                    if x >= self.scheduler_config.max_num_seqs
                ]
                if larger_sizes:
                    max_batchsize_to_capture = larger_sizes[0]
                else:
                    max_batchsize_to_capture = possible_sizes[-1]

                # filter out the sizes that are
                # larger than max_batchsize_to_capture
                batch_size_capture_list = [
                    size for size in possible_sizes
                    if size <= max_batchsize_to_capture
                ]
        else:
            batch_size_capture_list = []
            if self.model_config is not None and \
                not self.model_config.enforce_eager:
                batch_size_capture_list = [1, 2, 4
                                           ] + [i for i in range(8, 513, 8)]

        self.compilation_config.init_with_cudagraph_sizes(
            batch_size_capture_list)

3276
    def __str__(self):
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
        return (
            f"model={self.model_config.model!r},"
            f" speculative_config={self.speculative_config!r},"
            f" tokenizer={self.model_config.tokenizer!r}, "
            f"skip_tokenizer_init={self.model_config.skip_tokenizer_init},"
            f" tokenizer_mode={self.model_config.tokenizer_mode}, "
            f"revision={self.model_config.revision}, "
            f"override_neuron_config={self.model_config.override_neuron_config},"
            f" tokenizer_revision={self.model_config.tokenizer_revision}, "
            f"trust_remote_code={self.model_config.trust_remote_code}, "
            f"dtype={self.model_config.dtype}, "
            f"max_seq_len={self.model_config.max_model_len},"
            f" download_dir={self.load_config.download_dir!r}, "
            f"load_format={self.load_config.load_format}, "
            f"tensor_parallel_size={self.parallel_config.tensor_parallel_size},"
            f" pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, "  # noqa
            f"disable_custom_all_reduce={self.parallel_config.disable_custom_all_reduce}, "  # noqa
            f"quantization={self.model_config.quantization}, "
            f"enforce_eager={self.model_config.enforce_eager}, "
            f"kv_cache_dtype={self.cache_config.cache_dtype}, "
            f" device_config={self.device_config.device}, "
            f"decoding_config={self.decoding_config!r}, "
            f"observability_config={self.observability_config!r}, "
            f"seed={self.model_config.seed}, "
            f"served_model_name={self.model_config.served_model_name}, "
            f"num_scheduler_steps={self.scheduler_config.num_scheduler_steps}, "
            f"multi_step_stream_outputs={self.scheduler_config.multi_step_stream_outputs}, "  # noqa
            f"enable_prefix_caching={self.cache_config.enable_prefix_caching}, "
            f"chunked_prefill_enabled={self.scheduler_config.chunked_prefill_enabled}, "  # noqa
            f"use_async_output_proc={self.model_config.use_async_output_proc}, "
3307
            f"disable_mm_preprocessor_cache={self.model_config.disable_mm_preprocessor_cache!r}, "  # noqa
3308
            f"mm_processor_kwargs={self.model_config.mm_processor_kwargs}, "
3309
3310
            f"pooler_config={self.model_config.pooler_config!r}, "
            f"compilation_config={self.compilation_config!r}")
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360


_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