arg_utils.py 80.7 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
# yapf: disable
4
import argparse
5
import dataclasses
6
import json
7
import re
8
import threading
9
from dataclasses import MISSING, dataclass, fields
10
11
12
from typing import (TYPE_CHECKING, Any, Callable, Dict, List, Literal, Mapping,
                    Optional, Tuple, Type, TypeVar, Union, cast, get_args,
                    get_origin)
13

14
import torch
15
from typing_extensions import TypeIs
16

17
import vllm.envs as envs
18
from vllm import version
19
20
from vllm.config import (CacheConfig, CompilationConfig, Config, ConfigFormat,
                         DecodingConfig, Device, DeviceConfig,
21
                         DistributedExecutorBackend, HfOverrides,
22
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
23
                         ModelConfig, ModelImpl, ObservabilityConfig,
24
25
26
27
                         ParallelConfig, PoolerConfig, PoolType,
                         PromptAdapterConfig, SchedulerConfig, SchedulerPolicy,
                         SpeculativeConfig, TaskOption, TokenizerPoolConfig,
                         VllmConfig, get_attr_docs, get_field)
28
from vllm.executor.executor_base import ExecutorBase
29
from vllm.logger import init_logger
30
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
31
from vllm.plugins import load_general_plugins
32
from vllm.reasoning import ReasoningParserManager
33
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
34
from vllm.transformers_utils.utils import check_gguf_file
35
from vllm.usage.usage_lib import UsageContext
36
37
38
from vllm.utils import FlexibleArgumentParser, is_in_ray_actor

# yapf: enable
39

40
if TYPE_CHECKING:
41
    from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
42

43
44
logger = init_logger(__name__)

45
46
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]

47
48
49
50
51
# object is used to allow for special typing forms
T = TypeVar("T")
TypeHint = Union[type[Any], object]
TypeHintT = Union[type[T], object]

52

53
def optional_arg(val: str, return_type: Callable[[str], T]) -> Optional[T]:
54
    if val == "" or val == "None":
55
        return None
56
    try:
57
        return return_type(val)
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
    except ValueError as e:
        raise argparse.ArgumentTypeError(
            f"Value {val} cannot be converted to {return_type}.") from e


def optional_str(val: str) -> Optional[str]:
    return optional_arg(val, str)


def optional_int(val: str) -> Optional[int]:
    return optional_arg(val, int)


def optional_float(val: str) -> Optional[float]:
    return optional_arg(val, float)
73
74


75
76
77
78
79
def nullable_kvs(val: str) -> Optional[dict[str, int]]:
    """NOTE: This function is deprecated, args should be passed as JSON
    strings instead.
    
    Parses a string containing comma separate key [str] to value [int]
80
81
82
83
84
85
86
87
    pairs into a dictionary.

    Args:
        val: String value to be parsed.

    Returns:
        Dictionary with parsed values.
    """
88
89
90
91
92
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
93
94
95
96
97
        kv_parts = [part.lower().strip() for part in item.split("=")]
        if len(kv_parts) != 2:
            raise argparse.ArgumentTypeError(
                "Each item should be in the form KEY=VALUE")
        key, value = kv_parts
98
99

        try:
100
            parsed_value = int(value)
101
102
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
103
104
105
106
107
108
            raise argparse.ArgumentTypeError(msg) from exc

        if key in out_dict and out_dict[key] != parsed_value:
            raise argparse.ArgumentTypeError(
                f"Conflicting values specified for key: {key}")
        out_dict[key] = parsed_value
109
110
111
112

    return out_dict


113
114
115
116
117
118
119
120
121
122
123
def optional_dict(val: str) -> Optional[dict[str, int]]:
    try:
        return optional_arg(val, json.loads)
    except ValueError:
        logger.warning(
            "Failed to parse JSON string. Attempting to parse as "
            "comma-separated key=value pairs. This will be deprecated in a "
            "future release.")
        return nullable_kvs(val)


124
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
125
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
126
    """Arguments for vLLM engine."""
127
    model: str = 'facebook/opt-125m'
128
    served_model_name: Optional[Union[str, List[str]]] = None
129
    tokenizer: Optional[str] = None
130
    hf_config_path: Optional[str] = None
131
    task: TaskOption = "auto"
132
    skip_tokenizer_init: bool = False
133
    tokenizer_mode: str = 'auto'
134
    trust_remote_code: bool = False
135
    allowed_local_media_path: str = ""
136
137
    download_dir: Optional[str] = LoadConfig.download_dir
    load_format: str = LoadConfig.load_format
138
    config_format: ConfigFormat = ConfigFormat.AUTO
139
    dtype: str = 'auto'
140
    kv_cache_dtype: str = 'auto'
141
    seed: Optional[int] = None
142
    max_model_len: Optional[int] = None
143
144
145
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
146
    distributed_executor_backend: Optional[Union[
147
148
        DistributedExecutorBackend,
        Type[ExecutorBase]]] = ParallelConfig.distributed_executor_backend
149
    # number of P/D disaggregation (or other disaggregation) workers
150
151
152
153
154
155
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
    data_parallel_size: int = ParallelConfig.data_parallel_size
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
    max_parallel_loading_workers: Optional[
        int] = ParallelConfig.max_parallel_loading_workers
156
    block_size: Optional[int] = None
157
    enable_prefix_caching: Optional[bool] = None
158
    prefix_caching_hash_algo: str = "builtin"
159
    disable_sliding_window: bool = False
160
    disable_cascade_attn: bool = False
161
    use_v2_block_manager: bool = True
162
163
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
164
    gpu_memory_utilization: float = 0.90
165
166
167
168
169
170
171
    max_num_batched_tokens: Optional[
        int] = SchedulerConfig.max_num_batched_tokens
    max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
    max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
    long_prefill_token_threshold: int = \
        SchedulerConfig.long_prefill_token_threshold
    max_num_seqs: Optional[int] = SchedulerConfig.max_num_seqs
172
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
173
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
174
    revision: Optional[str] = None
175
    code_revision: Optional[str] = None
176
    rope_scaling: Optional[Dict[str, Any]] = None
177
    rope_theta: Optional[float] = None
178
    hf_token: Optional[Union[bool, str]] = None
179
    hf_overrides: Optional[HfOverrides] = None
180
    tokenizer_revision: Optional[str] = None
181
    quantization: Optional[str] = None
182
    enforce_eager: Optional[bool] = None
183
    max_seq_len_to_capture: int = 8192
184
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
185
    tokenizer_pool_size: int = TokenizerPoolConfig.pool_size
186
187
188
    # Note: Specifying a tokenizer pool by passing a class
    # is intended for expert use only. The API may change without
    # notice.
189
190
191
192
    tokenizer_pool_type: Union[PoolType, Type["BaseTokenizerGroup"]] = \
        TokenizerPoolConfig.pool_type
    tokenizer_pool_extra_config: dict[str, Any] = \
        get_field(TokenizerPoolConfig, "extra_config")
193
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
194
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
195
    disable_mm_preprocessor_cache: bool = False
196
    enable_lora: bool = False
197
    enable_lora_bias: bool = False
198
199
    max_loras: int = 1
    max_lora_rank: int = 16
200
201
202
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
203
    fully_sharded_loras: bool = False
204
    lora_extra_vocab_size: int = 256
205
    long_lora_scaling_factors: Optional[Tuple[float]] = None
206
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
207
    max_cpu_loras: Optional[int] = None
208
    device: Device = DeviceConfig.device
209
210
    num_scheduler_steps: int = SchedulerConfig.num_scheduler_steps
    multi_step_stream_outputs: bool = SchedulerConfig.multi_step_stream_outputs
211
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
212
    num_gpu_blocks_override: Optional[int] = None
213
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
214
215
    model_loader_extra_config: dict = \
        get_field(LoadConfig, "model_loader_extra_config")
216
217
    ignore_patterns: Optional[Union[str,
                                    List[str]]] = LoadConfig.ignore_patterns
218
    preemption_mode: Optional[str] = SchedulerConfig.preemption_mode
219

220
221
222
223
    scheduler_delay_factor: float = SchedulerConfig.delay_factor
    enable_chunked_prefill: Optional[
        bool] = SchedulerConfig.enable_chunked_prefill
    disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
224

225
    guided_decoding_backend: str = DecodingConfig.guided_decoding_backend
226
    logits_processor_pattern: Optional[str] = None
227

228
    speculative_config: Optional[Dict[str, Any]] = None
229

230
    qlora_adapter_name_or_path: Optional[str] = None
231
    show_hidden_metrics_for_version: Optional[str] = None
232
    otlp_traces_endpoint: Optional[str] = None
233
    collect_detailed_traces: Optional[str] = None
234
    disable_async_output_proc: bool = False
235
236
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
    scheduler_cls: Union[str, Type[object]] = SchedulerConfig.scheduler_cls
237

238
239
    override_neuron_config: Optional[Dict[str, Any]] = None
    override_pooler_config: Optional[PoolerConfig] = None
240
    compilation_config: Optional[CompilationConfig] = None
241
242
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
243

244
245
    kv_transfer_config: Optional[KVTransferConfig] = None

246
    generation_config: Optional[str] = "auto"
247
    override_generation_config: Optional[Dict[str, Any]] = None
248
    enable_sleep_mode: bool = False
249
    model_impl: str = "auto"
250

251
252
    calculate_kv_scales: Optional[bool] = None

253
    additional_config: Optional[Dict[str, Any]] = None
254
255
    enable_reasoning: Optional[bool] = None
    reasoning_parser: Optional[str] = None
256
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
257

258
    def __post_init__(self):
259
        if not self.tokenizer:
260
            self.tokenizer = self.model
261

262
263
264
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
265
        if isinstance(self.compilation_config, (int, dict)):
266
267
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
268

269
        # Setup plugins
270
271
        from vllm.plugins import load_general_plugins
        load_general_plugins()
272
273

    @staticmethod
274
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
275
        """Shared CLI arguments for vLLM engine."""
276

277
        def is_type_in_union(cls: TypeHint, type: TypeHint) -> bool:
278
            """Check if the class is a type in a union type."""
279
280
281
282
283
284
285
286
287
288
289
290
            is_union = get_origin(cls) is Union
            type_in_union = type in [get_origin(a) or a for a in get_args(cls)]
            return is_union and type_in_union

        def get_type_from_union(cls: TypeHint, type: TypeHintT) -> TypeHintT:
            """Get the type in a union type."""
            for arg in get_args(cls):
                if (get_origin(arg) or arg) is type:
                    return arg
            raise ValueError(f"Type {type} not found in union type {cls}.")

        def is_optional(cls: TypeHint) -> TypeIs[Union[Any, None]]:
291
            """Check if the class is an optional type."""
292
            return is_type_in_union(cls, type(None))
293

294
295
296
297
298
299
300
301
302
        def can_be_type(cls: TypeHint, type: TypeHintT) -> TypeIs[TypeHintT]:
            """Check if the class can be of type."""
            return cls is type or get_origin(cls) is type or is_type_in_union(
                cls, type)

        def is_custom_type(cls: TypeHint) -> bool:
            """Check if the class is a custom type."""
            return cls.__module__ != "builtins"

303
        def get_kwargs(cls: type[Config]) -> dict[str, Any]:
304
305
306
307
            cls_docs = get_attr_docs(cls)
            kwargs = {}
            for field in fields(cls):
                name = field.name
308
309
310
311
                default = field.default
                # This will only be True if default is MISSING
                if field.default_factory is not MISSING:
                    default = field.default_factory()
312
                kwargs[name] = {"default": default, "help": cls_docs[name]}
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340

                # Make note of if the field is optional and get the actual
                # type of the field if it is
                optional = is_optional(field.type)
                field_type = get_args(
                    field.type)[0] if optional else field.type

                if can_be_type(field_type, bool):
                    # Creates --no-<name> and --<name> flags
                    kwargs[name]["action"] = argparse.BooleanOptionalAction
                    kwargs[name]["type"] = bool
                elif can_be_type(field_type, Literal):
                    # Creates choices from Literal arguments
                    if is_type_in_union(field_type, Literal):
                        field_type = get_type_from_union(field_type, Literal)
                    choices = get_args(field_type)
                    kwargs[name]["choices"] = choices
                    choice_type = type(choices[0])
                    assert all(type(c) is choice_type for c in choices), (
                        f"All choices must be of the same type. "
                        f"Got {choices} with types {[type(c) for c in choices]}"
                    )
                    kwargs[name]["type"] = choice_type
                elif can_be_type(field_type, int):
                    kwargs[name]["type"] = optional_int if optional else int
                elif can_be_type(field_type, float):
                    kwargs[name][
                        "type"] = optional_float if optional else float
341
342
                elif can_be_type(field_type, dict):
                    kwargs[name]["type"] = optional_dict
343
344
345
346
347
348
                elif (can_be_type(field_type, str)
                      or is_custom_type(field_type)):
                    kwargs[name]["type"] = optional_str if optional else str
                else:
                    raise ValueError(
                        f"Unsupported type {field.type} for argument {name}. ")
349
350
            return kwargs

351
        # Model arguments
352
353
354
        parser.add_argument(
            '--model',
            type=str,
355
            default=EngineArgs.model,
356
            help='Name or path of the huggingface model to use.')
357
358
359
360
361
362
        parser.add_argument(
            '--task',
            default=EngineArgs.task,
            choices=get_args(TaskOption),
            help='The task to use the model for. Each vLLM instance only '
            'supports one task, even if the same model can be used for '
363
            'multiple tasks. When the model only supports one task, ``"auto"`` '
364
365
            'can be used to select it; otherwise, you must specify explicitly '
            'which task to use.')
366
367
        parser.add_argument(
            '--tokenizer',
368
            type=optional_str,
369
            default=EngineArgs.tokenizer,
370
371
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
372
373
        parser.add_argument(
            "--hf-config-path",
374
            type=optional_str,
375
376
377
            default=EngineArgs.hf_config_path,
            help='Name or path of the huggingface config to use. '
            'If unspecified, model name or path will be used.')
378
379
380
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
381
382
383
            help='Skip initialization of tokenizer and detokenizer. '
            'Expects valid prompt_token_ids and None for prompt from '
            'the input. The generated output will contain token ids.')
Jasmond L's avatar
Jasmond L committed
384
385
        parser.add_argument(
            '--revision',
386
            type=optional_str,
Jasmond L's avatar
Jasmond L committed
387
            default=None,
388
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
389
390
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
391
392
        parser.add_argument(
            '--code-revision',
393
            type=optional_str,
394
            default=None,
395
            help='The specific revision to use for the model code on '
396
397
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
398
399
        parser.add_argument(
            '--tokenizer-revision',
400
            type=optional_str,
401
            default=None,
402
403
404
            help='Revision of the huggingface tokenizer to use. '
            'It can be a branch name, a tag name, or a commit id. '
            'If unspecified, will use the default version.')
405
406
407
408
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
409
            choices=['auto', 'slow', 'mistral', 'custom'],
410
411
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
412
            'always use the slow tokenizer. \n* '
413
414
415
            '"mistral" will always use the `mistral_common` tokenizer. \n* '
            '"custom" will use --tokenizer to select the '
            'preregistered tokenizer.')
416
417
        parser.add_argument('--trust-remote-code',
                            action='store_true',
418
                            help='Trust remote code from huggingface.')
419
420
421
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
422
423
424
425
            help="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.")
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
        load_group.add_argument('--load-format',
                                choices=[f.value for f in LoadFormat],
                                **load_kwargs["load_format"])
        load_group.add_argument('--download-dir',
                                **load_kwargs["download_dir"])
        load_group.add_argument('--model-loader-extra-config',
                                **load_kwargs["model_loader_extra_config"])
        load_group.add_argument('--use-tqdm-on-load',
                                **load_kwargs["use_tqdm_on_load"])

442
443
444
445
446
447
448
        parser.add_argument(
            '--config-format',
            default=EngineArgs.config_format,
            choices=[f.value for f in ConfigFormat],
            help='The format of the model config to load.\n\n'
            '* "auto" will try to load the config in hf format '
            'if available else it will try to load in mistral format ')
449
450
451
452
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
453
454
455
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
456
457
458
459
460
461
462
463
            help='Data type for model weights and activations.\n\n'
            '* "auto" will use FP16 precision for FP32 and FP16 models, and '
            'BF16 precision for BF16 models.\n'
            '* "half" for FP16. Recommended for AWQ quantization.\n'
            '* "float16" is the same as "half".\n'
            '* "bfloat16" for a balance between precision and range.\n'
            '* "float" is shorthand for FP32 precision.\n'
            '* "float32" for FP32 precision.')
464
465
466
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
467
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
468
            default=EngineArgs.kv_cache_dtype,
469
            help='Data type for kv cache storage. If "auto", will use model '
470
471
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
472
        parser.add_argument('--max-model-len',
473
                            type=human_readable_int,
474
                            default=EngineArgs.max_model_len,
475
                            help='Model context length. If unspecified, will '
476
477
478
479
480
                            'be automatically derived from the model config. '
                            'Supports k/m/g/K/M/G in human-readable format.\n'
                            'Examples:\n'
                            '- 1k → 1000\n'
                            '- 1K → 1024\n')
481
482
483
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
484
            default=DecodingConfig.guided_decoding_backend,
485
            help='Which engine will be used for guided decoding'
486
            ' (JSON schema / regex etc) by default. Currently support '
487
488
489
            'https://github.com/mlc-ai/xgrammar and '
            'https://github.com/guidance-ai/llguidance.'
            'Valid backend values are "xgrammar", "guidance", and "auto". '
490
            'With "auto", we will make opinionated choices based on request '
491
            'contents and what the backend libraries currently support, so '
492
            'the behavior is subject to change in each release.')
493
494
        parser.add_argument(
            '--logits-processor-pattern',
495
            type=optional_str,
496
497
498
499
500
            default=None,
            help='Optional regex pattern specifying valid logits processor '
            'qualified names that can be passed with the `logits_processors` '
            'extra completion argument. Defaults to None, which allows no '
            'processors.')
501
502
503
504
505
506
507
508
509
510
511
512
        parser.add_argument(
            '--model-impl',
            type=str,
            default=EngineArgs.model_impl,
            choices=[f.value for f in ModelImpl],
            help='Which implementation of the model to use.\n\n'
            '* "auto" will try to use the vLLM implementation if it exists '
            'and fall back to the Transformers implementation if no vLLM '
            'implementation is available.\n'
            '* "vllm" will use the vLLM model implementation.\n'
            '* "transformers" will use the Transformers model '
            'implementation.\n')
513
        # Parallel arguments
514
515
516
517
518
519
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
520
            '--distributed-executor-backend',
521
522
523
524
525
526
527
528
529
            **parallel_kwargs["distributed_executor_backend"])
        parallel_group.add_argument(
            '--pipeline-parallel-size', '-pp',
            **parallel_kwargs["pipeline_parallel_size"])
        parallel_group.add_argument('--tensor-parallel-size', '-tp',
                                    **parallel_kwargs["tensor_parallel_size"])
        parallel_group.add_argument('--data-parallel-size', '-dp',
                                    **parallel_kwargs["data_parallel_size"])
        parallel_group.add_argument(
530
            '--enable-expert-parallel',
531
532
            **parallel_kwargs["enable_expert_parallel"])
        parallel_group.add_argument(
533
            '--max-parallel-loading-workers',
534
535
            **parallel_kwargs["max_parallel_loading_workers"])
        parallel_group.add_argument(
536
            '--ray-workers-use-nsight',
537
538
539
540
            **parallel_kwargs["ray_workers_use_nsight"])
        parallel_group.add_argument(
            '--disable-custom-all-reduce',
            **parallel_kwargs["disable_custom_all_reduce"])
541
        # KV cache arguments
542
543
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
544
                            default=EngineArgs.block_size,
545
                            choices=[8, 16, 32, 64, 128],
546
                            help='Token block size for contiguous chunks of '
547
                            'tokens. This is ignored on neuron devices and '
548
                            'set to ``--max-model-len``. On CUDA devices, '
549
550
                            'only block sizes up to 32 are supported. '
                            'On HPU devices, block size defaults to 128.')
551

552
553
554
555
556
        parser.add_argument(
            "--enable-prefix-caching",
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.enable_prefix_caching,
            help="Enables automatic prefix caching. "
557
            "Use ``--no-enable-prefix-caching`` to disable explicitly.",
558
        )
559
560
561
562
563
564
565
        parser.add_argument(
            "--prefix-caching-hash-algo",
            type=str,
            choices=["builtin", "sha256"],
            default=EngineArgs.prefix_caching_hash_algo,
            help="Set the hash algorithm for prefix caching. "
            "Options are 'builtin' (Python's built-in hash) or 'sha256' "
566
            "(collision resistant but with certain overheads).",
567
        )
568
569
570
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
571
                            'capping to sliding window size.')
572
573
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
574
                            default=True,
575
576
577
578
579
                            help='[DEPRECATED] block manager v1 has been '
                            'removed and SelfAttnBlockSpaceManager (i.e. '
                            'block manager v2) is now the default. '
                            'Setting this flag to True or False'
                            ' has no effect on vLLM behavior.')
580

581
582
583
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
584
                            help='Random seed for operations.')
585
        parser.add_argument('--swap-space',
586
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
587
                            default=EngineArgs.swap_space,
588
                            help='CPU swap space size (GiB) per GPU.')
589
590
591
592
593
594
595
596
597
        parser.add_argument(
            '--cpu-offload-gb',
            type=float,
            default=0,
            help='The space in GiB to offload to CPU, per GPU. '
            'Default is 0, which means no offloading. Intuitively, '
            'this argument can be seen as a virtual way to increase '
            'the GPU memory size. For example, if you have one 24 GB '
            'GPU and set this to 10, virtually you can think of it as '
598
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
599
            'which requires at least 26GB GPU memory. Note that this '
600
            'requires fast CPU-GPU interconnect, as part of the model is '
601
602
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
603
604
605
606
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
607
608
609
            help='The fraction of GPU memory to be used for the model '
            'executor, which can range from 0 to 1. For example, a value of '
            '0.5 would imply 50%% GPU memory utilization. If unspecified, '
610
611
612
613
614
615
            'will use the default value of 0.9. This is a per-instance '
            'limit, and only applies to the current vLLM instance.'
            'It does not matter if you have another vLLM instance running '
            'on the same GPU. For example, if you have two vLLM instances '
            'running on the same GPU, you can set the GPU memory utilization '
            'to 0.5 for each instance.')
616
        parser.add_argument(
617
            '--num-gpu-blocks-override',
618
619
620
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
621
            ' of GPU blocks. Used for testing preemption.')
622
623
624
625
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
626
627
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
628
629
        parser.add_argument('--disable-log-stats',
                            action='store_true',
630
                            help='Disable logging statistics.')
631
632
633
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
634
                            type=optional_str,
635
                            choices=[*QUANTIZATION_METHODS, None],
636
                            default=EngineArgs.quantization,
637
638
639
640
641
642
                            help='Method used to quantize the weights. If '
                            'None, we first check the `quantization_config` '
                            'attribute in the model config file. If that is '
                            'None, we assume the model weights are not '
                            'quantized and use `dtype` to determine the data '
                            'type of the weights.')
643
644
645
646
647
        parser.add_argument(
            '--rope-scaling',
            default=None,
            type=json.loads,
            help='RoPE scaling configuration in JSON format. '
648
            'For example, ``{"rope_type":"dynamic","factor":2.0}``')
649
650
651
652
653
654
        parser.add_argument('--rope-theta',
                            default=None,
                            type=float,
                            help='RoPE theta. Use with `rope_scaling`. In '
                            'some cases, changing the RoPE theta improves the '
                            'performance of the scaled model.')
655
656
657
658
659
660
661
662
663
664
        parser.add_argument(
            '--hf-token',
            type=str,
            nargs='?',
            const=True,
            default=None,
            help='The token to use as HTTP bearer authorization'
            ' for remote files. If `True`, will use the token '
            'generated when running `huggingface-cli login` '
            '(stored in `~/.huggingface`).')
665
666
667
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
668
                            help='Extra arguments for the HuggingFace config. '
669
670
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
671
672
673
674
675
        parser.add_argument('--enforce-eager',
                            action='store_true',
                            help='Always use eager-mode PyTorch. If False, '
                            'will use eager mode and CUDA graph in hybrid '
                            'for maximal performance and flexibility.')
676
        parser.add_argument('--max-seq-len-to-capture',
677
678
679
680
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
681
682
683
684
                            'larger than this, we fall back 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.')
685
686
687
688
689
690
691
692
693
694
695
696
697

        # Tokenizer arguments
        tokenizer_kwargs = get_kwargs(TokenizerPoolConfig)
        tokenizer_group = parser.add_argument_group(
            title="TokenizerPoolConfig",
            description=TokenizerPoolConfig.__doc__,
        )
        tokenizer_group.add_argument('--tokenizer-pool-size',
                                     **tokenizer_kwargs["pool_size"])
        tokenizer_group.add_argument('--tokenizer-pool-type',
                                     **tokenizer_kwargs["pool_type"])
        tokenizer_group.add_argument('--tokenizer-pool-extra-config',
                                     **tokenizer_kwargs["extra_config"])
698
699
700
701
702
703
704

        # Multimodal related configs
        parser.add_argument(
            '--limit-mm-per-prompt',
            type=nullable_kvs,
            default=EngineArgs.limit_mm_per_prompt,
            # The default value is given in
705
            # MultiModalConfig.get_default_limit_per_prompt
706
707
708
709
            help=('For each multimodal plugin, limit how many '
                  'input instances to allow for each prompt. '
                  'Expects a comma-separated list of items, '
                  'e.g.: `image=16,video=2` allows a maximum of 16 '
710
711
                  'images and 2 videos per prompt. Defaults to '
                  '1 (V0) or 999 (V1) for each modality.'))
712
713
714
715
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
716
            help=('Overrides for the multimodal input mapping/processing, '
717
                  'e.g., image processor. For example: ``{"num_crops": 4}``.'))
718
        parser.add_argument(
719
            '--disable-mm-preprocessor-cache',
720
            action='store_true',
721
722
            help='If true, then disables caching of the multi-modal '
            'preprocessor/mapper. (not recommended)')
723

724
725
726
727
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
728
729
730
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
        parser.add_argument('--max-loras',
                            type=int,
                            default=EngineArgs.max_loras,
                            help='Max number of LoRAs in a single batch.')
        parser.add_argument('--max-lora-rank',
                            type=int,
                            default=EngineArgs.max_lora_rank,
                            help='Max LoRA rank.')
        parser.add_argument(
            '--lora-extra-vocab-size',
            type=int,
            default=EngineArgs.lora_extra_vocab_size,
            help=('Maximum size of extra vocabulary that can be '
                  'present in a LoRA adapter (added to the base '
                  'model vocabulary).'))
        parser.add_argument(
            '--lora-dtype',
            type=str,
            default=EngineArgs.lora_dtype,
750
            choices=['auto', 'float16', 'bfloat16'],
751
752
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
753
754
        parser.add_argument(
            '--long-lora-scaling-factors',
755
            type=optional_str,
756
757
758
759
760
761
762
763
            default=EngineArgs.long_lora_scaling_factors,
            help=('Specify multiple scaling factors (which can '
                  'be different from base model scaling factor '
                  '- see eg. Long LoRA) to allow for multiple '
                  'LoRA adapters trained with those scaling '
                  'factors to be used at the same time. If not '
                  'specified, only adapters trained with the '
                  'base model scaling factor are allowed.'))
764
765
766
767
768
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
769
                  'Must be >= than max_loras.'))
770
771
772
773
774
775
776
777
        parser.add_argument(
            '--fully-sharded-loras',
            action='store_true',
            help=('By default, only half of the LoRA computation is '
                  'sharded with tensor parallelism. '
                  'Enabling this will use the fully sharded layers. '
                  'At high sequence length, max rank or '
                  'tensor parallel size, this is likely faster.'))
778
779
780
781
782
783
784
785
786
787
788
        parser.add_argument('--enable-prompt-adapter',
                            action='store_true',
                            help='If True, enable handling of PromptAdapters.')
        parser.add_argument('--max-prompt-adapters',
                            type=int,
                            default=EngineArgs.max_prompt_adapters,
                            help='Max number of PromptAdapters in a batch.')
        parser.add_argument('--max-prompt-adapter-token',
                            type=int,
                            default=EngineArgs.max_prompt_adapter_token,
                            help='Max number of PromptAdapters tokens')
789
790
791
792
793
794
795
796
797

        # Device arguments
        device_kwargs = get_kwargs(DeviceConfig)
        device_group = parser.add_argument_group(
            title="DeviceConfig",
            description=DeviceConfig.__doc__,
        )
        device_group.add_argument("--device", **device_kwargs["device"])

798
799
800
801
802
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
803

804
        parser.add_argument('--speculative-config',
805
                            type=json.loads,
806
807
808
                            default=None,
                            help='The configurations for speculative decoding.'
                            ' Should be a JSON string.')
809
810
811
812
813
814
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
815
            "Default to `original/**/*` to avoid repeated loading of llama's "
816
            "checkpoints.")
817
        parser.add_argument(
818
            '--preemption-mode',
819
820
            type=str,
            default=None,
821
822
823
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
824

825
826
827
828
829
830
831
832
833
834
        parser.add_argument(
            "--served-model-name",
            nargs="+",
            type=str,
            default=None,
            help="The model name(s) used in the API. If multiple "
            "names are provided, the server will respond to any "
            "of the provided names. The model name in the model "
            "field of a response will be the first name in this "
            "list. If not specified, the model name will be the "
835
            "same as the ``--model`` argument. Noted that this name(s) "
836
            "will also be used in `model_name` tag content of "
837
            "prometheus metrics, if multiple names provided, metrics "
838
            "tag will take the first one.")
839
840
841
842
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
843

844
845
846
847
848
849
850
851
852
853
854
855
        parser.add_argument('--show-hidden-metrics-for-version',
                            type=str,
                            default=None,
                            help='Enable deprecated Prometheus metrics that '
                            'have been hidden since the specified version. '
                            'For example, if a previously deprecated metric '
                            'has been hidden since the v0.7.0 release, you '
                            'use --show-hidden-metrics-for-version=0.7 as a '
                            'temporary escape hatch while you migrate to new '
                            'metrics. The metric is likely to be removed '
                            'completely in an upcoming release.')

856
857
858
859
860
        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
861
862
863
864
865
866
        parser.add_argument(
            '--collect-detailed-traces',
            type=str,
            default=None,
            help="Valid choices are " +
            ",".join(ALLOWED_DETAILED_TRACE_MODULES) +
867
            ". It makes sense to set this only if ``--otlp-traces-endpoint`` is"
868
869
870
            " set. If set, it will collect detailed traces for the specified "
            "modules. This involves use of possibly costly and or blocking "
            "operations and hence might have a performance impact.")
871

872
873
874
875
876
877
        parser.add_argument(
            '--disable-async-output-proc',
            action='store_true',
            default=EngineArgs.disable_async_output_proc,
            help="Disable async output processing. This may result in "
            "lower performance.")
878

879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
            '--max-num-batched-tokens',
            **scheduler_kwargs["max_num_batched_tokens"])
        scheduler_group.add_argument('--max-num-seqs',
                                     **scheduler_kwargs["max_num_seqs"])
        scheduler_group.add_argument(
            "--max-num-partial-prefills",
            **scheduler_kwargs["max_num_partial_prefills"])
        scheduler_group.add_argument(
            "--max-long-partial-prefills",
            **scheduler_kwargs["max_long_partial_prefills"])
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
            **scheduler_kwargs["long_prefill_token_threshold"])
        scheduler_group.add_argument('--num-lookahead-slots',
                                     **scheduler_kwargs["num_lookahead_slots"])
        scheduler_group.add_argument('--scheduler-delay-factor',
                                     **scheduler_kwargs["delay_factor"])
        scheduler_group.add_argument(
            '--enable-chunked-prefill',
            **scheduler_kwargs["enable_chunked_prefill"])
        scheduler_group.add_argument(
            '--multi-step-stream-outputs',
            **scheduler_kwargs["multi_step_stream_outputs"])
        scheduler_group.add_argument('--scheduling-policy',
                                     **scheduler_kwargs["policy"])
        scheduler_group.add_argument(
            "--disable-chunked-mm-input",
            **scheduler_kwargs["disable_chunked_mm_input"])
        parser.add_argument('--scheduler-cls',
                            **scheduler_kwargs["scheduler_cls"])
916

917
        parser.add_argument(
918
919
            '--override-neuron-config',
            type=json.loads,
920
            default=None,
921
            help="Override or set neuron device configuration. "
922
            "e.g. ``{\"cast_logits_dtype\": \"bloat16\"}``.")
923
        parser.add_argument(
924
925
            '--override-pooler-config',
            type=PoolerConfig.from_json,
926
            default=None,
927
            help="Override or set the pooling method for pooling models. "
928
            "e.g. ``{\"pooling_type\": \"mean\", \"normalize\": false}``.")
929

930
931
932
933
934
935
936
937
938
939
940
941
        parser.add_argument('--compilation-config',
                            '-O',
                            type=CompilationConfig.from_cli,
                            default=None,
                            help='torch.compile configuration for the model.'
                            'When it is a number (0, 1, 2, 3), it will be '
                            'interpreted as the optimization level.\n'
                            'NOTE: level 0 is the default level without '
                            'any optimization. level 1 and 2 are for internal '
                            'testing only. level 3 is the recommended level '
                            'for production.\n'
                            'To specify the full compilation config, '
942
943
                            'use a JSON string, e.g. ``{"level": 3, '
                            '"cudagraph_capture_sizes": [1, 2, 4, 8]}``\n'
944
                            'Following the convention of traditional '
945
946
                            'compilers, using ``-O`` without space is also '
                            'supported. ``-O3`` is equivalent to ``-O 3``.')
947

948
949
950
951
952
953
        parser.add_argument('--kv-transfer-config',
                            type=KVTransferConfig.from_cli,
                            default=None,
                            help='The configurations for distributed KV cache '
                            'transfer. Should be a JSON string.')

954
955
956
957
958
        parser.add_argument(
            '--worker-cls',
            type=str,
            default="auto",
            help='The worker class to use for distributed execution.')
959
960
961
962
963
964
965
        parser.add_argument(
            '--worker-extension-cls',
            type=str,
            default="",
            help='The worker extension class on top of the worker cls, '
            'it is useful if you just want to add new functions to the worker '
            'class without changing the existing functions.')
966
967
        parser.add_argument(
            "--generation-config",
968
            type=optional_str,
969
            default="auto",
970
            help="The folder path to the generation config. "
971
972
973
974
975
            "Defaults to 'auto', the generation config will be loaded from "
            "model path. If set to 'vllm', no generation config is loaded, "
            "vLLM defaults will be used. If set to a folder path, the "
            "generation config will be loaded from the specified folder path. "
            "If `max_new_tokens` is specified in generation config, then "
976
977
978
979
980
981
982
983
984
985
986
987
            "it sets a server-wide limit on the number of output tokens "
            "for all requests.")

        parser.add_argument(
            "--override-generation-config",
            type=json.loads,
            default=None,
            help="Overrides or sets generation config in JSON format. "
            "e.g. ``{\"temperature\": 0.5}``. If used with "
            "--generation-config=auto, the override parameters will be merged "
            "with the default config from the model. If generation-config is "
            "None, only the override parameters are used.")
988

989
990
991
992
993
994
        parser.add_argument("--enable-sleep-mode",
                            action="store_true",
                            default=False,
                            help="Enable sleep mode for the engine. "
                            "(only cuda platform is supported)")

995
996
997
998
999
1000
1001
1002
1003
        parser.add_argument(
            '--calculate-kv-scales',
            action='store_true',
            help='This enables dynamic calculation of '
            'k_scale and v_scale when kv-cache-dtype is fp8. '
            'If calculate-kv-scales is false, the scales will '
            'be loaded from the model checkpoint if available. '
            'Otherwise, the scales will default to 1.0.')

1004
1005
1006
1007
1008
1009
1010
1011
        parser.add_argument(
            "--additional-config",
            type=json.loads,
            default=None,
            help="Additional config for specified platform in JSON format. "
            "Different platforms may support different configs. Make sure the "
            "configs are valid for the platform you are using. The input format"
            " is like '{\"config_key\":\"config_value\"}'")
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023

        parser.add_argument(
            "--enable-reasoning",
            action="store_true",
            default=False,
            help="Whether to enable reasoning_content for the model. "
            "If enabled, the model will be able to generate reasoning content."
        )

        parser.add_argument(
            "--reasoning-parser",
            type=str,
1024
            choices=list(ReasoningParserManager.reasoning_parsers),
1025
1026
1027
1028
1029
1030
            default=None,
            help=
            "Select the reasoning parser depending on the model that you're "
            "using. This is used to parse the reasoning content into OpenAI "
            "API format. Required for ``--enable-reasoning``.")

1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
        parser.add_argument(
            "--disable-cascade-attn",
            action="store_true",
            default=False,
            help="Disable cascade attention for V1. While cascade attention "
            "does not change the mathematical correctness, disabling it "
            "could be useful for preventing potential numerical issues. "
            "Note that even if this is set to False, cascade attention will be "
            "only used when the heuristic tells that it's beneficial.")

1041
        return parser
1042
1043

    @classmethod
1044
    def from_cli_args(cls, args: argparse.Namespace):
1045
1046
1047
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
Zhuohan Li's avatar
Zhuohan Li committed
1048
1049
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
1050

1051
    def create_model_config(self) -> ModelConfig:
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
        # gguf file needs a specific model loader and doesn't use hf_repo
        if check_gguf_file(self.model):
            self.quantization = self.load_format = "gguf"

        # NOTE: This is to allow model loading from S3 in CI
        if (not isinstance(self, AsyncEngineArgs) and envs.VLLM_CI_USE_S3
                and self.model in MODELS_ON_S3
                and self.load_format == LoadFormat.AUTO):  # noqa: E501
            self.model = f"{MODEL_WEIGHTS_S3_BUCKET}/{self.model}"
            self.load_format = LoadFormat.RUNAI_STREAMER

1063
        return ModelConfig(
1064
            model=self.model,
1065
            hf_config_path=self.hf_config_path,
1066
            task=self.task,
1067
1068
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
1069
1070
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1071
            allowed_local_media_path=self.allowed_local_media_path,
1072
1073
1074
1075
1076
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
1077
            rope_theta=self.rope_theta,
1078
            hf_token=self.hf_token,
1079
            hf_overrides=self.hf_overrides,
1080
1081
1082
1083
1084
1085
1086
            tokenizer_revision=self.tokenizer_revision,
            max_model_len=self.max_model_len,
            quantization=self.quantization,
            enforce_eager=self.enforce_eager,
            max_seq_len_to_capture=self.max_seq_len_to_capture,
            max_logprobs=self.max_logprobs,
            disable_sliding_window=self.disable_sliding_window,
1087
            disable_cascade_attn=self.disable_cascade_attn,
1088
            skip_tokenizer_init=self.skip_tokenizer_init,
1089
            served_model_name=self.served_model_name,
1090
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1091
            use_async_output_proc=not self.disable_async_output_proc,
1092
            config_format=self.config_format,
1093
            mm_processor_kwargs=self.mm_processor_kwargs,
1094
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
1095
1096
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
1097
            logits_processor_pattern=self.logits_processor_pattern,
1098
            generation_config=self.generation_config,
1099
            override_generation_config=self.override_generation_config,
1100
            enable_sleep_mode=self.enable_sleep_mode,
1101
            model_impl=self.model_impl,
1102
        )
1103

1104
1105
    def create_load_config(self) -> LoadConfig:

1106
        if(self.qlora_adapter_name_or_path is not None) and \
1107
1108
            self.quantization != "bitsandbytes":
            raise ValueError(
1109
                "QLoRA adapter only support "
1110
1111
                f"'bitsandbytes' quantization, but got {self.quantization}")

1112
1113
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1114
1115
1116
1117
1118
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1119
            use_tqdm_on_load=self.use_tqdm_on_load,
1120
        )
1121

1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        enable_chunked_prefill: bool,
        disable_log_stats: bool,
    ) -> Optional["SpeculativeConfig"]:
        """Initializes and returns a SpeculativeConfig object based on
        `speculative_config`.

        This function utilizes `speculative_config` to create a
        SpeculativeConfig object. The `speculative_config` can either be
        provided as a JSON string input via CLI arguments or directly as a
1135
        dictionary from the engine.
1136
1137
        """
        if self.speculative_config is None:
1138
1139
            return None

1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.
        self.speculative_config.update({
            "target_model_config": target_model_config,
            "target_parallel_config": target_parallel_config,
            "enable_chunked_prefill": enable_chunked_prefill,
            "disable_log_stats": disable_log_stats,
        })
        speculative_config = SpeculativeConfig.from_dict(
            self.speculative_config)

        return speculative_config

1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
    def create_engine_config(
        self,
        usage_context: Optional[UsageContext] = None,
    ) -> VllmConfig:
        """
        Create the VllmConfig.

        NOTE: for autoselection of V0 vs V1 engine, we need to
        create the ModelConfig first, since ModelConfig's attrs
        (e.g. the model arch) are needed to make the decision.
Simon Mo's avatar
Simon Mo committed
1164

1165
1166
1167
1168
1169
1170
        This function set VLLM_USE_V1=X if VLLM_USE_V1 is
        unspecified by the user.

        If VLLM_USE_V1 is specified by the user but the VllmConfig
        is incompatible, we raise an error.
        """
1171
1172
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update()
1173

1174
        device_config = DeviceConfig(device=self.device)
1175
1176
        model_config = self.create_model_config()

1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
        # * If VLLM_USE_V1 is unset, we enable V1 for "supported features"
        #   and fall back to V0 for experimental or unsupported features.
        # * If VLLM_USE_V1=1, we enable V1 for supported + experimental
        #   features and raise error for unsupported features.
        # * If VLLM_USE_V1=0, we disable V1.
        use_v1 = False
        try_v1 = envs.VLLM_USE_V1 or not envs.is_set("VLLM_USE_V1")
        if try_v1 and self._is_v1_supported_oracle(model_config):
            use_v1 = True

        # If user explicitly set VLLM_USE_V1, sanity check we respect it.
        if envs.is_set("VLLM_USE_V1"):
            assert use_v1 == envs.VLLM_USE_V1
        # Otherwise, set the VLLM_USE_V1 variable globally.
        else:
            envs.set_vllm_use_v1(use_v1)

        # Set default arguments for V0 or V1 Engine.
        if use_v1:
            self._set_default_args_v1(usage_context)
        else:
            self._set_default_args_v0(model_config)
1199

1200
1201
        assert self.enable_chunked_prefill is not None

1202
        cache_config = CacheConfig(
1203
            block_size=self.block_size,
1204
1205
1206
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1207
            is_attention_free=model_config.is_attention_free,
1208
1209
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1210
            enable_prefix_caching=self.enable_prefix_caching,
1211
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1212
            cpu_offload_gb=self.cpu_offload_gb,
1213
            calculate_kv_scales=self.calculate_kv_scales,
1214
        )
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226

        # Get the current placement group if Ray is initialized and
        # we are in a Ray actor. If so, then the placement group will be
        # passed to spawned processes.
        placement_group = None
        if is_in_ray_actor():
            import ray

            # This call initializes Ray automatically if it is not initialized,
            # but we should not do this here.
            placement_group = ray.util.get_current_placement_group()

1227
        parallel_config = ParallelConfig(
1228
1229
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1230
            data_parallel_size=self.data_parallel_size,
1231
            enable_expert_parallel=self.enable_expert_parallel,
1232
1233
1234
            max_parallel_loading_workers=self.max_parallel_loading_workers,
            disable_custom_all_reduce=self.disable_custom_all_reduce,
            tokenizer_pool_config=TokenizerPoolConfig.create_config(
1235
1236
1237
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1238
            ),
1239
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1240
            placement_group=placement_group,
1241
1242
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1243
            worker_extension_cls=self.worker_extension_cls,
1244
        )
1245

1246
        speculative_config = self.create_speculative_config(
1247
1248
            target_model_config=model_config,
            target_parallel_config=parallel_config,
1249
            enable_chunked_prefill=self.enable_chunked_prefill,
1250
            disable_log_stats=self.disable_log_stats,
1251
1252
        )

1253
        # Reminder: Please update docs/source/features/compatibility_matrix.md
1254
        # If the feature combo become valid
1255
1256
1257
1258
        if self.num_scheduler_steps > 1:
            if speculative_config is not None:
                raise ValueError("Speculative decoding is not supported with "
                                 "multi-step (--num-scheduler-steps > 1)")
1259
1260
1261
            if self.enable_chunked_prefill and self.pipeline_parallel_size > 1:
                raise ValueError("Multi-Step Chunked-Prefill is not supported "
                                 "for pipeline-parallel-size > 1")
1262
1263
1264
1265
1266
1267
            from vllm.platforms import current_platform
            if current_platform.is_cpu():
                logger.warning("Multi-Step (--num-scheduler-steps > 1) is "
                               "currently not supported for CPUs and has been "
                               "disabled.")
                self.num_scheduler_steps = 1
1268
1269
1270
1271
1272
1273
1274
1275
1276

        # make sure num_lookahead_slots is set the higher value depending on
        # if we are using speculative decoding or multi-step
        num_lookahead_slots = max(self.num_lookahead_slots,
                                  self.num_scheduler_steps - 1)
        num_lookahead_slots = num_lookahead_slots \
            if speculative_config is None \
            else speculative_config.num_lookahead_slots

1277
        scheduler_config = SchedulerConfig(
1278
            runner_type=model_config.runner_type,
1279
1280
1281
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1282
            num_lookahead_slots=num_lookahead_slots,
1283
1284
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1285
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1286
            is_multimodal_model=model_config.is_multimodal_model,
1287
            preemption_mode=self.preemption_mode,
1288
            num_scheduler_steps=self.num_scheduler_steps,
1289
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1290
1291
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1292
            policy=self.scheduling_policy,
1293
            scheduler_cls=self.scheduler_cls,
1294
1295
1296
1297
            max_num_partial_prefills=self.max_num_partial_prefills,
            max_long_partial_prefills=self.max_long_partial_prefills,
            long_prefill_token_threshold=self.long_prefill_token_threshold,
        )
1298

1299
        lora_config = LoRAConfig(
1300
            bias_enabled=self.enable_lora_bias,
1301
1302
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1303
            fully_sharded_loras=self.fully_sharded_loras,
1304
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1305
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1306
1307
1308
            lora_dtype=self.lora_dtype,
            max_cpu_loras=self.max_cpu_loras if self.max_cpu_loras
            and self.max_cpu_loras > 0 else None) if self.enable_lora else None
1309

1310
1311
1312
1313
1314
        if self.qlora_adapter_name_or_path is not None and \
            self.qlora_adapter_name_or_path != "":
            self.model_loader_extra_config[
                "qlora_adapter_name_or_path"] = self.qlora_adapter_name_or_path

1315
1316
1317
1318
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1319
        load_config = self.create_load_config()
1320

1321
1322
1323
1324
1325
        prompt_adapter_config = PromptAdapterConfig(
            max_prompt_adapters=self.max_prompt_adapters,
            max_prompt_adapter_token=self.max_prompt_adapter_token) \
                                        if self.enable_prompt_adapter else None

1326
        decoding_config = DecodingConfig(
1327
1328
1329
1330
            guided_decoding_backend=self.guided_decoding_backend,
            reasoning_backend=self.reasoning_parser
            if self.enable_reasoning else None,
        )
1331

1332
1333
1334
1335
1336
        show_hidden_metrics = False
        if self.show_hidden_metrics_for_version is not None:
            show_hidden_metrics = version._prev_minor_version_was(
                self.show_hidden_metrics_for_version)

1337
1338
1339
1340
1341
1342
1343
1344
        detailed_trace_modules = []
        if self.collect_detailed_traces is not None:
            detailed_trace_modules = self.collect_detailed_traces.split(",")
        for m in detailed_trace_modules:
            if m not in ALLOWED_DETAILED_TRACE_MODULES:
                raise ValueError(
                    f"Invalid module {m} in collect_detailed_traces. "
                    f"Valid modules are {ALLOWED_DETAILED_TRACE_MODULES}")
1345
        observability_config = ObservabilityConfig(
1346
            show_hidden_metrics=show_hidden_metrics,
1347
1348
1349
1350
1351
1352
            otlp_traces_endpoint=self.otlp_traces_endpoint,
            collect_model_forward_time="model" in detailed_trace_modules
            or "all" in detailed_trace_modules,
            collect_model_execute_time="worker" in detailed_trace_modules
            or "all" in detailed_trace_modules,
        )
1353

1354
        config = VllmConfig(
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
            model_config=model_config,
            cache_config=cache_config,
            parallel_config=parallel_config,
            scheduler_config=scheduler_config,
            device_config=device_config,
            lora_config=lora_config,
            speculative_config=speculative_config,
            load_config=load_config,
            decoding_config=decoding_config,
            observability_config=observability_config,
1365
            prompt_adapter_config=prompt_adapter_config,
1366
            compilation_config=self.compilation_config,
1367
            kv_transfer_config=self.kv_transfer_config,
1368
            additional_config=self.additional_config,
1369
        )
1370

1371
1372
        return config

1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
    def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
        """Oracle for whether to use V0 or V1 Engine by default."""

        #############################################################
        # Unsupported Feature Flags on V1.

        if (self.load_format == LoadFormat.TENSORIZER.value
                or self.load_format == LoadFormat.SHARDED_STATE.value):
            _raise_or_fallback(
                feature_name=f"--load_format {self.load_format}",
                recommend_to_remove=False)
            return False

        if (self.logits_processor_pattern
                != EngineArgs.logits_processor_pattern):
            _raise_or_fallback(feature_name="--logits-processor-pattern",
                               recommend_to_remove=False)
            return False

1392
        if self.preemption_mode != SchedulerConfig.preemption_mode:
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
            _raise_or_fallback(feature_name="--preemption-mode",
                               recommend_to_remove=True)
            return False

        if (self.disable_async_output_proc
                != EngineArgs.disable_async_output_proc):
            _raise_or_fallback(feature_name="--disable-async-output-proc",
                               recommend_to_remove=True)
            return False

1403
        if self.scheduling_policy != SchedulerConfig.policy:
1404
1405
1406
1407
            _raise_or_fallback(feature_name="--scheduling-policy",
                               recommend_to_remove=False)
            return False

1408
        if self.num_scheduler_steps != SchedulerConfig.num_scheduler_steps:
1409
1410
1411
1412
            _raise_or_fallback(feature_name="--num-scheduler-steps",
                               recommend_to_remove=True)
            return False

1413
        if self.scheduler_delay_factor != SchedulerConfig.delay_factor:
1414
1415
1416
1417
1418
1419
1420
1421
1422
            _raise_or_fallback(feature_name="--scheduler-delay-factor",
                               recommend_to_remove=True)
            return False

        if self.additional_config != EngineArgs.additional_config:
            _raise_or_fallback(feature_name="--additional-config",
                               recommend_to_remove=False)
            return False

1423
        # Xgrammar and Guidance are supported.
1424
        SUPPORTED_GUIDED_DECODING = [
1425
1426
            "xgrammar", "xgrammar:disable-any-whitespace", "guidance",
            "guidance:disable-any-whitespace", "auto"
1427
        ]
1428
1429
1430
1431
1432
1433
        if self.guided_decoding_backend not in SUPPORTED_GUIDED_DECODING:
            _raise_or_fallback(feature_name="--guided-decoding-backend",
                               recommend_to_remove=False)
            return False

        # Need at least Ampere for now (FA support required).
1434
1435
1436
        # Skip this check if we are running on a non-GPU platform,
        # or if the device capability is not available
        # (e.g. in a Ray actor without GPUs).
1437
1438
        from vllm.platforms import current_platform
        if (current_platform.is_cuda()
1439
                and current_platform.get_device_capability()
1440
1441
1442
1443
1444
1445
1446
                and current_platform.get_device_capability().major < 8):
            _raise_or_fallback(feature_name="Compute Capability < 8.0",
                               recommend_to_remove=False)
            return False

        # No Fp8 KV cache so far.
        if self.kv_cache_dtype != "auto":
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
            fp8_attention = self.kv_cache_dtype.startswith("fp8")
            will_use_fa = (
                current_platform.is_cuda()
                and not envs.is_set("VLLM_ATTENTION_BACKEND")
            ) or envs.VLLM_ATTENTION_BACKEND == "FLASH_ATTN_VLLM_V1"
            supported = False
            if fp8_attention and will_use_fa:
                from vllm.vllm_flash_attn.fa_utils import (
                    flash_attn_supports_fp8)
                supported = flash_attn_supports_fp8()
            if not supported:
                _raise_or_fallback(feature_name="--kv-cache-dtype",
                                   recommend_to_remove=False)
                return False
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475

        # No Prompt Adapter so far.
        if self.enable_prompt_adapter:
            _raise_or_fallback(feature_name="--enable-prompt-adapter",
                               recommend_to_remove=False)
            return False

        # Only Fp16 and Bf16 dtypes since we only support FA.
        V1_SUPPORTED_DTYPES = [torch.bfloat16, torch.float16]
        if model_config.dtype not in V1_SUPPORTED_DTYPES:
            _raise_or_fallback(feature_name=f"--dtype {model_config.dtype}",
                               recommend_to_remove=False)
            return False

        # Some quantization is not compatible with torch.compile.
1476
        V1_UNSUPPORTED_QUANT = ["gguf"]
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
        if model_config.quantization in V1_UNSUPPORTED_QUANT:
            _raise_or_fallback(
                feature_name=f"--quantization {model_config.quantization}",
                recommend_to_remove=False)
            return False

        # No Embedding Models so far.
        if model_config.task not in ["generate"]:
            _raise_or_fallback(feature_name=f"--task {model_config.task}",
                               recommend_to_remove=False)
            return False

        # No Mamba or Encoder-Decoder so far.
        if not model_config.is_v1_compatible:
            _raise_or_fallback(feature_name=model_config.architectures,
                               recommend_to_remove=False)
            return False

        # No Concurrent Partial Prefills so far.
        if (self.max_num_partial_prefills
1497
                != SchedulerConfig.max_num_partial_prefills
1498
                or self.max_long_partial_prefills
1499
                != SchedulerConfig.max_long_partial_prefills):
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
            _raise_or_fallback(feature_name="Concurrent Partial Prefill",
                               recommend_to_remove=False)
            return False

        # No OTLP observability so far.
        if (self.otlp_traces_endpoint or self.collect_detailed_traces):
            _raise_or_fallback(feature_name="--otlp-traces-endpoint",
                               recommend_to_remove=False)
            return False

        # Only Ngram speculative decoding so far.
1511
        is_ngram_enabled = False
1512
        is_eagle_enabled = False
1513
        if self.speculative_config is not None:
1514
            # This is supported but experimental (handled below).
1515
1516
1517
1518
1519
1520
            speculative_method = self.speculative_config.get("method")
            if speculative_method:
                if speculative_method in ("ngram", "[ngram]"):
                    is_ngram_enabled = True
                elif speculative_method == "eagle":
                    is_eagle_enabled = True
1521
            else:
1522
1523
1524
1525
1526
                speculative_model = self.speculative_config.get("model")
                if speculative_model in ("ngram", "[ngram]"):
                    is_ngram_enabled = True
            if not (is_ngram_enabled or is_eagle_enabled):
                # Other speculative decoding methods are not supported yet.
1527
1528
1529
1530
1531
1532
1533
                _raise_or_fallback(feature_name="Speculative Decoding",
                                   recommend_to_remove=False)
                return False

        # No FlashInfer or XFormers so far.
        V1_BACKENDS = [
            "FLASH_ATTN_VLLM_V1", "FLASH_ATTN", "PALLAS", "PALLAS_VLLM_V1",
1534
            "TRITON_ATTN_VLLM_V1", "TRITON_MLA", "FLASHMLA"
1535
1536
1537
1538
1539
1540
1541
        ]
        if (envs.is_set("VLLM_ATTENTION_BACKEND")
                and envs.VLLM_ATTENTION_BACKEND not in V1_BACKENDS):
            name = f"VLLM_ATTENTION_BACKEND={envs.VLLM_ATTENTION_BACKEND}"
            _raise_or_fallback(feature_name=name, recommend_to_remove=True)
            return False

1542
1543
        # Platforms must decide if they can support v1 for this model
        if not current_platform.supports_v1(model_config=model_config):
1544
1545
1546
1547
            _raise_or_fallback(
                feature_name=f"device type={current_platform.device_type}",
                recommend_to_remove=False)
            return False
1548
1549
1550
        #############################################################
        # Experimental Features - allow users to opt in.

1551
1552
1553
1554
1555
        # Signal Handlers requires running in main thread.
        if (threading.current_thread() != threading.main_thread()
                and _warn_or_fallback("Engine in background thread")):
            return False

1556
1557
1558
        # PP is supported on V1 with Ray distributed executor,
        # but off for MP distributed executor for now.
        if (self.pipeline_parallel_size > 1
1559
1560
1561
                and self.distributed_executor_backend != "ray"):
            name = "Pipeline Parallelism without Ray distributed executor"
            _raise_or_fallback(feature_name=name, recommend_to_remove=False)
1562
1563
1564
            return False

        # ngram is supported on V1, but off by default for now.
1565
        if is_ngram_enabled and _warn_or_fallback("ngram"):
1566
1567
            return False

1568
1569
1570
1571
        # Eagle is under development, so we don't support it yet.
        if is_eagle_enabled and _warn_or_fallback("Eagle"):
            return False

1572
1573
1574
        # Non-CUDA is supported on V1, but off by default for now.
        not_cuda = not current_platform.is_cuda()
        if not_cuda and _warn_or_fallback(  # noqa: SIM103
1575
                current_platform.device_name):
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
            return False
        #############################################################

        return True

    def _set_default_args_v0(self, model_config: ModelConfig) -> None:
        """Set Default Arguments for V0 Engine."""

        max_model_len = model_config.max_model_len
        use_long_context = max_model_len > 32768
        if self.enable_chunked_prefill is None:
            # Chunked prefill not supported for Multimodal or MLA in V0.
            if model_config.is_multimodal_model or model_config.use_mla:
                self.enable_chunked_prefill = False

            # Enable chunked prefill by default for long context (> 32K)
            # models to avoid OOM errors in initial memory profiling phase.
            elif use_long_context:
                from vllm.platforms import current_platform
                is_gpu = current_platform.is_cuda()
                use_sliding_window = (model_config.get_sliding_window()
                                      is not None)
1598
                use_spec_decode = self.speculative_config is not None
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625

                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
                        and not self.enable_prompt_adapter
                        and model_config.runner_type != "pooling"):
                    self.enable_chunked_prefill = True
                    logger.warning(
                        "Chunked prefill is enabled by default for models "
                        "with max_model_len > 32K. Chunked prefill might "
                        "not work with some features or models. If you "
                        "encounter any issues, please disable by launching "
                        "with --enable-chunked-prefill=False.")

            if self.enable_chunked_prefill is None:
                self.enable_chunked_prefill = False

        if not self.enable_chunked_prefill and use_long_context:
            logger.warning(
                "The model has a long context length (%s). This may cause"
                "OOM during the initial memory profiling phase, or result "
                "in low performance due to small KV cache size. Consider "
                "setting --max-model-len to a smaller value.", max_model_len)
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
            msg = "Chunked prefill is not supported for pooling models"
            raise ValueError(msg)

1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
        # if using prefix caching, we must set a hash algo
        if self.enable_prefix_caching:
            # Disable prefix caching for multimodal models for VLLM_V0.
            if model_config.is_multimodal_model:
                logger.warning(
                    "--enable-prefix-caching is not supported for multimodal "
                    "models in V0 and has been disabled.")
                self.enable_prefix_caching = False

            # VLLM_V0 only supports builtin hash algo for prefix caching.
            if self.prefix_caching_hash_algo is None:
                self.prefix_caching_hash_algo = "builtin"
            elif self.prefix_caching_hash_algo == "sha256":
                raise ValueError(
                    "sha256 is not supported for prefix caching in V0 engine. "
                    "Please use 'builtin'.")
1642
1643
1644
1645
1646
1647
1648

        # Set max_num_seqs to 256 for VLLM_V0.
        if self.max_num_seqs is None:
            self.max_num_seqs = 256

    def _set_default_args_v1(self, usage_context: UsageContext) -> None:
        """Set Default Arguments for V1 Engine."""
1649

1650
1651
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
1652
1653
1654
1655
1656

        # V1 enables prefix caching by default.
        if self.enable_prefix_caching is None:
            self.enable_prefix_caching = True

1657
1658
1659
1660
        # if using prefix caching, we must set a hash algo
        if self.enable_prefix_caching and self.prefix_caching_hash_algo is None:
            self.prefix_caching_hash_algo = "builtin"

1661
1662
1663
        # V1 should use the new scheduler by default.
        # Swap it only if this arg is set to the original V0 default
        if self.scheduler_cls == EngineArgs.scheduler_cls:
1664
            self.scheduler_cls = "vllm.v1.core.sched.scheduler.Scheduler"
1665

1666
1667
        # When no user override, set the default values based on the usage
        # context.
1668
        # Use different default values for different hardware.
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681

        # Try to query the device name on the current platform. If it fails,
        # it may be because the platform that imports vLLM is not the same
        # as the platform that vLLM is running on (e.g. the case of scaling
        # vLLM with Ray) and has no GPUs. In this case we use the default
        # values for non-H100/H200 GPUs.
        try:
            from vllm.platforms import current_platform
            device_name = current_platform.get_device_name().lower()
        except Exception:
            # This is only used to set default_max_num_batched_tokens
            device_name = "no-device"

1682
1683
1684
1685
1686
1687
        if "h100" in device_name or "h200" in device_name:
            # For H100 and H200, we use larger default values.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1688
            default_max_num_seqs = 1024
1689
1690
1691
1692
1693
1694
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1695
            default_max_num_seqs = 256
1696

1697
        use_context_value = usage_context.value if usage_context else None
1698
1699
1700
1701
        if (self.max_num_batched_tokens is None
                and usage_context in default_max_num_batched_tokens):
            self.max_num_batched_tokens = default_max_num_batched_tokens[
                usage_context]
1702
            logger.debug(
1703
                "Setting max_num_batched_tokens to %d for %s usage context.",
1704
                self.max_num_batched_tokens, use_context_value)
1705

1706
1707
1708
1709
1710
        if self.max_num_seqs is None:
            self.max_num_seqs = default_max_num_seqs

            logger.debug("Setting max_num_seqs to %d for %s usage context.",
                         self.max_num_seqs, use_context_value)
1711

1712

1713
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1714
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1715
    """Arguments for asynchronous vLLM engine."""
1716
    disable_log_requests: bool = False
1717
1718

    @staticmethod
1719
1720
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1721
1722
1723
1724
        # Initialize plugin to update the parser, for example, The plugin may
        # adding a new kind of quantization method to --quantization argument or
        # a new device to --device argument.
        load_general_plugins()
1725
1726
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1727
1728
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1729
                            help='Disable logging requests.')
1730
1731
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update(parser)
1732
        return parser
1733
1734


1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
def _raise_or_fallback(feature_name: str, recommend_to_remove: bool):
    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
        raise NotImplementedError(
            f"VLLM_USE_V1=1 is not supported with {feature_name}.")
    msg = f"{feature_name} is not supported by the V1 Engine. "
    msg += "Falling back to V0. "
    if recommend_to_remove:
        msg += f"We recommend to remove {feature_name} from your config "
        msg += "in favor of the V1 Engine."
    logger.warning(msg)


def _warn_or_fallback(feature_name: str) -> bool:
    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
        logger.warning(
            "Detected VLLM_USE_V1=1 with %s. Usage should "
            "be considered experimental. Please report any "
            "issues on Github.", feature_name)
        should_exit = False
    else:
        logger.info(
            "%s is experimental on VLLM_USE_V1=1. "
            "Falling back to V0 Engine.", feature_name)
        should_exit = True
    return should_exit


1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
    
    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    """
    value = value.strip()
    match = re.fullmatch(r'(\d+(?:\.\d+)?)([kKmMgGtT])', value)
    if match:
        decimal_multiplier = {
            'k': 10**3,
            'm': 10**6,
            'g': 10**9,
        }
        binary_multiplier = {
            'K': 2**10,
            'M': 2**20,
            'G': 2**30,
        }

        number, suffix = match.groups()
        if suffix in decimal_multiplier:
            mult = decimal_multiplier[suffix]
            return int(float(number) * mult)
        elif suffix in binary_multiplier:
            mult = binary_multiplier[suffix]
            # Do not allow decimals with binary multipliers
            try:
                return int(number) * mult
            except ValueError as e:
                raise argparse.ArgumentTypeError("Decimals are not allowed " \
                f"with binary suffixes like {suffix}. Did you mean to use " \
                f"{number}{suffix.lower()} instead?") from e

    # Regular plain number.
    return int(value)


1803
1804
# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1805
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1806
1807
1808


def _async_engine_args_parser():
1809
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1810
                                        async_args_only=True)