arg_utils.py 69.3 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
from itertools import permutations
11
from typing import (Any, Callable, Dict, List, Literal, Optional, Type,
12
                    TypeVar, Union, cast, get_args, get_origin)
13

14
import torch
15
from typing_extensions import TypeIs, deprecated
16

17
import vllm.envs as envs
18
from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
19
20
21
22
23
24
                         ConfigFormat, ConfigType, DecodingConfig,
                         DetailedTraceModules, Device, DeviceConfig,
                         DistributedExecutorBackend, GuidedDecodingBackend,
                         GuidedDecodingBackendV1, HfOverrides, KVEventsConfig,
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
                         ModelConfig, ModelDType, ModelImpl, MultiModalConfig,
25
26
27
28
29
                         ObservabilityConfig, ParallelConfig, PoolerConfig,
                         PrefixCachingHashAlgo, PromptAdapterConfig,
                         SchedulerConfig, SchedulerPolicy, SpeculativeConfig,
                         TaskOption, TokenizerMode, TokenizerPoolConfig,
                         VllmConfig, get_attr_docs, get_field)
30
from vllm.executor.executor_base import ExecutorBase
31
from vllm.logger import init_logger
32
from vllm.model_executor.layers.quantization import QuantizationMethods
33
from vllm.plugins import load_general_plugins
34
from vllm.reasoning import ReasoningParserManager
35
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
36
from vllm.transformers_utils.utils import check_gguf_file
37
from vllm.usage.usage_lib import UsageContext
38
from vllm.utils import FlexibleArgumentParser, GiB_bytes, is_in_ray_actor
39
40

# yapf: enable
41

42
43
logger = init_logger(__name__)

44
45
46
47
48
# object is used to allow for special typing forms
T = TypeVar("T")
TypeHint = Union[type[Any], object]
TypeHintT = Union[type[T], object]

49

50
51
def optional_type(
        return_type: Callable[[str], T]) -> Callable[[str], Optional[T]]:
52

53
54
55
56
57
58
59
60
61
62
    def _optional_type(val: str) -> Optional[T]:
        if val == "" or val == "None":
            return None
        try:
            if return_type is json.loads and not re.match("^{.*}$", val):
                return cast(T, nullable_kvs(val))
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
                f"Value {val} cannot be converted to {return_type}.") from e
63

64
    return _optional_type
65
66


67
68
69
70
71
72
@deprecated(
    "Passing a JSON argument as a string containing comma separated key=value "
    "pairs is deprecated. This will be removed in v0.10.0. Please use a JSON "
    "string instead.")
def nullable_kvs(val: str) -> dict[str, int]:
    """Parses a string containing comma separate key [str] to value [int]
73
74
75
76
77
78
79
80
    pairs into a dictionary.

    Args:
        val: String value to be parsed.

    Returns:
        Dictionary with parsed values.
    """
81
    out_dict: dict[str, int] = {}
82
    for item in val.split(","):
83
84
85
86
87
        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
88
89

        try:
90
            parsed_value = int(value)
91
92
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
93
94
95
96
97
98
            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
99
100
101
102

    return out_dict


103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
def is_type(type_hint: TypeHint, type: TypeHintT) -> TypeIs[TypeHintT]:
    """Check if the type hint is a specific type."""
    return type_hint is type or get_origin(type_hint) is type


def contains_type(type_hints: set[TypeHint], type: TypeHintT) -> bool:
    """Check if the type hints contain a specific type."""
    return any(is_type(type_hint, type) for type_hint in type_hints)


def get_type(type_hints: set[TypeHint], type: TypeHintT) -> TypeHintT:
    """Get the specific type from the type hints."""
    return next((th for th in type_hints if is_type(th, type)), None)


118
119
120
121
122
123
124
125
126
127
128
129
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
    """Convert Literal type hints to argparse kwargs."""
    type_hint = get_type(type_hints, Literal)
    choices = get_args(type_hint)
    choice_type = type(choices[0])
    if not all(isinstance(choice, choice_type) for choice in choices):
        raise ValueError(
            "All choices must be of the same type. "
            f"Got {choices} with types {[type(c) for c in choices]}")
    return {"type": choice_type, "choices": sorted(choices)}


130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


def get_kwargs(cls: ConfigType) -> dict[str, Any]:
    cls_docs = get_attr_docs(cls)
    kwargs = {}
    for field in fields(cls):
        # Get the default value of the field
        default = field.default
        if field.default_factory is not MISSING:
            default = field.default_factory()

        # Get the help text for the field
        name = field.name
        help = cls_docs[name]
        # Escape % for argparse
        help = help.replace("%", "%%")

        # Initialise the kwargs dictionary for the field
        kwargs[name] = {"default": default, "help": help}

        # Get the set of possible types for the field
        type_hints: set[TypeHint] = set()
        if get_origin(field.type) is Union:
            type_hints.update(get_args(field.type))
        else:
            type_hints.add(field.type)

        # Set other kwargs based on the type hints
        if contains_type(type_hints, bool):
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
165
            kwargs[name].update(literal_to_kwargs(type_hints))
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
        elif contains_type(type_hints, tuple):
            type_hint = get_type(type_hints, tuple)
            types = get_args(type_hint)
            tuple_type = types[0]
            assert all(t is tuple_type for t in types if t is not Ellipsis), (
                "All non-Ellipsis tuple elements must be of the same "
                f"type. Got {types}.")
            kwargs[name]["type"] = tuple_type
            kwargs[name]["nargs"] = "+" if Ellipsis in types else len(types)
        elif contains_type(type_hints, list):
            type_hint = get_type(type_hints, list)
            types = get_args(type_hint)
            assert len(types) == 1, (
                "List type must have exactly one type. Got "
                f"{type_hint} with types {types}")
            kwargs[name]["type"] = types[0]
            kwargs[name]["nargs"] = "+"
        elif contains_type(type_hints, int):
            kwargs[name]["type"] = int
185
186
187
            # Special case for large integers
            if name in {"max_model_len"}:
                kwargs[name]["type"] = human_readable_int
188
189
190
191
192
193
194
195
196
197
198
199
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
        elif contains_type(type_hints, dict):
            # Dict arguments will always be optional
            kwargs[name]["type"] = optional_type(json.loads)
        elif (contains_type(type_hints, str)
              or any(is_not_builtin(th) for th in type_hints)):
            kwargs[name]["type"] = str
        else:
            raise ValueError(
                f"Unsupported type {type_hints} for argument {name}.")

200
201
202
203
204
        # If the type hint was a sequence of literals, use the helper function
        # to update the type and choices
        if get_origin(kwargs[name].get("type")) is Literal:
            kwargs[name].update(literal_to_kwargs({kwargs[name]["type"]}))

205
206
207
208
209
210
211
        # If None is in type_hints, make the argument optional.
        # But not if it's a bool, argparse will handle this better.
        if type(None) in type_hints and not contains_type(type_hints, bool):
            kwargs[name]["type"] = optional_type(kwargs[name]["type"])
            if kwargs[name].get("choices"):
                kwargs[name]["choices"].append("None")
    return kwargs
212
213


214
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
215
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
216
    """Arguments for vLLM engine."""
217
218
219
220
221
222
223
224
225
226
    model: str = ModelConfig.model
    served_model_name: Optional[Union[
        str, List[str]]] = ModelConfig.served_model_name
    tokenizer: Optional[str] = ModelConfig.tokenizer
    hf_config_path: Optional[str] = ModelConfig.hf_config_path
    task: TaskOption = ModelConfig.task
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
    tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
227
228
    download_dir: Optional[str] = LoadConfig.download_dir
    load_format: str = LoadConfig.load_format
229
230
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
231
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
232
233
    seed: Optional[int] = ModelConfig.seed
    max_model_len: Optional[int] = ModelConfig.max_model_len
234
235
236
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
237
    distributed_executor_backend: Optional[Union[
238
239
        DistributedExecutorBackend,
        Type[ExecutorBase]]] = ParallelConfig.distributed_executor_backend
240
    # number of P/D disaggregation (or other disaggregation) workers
241
242
243
244
245
246
    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
247
248
249
250
    block_size: Optional[BlockSize] = CacheConfig.block_size
    enable_prefix_caching: Optional[bool] = CacheConfig.enable_prefix_caching
    prefix_caching_hash_algo: PrefixCachingHashAlgo = \
        CacheConfig.prefix_caching_hash_algo
251
252
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
253
    use_v2_block_manager: bool = True
254
255
256
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
257
258
259
260
261
262
263
    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
264
    max_logprobs: int = ModelConfig.max_logprobs
265
    disable_log_stats: bool = False
266
267
268
269
270
271
272
273
274
275
276
    revision: Optional[str] = ModelConfig.revision
    code_revision: Optional[str] = ModelConfig.code_revision
    rope_scaling: dict[str, Any] = get_field(ModelConfig, "rope_scaling")
    rope_theta: Optional[float] = ModelConfig.rope_theta
    hf_token: Optional[Union[bool, str]] = ModelConfig.hf_token
    hf_overrides: Optional[HfOverrides] = \
        get_field(ModelConfig, "hf_overrides")
    tokenizer_revision: Optional[str] = ModelConfig.tokenizer_revision
    quantization: Optional[QuantizationMethods] = ModelConfig.quantization
    enforce_eager: bool = ModelConfig.enforce_eager
    max_seq_len_to_capture: int = ModelConfig.max_seq_len_to_capture
277
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
278
279
280
    # The following three fields are deprecated and will be removed in a future
    # release. Setting them will have no effect. Please remove them from your
    # configurations.
281
    tokenizer_pool_size: int = TokenizerPoolConfig.pool_size
282
283
    tokenizer_pool_type: str = TokenizerPoolConfig.pool_type
    tokenizer_pool_extra_config: dict = \
284
        get_field(TokenizerPoolConfig, "extra_config")
285
    limit_mm_per_prompt: dict[str, int] = \
286
        get_field(MultiModalConfig, "limit_per_prompt")
287
288
289
290
    mm_processor_kwargs: Optional[Dict[str, Any]] = \
        MultiModalConfig.mm_processor_kwargs
    disable_mm_preprocessor_cache: bool = \
        MultiModalConfig.disable_mm_preprocessor_cache
291
    # LoRA fields
292
    enable_lora: bool = False
293
294
295
296
297
298
299
300
301
302
    enable_lora_bias: bool = LoRAConfig.bias_enabled
    max_loras: int = LoRAConfig.max_loras
    max_lora_rank: int = LoRAConfig.max_lora_rank
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
    max_cpu_loras: Optional[int] = LoRAConfig.max_cpu_loras
    lora_dtype: Optional[Union[str, torch.dtype]] = LoRAConfig.lora_dtype
    lora_extra_vocab_size: int = LoRAConfig.lora_extra_vocab_size
    long_lora_scaling_factors: Optional[tuple[float, ...]] = \
        LoRAConfig.long_lora_scaling_factors
    # PromptAdapter fields
303
    enable_prompt_adapter: bool = False
304
305
306
307
    max_prompt_adapters: int = PromptAdapterConfig.max_prompt_adapters
    max_prompt_adapter_token: int = \
        PromptAdapterConfig.max_prompt_adapter_token

308
    device: Device = DeviceConfig.device
309
310
    num_scheduler_steps: int = SchedulerConfig.num_scheduler_steps
    multi_step_stream_outputs: bool = SchedulerConfig.multi_step_stream_outputs
311
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
312
313
    num_gpu_blocks_override: Optional[
        int] = CacheConfig.num_gpu_blocks_override
314
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
315
316
    model_loader_extra_config: dict = \
        get_field(LoadConfig, "model_loader_extra_config")
317
318
    ignore_patterns: Optional[Union[str,
                                    List[str]]] = LoadConfig.ignore_patterns
319
    preemption_mode: Optional[str] = SchedulerConfig.preemption_mode
320

321
322
323
324
    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
325

326
327
328
329
330
331
    guided_decoding_backend: GuidedDecodingBackend = DecodingConfig.backend
    guided_decoding_disable_fallback: bool = DecodingConfig.disable_fallback
    guided_decoding_disable_any_whitespace: bool = \
        DecodingConfig.disable_any_whitespace
    guided_decoding_disable_additional_properties: bool = \
        DecodingConfig.disable_additional_properties
332
333
    logits_processor_pattern: Optional[
        str] = ModelConfig.logits_processor_pattern
334

335
    speculative_config: Optional[Dict[str, Any]] = None
336

337
    qlora_adapter_name_or_path: Optional[str] = None
338
339
340
341
342
343
    show_hidden_metrics_for_version: Optional[str] = \
        ObservabilityConfig.show_hidden_metrics_for_version
    otlp_traces_endpoint: Optional[str] = \
        ObservabilityConfig.otlp_traces_endpoint
    collect_detailed_traces: Optional[list[DetailedTraceModules]] = \
        ObservabilityConfig.collect_detailed_traces
344
    disable_async_output_proc: bool = not ModelConfig.use_async_output_proc
345
346
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
    scheduler_cls: Union[str, Type[object]] = SchedulerConfig.scheduler_cls
347

348
349
350
351
    override_neuron_config: dict[str, Any] = \
        get_field(ModelConfig, "override_neuron_config")
    override_pooler_config: Optional[Union[dict, PoolerConfig]] = \
        ModelConfig.override_pooler_config
352
    compilation_config: Optional[CompilationConfig] = None
353
354
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
355

356
    kv_transfer_config: Optional[KVTransferConfig] = None
357
    kv_events_config: Optional[KVEventsConfig] = None
358

359
360
361
362
363
    generation_config: str = ModelConfig.generation_config
    enable_sleep_mode: bool = ModelConfig.enable_sleep_mode
    override_generation_config: dict[str, Any] = \
        get_field(ModelConfig, "override_generation_config")
    model_impl: str = ModelConfig.model_impl
364

365
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
366

367
    additional_config: Optional[Dict[str, Any]] = None
368
369
370
    enable_reasoning: Optional[bool] = None  # DEPRECATED
    reasoning_parser: str = DecodingConfig.reasoning_backend

371
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
372

373
    def __post_init__(self):
374
375
376
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
377
        if isinstance(self.compilation_config, (int, dict)):
378
379
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
380

381
        # Setup plugins
382
383
        from vllm.plugins import load_general_plugins
        load_general_plugins()
384
385

    @staticmethod
386
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
387
        """Shared CLI arguments for vLLM engine."""
388

389
        # Model arguments
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
        model_group.add_argument("--model", **model_kwargs["model"])
        model_group.add_argument("--task", **model_kwargs["task"])
        model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"])
        model_group.add_argument("--tokenizer-mode",
                                 **model_kwargs["tokenizer_mode"])
        model_group.add_argument("--trust-remote-code",
                                 **model_kwargs["trust_remote_code"])
        model_group.add_argument("--dtype", **model_kwargs["dtype"])
        model_group.add_argument("--seed", **model_kwargs["seed"])
        model_group.add_argument("--hf-config-path",
                                 **model_kwargs["hf_config_path"])
        model_group.add_argument("--allowed-local-media-path",
                                 **model_kwargs["allowed_local_media_path"])
        model_group.add_argument("--revision", **model_kwargs["revision"])
        model_group.add_argument("--code-revision",
                                 **model_kwargs["code_revision"])
        model_group.add_argument("--rope-scaling",
                                 **model_kwargs["rope_scaling"])
        model_group.add_argument("--rope-theta", **model_kwargs["rope_theta"])
        model_group.add_argument("--tokenizer-revision",
                                 **model_kwargs["tokenizer_revision"])
        model_group.add_argument("--max-model-len",
                                 **model_kwargs["max_model_len"])
        model_group.add_argument("--quantization", "-q",
                                 **model_kwargs["quantization"])
        model_group.add_argument("--enforce-eager",
                                 **model_kwargs["enforce_eager"])
        model_group.add_argument("--max-seq-len-to-capture",
                                 **model_kwargs["max_seq_len_to_capture"])
        model_group.add_argument("--max-logprobs",
                                 **model_kwargs["max_logprobs"])
        model_group.add_argument("--disable-sliding-window",
                                 **model_kwargs["disable_sliding_window"])
        model_group.add_argument("--disable-cascade-attn",
                                 **model_kwargs["disable_cascade_attn"])
        model_group.add_argument("--skip-tokenizer-init",
                                 **model_kwargs["skip_tokenizer_init"])
        model_group.add_argument("--served-model-name",
                                 **model_kwargs["served_model_name"])
        # This one is a special case because it is the
        # opposite of ModelConfig.use_async_output_proc
        model_group.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.")
        model_group.add_argument("--config-format",
                                 choices=[f.value for f in ConfigFormat],
                                 **model_kwargs["config_format"])
        # This one is a special case because it can bool
        # or str. TODO: Handle this in get_kwargs
        model_group.add_argument("--hf-token",
                                 type=str,
                                 nargs="?",
                                 const=True,
                                 default=model_kwargs["hf_token"]["default"],
                                 help=model_kwargs["hf_token"]["help"])
        model_group.add_argument("--hf-overrides",
                                 **model_kwargs["hf_overrides"])
        model_group.add_argument("--override-neuron-config",
                                 **model_kwargs["override_neuron_config"])
        model_group.add_argument("--override-pooler-config",
                                 **model_kwargs["override_pooler_config"])
        model_group.add_argument("--logits-processor-pattern",
                                 **model_kwargs["logits_processor_pattern"])
        model_group.add_argument("--generation-config",
                                 **model_kwargs["generation_config"])
        model_group.add_argument("--override-generation-config",
                                 **model_kwargs["override_generation_config"])
        model_group.add_argument("--enable-sleep-mode",
                                 **model_kwargs["enable_sleep_mode"])
        model_group.add_argument("--model-impl",
                                 choices=[f.value for f in ModelImpl],
                                 **model_kwargs["model_impl"])

471
472
473
474
475
476
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
477
        load_group.add_argument("--load-format",
478
479
                                choices=[f.value for f in LoadFormat],
                                **load_kwargs["load_format"])
480
        load_group.add_argument("--download-dir",
481
                                **load_kwargs["download_dir"])
482
        load_group.add_argument("--model-loader-extra-config",
483
                                **load_kwargs["model_loader_extra_config"])
484
485
486
        load_group.add_argument("--ignore-patterns",
                                **load_kwargs["ignore_patterns"])
        load_group.add_argument("--use-tqdm-on-load",
487
                                **load_kwargs["use_tqdm_on_load"])
488
489
490
491
        load_group.add_argument('--qlora-adapter-name-or-path',
                                type=str,
                                default=None,
                                help='Name or path of the QLoRA adapter.')
492

493
494
495
496
497
498
        # Guided decoding arguments
        guided_decoding_kwargs = get_kwargs(DecodingConfig)
        guided_decoding_group = parser.add_argument_group(
            title="DecodingConfig",
            description=DecodingConfig.__doc__,
        )
499
500
        guided_decoding_group.add_argument("--guided-decoding-backend",
                                           **guided_decoding_kwargs["backend"])
501
        guided_decoding_group.add_argument(
502
503
504
505
506
507
508
509
            "--guided-decoding-disable-fallback",
            **guided_decoding_kwargs["disable_fallback"])
        guided_decoding_group.add_argument(
            "--guided-decoding-disable-any-whitespace",
            **guided_decoding_kwargs["disable_any_whitespace"])
        guided_decoding_group.add_argument(
            "--guided-decoding-disable-additional-properties",
            **guided_decoding_kwargs["disable_additional_properties"])
510
511
512
513
514
515
516
517
        guided_decoding_group.add_argument(
            "--enable-reasoning",
            action=argparse.BooleanOptionalAction,
            help="[DEPRECATED] The `--enable-reasoning` flag is deprecated as "
            "of v0.8.6. Use `--reasoning-parser` to specify the reasoning "
            "parser backend insteadThis flag (`--enable-reasoning`) will be "
            "removed in v0.10.0. When `--reasoning-parser` is specified, "
            "reasoning mode is automatically enabled.")
518
519
520
521
522
523
        guided_decoding_group.add_argument(
            "--reasoning-parser",
            # This choices is a special case because it's not static
            choices=list(ReasoningParserManager.reasoning_parsers),
            **guided_decoding_kwargs["reasoning_backend"])

524
        # Parallel arguments
525
526
527
528
529
530
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
531
            "--distributed-executor-backend",
532
533
            **parallel_kwargs["distributed_executor_backend"])
        parallel_group.add_argument(
534
            "--pipeline-parallel-size", "-pp",
535
            **parallel_kwargs["pipeline_parallel_size"])
536
        parallel_group.add_argument("--tensor-parallel-size", "-tp",
537
                                    **parallel_kwargs["tensor_parallel_size"])
538
        parallel_group.add_argument("--data-parallel-size", "-dp",
539
540
                                    **parallel_kwargs["data_parallel_size"])
        parallel_group.add_argument(
541
            "--enable-expert-parallel",
542
543
            **parallel_kwargs["enable_expert_parallel"])
        parallel_group.add_argument(
544
            "--max-parallel-loading-workers",
545
546
            **parallel_kwargs["max_parallel_loading_workers"])
        parallel_group.add_argument(
547
            "--ray-workers-use-nsight",
548
549
            **parallel_kwargs["ray_workers_use_nsight"])
        parallel_group.add_argument(
550
            "--disable-custom-all-reduce",
551
            **parallel_kwargs["disable_custom_all_reduce"])
552
553
554
555
        parallel_group.add_argument("--worker-cls",
                                    **parallel_kwargs["worker_cls"])
        parallel_group.add_argument("--worker-extension-cls",
                                    **parallel_kwargs["worker_extension_cls"])
556

557
558
559
560
561
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
562
        )
563
564
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
        cache_group.add_argument("--gpu-memory-utilization",
565
                                 **cache_kwargs["gpu_memory_utilization"])
566
567
        cache_group.add_argument("--swap-space", **cache_kwargs["swap_space"])
        cache_group.add_argument("--kv-cache-dtype",
568
                                 **cache_kwargs["cache_dtype"])
569
        cache_group.add_argument("--num-gpu-blocks-override",
570
571
572
573
574
                                 **cache_kwargs["num_gpu_blocks_override"])
        cache_group.add_argument("--enable-prefix-caching",
                                 **cache_kwargs["enable_prefix_caching"])
        cache_group.add_argument("--prefix-caching-hash-algo",
                                 **cache_kwargs["prefix_caching_hash_algo"])
575
        cache_group.add_argument("--cpu-offload-gb",
576
                                 **cache_kwargs["cpu_offload_gb"])
577
        cache_group.add_argument("--calculate-kv-scales",
578
579
                                 **cache_kwargs["calculate_kv_scales"])

580
581
582
583
584
585
        # Tokenizer arguments
        tokenizer_kwargs = get_kwargs(TokenizerPoolConfig)
        tokenizer_group = parser.add_argument_group(
            title="TokenizerPoolConfig",
            description=TokenizerPoolConfig.__doc__,
        )
586
        tokenizer_group.add_argument("--tokenizer-pool-size",
587
                                     **tokenizer_kwargs["pool_size"])
588
        tokenizer_group.add_argument("--tokenizer-pool-type",
589
                                     **tokenizer_kwargs["pool_type"])
590
        tokenizer_group.add_argument("--tokenizer-pool-extra-config",
591
                                     **tokenizer_kwargs["extra_config"])
592
593

        # Multimodal related configs
594
595
596
597
598
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
599
        multimodal_group.add_argument("--limit-mm-per-prompt",
600
                                      **multimodal_kwargs["limit_per_prompt"])
601
        multimodal_group.add_argument(
602
            "--mm-processor-kwargs",
603
604
            **multimodal_kwargs["mm_processor_kwargs"])
        multimodal_group.add_argument(
605
            "--disable-mm-preprocessor-cache",
606
            **multimodal_kwargs["disable_mm_preprocessor_cache"])
607

608
        # LoRA related configs
609
610
611
612
613
614
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
615
            "--enable-lora",
616
            action=argparse.BooleanOptionalAction,
617
618
            help="If True, enable handling of LoRA adapters.")
        lora_group.add_argument("--enable-lora-bias",
619
                                **lora_kwargs["bias_enabled"])
620
621
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
        lora_group.add_argument("--max-lora-rank",
622
                                **lora_kwargs["max_lora_rank"])
623
        lora_group.add_argument("--lora-extra-vocab-size",
624
625
                                **lora_kwargs["lora_extra_vocab_size"])
        lora_group.add_argument(
626
            "--lora-dtype",
627
628
            **lora_kwargs["lora_dtype"],
        )
629
        lora_group.add_argument("--long-lora-scaling-factors",
630
                                **lora_kwargs["long_lora_scaling_factors"])
631
        lora_group.add_argument("--max-cpu-loras",
632
                                **lora_kwargs["max_cpu_loras"])
633
        lora_group.add_argument("--fully-sharded-loras",
634
635
636
637
638
639
640
641
642
                                **lora_kwargs["fully_sharded_loras"])

        # PromptAdapter related configs
        prompt_adapter_kwargs = get_kwargs(PromptAdapterConfig)
        prompt_adapter_group = parser.add_argument_group(
            title="PromptAdapterConfig",
            description=PromptAdapterConfig.__doc__,
        )
        prompt_adapter_group.add_argument(
643
            "--enable-prompt-adapter",
644
            action=argparse.BooleanOptionalAction,
645
            help="If True, enable handling of PromptAdapters.")
646
        prompt_adapter_group.add_argument(
647
            "--max-prompt-adapters",
648
649
            **prompt_adapter_kwargs["max_prompt_adapters"])
        prompt_adapter_group.add_argument(
650
            "--max-prompt-adapter-token",
651
            **prompt_adapter_kwargs["max_prompt_adapter_token"])
652
653
654
655
656
657
658
659
660

        # 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"])

661
662
663
664
665
666
        # Speculative arguments
        speculative_group = parser.add_argument_group(
            title="SpeculativeConfig",
            description=SpeculativeConfig.__doc__,
        )
        speculative_group.add_argument(
667
            "--speculative-config",
668
669
            type=json.loads,
            default=None,
670
671
            help="The configurations for speculative decoding. Should be a "
            "JSON string.")
672

673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
        # Observability arguments
        observability_kwargs = get_kwargs(ObservabilityConfig)
        observability_group = parser.add_argument_group(
            title="ObservabilityConfig",
            description=ObservabilityConfig.__doc__,
        )
        observability_group.add_argument(
            "--show-hidden-metrics-for-version",
            **observability_kwargs["show_hidden_metrics_for_version"])
        observability_group.add_argument(
            "--otlp-traces-endpoint",
            **observability_kwargs["otlp_traces_endpoint"])
        # TODO: generalise this special case
        choices = observability_kwargs["collect_detailed_traces"]["choices"]
        metavar = f"{{{','.join(choices)}}}"
        observability_kwargs["collect_detailed_traces"]["metavar"] = metavar
        observability_kwargs["collect_detailed_traces"]["choices"] += [
            ",".join(p)
            for p in permutations(get_args(DetailedTraceModules), r=2)
        ]
        observability_group.add_argument(
            "--collect-detailed-traces",
            **observability_kwargs["collect_detailed_traces"])
696

697
698
699
700
701
702
703
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
704
            "--max-num-batched-tokens",
705
            **scheduler_kwargs["max_num_batched_tokens"])
706
        scheduler_group.add_argument("--max-num-seqs",
707
708
709
710
711
712
713
714
715
716
                                     **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"])
717
        scheduler_group.add_argument("--num-lookahead-slots",
718
                                     **scheduler_kwargs["num_lookahead_slots"])
719
        scheduler_group.add_argument("--scheduler-delay-factor",
720
                                     **scheduler_kwargs["delay_factor"])
721
        scheduler_group.add_argument("--preemption-mode",
722
                                     **scheduler_kwargs["preemption_mode"])
723
        scheduler_group.add_argument("--num-scheduler-steps",
724
                                     **scheduler_kwargs["num_scheduler_steps"])
725
        scheduler_group.add_argument(
726
            "--multi-step-stream-outputs",
727
            **scheduler_kwargs["multi_step_stream_outputs"])
728
        scheduler_group.add_argument("--scheduling-policy",
729
                                     **scheduler_kwargs["policy"])
730
        scheduler_group.add_argument(
731
            "--enable-chunked-prefill",
732
            **scheduler_kwargs["enable_chunked_prefill"])
733
734
735
        scheduler_group.add_argument(
            "--disable-chunked-mm-input",
            **scheduler_kwargs["disable_chunked_mm_input"])
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
        scheduler_group.add_argument("--scheduler-cls",
                                     **scheduler_kwargs["scheduler_cls"])

        # Compilation arguments
        # compilation_kwargs = get_kwargs(CompilationConfig)
        compilation_group = parser.add_argument_group(
            title="CompilationConfig",
            description=CompilationConfig.__doc__,
        )
        compilation_group.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, "
            "use a JSON string, e.g. ``{\"level\": 3, "
            "\"cudagraph_capture_sizes\": [1, 2, 4, 8]}``\n"
            "Following the convention of traditional "
            "compilers, using ``-O`` without space is also "
            "supported. ``-O3`` is equivalent to ``-O 3``.")

        # KVTransfer arguments
        # kv_transfer_kwargs = get_kwargs(KVTransferConfig)
        kv_transfer_group = parser.add_argument_group(
            title="KVTransferConfig",
            description=KVTransferConfig.__doc__,
        )
        kv_transfer_group.add_argument(
            "--kv-transfer-config",
            type=KVTransferConfig.from_cli,
            default=None,
            help="The configurations for distributed KV cache "
            "transfer. Should be a JSON string.")
        kv_transfer_group.add_argument(
            '--kv-events-config',
            type=KVEventsConfig.from_cli,
            default=None,
            help='The configurations for event publishing.')

        # vLLM arguments
        # vllm_kwargs = get_kwargs(VllmConfig)
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
        vllm_group.add_argument(
789
790
791
792
793
794
795
            "--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\"}'")
796

797
798
799
800
801
802
803
804
805
806
807
808
        # Other arguments
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
                            default=True,
                            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.')
        parser.add_argument('--disable-log-stats',
                            action='store_true',
                            help='Disable logging statistics.')
809

810
        return parser
811
812

    @classmethod
813
    def from_cli_args(cls, args: argparse.Namespace):
814
815
816
        # 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
817
818
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
819

820
    def create_model_config(self) -> ModelConfig:
821
822
823
824
825
826
827
828
829
830
831
        # 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

832
        return ModelConfig(
833
            model=self.model,
834
            hf_config_path=self.hf_config_path,
835
            task=self.task,
836
            tokenizer=self.tokenizer,
837
838
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
839
            allowed_local_media_path=self.allowed_local_media_path,
840
841
842
843
844
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
845
            rope_theta=self.rope_theta,
846
            hf_token=self.hf_token,
847
            hf_overrides=self.hf_overrides,
848
849
850
851
852
853
854
            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,
855
            disable_cascade_attn=self.disable_cascade_attn,
856
            skip_tokenizer_init=self.skip_tokenizer_init,
857
            served_model_name=self.served_model_name,
858
            limit_mm_per_prompt=self.limit_mm_per_prompt,
859
            use_async_output_proc=not self.disable_async_output_proc,
860
            config_format=self.config_format,
861
            mm_processor_kwargs=self.mm_processor_kwargs,
862
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
863
864
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
865
            logits_processor_pattern=self.logits_processor_pattern,
866
            generation_config=self.generation_config,
867
            override_generation_config=self.override_generation_config,
868
            enable_sleep_mode=self.enable_sleep_mode,
869
            model_impl=self.model_impl,
870
        )
871

872
873
    def create_load_config(self) -> LoadConfig:

874
        if(self.qlora_adapter_name_or_path is not None) and \
875
876
            self.quantization != "bitsandbytes":
            raise ValueError(
877
                "QLoRA adapter only support "
878
879
                f"'bitsandbytes' quantization, but got {self.quantization}")

880
881
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
882
883
884
885
886
        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,
887
            use_tqdm_on_load=self.use_tqdm_on_load,
888
        )
889

890
891
892
893
894
895
896
897
898
899
900
901
902
    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
903
        dictionary from the engine.
904
905
        """
        if self.speculative_config is None:
906
907
            return None

908
909
910
911
912
913
914
915
916
917
918
919
920
921
        # 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

922
923
924
925
926
927
928
929
930
931
    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
932

933
934
935
936
937
938
        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.
        """
939
940
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update()
941

942
        device_config = DeviceConfig(device=self.device)
943
944
        model_config = self.create_model_config()

945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
        # * 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)
967

968
969
        assert self.enable_chunked_prefill is not None

970
        cache_config = CacheConfig(
971
            block_size=self.block_size,
972
973
974
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
975
            is_attention_free=model_config.is_attention_free,
976
977
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
978
            enable_prefix_caching=self.enable_prefix_caching,
979
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
980
            cpu_offload_gb=self.cpu_offload_gb,
981
            calculate_kv_scales=self.calculate_kv_scales,
982
        )
983
984
985
986
987
988
989
990
991
992
993
994

        # 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()

995
        parallel_config = ParallelConfig(
996
997
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
998
            data_parallel_size=self.data_parallel_size,
999
            enable_expert_parallel=self.enable_expert_parallel,
1000
1001
1002
            max_parallel_loading_workers=self.max_parallel_loading_workers,
            disable_custom_all_reduce=self.disable_custom_all_reduce,
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1003
            placement_group=placement_group,
1004
1005
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1006
            worker_extension_cls=self.worker_extension_cls,
1007
        )
1008

1009
        speculative_config = self.create_speculative_config(
1010
1011
            target_model_config=model_config,
            target_parallel_config=parallel_config,
1012
            enable_chunked_prefill=self.enable_chunked_prefill,
1013
            disable_log_stats=self.disable_log_stats,
1014
1015
        )

1016
        # Reminder: Please update docs/source/features/compatibility_matrix.md
1017
        # If the feature combo become valid
1018
1019
1020
1021
        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)")
1022
1023
1024
            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")
1025
1026
1027
1028
1029
1030
            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
1031
1032
1033
1034
1035
1036
1037
1038
1039

        # 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

1040
        scheduler_config = SchedulerConfig(
1041
            runner_type=model_config.runner_type,
1042
1043
1044
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1045
            num_lookahead_slots=num_lookahead_slots,
1046
1047
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1048
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1049
            is_multimodal_model=model_config.is_multimodal_model,
1050
            preemption_mode=self.preemption_mode,
1051
            num_scheduler_steps=self.num_scheduler_steps,
1052
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1053
1054
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1055
            policy=self.scheduling_policy,
1056
            scheduler_cls=self.scheduler_cls,
1057
1058
1059
1060
            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,
        )
1061

1062
        lora_config = LoRAConfig(
1063
            bias_enabled=self.enable_lora_bias,
1064
1065
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1066
            fully_sharded_loras=self.fully_sharded_loras,
1067
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1068
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1069
1070
1071
            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
1072

1073
1074
1075
1076
1077
        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

1078
1079
1080
1081
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1082
        load_config = self.create_load_config()
1083

1084
1085
1086
1087
1088
        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

1089
        decoding_config = DecodingConfig(
1090
1091
1092
1093
1094
            backend=self.guided_decoding_backend,
            disable_fallback=self.guided_decoding_disable_fallback,
            disable_any_whitespace=self.guided_decoding_disable_any_whitespace,
            disable_additional_properties=\
                self.guided_decoding_disable_additional_properties,
1095
1096
            reasoning_backend=self.reasoning_parser
        )
1097

1098
        observability_config = ObservabilityConfig(
1099
1100
            show_hidden_metrics_for_version=self.
            show_hidden_metrics_for_version,
1101
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1102
            collect_detailed_traces=self.collect_detailed_traces,
1103
        )
1104

1105
        config = VllmConfig(
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
            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,
1116
            prompt_adapter_config=prompt_adapter_config,
1117
            compilation_config=self.compilation_config,
1118
            kv_transfer_config=self.kv_transfer_config,
1119
            kv_events_config=self.kv_events_config,
1120
            additional_config=self.additional_config,
1121
        )
1122

1123
1124
        return config

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

1144
        if self.preemption_mode != SchedulerConfig.preemption_mode:
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
            _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

1155
        if self.scheduling_policy != SchedulerConfig.policy:
1156
1157
1158
1159
            _raise_or_fallback(feature_name="--scheduling-policy",
                               recommend_to_remove=False)
            return False

1160
        if self.num_scheduler_steps != SchedulerConfig.num_scheduler_steps:
1161
1162
1163
1164
            _raise_or_fallback(feature_name="--num-scheduler-steps",
                               recommend_to_remove=True)
            return False

1165
        if self.scheduler_delay_factor != SchedulerConfig.delay_factor:
1166
1167
1168
1169
            _raise_or_fallback(feature_name="--scheduler-delay-factor",
                               recommend_to_remove=True)
            return False

1170
1171
        if self.guided_decoding_backend not in get_args(
                GuidedDecodingBackendV1):
1172
1173
1174
1175
            _raise_or_fallback(
                feature_name=
                f"--guided-decoding-backend={self.guided_decoding_backend}",
                recommend_to_remove=False)
1176
1177
1178
            return False

        # Need at least Ampere for now (FA support required).
1179
1180
1181
        # 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).
1182
1183
        from vllm.platforms import current_platform
        if (current_platform.is_cuda()
1184
                and current_platform.get_device_capability()
1185
1186
1187
1188
1189
1190
1191
                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":
1192
1193
1194
1195
1196
1197
1198
            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:
1199
                from vllm.attention.utils.fa_utils import (
1200
1201
1202
1203
1204
1205
                    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
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220

        # 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.
1221
        V1_UNSUPPORTED_QUANT = ["gguf"]
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
        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
1242
                != SchedulerConfig.max_num_partial_prefills
1243
                or self.max_long_partial_prefills
1244
                != SchedulerConfig.max_long_partial_prefills):
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
            _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.
1256
        is_ngram_enabled = False
1257
        is_eagle_enabled = False
1258
        if self.speculative_config is not None:
1259
            # This is supported but experimental (handled below).
1260
1261
1262
1263
            speculative_method = self.speculative_config.get("method")
            if speculative_method:
                if speculative_method in ("ngram", "[ngram]"):
                    is_ngram_enabled = True
1264
                elif speculative_method in ("eagle", "eagle3"):
1265
                    is_eagle_enabled = True
1266
            else:
1267
1268
1269
1270
1271
                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.
1272
1273
1274
1275
                _raise_or_fallback(feature_name="Speculative Decoding",
                                   recommend_to_remove=False)
                return False

1276
        # No XFormers so far.
1277
        V1_BACKENDS = [
1278
1279
1280
1281
1282
1283
1284
1285
1286
            "FLASH_ATTN_VLLM_V1",
            "FLASH_ATTN",
            "PALLAS",
            "PALLAS_VLLM_V1",
            "TRITON_ATTN_VLLM_V1",
            "TRITON_MLA",
            "FLASHMLA",
            "FLASHINFER",
            "FLASHINFER_VLLM_V1",
1287
1288
1289
1290
1291
1292
1293
        ]
        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

1294
1295
        # Platforms must decide if they can support v1 for this model
        if not current_platform.supports_v1(model_config=model_config):
1296
1297
1298
1299
            _raise_or_fallback(
                feature_name=f"device type={current_platform.device_type}",
                recommend_to_remove=False)
            return False
1300
1301
1302
        #############################################################
        # Experimental Features - allow users to opt in.

1303
1304
1305
1306
1307
        # Signal Handlers requires running in main thread.
        if (threading.current_thread() != threading.main_thread()
                and _warn_or_fallback("Engine in background thread")):
            return False

1308
1309
1310
        # PP is supported on V1 with Ray distributed executor,
        # but off for MP distributed executor for now.
        if (self.pipeline_parallel_size > 1
1311
1312
1313
                and self.distributed_executor_backend != "ray"):
            name = "Pipeline Parallelism without Ray distributed executor"
            _raise_or_fallback(feature_name=name, recommend_to_remove=False)
1314
1315
1316
            return False

        # ngram is supported on V1, but off by default for now.
1317
        if is_ngram_enabled and _warn_or_fallback("ngram"):
1318
1319
            return False

1320
1321
1322
1323
        # Eagle is under development, so we don't support it yet.
        if is_eagle_enabled and _warn_or_fallback("Eagle"):
            return False

1324
1325
1326
        # 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
1327
                current_platform.device_name):
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
            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)
1350
                use_spec_decode = self.speculative_config is not None
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377

                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)

1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
        # 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.
1388
            if self.prefix_caching_hash_algo == "sha256":
1389
1390
1391
                raise ValueError(
                    "sha256 is not supported for prefix caching in V0 engine. "
                    "Please use 'builtin'.")
1392
1393
1394
1395
1396
1397
1398

        # 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."""
1399

1400
1401
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
1402
1403
1404
1405
1406

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

1407
1408
1409
        # 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:
1410
            self.scheduler_cls = "vllm.v1.core.sched.scheduler.Scheduler"
1411

1412
1413
        # When no user override, set the default values based on the usage
        # context.
1414
        # Use different default values for different hardware.
1415
1416
1417
1418
1419
1420
1421
1422

        # 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
1423
            device_memory = current_platform.get_device_total_memory()
1424
1425
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1426
            device_memory = 0
1427

1428
1429
        if device_memory >= 70 * GiB_bytes:
            # For GPUs like H100 and MI300x, use larger default values.
1430
1431
1432
1433
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1434
            default_max_num_seqs = 1024
1435
1436
1437
1438
1439
1440
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1441
            default_max_num_seqs = 256
1442

1443
        use_context_value = usage_context.value if usage_context else None
1444
1445
1446
1447
        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]
1448
            logger.debug(
1449
                "Setting max_num_batched_tokens to %d for %s usage context.",
1450
                self.max_num_batched_tokens, use_context_value)
1451

1452
1453
1454
1455
1456
        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)
1457

1458

1459
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1460
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1461
    """Arguments for asynchronous vLLM engine."""
1462
    disable_log_requests: bool = False
1463
1464

    @staticmethod
1465
1466
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1467
1468
1469
1470
        # 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()
1471
1472
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1473
1474
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1475
                            help='Disable logging requests.')
1476
1477
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update(parser)
1478
        return parser
1479
1480


1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
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


1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
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)


1549
1550
# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1551
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1552
1553
1554


def _async_engine_args_parser():
1555
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1556
                                        async_args_only=True)