arg_utils.py 84 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import argparse
5
import copy
6
import dataclasses
7
import functools
8
import json
9
import sys
10
from collections.abc import Callable
11
from dataclasses import MISSING, dataclass, fields, is_dataclass
12
from itertools import permutations
13
from types import UnionType
14
15
16
17
18
from typing import (
    TYPE_CHECKING,
    Annotated,
    Any,
    Literal,
19
    TypeAlias,
20
21
22
23
24
25
    TypeVar,
    Union,
    cast,
    get_args,
    get_origin,
)
26

27
import huggingface_hub
28
import regex as re
29
import torch
30
from pydantic import TypeAdapter, ValidationError
31
from pydantic.fields import FieldInfo
32
from typing_extensions import TypeIs, deprecated
33

34
import vllm.envs as envs
35
from vllm.attention.backends.registry import _Backend
36
37
38
39
40
41
42
43
44
45
46
from vllm.config import (
    CacheConfig,
    CompilationConfig,
    ConfigType,
    DeviceConfig,
    EPLBConfig,
    KVEventsConfig,
    KVTransferConfig,
    LoadConfig,
    LoRAConfig,
    ModelConfig,
47
    MultiModalConfig,
48
49
50
51
52
53
54
55
56
    ObservabilityConfig,
    ParallelConfig,
    PoolerConfig,
    SchedulerConfig,
    SpeculativeConfig,
    StructuredOutputsConfig,
    VllmConfig,
    get_attr_docs,
)
57
58
59
60
61
62
63
from vllm.config.cache import (
    BlockSize,
    CacheDType,
    KVOffloadingBackend,
    MambaDType,
    PrefixCachingHashAlgo,
)
64
65
66
67
68
69
70
71
72
73
74
75
76
77
from vllm.config.device import Device
from vllm.config.model import (
    ConvertOption,
    HfOverrides,
    LogprobsMode,
    ModelDType,
    RunnerOption,
    TaskOption,
    TokenizerMode,
)
from vllm.config.multimodal import MMCacheType, MMEncoderTPMode
from vllm.config.observability import DetailedTraceModules
from vllm.config.parallel import DistributedExecutorBackend, ExpertPlacementStrategy
from vllm.config.scheduler import SchedulerPolicy
78
from vllm.config.utils import get_field
79
from vllm.logger import init_logger
80
from vllm.platforms import CpuArchEnum, current_platform
81
from vllm.plugins import load_general_plugins
82
from vllm.ray.lazy_utils import is_in_ray_actor, is_ray_initialized
83
from vllm.reasoning import ReasoningParserManager
84
85
86
87
88
from vllm.transformers_utils.config import (
    get_model_path,
    is_interleaved,
    maybe_override_with_speculators,
)
89
from vllm.transformers_utils.utils import check_gguf_file, is_cloud_storage
90
from vllm.utils.argparse_utils import FlexibleArgumentParser
91
from vllm.utils.mem_constants import GiB_bytes
92
from vllm.utils.network_utils import get_ip
93
from vllm.v1.sample.logits_processor import LogitsProcessor
94

95
96
if TYPE_CHECKING:
    from vllm.model_executor.layers.quantization import QuantizationMethods
97
    from vllm.model_executor.model_loader import LoadFormats
98
    from vllm.usage.usage_lib import UsageContext
99
    from vllm.v1.executor import Executor
100
else:
101
    Executor = Any
102
    QuantizationMethods = Any
103
    LoadFormats = Any
104
105
    UsageContext = Any

106
107
logger = init_logger(__name__)

108
109
# object is used to allow for special typing forms
T = TypeVar("T")
110
111
TypeHint: TypeAlias = type[Any] | object
TypeHintT: TypeAlias = type[T] | object
112

113

114
115
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
    def _parse_type(val: str) -> T:
116
117
118
119
        try:
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
120
121
                f"Value {val} cannot be converted to {return_type}."
            ) from e
122

123
124
125
    return _parse_type


126
127
def optional_type(return_type: Callable[[str], T]) -> Callable[[str], T | None]:
    def _optional_type(val: str) -> T | None:
128
129
130
131
        if val == "" or val == "None":
            return None
        return parse_type(return_type)(val)

132
    return _optional_type
133
134


135
def union_dict_and_str(val: str) -> str | dict[str, str] | None:
136
    if not re.match(r"(?s)^\s*{.*}\s*$", val):
137
        return str(val)
138
    return optional_type(json.loads)(val)
139
140


141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
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)


156
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
157
158
159
160
    """Get the `type` and `choices` from a `Literal` type hint in `type_hints`.

    If `type_hints` also contains `str`, we use `metavar` instead of `choices`.
    """
161
    type_hint = get_type(type_hints, Literal)
162
163
164
    options = get_args(type_hint)
    option_type = type(options[0])
    if not all(isinstance(option, option_type) for option in options):
165
        raise ValueError(
166
            "All options must be of the same type. "
167
168
            f"Got {options} with types {[type(c) for c in options]}"
        )
169
170
    kwarg = "metavar" if contains_type(type_hints, str) else "choices"
    return {"type": option_type, kwarg: sorted(options)}
171
172


173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
def collection_to_kwargs(type_hints: set[TypeHint], type: TypeHint) -> dict[str, Any]:
    type_hint = get_type(type_hints, type)
    types = get_args(type_hint)
    elem_type = types[0]

    # Handle Ellipsis
    assert all(t is elem_type for t in types if t is not Ellipsis), (
        f"All non-Ellipsis elements must be of the same type. Got {types}."
    )

    # Handle Union types
    if get_origin(elem_type) in {Union, UnionType}:
        # Union for Union[X, Y] and UnionType for X | Y
        assert str in get_args(elem_type), (
            "If element can have multiple types, one must be 'str' "
            f"(i.e. 'list[int | str]'). Got {elem_type}."
        )
        elem_type = str

    return {
        "type": elem_type,
        "nargs": "+" if type is not tuple or Ellipsis in types else len(types),
    }


198
199
200
201
202
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


203
204
205
206
207
208
209
210
def get_type_hints(type_hint: TypeHint) -> set[TypeHint]:
    """Extract type hints from Annotated or Union type hints."""
    type_hints: set[TypeHint] = set()
    origin = get_origin(type_hint)
    args = get_args(type_hint)

    if origin is Annotated:
        type_hints.update(get_type_hints(args[0]))
211
212
    elif origin in {Union, UnionType}:
        # Union for Union[X, Y] and UnionType for X | Y
213
214
215
216
217
218
219
220
        for arg in args:
            type_hints.update(get_type_hints(arg))
    else:
        type_hints.add(type_hint)

    return type_hints


221
222
223
224
def is_online_quantization(quantization: Any) -> bool:
    return quantization in ["inc"]


225
NEEDS_HELP = (
226
227
    any("--help" in arg for arg in sys.argv)  # vllm SUBCOMMAND --help
    or (argv0 := sys.argv[0]).endswith("mkdocs")  # mkdocs SUBCOMMAND
228
229
230
231
    or argv0.endswith("mkdocs/__main__.py")  # python -m mkdocs SUBCOMMAND
)


232
@functools.lru_cache(maxsize=30)
233
def _compute_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
234
235
    # Save time only getting attr docs if we're generating help text
    cls_docs = get_attr_docs(cls) if NEEDS_HELP else {}
236
237
    kwargs = {}
    for field in fields(cls):
238
        # Get the set of possible types for the field
239
        type_hints: set[TypeHint] = get_type_hints(field.type)
240
241
242
243
244

        # If the field is a dataclass, we can use the model_validate_json
        generator = (th for th in type_hints if is_dataclass(th))
        dataclass_cls = next(generator, None)

245
        # Get the default value of the field
246
247
        if field.default is not MISSING:
            default = field.default
248
249
250
251
252
253
254
            # Handle pydantic.Field defaults
            if isinstance(default, FieldInfo):
                default = (
                    default.default
                    if default.default_factory is None
                    else default.default_factory()
                )
255
        elif field.default_factory is not MISSING:
256
            default = field.default_factory()
257
258
259

        # Get the help text for the field
        name = field.name
260
        help = cls_docs.get(name, "").strip()
261
262
263
264
265
266
267
        # Escape % for argparse
        help = help.replace("%", "%%")

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

        # Set other kwargs based on the type hints
268
269
270
        json_tip = (
            "Should either be a valid JSON string or JSON keys passed individually."
        )
271
        if dataclass_cls is not None:
272
273
274
275
276
277
278
279

            def parse_dataclass(val: str, cls=dataclass_cls) -> Any:
                try:
                    return TypeAdapter(cls).validate_json(val)
                except ValidationError as e:
                    raise argparse.ArgumentTypeError(repr(e)) from e

            kwargs[name]["type"] = parse_dataclass
280
            kwargs[name]["help"] += f"\n\n{json_tip}"
281
        elif contains_type(type_hints, bool):
282
283
284
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
285
            kwargs[name].update(literal_to_kwargs(type_hints))
286
        elif contains_type(type_hints, tuple):
287
            kwargs[name].update(collection_to_kwargs(type_hints, tuple))
288
        elif contains_type(type_hints, list):
289
290
291
            kwargs[name].update(collection_to_kwargs(type_hints, list))
        elif contains_type(type_hints, set):
            kwargs[name].update(collection_to_kwargs(type_hints, set))
292
293
        elif contains_type(type_hints, int):
            kwargs[name]["type"] = int
294
            # Special case for large integers
295
296
297
298
299
300
            human_readable_ints = {
                "max_model_len",
                "max_num_batched_tokens",
                "kv_cache_memory_bytes",
            }
            if name in human_readable_ints:
301
                kwargs[name]["type"] = human_readable_int
302
                kwargs[name]["help"] += f"\n\n{human_readable_int.__doc__}"
303
304
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
305
306
307
308
        elif contains_type(type_hints, dict) and (
            contains_type(type_hints, str)
            or any(is_not_builtin(th) for th in type_hints)
        ):
309
            kwargs[name]["type"] = union_dict_and_str
310
        elif contains_type(type_hints, dict):
311
            kwargs[name]["type"] = parse_type(json.loads)
312
            kwargs[name]["help"] += f"\n\n{json_tip}"
313
314
315
        elif contains_type(type_hints, str) or any(
            is_not_builtin(th) for th in type_hints
        ):
316
317
            kwargs[name]["type"] = str
        else:
318
            raise ValueError(f"Unsupported type {type_hints} for argument {name}.")
319

320
321
322
323
324
        # 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"]}))

325
326
327
328
329
330
331
        # 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
332
333


334
def get_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
335
336
    """Return argparse kwargs for the given Config dataclass.

337
338
339
    If `--help` or `mkdocs` are not present in the command line command, the
    attribute documentation will not be included in the help output.

340
341
342
343
344
345
346
    The heavy computation is cached via functools.lru_cache, and a deep copy
    is returned so callers can mutate the dictionary without affecting the
    cached version.
    """
    return copy.deepcopy(_compute_kwargs(cls))


347
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
348
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
349
    """Arguments for vLLM engine."""
350

351
    model: str = ModelConfig.model
352
353
354
    served_model_name: str | list[str] | None = ModelConfig.served_model_name
    tokenizer: str | None = ModelConfig.tokenizer
    hf_config_path: str | None = ModelConfig.hf_config_path
355
356
    runner: RunnerOption = ModelConfig.runner
    convert: ConvertOption = ModelConfig.convert
357
    task: TaskOption | None = ModelConfig.task
358
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
359
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
360
361
362
    tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
363
364
    allowed_media_domains: list[str] | None = ModelConfig.allowed_media_domains
    download_dir: str | None = LoadConfig.download_dir
365
    safetensors_load_strategy: str = LoadConfig.safetensors_load_strategy
366
    load_format: str | LoadFormats = LoadConfig.load_format
367
368
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
369
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
370
371
    seed: int | None = ModelConfig.seed
    max_model_len: int | None = ModelConfig.max_model_len
372
373
374
375
376
377
378
    cuda_graph_sizes: list[int] | None = CompilationConfig.cudagraph_capture_sizes
    cudagraph_capture_sizes: list[int] | None = (
        CompilationConfig.cudagraph_capture_sizes
    )
    max_cudagraph_capture_size: int | None = get_field(
        CompilationConfig, "max_cudagraph_capture_size"
    )
379
380
381
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
382
    distributed_executor_backend: (
383
        str | DistributedExecutorBackend | type[Executor] | None
384
    ) = ParallelConfig.distributed_executor_backend
385
    # number of P/D disaggregation (or other disaggregation) workers
386
387
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
388
    decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
389
    data_parallel_size: int = ParallelConfig.data_parallel_size
390
391
392
393
394
    data_parallel_rank: int | None = None
    data_parallel_start_rank: int | None = None
    data_parallel_size_local: int | None = None
    data_parallel_address: str | None = None
    data_parallel_rpc_port: int | None = None
395
    data_parallel_hybrid_lb: bool = False
Rui Qiao's avatar
Rui Qiao committed
396
    data_parallel_backend: str = ParallelConfig.data_parallel_backend
397
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
398
    all2all_backend: str | None = ParallelConfig.all2all_backend
399
    enable_dbo: bool = ParallelConfig.enable_dbo
400
401
    dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold
    dbo_prefill_token_threshold: int = ParallelConfig.dbo_prefill_token_threshold
402
403
404
    disable_nccl_for_dp_synchronization: bool = (
        ParallelConfig.disable_nccl_for_dp_synchronization
    )
405
    eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config")
406
    enable_eplb: bool = ParallelConfig.enable_eplb
407
    expert_placement_strategy: ExpertPlacementStrategy = (
408
        ParallelConfig.expert_placement_strategy
409
    )
410
411
    _api_process_count: int = ParallelConfig._api_process_count
    _api_process_rank: int = ParallelConfig._api_process_rank
412
413
414
415
    num_redundant_experts: int = EPLBConfig.num_redundant_experts
    eplb_window_size: int = EPLBConfig.window_size
    eplb_step_interval: int = EPLBConfig.step_interval
    eplb_log_balancedness: bool = EPLBConfig.log_balancedness
416
    max_parallel_loading_workers: int | None = (
417
418
        ParallelConfig.max_parallel_loading_workers
    )
419
420
    block_size: BlockSize | None = CacheConfig.block_size
    enable_prefix_caching: bool | None = CacheConfig.enable_prefix_caching
421
    prefix_caching_hash_algo: PrefixCachingHashAlgo = (
422
        CacheConfig.prefix_caching_hash_algo
423
    )
424
425
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
426
427
428
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
429
430
    kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
    max_num_batched_tokens: int | None = SchedulerConfig.max_num_batched_tokens
431
432
    max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
    max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
433
    long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold
434
    max_num_seqs: int | None = SchedulerConfig.max_num_seqs
435
    max_logprobs: int = ModelConfig.max_logprobs
436
    logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode
437
    disable_log_stats: bool = False
438
    aggregate_engine_logging: bool = False
439
440
    revision: str | None = ModelConfig.revision
    code_revision: str | None = ModelConfig.code_revision
441
    rope_scaling: dict[str, Any] = get_field(ModelConfig, "rope_scaling")
442
443
    rope_theta: float | None = ModelConfig.rope_theta
    hf_token: bool | str | None = ModelConfig.hf_token
444
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
445
446
    tokenizer_revision: str | None = ModelConfig.tokenizer_revision
    quantization: QuantizationMethods | None = ModelConfig.quantization
447
    enforce_eager: bool = ModelConfig.enforce_eager
448
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
449
    limit_mm_per_prompt: dict[str, int | dict[str, int]] = get_field(
450
451
        MultiModalConfig, "limit_per_prompt"
    )
452
    enable_mm_embeds: bool = MultiModalConfig.enable_mm_embeds
453
    interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings
454
455
456
    media_io_kwargs: dict[str, dict[str, Any]] = get_field(
        MultiModalConfig, "media_io_kwargs"
    )
457
    mm_processor_kwargs: dict[str, Any] | None = MultiModalConfig.mm_processor_kwargs
458
    disable_mm_preprocessor_cache: bool = False  # DEPRECATED
459
    mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb
460
    mm_processor_cache_type: MMCacheType | None = (
461
        MultiModalConfig.mm_processor_cache_type
462
463
    )
    mm_shm_cache_max_object_size_mb: int = (
464
        MultiModalConfig.mm_shm_cache_max_object_size_mb
465
    )
466
    mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
467
468
469
    mm_encoder_attn_backend: _Backend | str | None = (
        MultiModalConfig.mm_encoder_attn_backend
    )
470
    io_processor_plugin: str | None = None
471
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
472
    video_pruning_rate: float = MultiModalConfig.video_pruning_rate
473
    # LoRA fields
474
    enable_lora: bool = False
475
476
    max_loras: int = LoRAConfig.max_loras
    max_lora_rank: int = LoRAConfig.max_lora_rank
477
    default_mm_loras: dict[str, str] | None = LoRAConfig.default_mm_loras
478
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
479
480
    max_cpu_loras: int | None = LoRAConfig.max_cpu_loras
    lora_dtype: str | torch.dtype | None = LoRAConfig.lora_dtype
481
482
    lora_extra_vocab_size: int = LoRAConfig.lora_extra_vocab_size

483
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
484
    num_gpu_blocks_override: int | None = CacheConfig.num_gpu_blocks_override
485
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
486
    model_loader_extra_config: dict = get_field(LoadConfig, "model_loader_extra_config")
487
    ignore_patterns: str | list[str] = get_field(LoadConfig, "ignore_patterns")
488

489
    enable_chunked_prefill: bool | None = SchedulerConfig.enable_chunked_prefill
490
    disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
491

492
    disable_hybrid_kv_cache_manager: bool = (
493
494
        SchedulerConfig.disable_hybrid_kv_cache_manager
    )
495

496
    structured_outputs_config: StructuredOutputsConfig = get_field(
497
498
        VllmConfig, "structured_outputs_config"
    )
499
    reasoning_parser: str = StructuredOutputsConfig.reasoning_parser
500

501
    # Deprecated guided decoding fields
502
503
504
505
    guided_decoding_backend: str | None = None
    guided_decoding_disable_fallback: bool | None = None
    guided_decoding_disable_any_whitespace: bool | None = None
    guided_decoding_disable_additional_properties: bool | None = None
506

507
    logits_processor_pattern: str | None = ModelConfig.logits_processor_pattern
508

509
    speculative_config: dict[str, Any] | None = None
510

511
    show_hidden_metrics_for_version: str | None = (
512
        ObservabilityConfig.show_hidden_metrics_for_version
513
    )
514
515
    otlp_traces_endpoint: str | None = ObservabilityConfig.otlp_traces_endpoint
    collect_detailed_traces: list[DetailedTraceModules] | None = (
516
        ObservabilityConfig.collect_detailed_traces
517
    )
518
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
519
    scheduler_cls: str | type[object] = SchedulerConfig.scheduler_cls
520

521
522
    pooler_config: PoolerConfig | None = ModelConfig.pooler_config
    override_pooler_config: dict | PoolerConfig | None = (
523
        ModelConfig.override_pooler_config
524
525
    )
    compilation_config: CompilationConfig = get_field(VllmConfig, "compilation_config")
526
527
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
528

529
530
    kv_transfer_config: KVTransferConfig | None = None
    kv_events_config: KVEventsConfig | None = None
531

532
533
    generation_config: str = ModelConfig.generation_config
    enable_sleep_mode: bool = ModelConfig.enable_sleep_mode
534
535
536
    override_generation_config: dict[str, Any] = get_field(
        ModelConfig, "override_generation_config"
    )
537
    model_impl: str = ModelConfig.model_impl
538
    override_attention_dtype: str = ModelConfig.override_attention_dtype
539

540
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
541
542
    mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype
    mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype
543
    mamba_block_size: int | None = get_field(CacheConfig, "mamba_block_size")
544

545
    additional_config: dict[str, Any] = get_field(VllmConfig, "additional_config")
546

547
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
548
    pt_load_map_location: str = LoadConfig.pt_load_map_location
549

550
551
    # DEPRECATED
    enable_multimodal_encoder_data_parallel: bool = False
552

553
    logits_processors: list[str | type[LogitsProcessor]] | None = (
554
555
        ModelConfig.logits_processors
    )
556
557
    """Custom logitproc types"""

558
559
    async_scheduling: bool = SchedulerConfig.async_scheduling

560
    kv_sharing_fast_prefill: bool = CacheConfig.kv_sharing_fast_prefill
561

562
563
564
565
566
    kv_offloading_size: float | None = CacheConfig.kv_offloading_size
    kv_offloading_backend: KVOffloadingBackend | None = (
        CacheConfig.kv_offloading_backend
    )

567
    def __post_init__(self):
568
569
570
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
571
        if isinstance(self.compilation_config, dict):
572
            self.compilation_config = CompilationConfig(**self.compilation_config)
573
        if isinstance(self.eplb_config, dict):
574
            self.eplb_config = EPLBConfig(**self.eplb_config)
575
        # Setup plugins
576
        from vllm.plugins import load_general_plugins
577

578
        load_general_plugins()
579
580
581
582
583
        # when use hf offline,replace model id to local model path
        if huggingface_hub.constants.HF_HUB_OFFLINE:
            model_id = self.model
            self.model = get_model_path(self.model, self.revision)
            logger.info(
584
585
586
587
                "HF_HUB_OFFLINE is True, replace model_id [%s] to model_path [%s]",
                model_id,
                self.model,
            )
588
589

    @staticmethod
590
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
591
        """Shared CLI arguments for vLLM engine."""
592

593
        # Model arguments
594
595
596
597
598
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
599
        if not ("serve" in sys.argv[1:] and "--help" in sys.argv[1:]):
600
            model_group.add_argument("--model", **model_kwargs["model"])
601
602
        model_group.add_argument("--runner", **model_kwargs["runner"])
        model_group.add_argument("--convert", **model_kwargs["convert"])
603
        model_group.add_argument("--task", **model_kwargs["task"], deprecated=True)
604
        model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"])
605
606
607
608
        model_group.add_argument("--tokenizer-mode", **model_kwargs["tokenizer_mode"])
        model_group.add_argument(
            "--trust-remote-code", **model_kwargs["trust_remote_code"]
        )
609
610
        model_group.add_argument("--dtype", **model_kwargs["dtype"])
        model_group.add_argument("--seed", **model_kwargs["seed"])
611
612
613
614
615
616
617
        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(
            "--allowed-media-domains", **model_kwargs["allowed_media_domains"]
        )
618
        model_group.add_argument("--revision", **model_kwargs["revision"])
619
620
        model_group.add_argument("--code-revision", **model_kwargs["code_revision"])
        model_group.add_argument("--rope-scaling", **model_kwargs["rope_scaling"])
621
        model_group.add_argument("--rope-theta", **model_kwargs["rope_theta"])
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
        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-logprobs", **model_kwargs["max_logprobs"])
        model_group.add_argument("--logprobs-mode", **model_kwargs["logprobs_mode"])
        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(
            "--enable-prompt-embeds", **model_kwargs["enable_prompt_embeds"]
        )
        model_group.add_argument(
            "--served-model-name", **model_kwargs["served_model_name"]
        )
        model_group.add_argument("--config-format", **model_kwargs["config_format"])
646
647
        # This one is a special case because it can bool
        # or str. TODO: Handle this in get_kwargs
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
        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("--pooler-config", **model_kwargs["pooler_config"])
        model_group.add_argument(
            "--override-pooler-config",
            **model_kwargs["override_pooler_config"],
            deprecated=True,
        )
        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"]
        )
675
        model_group.add_argument("--model-impl", **model_kwargs["model_impl"])
676
677
678
679
680
681
682
683
684
        model_group.add_argument(
            "--override-attention-dtype", **model_kwargs["override_attention_dtype"]
        )
        model_group.add_argument(
            "--logits-processors", **model_kwargs["logits_processors"]
        )
        model_group.add_argument(
            "--io-processor-plugin", **model_kwargs["io_processor_plugin"]
        )
685

686
687
688
689
690
691
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
692
        load_group.add_argument("--load-format", **load_kwargs["load_format"])
693
694
695
696
697
698
699
700
701
702
703
704
        load_group.add_argument("--download-dir", **load_kwargs["download_dir"])
        load_group.add_argument(
            "--safetensors-load-strategy", **load_kwargs["safetensors_load_strategy"]
        )
        load_group.add_argument(
            "--model-loader-extra-config", **load_kwargs["model_loader_extra_config"]
        )
        load_group.add_argument("--ignore-patterns", **load_kwargs["ignore_patterns"])
        load_group.add_argument("--use-tqdm-on-load", **load_kwargs["use_tqdm_on_load"])
        load_group.add_argument(
            "--pt-load-map-location", **load_kwargs["pt_load_map_location"]
        )
705

706
707
708
709
710
        # Structured outputs arguments
        structured_outputs_kwargs = get_kwargs(StructuredOutputsConfig)
        structured_outputs_group = parser.add_argument_group(
            title="StructuredOutputsConfig",
            description=StructuredOutputsConfig.__doc__,
711
        )
712
        structured_outputs_group.add_argument(
713
            "--reasoning-parser",
714
            # This choice is a special case because it's not static
715
            choices=list(ReasoningParserManager.reasoning_parsers),
716
717
            **structured_outputs_kwargs["reasoning_parser"],
        )
718
719
720
721
722
723
724
725
726
727
728
        # Deprecated guided decoding arguments
        for arg, type in [
            ("--guided-decoding-backend", str),
            ("--guided-decoding-disable-fallback", bool),
            ("--guided-decoding-disable-any-whitespace", bool),
            ("--guided-decoding-disable-additional-properties", bool),
        ]:
            structured_outputs_group.add_argument(
                arg,
                type=type,
                help=(f"[DEPRECATED] {arg} will be removed in v0.12.0."),
729
730
                deprecated=True,
            )
731

732
        # Parallel arguments
733
734
735
736
737
738
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
739
            "--distributed-executor-backend",
740
741
            **parallel_kwargs["distributed_executor_backend"],
        )
742
        parallel_group.add_argument(
743
744
745
746
            "--pipeline-parallel-size",
            "-pp",
            **parallel_kwargs["pipeline_parallel_size"],
        )
747
        parallel_group.add_argument(
748
749
            "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
        )
750
        parallel_group.add_argument(
751
752
753
754
755
756
757
758
759
760
            "--decode-context-parallel-size",
            "-dcp",
            **parallel_kwargs["decode_context_parallel_size"],
        )
        parallel_group.add_argument(
            "--data-parallel-size", "-dp", **parallel_kwargs["data_parallel_size"]
        )
        parallel_group.add_argument(
            "--data-parallel-rank",
            "-dpn",
761
            type=int,
762
763
764
            help="Data parallel rank of this instance. "
            "When set, enables external load balancer mode.",
        )
765
        parallel_group.add_argument(
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
            "--data-parallel-start-rank",
            "-dpr",
            type=int,
            help="Starting data parallel rank for secondary nodes.",
        )
        parallel_group.add_argument(
            "--data-parallel-size-local",
            "-dpl",
            type=int,
            help="Number of data parallel replicas to run on this node.",
        )
        parallel_group.add_argument(
            "--data-parallel-address",
            "-dpa",
            type=str,
            help="Address of data parallel cluster head-node.",
        )
        parallel_group.add_argument(
            "--data-parallel-rpc-port",
            "-dpp",
            type=int,
            help="Port for data parallel RPC communication.",
        )
        parallel_group.add_argument(
            "--data-parallel-backend",
            "-dpb",
            type=str,
            default="mp",
            help='Backend for data parallel, either "mp" or "ray".',
        )
796
        parallel_group.add_argument(
797
798
799
800
801
            "--data-parallel-hybrid-lb", **parallel_kwargs["data_parallel_hybrid_lb"]
        )
        parallel_group.add_argument(
            "--enable-expert-parallel", **parallel_kwargs["enable_expert_parallel"]
        )
802
803
804
        parallel_group.add_argument(
            "--all2all-backend", **parallel_kwargs["all2all_backend"]
        )
805
        parallel_group.add_argument("--enable-dbo", **parallel_kwargs["enable_dbo"])
806
807
        parallel_group.add_argument(
            "--dbo-decode-token-threshold",
808
809
            **parallel_kwargs["dbo_decode_token_threshold"],
        )
810
811
        parallel_group.add_argument(
            "--dbo-prefill-token-threshold",
812
813
            **parallel_kwargs["dbo_prefill_token_threshold"],
        )
814
815
816
817
        parallel_group.add_argument(
            "--disable-nccl-for-dp-synchronization",
            **parallel_kwargs["disable_nccl_for_dp_synchronization"],
        )
818
819
        parallel_group.add_argument("--enable-eplb", **parallel_kwargs["enable_eplb"])
        parallel_group.add_argument("--eplb-config", **parallel_kwargs["eplb_config"])
820
821
        parallel_group.add_argument(
            "--expert-placement-strategy",
822
823
            **parallel_kwargs["expert_placement_strategy"],
        )
824
825
826
        parallel_group.add_argument(
            "--num-redundant-experts",
            type=int,
827
828
829
            help="[DEPRECATED] --num-redundant-experts will be removed in v0.12.0.",
            deprecated=True,
        )
830
831
832
833
        parallel_group.add_argument(
            "--eplb-window-size",
            type=int,
            help="[DEPRECATED] --eplb-window-size will be removed in v0.12.0.",
834
835
            deprecated=True,
        )
836
837
838
        parallel_group.add_argument(
            "--eplb-step-interval",
            type=int,
839
840
841
            help="[DEPRECATED] --eplb-step-interval will be removed in v0.12.0.",
            deprecated=True,
        )
842
843
844
        parallel_group.add_argument(
            "--eplb-log-balancedness",
            action=argparse.BooleanOptionalAction,
845
846
847
            help="[DEPRECATED] --eplb-log-balancedness will be removed in v0.12.0.",
            deprecated=True,
        )
848

849
        parallel_group.add_argument(
850
            "--max-parallel-loading-workers",
851
852
            **parallel_kwargs["max_parallel_loading_workers"],
        )
853
        parallel_group.add_argument(
854
855
            "--ray-workers-use-nsight", **parallel_kwargs["ray_workers_use_nsight"]
        )
856
        parallel_group.add_argument(
857
            "--disable-custom-all-reduce",
858
859
860
861
862
863
            **parallel_kwargs["disable_custom_all_reduce"],
        )
        parallel_group.add_argument("--worker-cls", **parallel_kwargs["worker_cls"])
        parallel_group.add_argument(
            "--worker-extension-cls", **parallel_kwargs["worker_extension_cls"]
        )
864
865
        parallel_group.add_argument(
            "--enable-multimodal-encoder-data-parallel",
866
            action="store_true",
867
868
            deprecated=True,
        )
869

870
871
872
873
874
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
875
        )
876
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
877
878
879
880
881
882
        cache_group.add_argument(
            "--gpu-memory-utilization", **cache_kwargs["gpu_memory_utilization"]
        )
        cache_group.add_argument(
            "--kv-cache-memory-bytes", **cache_kwargs["kv_cache_memory_bytes"]
        )
883
        cache_group.add_argument("--swap-space", **cache_kwargs["swap_space"])
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
        cache_group.add_argument("--kv-cache-dtype", **cache_kwargs["cache_dtype"])
        cache_group.add_argument(
            "--num-gpu-blocks-override", **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"]
        )
        cache_group.add_argument("--cpu-offload-gb", **cache_kwargs["cpu_offload_gb"])
        cache_group.add_argument(
            "--calculate-kv-scales", **cache_kwargs["calculate_kv_scales"]
        )
        cache_group.add_argument(
            "--kv-sharing-fast-prefill", **cache_kwargs["kv_sharing_fast_prefill"]
        )
        cache_group.add_argument(
            "--mamba-cache-dtype", **cache_kwargs["mamba_cache_dtype"]
        )
        cache_group.add_argument(
            "--mamba-ssm-cache-dtype", **cache_kwargs["mamba_ssm_cache_dtype"]
        )
907
908
909
        cache_group.add_argument(
            "--mamba-block-size", **cache_kwargs["mamba_block_size"]
        )
910
911
912
913
914
915
        cache_group.add_argument(
            "--kv-offloading-size", **cache_kwargs["kv_offloading_size"]
        )
        cache_group.add_argument(
            "--kv-offloading-backend", **cache_kwargs["kv_offloading_backend"]
        )
916

917
        # Multimodal related configs
918
919
920
921
922
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
923
        multimodal_group.add_argument(
924
925
            "--limit-mm-per-prompt", **multimodal_kwargs["limit_per_prompt"]
        )
926
927
928
        multimodal_group.add_argument(
            "--enable-mm-embeds", **multimodal_kwargs["enable_mm_embeds"]
        )
929
930
931
932
933
934
935
936
937
        multimodal_group.add_argument(
            "--media-io-kwargs", **multimodal_kwargs["media_io_kwargs"]
        )
        multimodal_group.add_argument(
            "--mm-processor-kwargs", **multimodal_kwargs["mm_processor_kwargs"]
        )
        multimodal_group.add_argument(
            "--mm-processor-cache-gb", **multimodal_kwargs["mm_processor_cache_gb"]
        )
938
        multimodal_group.add_argument(
939
940
            "--disable-mm-preprocessor-cache", action="store_true", deprecated=True
        )
941
        multimodal_group.add_argument(
942
943
            "--mm-processor-cache-type", **multimodal_kwargs["mm_processor_cache_type"]
        )
944
945
        multimodal_group.add_argument(
            "--mm-shm-cache-max-object-size-mb",
946
947
            **multimodal_kwargs["mm_shm_cache_max_object_size_mb"],
        )
948
        multimodal_group.add_argument(
949
950
            "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"]
        )
951
952
953
954
        multimodal_group.add_argument(
            "--mm-encoder-attn-backend",
            **multimodal_kwargs["mm_encoder_attn_backend"],
        )
955
956
957
        multimodal_group.add_argument(
            "--interleave-mm-strings", **multimodal_kwargs["interleave_mm_strings"]
        )
958
        multimodal_group.add_argument(
959
960
            "--skip-mm-profiling", **multimodal_kwargs["skip_mm_profiling"]
        )
961

962
        multimodal_group.add_argument(
963
964
            "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"]
        )
965

966
        # LoRA related configs
967
968
969
970
971
972
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
973
            "--enable-lora",
974
            action=argparse.BooleanOptionalAction,
975
976
            help="If True, enable handling of LoRA adapters.",
        )
977
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
978
979
980
981
        lora_group.add_argument("--max-lora-rank", **lora_kwargs["max_lora_rank"])
        lora_group.add_argument(
            "--lora-extra-vocab-size", **lora_kwargs["lora_extra_vocab_size"]
        )
982
        lora_group.add_argument(
983
            "--lora-dtype",
984
985
            **lora_kwargs["lora_dtype"],
        )
986
987
988
989
990
        lora_group.add_argument("--max-cpu-loras", **lora_kwargs["max_cpu_loras"])
        lora_group.add_argument(
            "--fully-sharded-loras", **lora_kwargs["fully_sharded_loras"]
        )
        lora_group.add_argument("--default-mm-loras", **lora_kwargs["default_mm_loras"])
991

992
993
994
995
996
997
998
999
        # 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",
1000
1001
            **observability_kwargs["show_hidden_metrics_for_version"],
        )
1002
        observability_group.add_argument(
1003
1004
            "--otlp-traces-endpoint", **observability_kwargs["otlp_traces_endpoint"]
        )
1005
1006
1007
1008
1009
        # 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"] += [
1010
            ",".join(p) for p in permutations(get_args(DetailedTraceModules), r=2)
1011
1012
1013
        ]
        observability_group.add_argument(
            "--collect-detailed-traces",
1014
1015
            **observability_kwargs["collect_detailed_traces"],
        )
1016

1017
1018
1019
1020
1021
1022
1023
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
1024
1025
            "--max-num-batched-tokens", **scheduler_kwargs["max_num_batched_tokens"]
        )
1026
        scheduler_group.add_argument(
1027
1028
1029
1030
1031
            "--max-num-seqs", **scheduler_kwargs["max_num_seqs"]
        )
        scheduler_group.add_argument(
            "--max-num-partial-prefills", **scheduler_kwargs["max_num_partial_prefills"]
        )
1032
1033
        scheduler_group.add_argument(
            "--max-long-partial-prefills",
1034
1035
            **scheduler_kwargs["max_long_partial_prefills"],
        )
1036
1037
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
1038
1039
1040
1041
1042
            **scheduler_kwargs["long_prefill_token_threshold"],
        )
        scheduler_group.add_argument(
            "--num-lookahead-slots", **scheduler_kwargs["num_lookahead_slots"]
        )
1043
1044
        # multi-step scheduling has been removed; corresponding arguments
        # are no longer supported.
1045
        scheduler_group.add_argument(
1046
1047
            "--scheduling-policy", **scheduler_kwargs["policy"]
        )
1048
        scheduler_group.add_argument(
1049
1050
1051
1052
1053
1054
1055
1056
            "--enable-chunked-prefill", **scheduler_kwargs["enable_chunked_prefill"]
        )
        scheduler_group.add_argument(
            "--disable-chunked-mm-input", **scheduler_kwargs["disable_chunked_mm_input"]
        )
        scheduler_group.add_argument(
            "--scheduler-cls", **scheduler_kwargs["scheduler_cls"]
        )
1057
1058
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
1059
1060
1061
1062
1063
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"],
        )
        scheduler_group.add_argument(
            "--async-scheduling", **scheduler_kwargs["async_scheduling"]
        )
1064

1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
        # Compilation arguments
        compilation_kwargs = get_kwargs(CompilationConfig)
        compilation_group = parser.add_argument_group(
            title="CompilationConfig",
            description=CompilationConfig.__doc__,
        )
        compilation_group.add_argument(
            "--cudagraph-capture-sizes", **compilation_kwargs["cudagraph_capture_sizes"]
        )
        compilation_kwargs["cudagraph_capture_sizes"]["help"] = (
            "--cuda-graph-sizes is deprecated and will be removed in v0.13.0 or v1.0.0,"
            " whichever is soonest. Please use --cudagraph-capture-sizes instead."
        )
        compilation_group.add_argument(
            "--cuda-graph-sizes",
            **compilation_kwargs["cudagraph_capture_sizes"],
            deprecated=True,
        )
        compilation_group.add_argument(
            "--max-cudagraph-capture-size",
            **compilation_kwargs["max_cudagraph_capture_size"],
        )

1088
        # vLLM arguments
1089
        vllm_kwargs = get_kwargs(VllmConfig)
1090
1091
1092
1093
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
1094
1095
1096
1097
        # We construct SpeculativeConfig using fields from other configs in
        # create_engine_config. So we set the type to a JSON string here to
        # delay the Pydantic validation that comes with SpeculativeConfig.
        vllm_kwargs["speculative_config"]["type"] = optional_type(json.loads)
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
        vllm_group.add_argument(
            "--speculative-config", **vllm_kwargs["speculative_config"]
        )
        vllm_group.add_argument(
            "--kv-transfer-config", **vllm_kwargs["kv_transfer_config"]
        )
        vllm_group.add_argument("--kv-events-config", **vllm_kwargs["kv_events_config"])
        vllm_group.add_argument(
            "--compilation-config", "-O", **vllm_kwargs["compilation_config"]
        )
        vllm_group.add_argument(
            "--additional-config", **vllm_kwargs["additional_config"]
        )
        vllm_group.add_argument(
            "--structured-outputs-config", **vllm_kwargs["structured_outputs_config"]
        )
1114

1115
        # Other arguments
1116
1117
1118
1119
1120
        parser.add_argument(
            "--disable-log-stats",
            action="store_true",
            help="Disable logging statistics.",
        )
1121

1122
1123
1124
1125
1126
1127
        parser.add_argument(
            "--aggregate-engine-logging",
            action="store_true",
            help="Log aggregate rather than per-engine statistics "
            "when using data parallelism.",
        )
1128
        return parser
1129
1130

    @classmethod
1131
    def from_cli_args(cls, args: argparse.Namespace):
1132
1133
1134
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
1135
1136
1137
        engine_args = cls(
            **{attr: getattr(args, attr) for attr in attrs if hasattr(args, attr)}
        )
Zhuohan Li's avatar
Zhuohan Li committed
1138
        return engine_args
1139

1140
    def create_model_config(self) -> ModelConfig:
1141
1142
1143
1144
        # 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"

1145
1146
1147
1148
        if self.disable_mm_preprocessor_cache:
            logger.warning(
                "`--disable-mm-preprocessor-cache` is deprecated "
                "and will be removed in v0.13. "
1149
1150
                "Please use `--mm-processor-cache-gb 0` instead.",
            )
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162

            self.mm_processor_cache_gb = 0
        elif envs.VLLM_MM_INPUT_CACHE_GIB != 4:
            logger.warning(
                "VLLM_MM_INPUT_CACHE_GIB` is deprecated "
                "and will be removed in v0.13. "
                "Please use `--mm-processor-cache-gb %d` instead.",
                envs.VLLM_MM_INPUT_CACHE_GIB,
            )

            self.mm_processor_cache_gb = envs.VLLM_MM_INPUT_CACHE_GIB

1163
1164
1165
1166
        if self.enable_multimodal_encoder_data_parallel:
            logger.warning(
                "--enable-multimodal-encoder-data-parallel` is deprecated "
                "and will be removed in v0.13. "
1167
1168
                "Please use `--mm-encoder-tp-mode data` instead."
            )
1169
1170
1171

            self.mm_encoder_tp_mode = "data"

1172
        return ModelConfig(
1173
            model=self.model,
1174
            hf_config_path=self.hf_config_path,
1175
1176
            runner=self.runner,
            convert=self.convert,
1177
            task=self.task,
1178
            tokenizer=self.tokenizer,
1179
1180
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1181
            allowed_local_media_path=self.allowed_local_media_path,
1182
            allowed_media_domains=self.allowed_media_domains,
1183
1184
1185
1186
1187
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
1188
            rope_theta=self.rope_theta,
1189
            hf_token=self.hf_token,
1190
            hf_overrides=self.hf_overrides,
1191
1192
1193
1194
1195
            tokenizer_revision=self.tokenizer_revision,
            max_model_len=self.max_model_len,
            quantization=self.quantization,
            enforce_eager=self.enforce_eager,
            max_logprobs=self.max_logprobs,
1196
            logprobs_mode=self.logprobs_mode,
1197
            disable_sliding_window=self.disable_sliding_window,
1198
            disable_cascade_attn=self.disable_cascade_attn,
1199
            skip_tokenizer_init=self.skip_tokenizer_init,
1200
            enable_prompt_embeds=self.enable_prompt_embeds,
1201
            served_model_name=self.served_model_name,
1202
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1203
            enable_mm_embeds=self.enable_mm_embeds,
1204
            interleave_mm_strings=self.interleave_mm_strings,
1205
            media_io_kwargs=self.media_io_kwargs,
1206
            skip_mm_profiling=self.skip_mm_profiling,
1207
            config_format=self.config_format,
1208
            mm_processor_kwargs=self.mm_processor_kwargs,
1209
            mm_processor_cache_gb=self.mm_processor_cache_gb,
1210
            mm_processor_cache_type=self.mm_processor_cache_type,
1211
            mm_shm_cache_max_object_size_mb=self.mm_shm_cache_max_object_size_mb,
1212
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
1213
            mm_encoder_attn_backend=self.mm_encoder_attn_backend,
1214
            pooler_config=self.pooler_config,
1215
            override_pooler_config=self.override_pooler_config,
1216
            logits_processor_pattern=self.logits_processor_pattern,
1217
            generation_config=self.generation_config,
1218
            override_generation_config=self.override_generation_config,
1219
            enable_sleep_mode=self.enable_sleep_mode,
1220
            model_impl=self.model_impl,
1221
            override_attention_dtype=self.override_attention_dtype,
1222
            logits_processors=self.logits_processors,
1223
            video_pruning_rate=self.video_pruning_rate,
1224
            io_processor_plugin=self.io_processor_plugin,
1225
        )
1226

1227
    def validate_tensorizer_args(self):
1228
1229
        from vllm.model_executor.model_loader.tensorizer import TensorizerConfig

1230
1231
        for key in self.model_loader_extra_config:
            if key in TensorizerConfig._fields:
1232
1233
1234
                self.model_loader_extra_config["tensorizer_config"][key] = (
                    self.model_loader_extra_config[key]
                )
1235

1236
    def create_load_config(self) -> LoadConfig:
1237
1238
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1239

1240
1241
1242
        if self.load_format == "tensorizer":
            if hasattr(self.model_loader_extra_config, "to_serializable"):
                self.model_loader_extra_config = (
1243
1244
                    self.model_loader_extra_config.to_serializable()
                )
1245
            self.model_loader_extra_config["tensorizer_config"] = {}
1246
1247
1248
            self.model_loader_extra_config["tensorizer_config"]["tensorizer_dir"] = (
                self.model
            )
1249
            self.validate_tensorizer_args()
1250

1251
1252
1253
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
1254
            safetensors_load_strategy=self.safetensors_load_strategy,
1255
            device="cpu" if is_online_quantization(self.quantization) else None,
1256
1257
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1258
            use_tqdm_on_load=self.use_tqdm_on_load,
1259
            pt_load_map_location=self.pt_load_map_location,
1260
        )
1261

1262
1263
1264
1265
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
1266
    ) -> SpeculativeConfig | None:
1267
1268
1269
1270
1271
1272
        """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
1273
        dictionary from the engine.
1274
1275
        """
        if self.speculative_config is None:
1276
            return None
1277

1278
1279
1280
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.
1281
1282
1283
1284
1285
1286
        self.speculative_config.update(
            {
                "target_model_config": target_model_config,
                "target_parallel_config": target_parallel_config,
            }
        )
1287
        return SpeculativeConfig(**self.speculative_config)
1288

1289
1290
    def create_engine_config(
        self,
1291
        usage_context: UsageContext | None = None,
1292
        headless: bool = False,
1293
1294
1295
1296
1297
1298
1299
    ) -> 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
1300

1301
1302
1303
1304
1305
1306
        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.
        """
1307
        current_platform.pre_register_and_update()
1308

1309
        device_config = DeviceConfig(device=cast(Device, current_platform.device_type))
1310

1311
1312
        # Check if the model is a speculator and override model/tokenizer/config
        # BEFORE creating ModelConfig, so the config is created with the target model
1313
1314
1315
1316
        # Skip speculator detection for cloud storage models (eg: S3, GCS) since
        # HuggingFace cannot load configs directly from S3 URLs. S3 models can still
        # use speculators with explicit --speculative-config.
        if not is_cloud_storage(self.model):
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
            (self.model, self.tokenizer, self.speculative_config) = (
                maybe_override_with_speculators(
                    model=self.model,
                    tokenizer=self.tokenizer,
                    revision=self.revision,
                    trust_remote_code=self.trust_remote_code,
                    vllm_speculative_config=self.speculative_config,
                )
            )

1327
1328
1329
1330
        model_config = self.create_model_config()
        self.model = model_config.model
        self.tokenizer = model_config.tokenizer

1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
        # * 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)

1348
1349
        # Set default arguments for V1 Engine.
        self._set_default_args(usage_context, model_config)
1350
1351
        # Disable chunked prefill and prefix caching for:
        # POWER (ppc64le)/ARM/s390x/RISCV CPUs in V1
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
        if current_platform.is_cpu() and current_platform.get_cpu_architecture() in (
            CpuArchEnum.POWERPC,
            CpuArchEnum.S390X,
            CpuArchEnum.ARM,
            CpuArchEnum.RISCV,
        ):
            logger.info(
                "Chunked prefill is not supported for ARM and POWER, "
                "S390X and RISC-V CPUs; "
                "disabling it for V1 backend."
            )
1363
            self.enable_chunked_prefill = False
1364
1365
1366
1367
1368
1369
1370
            logger.info(
                "Prefix caching is not supported for ARM and POWER, "
                "S390X and RISC-V CPUs; "
                "disabling it for V1 backend."
            )
            self.enable_prefix_caching = False

1371
1372
        assert self.enable_chunked_prefill is not None

1373
        sliding_window: int | None = None
1374
1375
1376
1377
1378
1379
        if not is_interleaved(model_config.hf_text_config):
            # Only set CacheConfig.sliding_window if the model is all sliding
            # window. Otherwise CacheConfig.sliding_window will override the
            # global layers in interleaved sliding window models.
            sliding_window = model_config.get_sliding_window()

1380
1381
1382
        # Note(hc): In the current implementation of decode context
        # parallel(DCP), tp_size needs to be divisible by dcp_size,
        # because the world size does not change by dcp, it simply
1383
        # reuses the GPUs of TP group, and split one TP group into
1384
        # tp_size//dcp_size DCP groups.
1385
        assert self.tensor_parallel_size % self.decode_context_parallel_size == 0, (
1386
1387
1388
1389
            f"tp_size={self.tensor_parallel_size} must be divisible by"
            f"dcp_size={self.decode_context_parallel_size}."
        )

1390
        cache_config = CacheConfig(
1391
            block_size=self.block_size,
1392
            gpu_memory_utilization=self.gpu_memory_utilization,
1393
            kv_cache_memory_bytes=self.kv_cache_memory_bytes,
1394
1395
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1396
            is_attention_free=model_config.is_attention_free,
1397
            num_gpu_blocks_override=self.num_gpu_blocks_override,
1398
            sliding_window=sliding_window,
1399
            enable_prefix_caching=self.enable_prefix_caching,
1400
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1401
            cpu_offload_gb=self.cpu_offload_gb,
1402
            calculate_kv_scales=self.calculate_kv_scales,
1403
            kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
1404
1405
            mamba_cache_dtype=self.mamba_cache_dtype,
            mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype,
1406
            mamba_block_size=self.mamba_block_size,
1407
1408
            kv_offloading_size=self.kv_offloading_size,
            kv_offloading_backend=self.kv_offloading_backend,
1409
        )
1410

1411
1412
1413
1414
1415
1416
        ray_runtime_env = None
        if is_ray_initialized():
            # Ray Serve LLM calls `create_engine_config` in the context
            # of a Ray task, therefore we check is_ray_initialized()
            # as opposed to is_in_ray_actor().
            import ray
1417

1418
            ray_runtime_env = ray.get_runtime_context().runtime_env
1419
1420
1421
1422
1423
1424
1425
            # Avoid logging sensitive environment variables
            sanitized_env = ray_runtime_env.to_dict() if ray_runtime_env else {}
            if "env_vars" in sanitized_env:
                sanitized_env["env_vars"] = {
                    k: "***" for k in sanitized_env["env_vars"]
                }
            logger.info("Using ray runtime env (env vars redacted): %s", sanitized_env)
1426

1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
        # 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()

1438
        assert not headless or not self.data_parallel_hybrid_lb, (
1439
1440
            "data_parallel_hybrid_lb is not applicable in headless mode"
        )
1441

1442
        data_parallel_external_lb = self.data_parallel_rank is not None
1443
        # Local DP rank = 1, use pure-external LB.
1444
1445
        if data_parallel_external_lb:
            assert self.data_parallel_size_local in (1, None), (
1446
1447
                "data_parallel_size_local must be 1 when data_parallel_rank is set"
            )
1448
            data_parallel_size_local = 1
1449
1450
            # Use full external lb if we have local_size of 1.
            self.data_parallel_hybrid_lb = False
1451
1452
        elif self.data_parallel_size_local is not None:
            data_parallel_size_local = self.data_parallel_size_local
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467

            if self.data_parallel_start_rank and not headless:
                # Infer hybrid LB mode.
                self.data_parallel_hybrid_lb = True

            if self.data_parallel_hybrid_lb and data_parallel_size_local == 1:
                # Use full external lb if we have local_size of 1.
                data_parallel_external_lb = True
                self.data_parallel_hybrid_lb = False

            if data_parallel_size_local == self.data_parallel_size:
                # Disable hybrid LB mode if set for a single node
                self.data_parallel_hybrid_lb = False

            self.data_parallel_rank = self.data_parallel_start_rank or 0
1468
        else:
1469
            assert not self.data_parallel_hybrid_lb, (
1470
1471
                "data_parallel_size_local must be set to use data_parallel_hybrid_lb."
            )
1472

1473
1474
1475
1476
1477
1478
1479
1480
1481
            if self.data_parallel_backend == "ray" and (
                envs.VLLM_RAY_DP_PACK_STRATEGY == "span"
            ):
                # Data parallel size defaults to 1 if DP ranks are spanning
                # multiple nodes
                data_parallel_size_local = 1
            else:
                # Otherwise local DP size defaults to global DP size if not set
                data_parallel_size_local = self.data_parallel_size
1482
1483
1484

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1485
1486
1487
1488
        if self.data_parallel_address is None:
            if self.data_parallel_backend == "ray":
                host_ip = get_ip()
                logger.info(
1489
1490
                    "Using host IP %s as ray-based data parallel address", host_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1491
1492
1493
1494
                data_parallel_address = host_ip
            else:
                assert self.data_parallel_backend == "mp", (
                    "data_parallel_backend can only be ray or mp, got %s",
1495
1496
                    self.data_parallel_backend,
                )
Rui Qiao's avatar
Rui Qiao committed
1497
1498
1499
                data_parallel_address = ParallelConfig.data_parallel_master_ip
        else:
            data_parallel_address = self.data_parallel_address
1500
1501
1502

        # This port is only used when there are remote data parallel engines,
        # otherwise the local IPC transport is used.
1503
        data_parallel_rpc_port = (
1504
            self.data_parallel_rpc_port
1505
1506
1507
            if (self.data_parallel_rpc_port is not None)
            else ParallelConfig.data_parallel_rpc_port
        )
1508

1509
1510
        if self.async_scheduling:
            if self.pipeline_parallel_size > 1:
1511
1512
1513
                raise ValueError(
                    "Async scheduling is not supported with pipeline-parallel-size > 1."
                )
1514
1515
1516
1517
1518
1519

            # Currently, async scheduling does not support speculative decoding.
            # TODO(woosuk): Support it.
            if self.speculative_config is not None:
                raise ValueError(
                    "Currently, speculative decoding is not supported with "
1520
1521
                    "async scheduling."
                )
1522

1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
        # Forward the deprecated CLI args to the EPLB config.
        if self.num_redundant_experts is not None:
            self.eplb_config.num_redundant_experts = self.num_redundant_experts
        if self.eplb_window_size is not None:
            self.eplb_config.window_size = self.eplb_window_size
        if self.eplb_step_interval is not None:
            self.eplb_config.step_interval = self.eplb_step_interval
        if self.eplb_log_balancedness is not None:
            self.eplb_config.log_balancedness = self.eplb_log_balancedness

1533
        parallel_config = ParallelConfig(
1534
1535
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1536
            data_parallel_size=self.data_parallel_size,
1537
1538
            data_parallel_rank=self.data_parallel_rank or 0,
            data_parallel_external_lb=data_parallel_external_lb,
1539
1540
1541
            data_parallel_size_local=data_parallel_size_local,
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
1542
            data_parallel_backend=self.data_parallel_backend,
1543
            data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
1544
            enable_expert_parallel=self.enable_expert_parallel,
1545
            all2all_backend=self.all2all_backend,
1546
1547
            enable_dbo=self.enable_dbo,
            dbo_decode_token_threshold=self.dbo_decode_token_threshold,
1548
            dbo_prefill_token_threshold=self.dbo_prefill_token_threshold,
1549
            disable_nccl_for_dp_synchronization=self.disable_nccl_for_dp_synchronization,
1550
            enable_eplb=self.enable_eplb,
1551
            eplb_config=self.eplb_config,
1552
            expert_placement_strategy=self.expert_placement_strategy,
1553
1554
1555
            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,
1556
            ray_runtime_env=ray_runtime_env,
1557
            placement_group=placement_group,
1558
1559
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1560
            worker_extension_cls=self.worker_extension_cls,
1561
            decode_context_parallel_size=self.decode_context_parallel_size,
1562
1563
            _api_process_count=self._api_process_count,
            _api_process_rank=self._api_process_rank,
1564
        )
1565

1566
        if self.async_scheduling and (
1567
1568
            parallel_config.distributed_executor_backend
            not in ("mp", "uni", "external_launcher")
1569
1570
        ):
            raise ValueError(
1571
1572
                "Currently, async scheduling only supports `mp`, `uni` or "
                "`external_launcher` distributed executor backend, but you choose "
1573
1574
1575
                f"`{parallel_config.distributed_executor_backend}`."
            )

1576
        speculative_config = self.create_speculative_config(
1577
1578
1579
1580
            target_model_config=model_config,
            target_parallel_config=parallel_config,
        )

1581
1582
1583
1584
1585
        # make sure num_lookahead_slots is set appropriately depending on
        # whether speculative decoding is enabled
        num_lookahead_slots = self.num_lookahead_slots
        if speculative_config is not None:
            num_lookahead_slots = speculative_config.num_lookahead_slots
1586

1587
        scheduler_config = SchedulerConfig(
1588
            runner_type=model_config.runner_type,
1589
1590
1591
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1592
            num_lookahead_slots=num_lookahead_slots,
1593
            enable_chunked_prefill=self.enable_chunked_prefill,
1594
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1595
            is_multimodal_model=model_config.is_multimodal_model,
1596
            is_encoder_decoder=model_config.is_encoder_decoder,
1597
            policy=self.scheduling_policy,
1598
            scheduler_cls=self.scheduler_cls,
1599
1600
1601
            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,
1602
            disable_hybrid_kv_cache_manager=self.disable_hybrid_kv_cache_manager,
1603
            async_scheduling=self.async_scheduling,
1604
        )
1605

1606
1607
1608
        if not model_config.is_multimodal_model and self.default_mm_loras:
            raise ValueError(
                "Default modality-specific LoRA(s) were provided for a "
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
                "non multimodal model"
            )

        lora_config = (
            LoRAConfig(
                max_lora_rank=self.max_lora_rank,
                max_loras=self.max_loras,
                default_mm_loras=self.default_mm_loras,
                fully_sharded_loras=self.fully_sharded_loras,
                lora_extra_vocab_size=self.lora_extra_vocab_size,
                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
        )
1627

1628
1629
1630
1631
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1632
        load_config = self.create_load_config()
1633

1634
1635
        # Pass reasoning_parser into StructuredOutputsConfig
        if self.reasoning_parser:
1636
            self.structured_outputs_config.reasoning_parser = self.reasoning_parser
1637
1638
1639
1640

        # Forward the deprecated CLI args to the StructuredOutputsConfig
        so_config = self.structured_outputs_config
        if self.guided_decoding_backend is not None:
1641
            so_config.guided_decoding_backend = self.guided_decoding_backend
1642
        if self.guided_decoding_disable_fallback is not None:
1643
            so_config.disable_fallback = self.guided_decoding_disable_fallback
1644
        if self.guided_decoding_disable_any_whitespace is not None:
1645
            so_config.disable_any_whitespace = (
1646
1647
                self.guided_decoding_disable_any_whitespace
            )
1648
        if self.guided_decoding_disable_additional_properties is not None:
1649
            so_config.disable_additional_properties = (
1650
1651
                self.guided_decoding_disable_additional_properties
            )
1652

1653
        observability_config = ObservabilityConfig(
1654
            show_hidden_metrics_for_version=(self.show_hidden_metrics_for_version),
1655
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1656
            collect_detailed_traces=self.collect_detailed_traces,
1657
        )
1658

1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
        # Compilation config overrides
        if self.cuda_graph_sizes is not None:
            logger.warning(
                "--cuda-graph-sizes is deprecated and will be removed in v0.13.0 or "
                "v1.0.0, whichever is soonest. Please use --cudagraph-capture-sizes "
                "instead."
            )
            if self.compilation_config.cudagraph_capture_sizes is not None:
                raise ValueError(
                    "cuda_graph_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
            self.compilation_config.cudagraph_capture_sizes = self.cuda_graph_sizes
        if self.cudagraph_capture_sizes is not None:
            if self.compilation_config.cudagraph_capture_sizes is not None:
                raise ValueError(
                    "cudagraph_capture_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
            self.compilation_config.cudagraph_capture_sizes = (
                self.cudagraph_capture_sizes
            )
        if self.max_cudagraph_capture_size is not None:
            if self.compilation_config.max_cudagraph_capture_size is not None:
                raise ValueError(
                    "max_cudagraph_capture_size and compilation_config."
                    "max_cudagraph_capture_size are mutually exclusive"
                )
            self.compilation_config.max_cudagraph_capture_size = (
                self.max_cudagraph_capture_size
            )

1691
        config = VllmConfig(
1692
1693
1694
1695
1696
1697
1698
1699
            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,
1700
            structured_outputs_config=self.structured_outputs_config,
1701
            observability_config=observability_config,
1702
            compilation_config=self.compilation_config,
1703
            kv_transfer_config=self.kv_transfer_config,
1704
            kv_events_config=self.kv_events_config,
1705
            additional_config=self.additional_config,
1706
        )
1707

1708
1709
        return config

1710
1711
1712
1713
1714
1715
    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.

1716
1717
1718
1719
        if self.logits_processor_pattern != EngineArgs.logits_processor_pattern:
            _raise_or_fallback(
                feature_name="--logits-processor-pattern", recommend_to_remove=False
            )
1720
1721
1722
            return False

        # No Concurrent Partial Prefills so far.
1723
1724
1725
1726
1727
1728
1729
1730
        if (
            self.max_num_partial_prefills != SchedulerConfig.max_num_partial_prefills
            or self.max_long_partial_prefills
            != SchedulerConfig.max_long_partial_prefills
        ):
            _raise_or_fallback(
                feature_name="Concurrent Partial Prefill", recommend_to_remove=False
            )
1731
1732
            return False

1733
        # V1 supports N-gram, Medusa, and Eagle speculative decoding.
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
        if self.speculative_config is not None:
            # speculative_config could still be a dict at this point
            if isinstance(self.speculative_config, dict):
                method = self.speculative_config.get("method", None)
            else:
                method = self.speculative_config.method

            if method == "draft_model":
                raise NotImplementedError(
                    "Draft model speculative decoding is not supported yet. "
                    "Please consider using other speculative decoding methods "
1745
1746
                    "such as ngram, medusa, eagle, or mtp."
                )
1747
1748

        V1_BACKENDS = [
1749
1750
            "FLASH_ATTN",
            "PALLAS",
1751
            "TRITON_ATTN",
1752
            "TRITON_MLA",
1753
            "CUTLASS_MLA",
1754
            "FLASHMLA",
1755
            "FLASH_ATTN_MLA",
1756
            "FLASHINFER",
1757
            "FLASHINFER_MLA",
1758
            "ROCM_AITER_MLA",
1759
            "TORCH_SDPA",
1760
            "FLEX_ATTENTION",
1761
            "TREE_ATTN",
1762
1763
            "XFORMERS",
            "ROCM_ATTN",
1764
            "ROCM_AITER_UNIFIED_ATTN",
1765
        ]
1766
1767
1768
1769
        if (
            envs.is_set("VLLM_ATTENTION_BACKEND")
            and envs.VLLM_ATTENTION_BACKEND not in V1_BACKENDS
        ):
1770
1771
1772
1773
1774
1775
1776
            name = f"VLLM_ATTENTION_BACKEND={envs.VLLM_ATTENTION_BACKEND}"
            _raise_or_fallback(feature_name=name, recommend_to_remove=True)
            return False

        #############################################################
        # Experimental Features - allow users to opt in.

1777
        if self.pipeline_parallel_size > 1:
1778
1779
1780
            supports_pp = getattr(
                self.distributed_executor_backend, "supports_pp", False
            )
1781
            if not supports_pp and self.distributed_executor_backend not in (
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
                ParallelConfig.distributed_executor_backend,
                "ray",
                "mp",
                "external_launcher",
            ):
                name = (
                    "Pipeline Parallelism without Ray distributed "
                    "executor or multiprocessing executor or external "
                    "launcher"
                )
                _raise_or_fallback(feature_name=name, recommend_to_remove=False)
1793
                return False
1794

1795
1796
1797
1798
        if current_platform.is_cpu() and model_config.get_sliding_window() is not None:
            _raise_or_fallback(
                feature_name="sliding window (CPU backend)", recommend_to_remove=False
            )
1799
1800
            return False

1801
1802
1803
1804
        #############################################################

        return True

1805
1806
1807
    def _set_default_args(
        self, usage_context: UsageContext, model_config: ModelConfig
    ) -> None:
1808
        """Set Default Arguments for V1 Engine."""
1809

1810
        # V1 uses chunked prefills and prefix caching by default
1811
1812
1813
1814
        # for non-pooling tasks.
        # For pooling tasks the default is False
        if model_config.runner_type != "pooling":
            self.enable_chunked_prefill = True
1815

1816
            if self.enable_prefix_caching is None:
1817
1818
1819
1820
1821
1822
                # Disable prefix caching default for hybrid models
                # since the feature is still experimental.
                if model_config.is_hybrid:
                    self.enable_prefix_caching = False
                else:
                    self.enable_prefix_caching = True
1823
1824
        else:
            pooling_type = model_config.pooler_config.pooling_type
1825
            is_causal = getattr(model_config.hf_config, "is_causal", True)
1826
1827
1828
            incremental_prefill_supported = (
                pooling_type is not None
                and pooling_type.lower() == "last"
1829
                and bool(is_causal)
1830
            )
1831

1832
            action = "Enabling" if incremental_prefill_supported else "Disabling"
1833
1834
1835
1836
1837
1838
1839
1840

            if self.enable_chunked_prefill is None:
                self.enable_chunked_prefill = incremental_prefill_supported
                logger.info("(%s) chunked prefill by default", action)
            if self.enable_prefix_caching is None:
                self.enable_prefix_caching = incremental_prefill_supported
                logger.info("(%s) prefix caching by default", action)

1841
1842
        # When no user override, set the default values based on the usage
        # context.
1843
        # Use different default values for different hardware.
1844
1845
1846
1847
1848
1849
1850

        # 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:
1851
            device_memory = current_platform.get_device_total_memory()
1852
            device_name = current_platform.get_device_name().lower()
1853
1854
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1855
            device_memory = 0
1856

1857
1858
1859
        # NOTE(Kuntai): Setting large `max_num_batched_tokens` for A100 reduces
        # throughput, see PR #17885 for more details.
        # So here we do an extra device name check to prevent such regression.
1860
        from vllm.usage.usage_lib import UsageContext
1861

1862
        if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
1863
            # For GPUs like H100 and MI300x, use larger default values.
1864
1865
1866
1867
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1868
1869
1870
1871
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
1872
1873
1874
1875
1876
1877
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1878
1879
1880
1881
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
1882

1883
1884
1885
1886
        # tpu specific default values.
        if current_platform.is_tpu():
            default_max_num_batched_tokens_tpu = {
                UsageContext.LLM_CLASS: {
1887
1888
1889
                    "V6E": 2048,
                    "V5E": 1024,
                    "V5P": 512,
1890
1891
                },
                UsageContext.OPENAI_API_SERVER: {
1892
1893
1894
1895
                    "V6E": 1024,
                    "V5E": 512,
                    "V5P": 256,
                },
1896
1897
            }

1898
1899
        # cpu specific default values.
        if current_platform.is_cpu():
1900
            world_size = self.pipeline_parallel_size * self.tensor_parallel_size
1901
            default_max_num_batched_tokens = {
1902
1903
                UsageContext.LLM_CLASS: 4096 * world_size,
                UsageContext.OPENAI_API_SERVER: 2048 * world_size,
1904
1905
            }
            default_max_num_seqs = {
1906
1907
                UsageContext.LLM_CLASS: 256 * world_size,
                UsageContext.OPENAI_API_SERVER: 128 * world_size,
1908
1909
            }

1910
        use_context_value = usage_context.value if usage_context else None
1911
1912
1913
1914
        if (
            self.max_num_batched_tokens is None
            and usage_context in default_max_num_batched_tokens
        ):
1915
1916
            if current_platform.is_tpu():
                chip_name = current_platform.get_device_name()
1917
1918
1919
1920
                if chip_name in default_max_num_batched_tokens_tpu[usage_context]:
                    self.max_num_batched_tokens = default_max_num_batched_tokens_tpu[
                        usage_context
                    ][chip_name]
1921
                else:
1922
1923
1924
                    self.max_num_batched_tokens = default_max_num_batched_tokens[
                        usage_context
                    ]
1925
            else:
1926
1927
1928
                if not self.enable_chunked_prefill:
                    self.max_num_batched_tokens = model_config.max_model_len
                else:
1929
1930
1931
                    self.max_num_batched_tokens = default_max_num_batched_tokens[
                        usage_context
                    ]
1932
            logger.debug(
1933
                "Setting max_num_batched_tokens to %d for %s usage context.",
1934
1935
1936
                self.max_num_batched_tokens,
                use_context_value,
            )
1937

1938
1939
1940
1941
1942
        if self.max_num_seqs is None and usage_context in default_max_num_seqs:
            self.max_num_seqs = min(
                default_max_num_seqs[usage_context],
                self.max_num_batched_tokens or sys.maxsize,
            )
1943

1944
1945
1946
1947
1948
            logger.debug(
                "Setting max_num_seqs to %d for %s usage context.",
                self.max_num_seqs,
                use_context_value,
            )
1949

1950

1951
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1952
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1953
    """Arguments for asynchronous vLLM engine."""
1954

1955
1956
1957
1958
1959
1960
    enable_log_requests: bool = False

    @property
    @deprecated(
        "`disable_log_requests` is deprecated and has been replaced with "
        "`enable_log_requests`. This will be removed in v0.12.0. Please use "
1961
1962
        "`enable_log_requests` instead."
    )
1963
1964
1965
1966
1967
1968
1969
    def disable_log_requests(self) -> bool:
        return not self.enable_log_requests

    @disable_log_requests.setter
    @deprecated(
        "`disable_log_requests` is deprecated and has been replaced with "
        "`enable_log_requests`. This will be removed in v0.12.0. Please use "
1970
1971
        "`enable_log_requests` instead."
    )
1972
1973
    def disable_log_requests(self, value: bool):
        self.enable_log_requests = not value
1974
1975

    @staticmethod
1976
1977
1978
    def add_cli_args(
        parser: FlexibleArgumentParser, async_args_only: bool = False
    ) -> FlexibleArgumentParser:
1979
        # Initialize plugin to update the parser, for example, The plugin may
1980
        # add a new kind of quantization method to --quantization argument or
1981
1982
        # a new device to --device argument.
        load_general_plugins()
1983
1984
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
        parser.add_argument(
            "--enable-log-requests",
            action=argparse.BooleanOptionalAction,
            default=AsyncEngineArgs.enable_log_requests,
            help="Enable logging requests.",
        )
        parser.add_argument(
            "--disable-log-requests",
            action=argparse.BooleanOptionalAction,
            default=not AsyncEngineArgs.enable_log_requests,
            help="[DEPRECATED] Disable logging requests.",
            deprecated=True,
        )
1998
        current_platform.pre_register_and_update(parser)
1999
        return parser
2000
2001


2002
2003
2004
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(
2005
2006
            f"VLLM_USE_V1=1 is not supported with {feature_name}."
        )
2007
2008
2009
2010
2011
2012
2013
2014
    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)


2015
2016
2017
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
2018

2019
2020
2021
2022
2023
2024
    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    """
    value = value.strip()
2025
    match = re.fullmatch(r"(\d+(?:\.\d+)?)([kKmMgGtT])", value)
2026
2027
    if match:
        decimal_multiplier = {
2028
2029
2030
            "k": 10**3,
            "m": 10**6,
            "g": 10**9,
2031
2032
        }
        binary_multiplier = {
2033
2034
2035
            "K": 2**10,
            "M": 2**20,
            "G": 2**30,
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
        }

        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:
2048
2049
2050
2051
2052
                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
2053
2054
2055

    # Regular plain number.
    return int(value)