arg_utils.py 85.9 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
33

34
import vllm.envs as envs
35
from vllm.attention.backends.registry import AttentionBackendEnum
36
37
38
39
40
from vllm.config import (
    CacheConfig,
    CompilationConfig,
    ConfigType,
    DeviceConfig,
41
    ECTransferConfig,
42
43
44
45
46
47
    EPLBConfig,
    KVEventsConfig,
    KVTransferConfig,
    LoadConfig,
    LoRAConfig,
    ModelConfig,
48
    MultiModalConfig,
49
50
51
52
53
54
55
56
57
    ObservabilityConfig,
    ParallelConfig,
    PoolerConfig,
    SchedulerConfig,
    SpeculativeConfig,
    StructuredOutputsConfig,
    VllmConfig,
    get_attr_docs,
)
58
59
60
61
62
63
64
from vllm.config.cache import (
    BlockSize,
    CacheDType,
    KVOffloadingBackend,
    MambaDType,
    PrefixCachingHashAlgo,
)
65
66
67
68
69
70
71
72
73
74
75
76
77
78
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
79
from vllm.config.utils import get_field
80
from vllm.config.vllm import OptimizationLevel
81
from vllm.logger import init_logger, suppress_logging
82
from vllm.platforms import CpuArchEnum, current_platform
83
from vllm.plugins import load_general_plugins
84
from vllm.ray.lazy_utils import is_in_ray_actor, is_ray_initialized
85
86
87
88
from vllm.transformers_utils.config import (
    is_interleaved,
    maybe_override_with_speculators,
)
89
from vllm.transformers_utils.gguf_utils import is_gguf
90
from vllm.transformers_utils.repo_utils import get_model_path
91
from vllm.transformers_utils.utils import is_cloud_storage
92
from vllm.utils.argparse_utils import FlexibleArgumentParser
93
from vllm.utils.mem_constants import GiB_bytes
94
from vllm.utils.network_utils import get_ip
95
from vllm.v1.sample.logits_processor import LogitsProcessor
96

97
98
if TYPE_CHECKING:
    from vllm.model_executor.layers.quantization import QuantizationMethods
99
    from vllm.model_executor.model_loader import LoadFormats
100
    from vllm.usage.usage_lib import UsageContext
101
    from vllm.v1.executor import Executor
102
else:
103
    Executor = Any
104
    QuantizationMethods = Any
105
    LoadFormats = Any
106
107
    UsageContext = Any

108
109
logger = init_logger(__name__)

110
111
# object is used to allow for special typing forms
T = TypeVar("T")
112
113
TypeHint: TypeAlias = type[Any] | object
TypeHintT: TypeAlias = type[T] | object
114

115

116
117
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
    def _parse_type(val: str) -> T:
118
119
120
121
        try:
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
122
123
                f"Value {val} cannot be converted to {return_type}."
            ) from e
124

125
126
127
    return _parse_type


128
129
def optional_type(return_type: Callable[[str], T]) -> Callable[[str], T | None]:
    def _optional_type(val: str) -> T | None:
130
131
132
133
        if val == "" or val == "None":
            return None
        return parse_type(return_type)(val)

134
    return _optional_type
135
136


137
def union_dict_and_str(val: str) -> str | dict[str, str] | None:
138
    if not re.match(r"(?s)^\s*{.*}\s*$", val):
139
        return str(val)
140
    return optional_type(json.loads)(val)
141
142


143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
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)


158
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
159
160
161
162
    """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`.
    """
163
    type_hint = get_type(type_hints, Literal)
164
165
166
    options = get_args(type_hint)
    option_type = type(options[0])
    if not all(isinstance(option, option_type) for option in options):
167
        raise ValueError(
168
            "All options must be of the same type. "
169
170
            f"Got {options} with types {[type(c) for c in options]}"
        )
171
172
    kwarg = "metavar" if contains_type(type_hints, str) else "choices"
    return {"type": option_type, kwarg: sorted(options)}
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
198
199
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),
    }


200
201
202
203
204
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


205
206
207
208
209
210
211
212
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]))
213
214
    elif origin in {Union, UnionType}:
        # Union for Union[X, Y] and UnionType for X | Y
215
216
217
218
219
220
221
222
        for arg in args:
            type_hints.update(get_type_hints(arg))
    else:
        type_hints.add(type_hint)

    return type_hints


223
224
225
226
def is_online_quantization(quantization: Any) -> bool:
    return quantization in ["inc"]


227
NEEDS_HELP = (
228
229
    any("--help" in arg for arg in sys.argv)  # vllm SUBCOMMAND --help
    or (argv0 := sys.argv[0]).endswith("mkdocs")  # mkdocs SUBCOMMAND
230
231
232
233
    or argv0.endswith("mkdocs/__main__.py")  # python -m mkdocs SUBCOMMAND
)


234
@functools.lru_cache(maxsize=30)
235
def _compute_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
236
237
    # Save time only getting attr docs if we're generating help text
    cls_docs = get_attr_docs(cls) if NEEDS_HELP else {}
238
239
    kwargs = {}
    for field in fields(cls):
240
        # Get the set of possible types for the field
241
        type_hints: set[TypeHint] = get_type_hints(field.type)
242
243
244
245
246

        # 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)

247
        # Get the default value of the field
248
249
        if field.default is not MISSING:
            default = field.default
250
251
            # Handle pydantic.Field defaults
            if isinstance(default, FieldInfo):
252
253
254
255
256
257
258
                if default.default_factory is None:
                    default = default.default
                else:
                    # VllmConfig's Fields have default_factory set to config classes.
                    # These could emit logs on init, which would be confusing.
                    with suppress_logging():
                        default = default.default_factory()
259
        elif field.default_factory is not MISSING:
260
            default = field.default_factory()
261
262
263

        # Get the help text for the field
        name = field.name
264
        help = cls_docs.get(name, "").strip()
265
266
267
268
269
270
271
        # 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
272
273
274
        json_tip = (
            "Should either be a valid JSON string or JSON keys passed individually."
        )
275
        if dataclass_cls is not None:
276
277
278
279
280
281
282
283

            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
284
            kwargs[name]["help"] += f"\n\n{json_tip}"
285
        elif contains_type(type_hints, bool):
286
287
288
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
289
            kwargs[name].update(literal_to_kwargs(type_hints))
290
        elif contains_type(type_hints, tuple):
291
            kwargs[name].update(collection_to_kwargs(type_hints, tuple))
292
        elif contains_type(type_hints, list):
293
294
295
            kwargs[name].update(collection_to_kwargs(type_hints, list))
        elif contains_type(type_hints, set):
            kwargs[name].update(collection_to_kwargs(type_hints, set))
296
297
        elif contains_type(type_hints, int):
            kwargs[name]["type"] = int
298
            # Special case for large integers
299
300
301
302
303
304
            human_readable_ints = {
                "max_model_len",
                "max_num_batched_tokens",
                "kv_cache_memory_bytes",
            }
            if name in human_readable_ints:
305
                kwargs[name]["type"] = human_readable_int
306
                kwargs[name]["help"] += f"\n\n{human_readable_int.__doc__}"
307
308
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
309
310
311
312
        elif contains_type(type_hints, dict) and (
            contains_type(type_hints, str)
            or any(is_not_builtin(th) for th in type_hints)
        ):
313
            kwargs[name]["type"] = union_dict_and_str
314
        elif contains_type(type_hints, dict):
315
            kwargs[name]["type"] = parse_type(json.loads)
316
            kwargs[name]["help"] += f"\n\n{json_tip}"
317
318
319
        elif contains_type(type_hints, str) or any(
            is_not_builtin(th) for th in type_hints
        ):
320
321
            kwargs[name]["type"] = str
        else:
322
            raise ValueError(f"Unsupported type {type_hints} for argument {name}.")
323

324
325
326
327
328
        # 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"]}))

329
330
331
332
333
334
335
        # 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
336
337


338
def get_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
339
340
    """Return argparse kwargs for the given Config dataclass.

341
342
343
    If `--help` or `mkdocs` are not present in the command line command, the
    attribute documentation will not be included in the help output.

344
345
346
347
348
349
350
    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))


351
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
352
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
353
    """Arguments for vLLM engine."""
354

355
    model: str = ModelConfig.model
356
357
358
    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
359
360
    runner: RunnerOption = ModelConfig.runner
    convert: ConvertOption = ModelConfig.convert
361
    task: TaskOption | None = ModelConfig.task
362
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
363
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
364
    tokenizer_mode: TokenizerMode | str = ModelConfig.tokenizer_mode
365
366
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
367
368
    allowed_media_domains: list[str] | None = ModelConfig.allowed_media_domains
    download_dir: str | None = LoadConfig.download_dir
369
    safetensors_load_strategy: str = LoadConfig.safetensors_load_strategy
370
    load_format: str | LoadFormats = LoadConfig.load_format
371
372
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
373
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
374
    seed: int | None = 0
375
    max_model_len: int | None = ModelConfig.max_model_len
376
377
378
379
380
381
382
    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"
    )
383
384
385
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
386
    distributed_executor_backend: (
387
        str | DistributedExecutorBackend | type[Executor] | None
388
    ) = ParallelConfig.distributed_executor_backend
389
    # number of P/D disaggregation (or other disaggregation) workers
390
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
391
392
393
394
    master_addr: str = ParallelConfig.master_addr
    master_port: int = ParallelConfig.master_port
    nnodes: int = ParallelConfig.nnodes
    node_rank: int = ParallelConfig.node_rank
395
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
396
    prefill_context_parallel_size: int = ParallelConfig.prefill_context_parallel_size
397
    decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
398
    dcp_kv_cache_interleave_size: int = ParallelConfig.dcp_kv_cache_interleave_size
399
    cp_kv_cache_interleave_size: int = ParallelConfig.cp_kv_cache_interleave_size
400
    data_parallel_size: int = ParallelConfig.data_parallel_size
401
402
403
404
405
    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
406
    data_parallel_hybrid_lb: bool = False
407
    data_parallel_external_lb: bool = False
Rui Qiao's avatar
Rui Qiao committed
408
    data_parallel_backend: str = ParallelConfig.data_parallel_backend
409
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
410
    all2all_backend: str | None = ParallelConfig.all2all_backend
411
    enable_dbo: bool = ParallelConfig.enable_dbo
412
413
    dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold
    dbo_prefill_token_threshold: int = ParallelConfig.dbo_prefill_token_threshold
414
415
416
    disable_nccl_for_dp_synchronization: bool = (
        ParallelConfig.disable_nccl_for_dp_synchronization
    )
417
    eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config")
418
    enable_eplb: bool = ParallelConfig.enable_eplb
419
    expert_placement_strategy: ExpertPlacementStrategy = (
420
        ParallelConfig.expert_placement_strategy
421
    )
422
423
    _api_process_count: int = ParallelConfig._api_process_count
    _api_process_rank: int = ParallelConfig._api_process_rank
424
    max_parallel_loading_workers: int | None = (
425
426
        ParallelConfig.max_parallel_loading_workers
    )
427
    block_size: BlockSize | None = CacheConfig.block_size
428
    enable_prefix_caching: bool | None = None
429
    prefix_caching_hash_algo: PrefixCachingHashAlgo = (
430
        CacheConfig.prefix_caching_hash_algo
431
    )
432
433
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
434
435
436
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
437
    kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
438
    max_num_batched_tokens: int | None = None
439
440
    max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
    max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
441
    long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold
442
    max_num_seqs: int | None = None
443
    max_logprobs: int = ModelConfig.max_logprobs
444
    logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode
445
    disable_log_stats: bool = False
446
    aggregate_engine_logging: bool = False
447
448
449
    revision: str | None = ModelConfig.revision
    code_revision: str | None = ModelConfig.code_revision
    hf_token: bool | str | None = ModelConfig.hf_token
450
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
451
452
    tokenizer_revision: str | None = ModelConfig.tokenizer_revision
    quantization: QuantizationMethods | None = ModelConfig.quantization
453
    enforce_eager: bool = ModelConfig.enforce_eager
454
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
455
    limit_mm_per_prompt: dict[str, int | dict[str, int]] = get_field(
456
457
        MultiModalConfig, "limit_per_prompt"
    )
458
    enable_mm_embeds: bool = MultiModalConfig.enable_mm_embeds
459
    interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings
460
461
462
    media_io_kwargs: dict[str, dict[str, Any]] = get_field(
        MultiModalConfig, "media_io_kwargs"
    )
463
    mm_processor_kwargs: dict[str, Any] | None = MultiModalConfig.mm_processor_kwargs
464
    disable_mm_preprocessor_cache: bool = False  # DEPRECATED
465
    mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb
466
    mm_processor_cache_type: MMCacheType | None = (
467
        MultiModalConfig.mm_processor_cache_type
468
469
    )
    mm_shm_cache_max_object_size_mb: int = (
470
        MultiModalConfig.mm_shm_cache_max_object_size_mb
471
    )
472
    mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
473
    mm_encoder_attn_backend: AttentionBackendEnum | str | None = (
474
475
        MultiModalConfig.mm_encoder_attn_backend
    )
476
    io_processor_plugin: str | None = None
477
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
478
    video_pruning_rate: float = MultiModalConfig.video_pruning_rate
479
    # LoRA fields
480
    enable_lora: bool = False
481
482
    max_loras: int = LoRAConfig.max_loras
    max_lora_rank: int = LoRAConfig.max_lora_rank
483
    default_mm_loras: dict[str, str] | None = LoRAConfig.default_mm_loras
484
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
485
486
    max_cpu_loras: int | None = LoRAConfig.max_cpu_loras
    lora_dtype: str | torch.dtype | None = LoRAConfig.lora_dtype
487

488
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
489
    num_gpu_blocks_override: int | None = CacheConfig.num_gpu_blocks_override
490
    model_loader_extra_config: dict = get_field(LoadConfig, "model_loader_extra_config")
491
    ignore_patterns: str | list[str] = get_field(LoadConfig, "ignore_patterns")
492

493
    enable_chunked_prefill: bool | None = None
494
    disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
495

496
    disable_hybrid_kv_cache_manager: bool = (
497
498
        SchedulerConfig.disable_hybrid_kv_cache_manager
    )
499

500
    structured_outputs_config: StructuredOutputsConfig = get_field(
501
502
        VllmConfig, "structured_outputs_config"
    )
503
    reasoning_parser: str = StructuredOutputsConfig.reasoning_parser
504
    reasoning_parser_plugin: str | None = None
505

506
    logits_processor_pattern: str | None = ModelConfig.logits_processor_pattern
507

508
    speculative_config: dict[str, Any] | None = None
509

510
    show_hidden_metrics_for_version: str | None = (
511
        ObservabilityConfig.show_hidden_metrics_for_version
512
    )
513
514
    otlp_traces_endpoint: str | None = ObservabilityConfig.otlp_traces_endpoint
    collect_detailed_traces: list[DetailedTraceModules] | None = (
515
        ObservabilityConfig.collect_detailed_traces
516
    )
517
518
519
520
    kv_cache_metrics: bool = ObservabilityConfig.kv_cache_metrics
    kv_cache_metrics_sample: float = get_field(
        ObservabilityConfig, "kv_cache_metrics_sample"
    )
521
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
522
    scheduler_cls: str | type[object] | None = SchedulerConfig.scheduler_cls
523

524
    pooler_config: PoolerConfig | None = ModelConfig.pooler_config
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
    ec_transfer_config: ECTransferConfig | None = None

534
535
    generation_config: str = ModelConfig.generation_config
    enable_sleep_mode: bool = ModelConfig.enable_sleep_mode
536
537
538
    override_generation_config: dict[str, Any] = get_field(
        ModelConfig, "override_generation_config"
    )
539
    model_impl: str = ModelConfig.model_impl
540
    override_attention_dtype: str = ModelConfig.override_attention_dtype
541

542
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
543
544
    mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype
    mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype
545
    mamba_block_size: int | None = get_field(CacheConfig, "mamba_block_size")
546

547
    additional_config: dict[str, Any] = get_field(VllmConfig, "additional_config")
548

549
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
550
    pt_load_map_location: str = LoadConfig.pt_load_map_location
551

552
553
    # DEPRECATED
    enable_multimodal_encoder_data_parallel: bool = False
554

555
    logits_processors: list[str | type[LogitsProcessor]] | None = (
556
557
        ModelConfig.logits_processors
    )
558
559
    """Custom logitproc types"""

560
    async_scheduling: bool | None = SchedulerConfig.async_scheduling
561

562
563
    stream_interval: int = SchedulerConfig.stream_interval

564
    kv_sharing_fast_prefill: bool = CacheConfig.kv_sharing_fast_prefill
565
    optimization_level: OptimizationLevel = VllmConfig.optimization_level
566

567
568
569
570
    kv_offloading_size: float | None = CacheConfig.kv_offloading_size
    kv_offloading_backend: KVOffloadingBackend | None = (
        CacheConfig.kv_offloading_backend
    )
571
    tokens_only: bool = False
572

573
    def __post_init__(self):
574
575
576
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
577
        if isinstance(self.compilation_config, dict):
578
            self.compilation_config = CompilationConfig(**self.compilation_config)
579
        if isinstance(self.eplb_config, dict):
580
            self.eplb_config = EPLBConfig(**self.eplb_config)
581
        # Setup plugins
582
        from vllm.plugins import load_general_plugins
583

584
        load_general_plugins()
585
        # when use hf offline,replace model and tokenizer id to local model path
586
587
588
        if huggingface_hub.constants.HF_HUB_OFFLINE:
            model_id = self.model
            self.model = get_model_path(self.model, self.revision)
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
            if model_id is not self.model:
                logger.info(
                    "HF_HUB_OFFLINE is True, replace model_id [%s] to model_path [%s]",
                    model_id,
                    self.model,
                )
            if self.tokenizer is not None:
                tokenizer_id = self.tokenizer
                self.tokenizer = get_model_path(self.tokenizer, self.tokenizer_revision)
                if tokenizer_id is not self.tokenizer:
                    logger.info(
                        "HF_HUB_OFFLINE is True, replace tokenizer_id [%s] "
                        "to tokenizer_path [%s]",
                        tokenizer_id,
                        self.tokenizer,
                    )
605
606

    @staticmethod
607
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
608
        """Shared CLI arguments for vLLM engine."""
609

610
        # Model arguments
611
612
613
614
615
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
616
        if not ("serve" in sys.argv[1:] and "--help" in sys.argv[1:]):
617
            model_group.add_argument("--model", **model_kwargs["model"])
618
619
        model_group.add_argument("--runner", **model_kwargs["runner"])
        model_group.add_argument("--convert", **model_kwargs["convert"])
620
        model_group.add_argument("--task", **model_kwargs["task"], deprecated=True)
621
        model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"])
622
623
624
625
        model_group.add_argument("--tokenizer-mode", **model_kwargs["tokenizer_mode"])
        model_group.add_argument(
            "--trust-remote-code", **model_kwargs["trust_remote_code"]
        )
626
627
        model_group.add_argument("--dtype", **model_kwargs["dtype"])
        model_group.add_argument("--seed", **model_kwargs["seed"])
628
629
630
631
632
633
634
        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"]
        )
635
        model_group.add_argument("--revision", **model_kwargs["revision"])
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
        model_group.add_argument("--code-revision", **model_kwargs["code_revision"])
        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"])
661
662
        # This one is a special case because it can bool
        # or str. TODO: Handle this in get_kwargs
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
        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(
            "--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"]
        )
685
        model_group.add_argument("--model-impl", **model_kwargs["model_impl"])
686
687
688
689
690
691
692
693
694
        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"]
        )
695

696
697
698
699
700
701
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
702
        load_group.add_argument("--load-format", **load_kwargs["load_format"])
703
704
705
706
707
708
709
710
711
712
713
714
        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"]
        )
715

716
717
718
719
720
        # Structured outputs arguments
        structured_outputs_kwargs = get_kwargs(StructuredOutputsConfig)
        structured_outputs_group = parser.add_argument_group(
            title="StructuredOutputsConfig",
            description=StructuredOutputsConfig.__doc__,
721
        )
722
        structured_outputs_group.add_argument(
723
            "--reasoning-parser",
724
            # Choices need to be validated after parsing to include plugins
725
726
            **structured_outputs_kwargs["reasoning_parser"],
        )
727
728
729
730
        structured_outputs_group.add_argument(
            "--reasoning-parser-plugin",
            **structured_outputs_kwargs["reasoning_parser_plugin"],
        )
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
748
749
750
        parallel_group.add_argument("--master-addr", **parallel_kwargs["master_addr"])
        parallel_group.add_argument("--master-port", **parallel_kwargs["master_port"])
        parallel_group.add_argument("--nnodes", "-n", **parallel_kwargs["nnodes"])
        parallel_group.add_argument("--node-rank", "-r", **parallel_kwargs["node_rank"])
751
        parallel_group.add_argument(
752
753
            "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
        )
754
        parallel_group.add_argument(
755
756
757
758
            "--decode-context-parallel-size",
            "-dcp",
            **parallel_kwargs["decode_context_parallel_size"],
        )
759
760
761
762
        parallel_group.add_argument(
            "--dcp-kv-cache-interleave-size",
            **parallel_kwargs["dcp_kv_cache_interleave_size"],
        )
763
764
765
766
767
768
769
770
771
        parallel_group.add_argument(
            "--cp-kv-cache-interleave-size",
            **parallel_kwargs["cp_kv_cache_interleave_size"],
        )
        parallel_group.add_argument(
            "--prefill-context-parallel-size",
            "-pcp",
            **parallel_kwargs["prefill_context_parallel_size"],
        )
772
773
774
775
776
777
        parallel_group.add_argument(
            "--data-parallel-size", "-dp", **parallel_kwargs["data_parallel_size"]
        )
        parallel_group.add_argument(
            "--data-parallel-rank",
            "-dpn",
778
            type=int,
779
780
781
            help="Data parallel rank of this instance. "
            "When set, enables external load balancer mode.",
        )
782
        parallel_group.add_argument(
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
            "--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".',
        )
813
        parallel_group.add_argument(
814
815
816
817
818
819
820
821
            "--data-parallel-hybrid-lb",
            "-dph",
            **parallel_kwargs["data_parallel_hybrid_lb"],
        )
        parallel_group.add_argument(
            "--data-parallel-external-lb",
            "-dpe",
            **parallel_kwargs["data_parallel_external_lb"],
822
823
824
825
        )
        parallel_group.add_argument(
            "--enable-expert-parallel", **parallel_kwargs["enable_expert_parallel"]
        )
826
827
828
        parallel_group.add_argument(
            "--all2all-backend", **parallel_kwargs["all2all_backend"]
        )
829
        parallel_group.add_argument("--enable-dbo", **parallel_kwargs["enable_dbo"])
830
831
        parallel_group.add_argument(
            "--dbo-decode-token-threshold",
832
833
            **parallel_kwargs["dbo_decode_token_threshold"],
        )
834
835
        parallel_group.add_argument(
            "--dbo-prefill-token-threshold",
836
837
            **parallel_kwargs["dbo_prefill_token_threshold"],
        )
838
839
840
841
        parallel_group.add_argument(
            "--disable-nccl-for-dp-synchronization",
            **parallel_kwargs["disable_nccl_for_dp_synchronization"],
        )
842
843
        parallel_group.add_argument("--enable-eplb", **parallel_kwargs["enable_eplb"])
        parallel_group.add_argument("--eplb-config", **parallel_kwargs["eplb_config"])
844
845
        parallel_group.add_argument(
            "--expert-placement-strategy",
846
847
            **parallel_kwargs["expert_placement_strategy"],
        )
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
        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(
889
890
891
892
893
            "--enable-prefix-caching",
            **{
                **cache_kwargs["enable_prefix_caching"],
                "default": None,
            },
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
        )
        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"]
        )
911
912
913
        cache_group.add_argument(
            "--mamba-block-size", **cache_kwargs["mamba_block_size"]
        )
914
915
916
917
918
919
        cache_group.add_argument(
            "--kv-offloading-size", **cache_kwargs["kv_offloading_size"]
        )
        cache_group.add_argument(
            "--kv-offloading-backend", **cache_kwargs["kv_offloading_backend"]
        )
920

921
        # Multimodal related configs
922
923
924
925
926
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
927
        multimodal_group.add_argument(
928
929
            "--limit-mm-per-prompt", **multimodal_kwargs["limit_per_prompt"]
        )
930
931
932
        multimodal_group.add_argument(
            "--enable-mm-embeds", **multimodal_kwargs["enable_mm_embeds"]
        )
933
934
935
936
937
938
939
940
941
        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"]
        )
942
        multimodal_group.add_argument(
943
944
            "--disable-mm-preprocessor-cache", action="store_true", deprecated=True
        )
945
        multimodal_group.add_argument(
946
947
            "--mm-processor-cache-type", **multimodal_kwargs["mm_processor_cache_type"]
        )
948
949
        multimodal_group.add_argument(
            "--mm-shm-cache-max-object-size-mb",
950
951
            **multimodal_kwargs["mm_shm_cache_max_object_size_mb"],
        )
952
        multimodal_group.add_argument(
953
954
            "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"]
        )
955
956
957
958
        multimodal_group.add_argument(
            "--mm-encoder-attn-backend",
            **multimodal_kwargs["mm_encoder_attn_backend"],
        )
959
960
961
        multimodal_group.add_argument(
            "--interleave-mm-strings", **multimodal_kwargs["interleave_mm_strings"]
        )
962
        multimodal_group.add_argument(
963
964
            "--skip-mm-profiling", **multimodal_kwargs["skip_mm_profiling"]
        )
965

966
        multimodal_group.add_argument(
967
968
            "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"]
        )
969

970
        # LoRA related configs
971
972
973
974
975
976
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
977
            "--enable-lora",
978
            action=argparse.BooleanOptionalAction,
979
980
            help="If True, enable handling of LoRA adapters.",
        )
981
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
982
        lora_group.add_argument("--max-lora-rank", **lora_kwargs["max_lora_rank"])
983
        lora_group.add_argument(
984
            "--lora-dtype",
985
986
            **lora_kwargs["lora_dtype"],
        )
987
988
989
990
991
        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"])
992

993
994
995
996
997
998
999
1000
        # 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",
1001
1002
            **observability_kwargs["show_hidden_metrics_for_version"],
        )
1003
        observability_group.add_argument(
1004
1005
            "--otlp-traces-endpoint", **observability_kwargs["otlp_traces_endpoint"]
        )
1006
1007
1008
1009
1010
        # 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"] += [
1011
            ",".join(p) for p in permutations(get_args(DetailedTraceModules), r=2)
1012
1013
1014
        ]
        observability_group.add_argument(
            "--collect-detailed-traces",
1015
1016
            **observability_kwargs["collect_detailed_traces"],
        )
1017
1018
1019
1020
1021
1022
1023
        observability_group.add_argument(
            "--kv-cache-metrics", **observability_kwargs["kv_cache_metrics"]
        )
        observability_group.add_argument(
            "--kv-cache-metrics-sample",
            **observability_kwargs["kv_cache_metrics_sample"],
        )
1024

1025
1026
1027
1028
1029
1030
1031
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
1032
1033
1034
1035
1036
            "--max-num-batched-tokens",
            **{
                **scheduler_kwargs["max_num_batched_tokens"],
                "default": None,
            },
1037
        )
1038
        scheduler_group.add_argument(
1039
1040
1041
1042
1043
            "--max-num-seqs",
            **{
                **scheduler_kwargs["max_num_seqs"],
                "default": None,
            },
1044
1045
1046
1047
        )
        scheduler_group.add_argument(
            "--max-num-partial-prefills", **scheduler_kwargs["max_num_partial_prefills"]
        )
1048
1049
        scheduler_group.add_argument(
            "--max-long-partial-prefills",
1050
1051
            **scheduler_kwargs["max_long_partial_prefills"],
        )
1052
1053
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
1054
1055
            **scheduler_kwargs["long_prefill_token_threshold"],
        )
1056
1057
        # multi-step scheduling has been removed; corresponding arguments
        # are no longer supported.
1058
        scheduler_group.add_argument(
1059
1060
            "--scheduling-policy", **scheduler_kwargs["policy"]
        )
1061
        scheduler_group.add_argument(
1062
1063
1064
1065
1066
            "--enable-chunked-prefill",
            **{
                **scheduler_kwargs["enable_chunked_prefill"],
                "default": None,
            },
1067
1068
1069
1070
1071
1072
1073
        )
        scheduler_group.add_argument(
            "--disable-chunked-mm-input", **scheduler_kwargs["disable_chunked_mm_input"]
        )
        scheduler_group.add_argument(
            "--scheduler-cls", **scheduler_kwargs["scheduler_cls"]
        )
1074
1075
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
1076
1077
1078
1079
1080
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"],
        )
        scheduler_group.add_argument(
            "--async-scheduling", **scheduler_kwargs["async_scheduling"]
        )
1081
1082
1083
        scheduler_group.add_argument(
            "--stream-interval", **scheduler_kwargs["stream_interval"]
        )
1084

1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
        # 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"],
        )

1108
        # vLLM arguments
1109
        vllm_kwargs = get_kwargs(VllmConfig)
1110
1111
1112
1113
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
1114
1115
1116
1117
        # 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)
1118
1119
1120
1121
1122
1123
1124
        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"])
1125
1126
1127
        vllm_group.add_argument(
            "--ec-transfer-config", **vllm_kwargs["ec_transfer_config"]
        )
1128
        vllm_group.add_argument(
1129
            "--compilation-config", "-cc", **vllm_kwargs["compilation_config"]
1130
1131
1132
1133
1134
1135
1136
        )
        vllm_group.add_argument(
            "--additional-config", **vllm_kwargs["additional_config"]
        )
        vllm_group.add_argument(
            "--structured-outputs-config", **vllm_kwargs["structured_outputs_config"]
        )
1137

1138
1139
1140
1141
        vllm_group.add_argument(
            "--optimization-level", **vllm_kwargs["optimization_level"]
        )

1142
        # Other arguments
1143
1144
1145
1146
1147
        parser.add_argument(
            "--disable-log-stats",
            action="store_true",
            help="Disable logging statistics.",
        )
1148

1149
1150
1151
1152
1153
1154
        parser.add_argument(
            "--aggregate-engine-logging",
            action="store_true",
            help="Log aggregate rather than per-engine statistics "
            "when using data parallelism.",
        )
1155
        return parser
1156
1157

    @classmethod
1158
    def from_cli_args(cls, args: argparse.Namespace):
1159
1160
1161
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
1162
1163
1164
        engine_args = cls(
            **{attr: getattr(args, attr) for attr in attrs if hasattr(args, attr)}
        )
Zhuohan Li's avatar
Zhuohan Li committed
1165
        return engine_args
1166

1167
    def create_model_config(self) -> ModelConfig:
1168
1169
        # gguf file needs a specific model loader
        if is_gguf(self.model):
1170
1171
            self.quantization = self.load_format = "gguf"

1172
1173
1174
1175
        # NOTE(woosuk): In V1, we use separate processes for workers (unless
        # VLLM_ENABLE_V1_MULTIPROCESSING=0), so setting a seed here
        # doesn't affect the user process.
        if self.seed is None:
1176
1177
1178
1179
1180
1181
            logger.warning_once(
                "`seed=None` is equivalent to `seed=0` in V1 Engine. "
                "You will no longer be allowed to pass `None` in v0.13.",
                scope="local",
            )

1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
            self.seed = 0
            if not envs.VLLM_ENABLE_V1_MULTIPROCESSING:
                logger.warning(
                    "The global random seed is set to %d. Since "
                    "VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may "
                    "affect the random state of the Python process that "
                    "launched vLLM.",
                    self.seed,
                )

1192
        if self.disable_mm_preprocessor_cache:
1193
            logger.warning_once(
1194
1195
                "`--disable-mm-preprocessor-cache` is deprecated "
                "and will be removed in v0.13. "
1196
                "Please use `--mm-processor-cache-gb 0` instead.",
1197
                scope="local",
1198
            )
1199
1200
1201

            self.mm_processor_cache_gb = 0
        elif envs.VLLM_MM_INPUT_CACHE_GIB != 4:
1202
            logger.warning_once(
1203
1204
1205
1206
                "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,
1207
                scope="local",
1208
1209
1210
1211
            )

            self.mm_processor_cache_gb = envs.VLLM_MM_INPUT_CACHE_GIB

1212
        if self.enable_multimodal_encoder_data_parallel:
1213
            logger.warning_once(
1214
1215
                "--enable-multimodal-encoder-data-parallel` is deprecated "
                "and will be removed in v0.13. "
1216
1217
                "Please use `--mm-encoder-tp-mode data` instead.",
                scope="local",
1218
            )
1219
1220
1221

            self.mm_encoder_tp_mode = "data"

1222
        return ModelConfig(
1223
            model=self.model,
1224
            hf_config_path=self.hf_config_path,
1225
1226
            runner=self.runner,
            convert=self.convert,
1227
            task=self.task,
1228
            tokenizer=self.tokenizer,
1229
1230
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1231
            allowed_local_media_path=self.allowed_local_media_path,
1232
            allowed_media_domains=self.allowed_media_domains,
1233
1234
1235
1236
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
1237
            hf_token=self.hf_token,
1238
            hf_overrides=self.hf_overrides,
1239
1240
1241
1242
1243
            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,
1244
            logprobs_mode=self.logprobs_mode,
1245
            disable_sliding_window=self.disable_sliding_window,
1246
            disable_cascade_attn=self.disable_cascade_attn,
1247
            skip_tokenizer_init=self.skip_tokenizer_init,
1248
            enable_prompt_embeds=self.enable_prompt_embeds,
1249
            served_model_name=self.served_model_name,
1250
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1251
            enable_mm_embeds=self.enable_mm_embeds,
1252
            interleave_mm_strings=self.interleave_mm_strings,
1253
            media_io_kwargs=self.media_io_kwargs,
1254
            skip_mm_profiling=self.skip_mm_profiling,
1255
            config_format=self.config_format,
1256
            mm_processor_kwargs=self.mm_processor_kwargs,
1257
            mm_processor_cache_gb=self.mm_processor_cache_gb,
1258
            mm_processor_cache_type=self.mm_processor_cache_type,
1259
            mm_shm_cache_max_object_size_mb=self.mm_shm_cache_max_object_size_mb,
1260
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
1261
            mm_encoder_attn_backend=self.mm_encoder_attn_backend,
1262
            pooler_config=self.pooler_config,
1263
            logits_processor_pattern=self.logits_processor_pattern,
1264
            generation_config=self.generation_config,
1265
            override_generation_config=self.override_generation_config,
1266
            enable_sleep_mode=self.enable_sleep_mode,
1267
            model_impl=self.model_impl,
1268
            override_attention_dtype=self.override_attention_dtype,
1269
            logits_processors=self.logits_processors,
1270
            video_pruning_rate=self.video_pruning_rate,
1271
            io_processor_plugin=self.io_processor_plugin,
1272
        )
1273

1274
    def validate_tensorizer_args(self):
1275
1276
        from vllm.model_executor.model_loader.tensorizer import TensorizerConfig

1277
1278
        for key in self.model_loader_extra_config:
            if key in TensorizerConfig._fields:
1279
1280
1281
                self.model_loader_extra_config["tensorizer_config"][key] = (
                    self.model_loader_extra_config[key]
                )
1282

1283
    def create_load_config(self) -> LoadConfig:
1284
1285
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1286

1287
1288
1289
        if self.load_format == "tensorizer":
            if hasattr(self.model_loader_extra_config, "to_serializable"):
                self.model_loader_extra_config = (
1290
1291
                    self.model_loader_extra_config.to_serializable()
                )
1292
            self.model_loader_extra_config["tensorizer_config"] = {}
1293
1294
1295
            self.model_loader_extra_config["tensorizer_config"]["tensorizer_dir"] = (
                self.model
            )
1296
            self.validate_tensorizer_args()
1297

1298
1299
1300
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
1301
            safetensors_load_strategy=self.safetensors_load_strategy,
1302
            device="cpu" if is_online_quantization(self.quantization) else None,
1303
1304
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1305
            use_tqdm_on_load=self.use_tqdm_on_load,
1306
            pt_load_map_location=self.pt_load_map_location,
1307
        )
1308

1309
1310
1311
1312
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
1313
    ) -> SpeculativeConfig | None:
1314
1315
1316
1317
1318
1319
        """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
1320
        dictionary from the engine.
1321
1322
        """
        if self.speculative_config is None:
1323
            return None
1324

1325
1326
1327
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.
1328
1329
1330
1331
1332
1333
        self.speculative_config.update(
            {
                "target_model_config": target_model_config,
                "target_parallel_config": target_parallel_config,
            }
        )
1334
        return SpeculativeConfig(**self.speculative_config)
1335

1336
1337
    def create_engine_config(
        self,
1338
        usage_context: UsageContext | None = None,
1339
        headless: bool = False,
1340
1341
1342
1343
    ) -> VllmConfig:
        """
        Create the VllmConfig.

1344
        NOTE: If VllmConfig is incompatible, we raise an error.
1345
        """
1346
        current_platform.pre_register_and_update()
1347

1348
        device_config = DeviceConfig(device=cast(Device, current_platform.device_type))
1349

1350
1351
        # Check if the model is a speculator and override model/tokenizer/config
        # BEFORE creating ModelConfig, so the config is created with the target model
1352
1353
1354
1355
        # 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):
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
            (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,
                )
            )

1366
1367
1368
1369
        model_config = self.create_model_config()
        self.model = model_config.model
        self.tokenizer = model_config.tokenizer

1370
        self._check_feature_supported(model_config)
1371
1372
1373
1374
        self._set_default_chunked_prefill_and_prefix_caching_args(model_config)
        self._set_default_max_num_seqs_and_batched_tokens_args(
            usage_context, model_config
        )
1375

1376
        sliding_window: int | None = None
1377
1378
1379
1380
1381
1382
        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()

1383
1384
1385
        # 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
1386
        # reuses the GPUs of TP group, and split one TP group into
1387
        # tp_size//dcp_size DCP groups.
1388
        assert self.tensor_parallel_size % self.decode_context_parallel_size == 0, (
1389
1390
1391
1392
            f"tp_size={self.tensor_parallel_size} must be divisible by"
            f"dcp_size={self.decode_context_parallel_size}."
        )

1393
        cache_config = CacheConfig(
1394
            block_size=self.block_size,
1395
            gpu_memory_utilization=self.gpu_memory_utilization,
1396
            kv_cache_memory_bytes=self.kv_cache_memory_bytes,
1397
1398
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1399
            is_attention_free=model_config.is_attention_free,
1400
            num_gpu_blocks_override=self.num_gpu_blocks_override,
1401
            sliding_window=sliding_window,
1402
            enable_prefix_caching=self.enable_prefix_caching,
1403
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1404
            cpu_offload_gb=self.cpu_offload_gb,
1405
            calculate_kv_scales=self.calculate_kv_scales,
1406
            kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
1407
1408
            mamba_cache_dtype=self.mamba_cache_dtype,
            mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype,
1409
            mamba_block_size=self.mamba_block_size,
1410
1411
            kv_offloading_size=self.kv_offloading_size,
            kv_offloading_backend=self.kv_offloading_backend,
1412
        )
1413

1414
1415
1416
1417
1418
1419
        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
1420

1421
            ray_runtime_env = ray.get_runtime_context().runtime_env
1422
1423
1424
1425
1426
1427
1428
            # 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)
1429

1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
        # 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()

1441
        assert not headless or not self.data_parallel_hybrid_lb, (
1442
1443
            "data_parallel_hybrid_lb is not applicable in headless mode"
        )
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
        assert not (self.data_parallel_hybrid_lb and self.data_parallel_external_lb), (
            "data_parallel_hybrid_lb and data_parallel_external_lb cannot both be True."
        )
        assert self.data_parallel_backend == "mp" or self.nnodes == 1, (
            "nnodes > 1 is only supported with data_parallel_backend=mp"
        )
        inferred_data_parallel_rank = 0
        if self.nnodes > 1:
            world_size = (
                self.data_parallel_size
                * self.pipeline_parallel_size
                * self.tensor_parallel_size
            )
            world_size_within_dp = (
                self.pipeline_parallel_size * self.tensor_parallel_size
            )
            local_world_size = world_size // self.nnodes
            assert world_size % self.nnodes == 0, (
                f"world_size={world_size} must be divisible by nnodes={self.nnodes}."
            )
            assert self.node_rank < self.nnodes, (
                f"node_rank={self.node_rank} must be less than nnodes={self.nnodes}."
            )
            inferred_data_parallel_rank = (
                self.node_rank * local_world_size
            ) // world_size_within_dp
            if self.data_parallel_size > 1 and self.data_parallel_external_lb:
                self.data_parallel_rank = inferred_data_parallel_rank
                logger.info(
                    "Inferred data_parallel_rank %d from node_rank %d for external lb",
                    self.data_parallel_rank,
                    self.node_rank,
                )
            elif self.data_parallel_size_local is None:
                # Infer data parallel size local for internal dplb:
                self.data_parallel_size_local = max(
                    local_world_size // world_size_within_dp, 1
                )
        data_parallel_external_lb = (
            self.data_parallel_external_lb or self.data_parallel_rank is not None
        )
1485
        # Local DP rank = 1, use pure-external LB.
1486
        if data_parallel_external_lb:
1487
            assert self.data_parallel_rank is not None, (
1488
                "data_parallel_rank or node_rank must be specified if "
1489
1490
                "data_parallel_external_lb is enable."
            )
1491
            assert self.data_parallel_size_local in (1, None), (
1492
1493
                "data_parallel_size_local must be 1 or None when data_parallel_rank "
                "is set"
1494
            )
1495
            data_parallel_size_local = 1
1496
1497
            # Use full external lb if we have local_size of 1.
            self.data_parallel_hybrid_lb = False
1498
1499
        elif self.data_parallel_size_local is not None:
            data_parallel_size_local = self.data_parallel_size_local
1500
1501
1502
1503
1504
1505
1506

            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.
1507
1508
1509
1510
1511
                logger.warning(
                    "data_parallel_hybrid_lb is not eligible when "
                    "data_parallel_size_local = 1, autoswitch to "
                    "data_parallel_external_lb."
                )
1512
1513
1514
1515
1516
1517
1518
                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

1519
1520
1521
1522
1523
1524
1525
1526
1527
            self.data_parallel_rank = (
                self.data_parallel_start_rank or inferred_data_parallel_rank
            )
            if self.nnodes > 1:
                logger.info(
                    "Inferred data_parallel_rank %d from node_rank %d",
                    self.data_parallel_rank,
                    self.node_rank,
                )
1528
        else:
1529
            assert not self.data_parallel_hybrid_lb, (
1530
1531
                "data_parallel_size_local must be set to use data_parallel_hybrid_lb."
            )
1532

1533
1534
1535
1536
1537
1538
1539
1540
1541
            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
1542
1543
1544

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1545
1546
1547
1548
        if self.data_parallel_address is None:
            if self.data_parallel_backend == "ray":
                host_ip = get_ip()
                logger.info(
1549
1550
                    "Using host IP %s as ray-based data parallel address", host_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1551
1552
1553
1554
                data_parallel_address = host_ip
            else:
                assert self.data_parallel_backend == "mp", (
                    "data_parallel_backend can only be ray or mp, got %s",
1555
1556
                    self.data_parallel_backend,
                )
1557
1558
1559
                data_parallel_address = (
                    self.master_addr or ParallelConfig.data_parallel_master_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1560
1561
        else:
            data_parallel_address = self.data_parallel_address
1562
1563
1564

        # This port is only used when there are remote data parallel engines,
        # otherwise the local IPC transport is used.
1565
        data_parallel_rpc_port = (
1566
            self.data_parallel_rpc_port
1567
1568
1569
            if (self.data_parallel_rpc_port is not None)
            else ParallelConfig.data_parallel_rpc_port
        )
1570

1571
1572
1573
1574
        if self.tokens_only and not model_config.skip_tokenizer_init:
            model_config.skip_tokenizer_init = True
            logger.info("Skipping tokenizer initialization for tokens-only mode.")

1575
1576
1577
1578
1579
1580
        if self.async_scheduling and not self.disable_nccl_for_dp_synchronization:
            logger.info(
                "Disabling NCCL for DP synchronization when using async scheduling."
            )
            self.disable_nccl_for_dp_synchronization = True

1581
        parallel_config = ParallelConfig(
1582
1583
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1584
            prefill_context_parallel_size=self.prefill_context_parallel_size,
1585
            data_parallel_size=self.data_parallel_size,
1586
1587
            data_parallel_rank=self.data_parallel_rank or 0,
            data_parallel_external_lb=data_parallel_external_lb,
1588
            data_parallel_size_local=data_parallel_size_local,
1589
1590
1591
1592
            master_addr=self.master_addr,
            master_port=self.master_port,
            nnodes=self.nnodes,
            node_rank=self.node_rank,
1593
1594
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
1595
            data_parallel_backend=self.data_parallel_backend,
1596
            data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
1597
            enable_expert_parallel=self.enable_expert_parallel,
1598
            all2all_backend=self.all2all_backend,
1599
1600
            enable_dbo=self.enable_dbo,
            dbo_decode_token_threshold=self.dbo_decode_token_threshold,
1601
            dbo_prefill_token_threshold=self.dbo_prefill_token_threshold,
1602
            disable_nccl_for_dp_synchronization=self.disable_nccl_for_dp_synchronization,
1603
            enable_eplb=self.enable_eplb,
1604
            eplb_config=self.eplb_config,
1605
            expert_placement_strategy=self.expert_placement_strategy,
1606
1607
1608
            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,
1609
            ray_runtime_env=ray_runtime_env,
1610
            placement_group=placement_group,
1611
1612
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1613
            worker_extension_cls=self.worker_extension_cls,
1614
            decode_context_parallel_size=self.decode_context_parallel_size,
1615
            dcp_kv_cache_interleave_size=self.dcp_kv_cache_interleave_size,
1616
            cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size,
1617
1618
            _api_process_count=self._api_process_count,
            _api_process_rank=self._api_process_rank,
1619
        )
1620

1621
        speculative_config = self.create_speculative_config(
1622
1623
1624
1625
            target_model_config=model_config,
            target_parallel_config=parallel_config,
        )

1626
        scheduler_config = SchedulerConfig(
1627
            runner_type=model_config.runner_type,
1628
1629
1630
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1631
            enable_chunked_prefill=self.enable_chunked_prefill,
1632
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1633
            is_multimodal_model=model_config.is_multimodal_model,
1634
            is_encoder_decoder=model_config.is_encoder_decoder,
1635
            policy=self.scheduling_policy,
1636
            scheduler_cls=self.scheduler_cls,
1637
1638
1639
            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,
1640
            disable_hybrid_kv_cache_manager=self.disable_hybrid_kv_cache_manager,
1641
            async_scheduling=self.async_scheduling,
1642
            stream_interval=self.stream_interval,
1643
        )
1644

1645
1646
1647
        if not model_config.is_multimodal_model and self.default_mm_loras:
            raise ValueError(
                "Default modality-specific LoRA(s) were provided for a "
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
                "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_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
        )
1665

1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
        if (
            lora_config is not None
            and speculative_config is not None
            and scheduler_config.max_num_batched_tokens
            < (
                scheduler_config.max_num_seqs
                * (speculative_config.num_speculative_tokens + 1)
            )
        ):
            raise ValueError(
                "Consider increasing max_num_batched_tokens or "
                "decreasing num_speculative_tokens"
            )

1680
1681
1682
1683
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1684
        load_config = self.create_load_config()
1685

1686
1687
        # Pass reasoning_parser into StructuredOutputsConfig
        if self.reasoning_parser:
1688
            self.structured_outputs_config.reasoning_parser = self.reasoning_parser
1689

1690
1691
1692
1693
1694
        if self.reasoning_parser_plugin:
            self.structured_outputs_config.reasoning_parser_plugin = (
                self.reasoning_parser_plugin
            )

1695
        observability_config = ObservabilityConfig(
1696
            show_hidden_metrics_for_version=self.show_hidden_metrics_for_version,
1697
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1698
            collect_detailed_traces=self.collect_detailed_traces,
1699
1700
            kv_cache_metrics=self.kv_cache_metrics,
            kv_cache_metrics_sample=self.kv_cache_metrics_sample,
1701
        )
1702

1703
        # Compilation config overrides
1704
        compilation_config = copy.deepcopy(self.compilation_config)
1705
1706
1707
1708
1709
1710
        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."
            )
1711
            if compilation_config.cudagraph_capture_sizes is not None:
1712
1713
1714
1715
                raise ValueError(
                    "cuda_graph_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
1716
            compilation_config.cudagraph_capture_sizes = self.cuda_graph_sizes
1717
        if self.cudagraph_capture_sizes is not None:
1718
            if compilation_config.cudagraph_capture_sizes is not None:
1719
1720
1721
1722
                raise ValueError(
                    "cudagraph_capture_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
1723
            compilation_config.cudagraph_capture_sizes = self.cudagraph_capture_sizes
1724
        if self.max_cudagraph_capture_size is not None:
1725
            if compilation_config.max_cudagraph_capture_size is not None:
1726
1727
1728
1729
                raise ValueError(
                    "max_cudagraph_capture_size and compilation_config."
                    "max_cudagraph_capture_size are mutually exclusive"
                )
1730
            compilation_config.max_cudagraph_capture_size = (
1731
1732
                self.max_cudagraph_capture_size
            )
1733
        config = VllmConfig(
1734
1735
1736
1737
1738
1739
1740
1741
            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,
1742
            structured_outputs_config=self.structured_outputs_config,
1743
            observability_config=observability_config,
1744
            compilation_config=compilation_config,
1745
            kv_transfer_config=self.kv_transfer_config,
1746
            kv_events_config=self.kv_events_config,
1747
            ec_transfer_config=self.ec_transfer_config,
1748
            additional_config=self.additional_config,
1749
            optimization_level=self.optimization_level,
1750
        )
1751

1752
1753
        return config

1754
1755
    def _check_feature_supported(self, model_config: ModelConfig):
        """Raise an error if the feature is not supported."""
1756
        if self.logits_processor_pattern != EngineArgs.logits_processor_pattern:
1757
            _raise_unsupported_error(feature_name="--logits-processor-pattern")
1758
1759

        # No Concurrent Partial Prefills so far.
1760
1761
1762
1763
1764
        if (
            self.max_num_partial_prefills != SchedulerConfig.max_num_partial_prefills
            or self.max_long_partial_prefills
            != SchedulerConfig.max_long_partial_prefills
        ):
1765
            _raise_unsupported_error(feature_name="Concurrent Partial Prefill")
1766

1767
        # N-gram, Medusa, and Eagle are supported for speculative decoding.
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
        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 "
1779
1780
                    "such as ngram, medusa, eagle, or mtp."
                )
1781

1782
        if self.pipeline_parallel_size > 1:
1783
1784
1785
            supports_pp = getattr(
                self.distributed_executor_backend, "supports_pp", False
            )
1786
            if not supports_pp and self.distributed_executor_backend not in (
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
                ParallelConfig.distributed_executor_backend,
                "ray",
                "mp",
                "external_launcher",
            ):
                name = (
                    "Pipeline Parallelism without Ray distributed "
                    "executor or multiprocessing executor or external "
                    "launcher"
                )
1797
                _raise_unsupported_error(feature_name=name)
1798

1799
1800
1801
1802
1803
1804
    @classmethod
    def get_batch_defaults(
        cls,
        world_size: int,
    ) -> tuple[dict[UsageContext | None, int], dict[UsageContext | None, int]]:
        from vllm.usage.usage_lib import UsageContext
1805

1806
1807
        default_max_num_batched_tokens: dict[UsageContext | None, int]
        default_max_num_seqs: dict[UsageContext | None, int]
1808

1809
1810
        # When no user override, set the default values based on the usage
        # context.
1811
        # Use different default values for different hardware.
1812
1813
1814
1815
1816
1817
1818

        # 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:
1819
            device_memory = current_platform.get_device_total_memory()
1820
            device_name = current_platform.get_device_name().lower()
1821
1822
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1823
            device_memory = 0
1824

1825
1826
1827
1828
        # 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.
        if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
1829
            # For GPUs like H100 and MI300x, use larger default values.
1830
1831
1832
1833
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1834
1835
1836
1837
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
1838
1839
1840
1841
1842
1843
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1844
1845
1846
1847
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
1848

1849
1850
        # tpu specific default values.
        if current_platform.is_tpu():
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
            chip_name = current_platform.get_device_name()

            if chip_name == "V6E":
                default_max_num_batched_tokens = {
                    UsageContext.LLM_CLASS: 2048,
                    UsageContext.OPENAI_API_SERVER: 1024,
                }
            elif chip_name == "V5E":
                default_max_num_batched_tokens = {
                    UsageContext.LLM_CLASS: 1024,
                    UsageContext.OPENAI_API_SERVER: 512,
                }
            elif chip_name == "V5P":
                default_max_num_batched_tokens = {
                    UsageContext.LLM_CLASS: 512,
                    UsageContext.OPENAI_API_SERVER: 256,
                }
1868

1869
1870
1871
        # cpu specific default values.
        if current_platform.is_cpu():
            default_max_num_batched_tokens = {
1872
1873
                UsageContext.LLM_CLASS: 4096 * world_size,
                UsageContext.OPENAI_API_SERVER: 2048 * world_size,
1874
1875
            }
            default_max_num_seqs = {
1876
1877
                UsageContext.LLM_CLASS: 256 * world_size,
                UsageContext.OPENAI_API_SERVER: 128 * world_size,
1878
1879
            }

1880
1881
        return default_max_num_batched_tokens, default_max_num_seqs

1882
1883
    def _set_default_chunked_prefill_and_prefix_caching_args(
        self, model_config: ModelConfig
1884
    ) -> None:
1885
1886
        default_chunked_prefill = model_config.is_chunked_prefill_supported
        default_prefix_caching = model_config.is_prefix_caching_supported
1887

1888
1889
1890
        if self.prefill_context_parallel_size > 1:
            default_chunked_prefill = False
            default_prefix_caching = False
1891
            logger.warning_once(
1892
1893
                "--prefill-context-parallel-size > 1 is not compatible with "
                "chunked prefill and prefix caching now. Chunked prefill "
1894
1895
                "and prefix caching have been disabled by default.",
                scope="local",
1896
1897
            )

1898
1899
1900
1901
1902
1903
1904
        if self.enable_chunked_prefill is None:
            self.enable_chunked_prefill = default_chunked_prefill

            logger.debug(
                "%s chunked prefill by default",
                "Enabling" if default_chunked_prefill else "Disabling",
            )
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
        elif (
            model_config.runner_type == "generate"
            and not self.enable_chunked_prefill
            and default_chunked_prefill
        ):
            logger.warning_once(
                "This model does not officially support disabling chunked prefill. "
                "Disabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
                scope="local",
            )
1916
1917
1918
1919
        elif (
            model_config.runner_type == "pooling"
            and self.enable_chunked_prefill
            and not default_chunked_prefill
1920
        ):
1921
            logger.warning_once(
1922
1923
1924
                "This model does not officially support chunked prefill. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
1925
                scope="local",
1926
1927
1928
1929
1930
            )

        if self.enable_prefix_caching is None:
            self.enable_prefix_caching = default_prefix_caching

1931
            logger.debug(
1932
1933
1934
1935
1936
1937
1938
1939
                "%s prefix caching by default",
                "Enabling" if default_prefix_caching else "Disabling",
            )
        elif (
            model_config.runner_type == "pooling"
            and self.enable_prefix_caching
            and not default_prefix_caching
        ):
1940
            logger.warning_once(
1941
1942
1943
                "This model does not officially support prefix caching. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
1944
                scope="local",
1945
1946
            )

1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
        # Disable chunked prefill and prefix caching for:
        # POWER (ppc64le)/s390x/RISCV CPUs in V1
        if current_platform.is_cpu() and current_platform.get_cpu_architecture() in (
            CpuArchEnum.POWERPC,
            CpuArchEnum.S390X,
            CpuArchEnum.RISCV,
        ):
            logger.info(
                "Chunked prefill is not supported for ARM and POWER, "
                "S390X and RISC-V CPUs; "
                "disabling it for V1 backend."
            )
            self.enable_chunked_prefill = False
            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

    def _set_default_max_num_seqs_and_batched_tokens_args(
1968
1969
1970
        self,
        usage_context: UsageContext | None,
        model_config: ModelConfig,
1971
    ):
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
        world_size = self.pipeline_parallel_size * self.tensor_parallel_size
        (
            default_max_num_batched_tokens,
            default_max_num_seqs,
        ) = self.get_batch_defaults(world_size)

        orig_max_num_batched_tokens = self.max_num_batched_tokens
        orig_max_num_seqs = self.max_num_seqs

        if self.max_num_batched_tokens is None:
            self.max_num_batched_tokens = default_max_num_batched_tokens.get(
                usage_context,
                SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS,
            )

        if self.max_num_seqs is None:
            self.max_num_seqs = default_max_num_seqs.get(
                usage_context,
                SchedulerConfig.DEFAULT_MAX_NUM_SEQS,
            )

        if orig_max_num_batched_tokens is None:
            if not self.enable_chunked_prefill:
                # If max_model_len is too short, use the default for higher throughput.
                self.max_num_batched_tokens = max(
                    model_config.max_model_len,
                    self.max_num_batched_tokens,
                )

            # When using default settings,
            # Ensure max_num_batched_tokens does not exceed model limit.
            # Some models (e.g., Whisper) have embeddings tied to max length.
            self.max_num_batched_tokens = min(
                self.max_num_seqs * model_config.max_model_len,
2006
2007
                self.max_num_batched_tokens,
            )
2008

2009
2010
2011
2012
            logger.debug(
                "Defaulting max_num_batched_tokens to %d for %s usage context.",
                self.max_num_batched_tokens,
                usage_context.value if usage_context else None,
2013
            )
2014

2015
2016
2017
2018
        if orig_max_num_seqs is None:
            assert self.max_num_batched_tokens is not None  # For type checking
            self.max_num_seqs = min(self.max_num_seqs, self.max_num_batched_tokens)

2019
            logger.debug(
2020
                "Defaulting max_num_seqs to %d for %s usage context.",
2021
                self.max_num_seqs,
2022
                usage_context.value if usage_context else None,
2023
            )
2024

2025

2026
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
2027
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
2028
    """Arguments for asynchronous vLLM engine."""
2029

2030
2031
    enable_log_requests: bool = False

2032
    @staticmethod
2033
2034
2035
    def add_cli_args(
        parser: FlexibleArgumentParser, async_args_only: bool = False
    ) -> FlexibleArgumentParser:
2036
        # Initialize plugin to update the parser, for example, The plugin may
2037
        # add a new kind of quantization method to --quantization argument or
2038
2039
        # a new device to --device argument.
        load_general_plugins()
2040
2041
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
        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,
        )
2055
        current_platform.pre_register_and_update(parser)
2056
        return parser
2057
2058


2059
2060
2061
2062
2063
2064
def _raise_unsupported_error(feature_name: str):
    msg = (
        f"{feature_name} is not supported. We recommend to "
        f"remove {feature_name} from your config."
    )
    raise NotImplementedError(msg)
2065
2066


2067
2068
2069
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
2070

2071
2072
2073
2074
2075
2076
    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    """
    value = value.strip()
2077
    match = re.fullmatch(r"(\d+(?:\.\d+)?)([kKmMgGtT])", value)
2078
2079
    if match:
        decimal_multiplier = {
2080
2081
2082
            "k": 10**3,
            "m": 10**6,
            "g": 10**9,
2083
2084
        }
        binary_multiplier = {
2085
2086
2087
            "K": 2**10,
            "M": 2**20,
            "G": 2**30,
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
        }

        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:
2100
2101
2102
2103
2104
                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
2105
2106
2107

    # Regular plain number.
    return int(value)