"vllm/vscode:/vscode.git/clone" did not exist on "9b8c8ba1198cbcd311d28b7647f0f8d5dcdc9212"
arg_utils.py 99.2 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, asdict, 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.config import (
36
    AttentionConfig,
37
38
39
40
    CacheConfig,
    CompilationConfig,
    ConfigType,
    DeviceConfig,
41
    ECTransferConfig,
42
    EPLBConfig,
43
    KernelConfig,
44
45
46
47
48
    KVEventsConfig,
    KVTransferConfig,
    LoadConfig,
    LoRAConfig,
    ModelConfig,
49
    MultiModalConfig,
50
    ObservabilityConfig,
51
    OffloadConfig,
52
53
    ParallelConfig,
    PoolerConfig,
54
    PrefetchOffloadConfig,
55
    ProfilerConfig,
56
    ReasoningConfig,
57
58
59
    SchedulerConfig,
    SpeculativeConfig,
    StructuredOutputsConfig,
60
    UVAOffloadConfig,
61
    VllmConfig,
62
    WeightTransferConfig,
63
64
    get_attr_docs,
)
65
66
67
from vllm.config.cache import (
    CacheDType,
    KVOffloadingBackend,
68
    MambaCacheMode,
69
70
71
    MambaDType,
    PrefixCachingHashAlgo,
)
72
from vllm.config.device import Device
73
from vllm.config.kernel import IrOpPriorityConfig, MoEBackend
74
from vllm.config.lora import MaxLoRARanks
75
76
77
78
79
80
from vllm.config.model import (
    ConvertOption,
    HfOverrides,
    LogprobsMode,
    ModelDType,
    RunnerOption,
81
    TokenizerMode,
82
)
83
from vllm.config.multimodal import MMCacheType, MMEncoderTPMode, MMTensorIPC
84
from vllm.config.observability import DetailedTraceModules
85
86
87
from vllm.config.parallel import (
    All2AllBackend,
    DataParallelBackend,
88
    DCPCommBackend,
89
90
91
    DistributedExecutorBackend,
    ExpertPlacementStrategy,
)
92
from vllm.config.scheduler import SchedulerPolicy
93
from vllm.config.utils import get_field
94
from vllm.config.vllm import OptimizationLevel, PerformanceMode
95
from vllm.logger import init_logger, suppress_logging
96
from vllm.platforms import CpuArchEnum, current_platform
97
from vllm.plugins import load_general_plugins
98
from vllm.ray.lazy_utils import is_in_ray_actor, is_ray_initialized
99
100
101
102
from vllm.transformers_utils.config import (
    is_interleaved,
    maybe_override_with_speculators,
)
103
from vllm.transformers_utils.gguf_utils import is_gguf
104
from vllm.transformers_utils.repo_utils import get_model_path
105
from vllm.transformers_utils.utils import is_cloud_storage
106
from vllm.utils.argparse_utils import FlexibleArgumentParser
107
from vllm.utils.mem_constants import GiB_bytes
108
from vllm.utils.network_utils import get_ip
109
from vllm.utils.torch_utils import resolve_kv_cache_dtype_string
110
from vllm.v1.attention.backends.registry import AttentionBackendEnum
111
from vllm.v1.sample.logits_processor import LogitsProcessor
112
from vllm.version import __version__ as VLLM_VERSION
113

114
if TYPE_CHECKING:
115
    from vllm.config.quantization import OnlineQuantizationConfigArgs
116
    from vllm.model_executor.layers.quantization import QuantizationMethods
117
    from vllm.model_executor.model_loader import LoadFormats
118
    from vllm.usage.usage_lib import UsageContext
119
    from vllm.v1.executor import Executor
120
else:
121
    Executor = Any
122
    QuantizationMethods = Any
123
    LoadFormats = Any
124
125
    UsageContext = Any

126

127
128
logger = init_logger(__name__)

129
130
# object is used to allow for special typing forms
T = TypeVar("T")
131
132
TypeHint: TypeAlias = type[Any] | object
TypeHintT: TypeAlias = type[T] | object
133

134

135
136
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
    def _parse_type(val: str) -> T:
137
138
139
140
        try:
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
141
142
                f"Value {val} cannot be converted to {return_type}."
            ) from e
143

144
145
146
    return _parse_type


147
148
def optional_type(return_type: Callable[[str], T]) -> Callable[[str], T | None]:
    def _optional_type(val: str) -> T | None:
149
150
151
152
        if val == "" or val == "None":
            return None
        return parse_type(return_type)(val)

153
    return _optional_type
154
155


156
def union_dict_and_str(val: str) -> str | dict[str, str] | None:
157
    if not re.match(r"(?s)^\s*{.*}\s*$", val):
158
        return str(val)
159
    return optional_type(json.loads)(val)
160
161


162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
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)


177
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
178
179
180
181
    """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`.
    """
182
    type_hint = get_type(type_hints, Literal)
183
184
185
    options = get_args(type_hint)
    option_type = type(options[0])
    if not all(isinstance(option, option_type) for option in options):
186
        raise ValueError(
187
            "All options must be of the same type. "
188
189
            f"Got {options} with types {[type(c) for c in options]}"
        )
190
191
    kwarg = "metavar" if contains_type(type_hints, str) else "choices"
    return {"type": option_type, kwarg: sorted(options)}
192
193


194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
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),
    }


219
220
221
222
223
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


224
225
226
227
228
229
230
231
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]))
232
233
    elif origin in {Union, UnionType}:
        # Union for Union[X, Y] and UnionType for X | Y
234
235
236
237
238
239
240
241
        for arg in args:
            type_hints.update(get_type_hints(arg))
    else:
        type_hints.add(type_hint)

    return type_hints


242
NEEDS_HELP = (
243
244
    any("--help" in arg for arg in sys.argv)  # vllm SUBCOMMAND --help
    or (argv0 := sys.argv[0]).endswith("mkdocs")  # mkdocs SUBCOMMAND
245
246
247
248
    or argv0.endswith("mkdocs/__main__.py")  # python -m mkdocs SUBCOMMAND
)


249
250
251
252
253
254
255
256
def _maybe_add_docs_url(cls: Any) -> str:
    """Generate API docs URL for a vllm config class."""
    if not cls.__module__.startswith("vllm.config"):
        return ""
    version = f"v{VLLM_VERSION}" if "dev" not in VLLM_VERSION else "latest"
    return f"\n\nAPI docs: https://docs.vllm.ai/en/{version}/api/vllm/config/#vllm.config.{cls.__name__}"


257
@functools.lru_cache(maxsize=30)
258
def _compute_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
259
260
    # Save time only getting attr docs if we're generating help text
    cls_docs = get_attr_docs(cls) if NEEDS_HELP else {}
261
262
    kwargs = {}
    for field in fields(cls):
263
        # Get the set of possible types for the field
264
        type_hints: set[TypeHint] = get_type_hints(field.type)
265
266
267
268
269

        # 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)

270
        # Get the default value of the field
271
272
        if field.default is not MISSING:
            default = field.default
273
274
            # Handle pydantic.Field defaults
            if isinstance(default, FieldInfo):
275
276
277
278
279
280
                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():
281
                        default = default.default_factory()  # type: ignore[call-arg]
282
        elif field.default_factory is not MISSING:
283
            default = field.default_factory()
284
285
286

        # Get the help text for the field
        name = field.name
287
        help = cls_docs.get(name, "").strip()
288
289
290
291
292
293
294
        # 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
295
296
297
        json_tip = (
            "Should either be a valid JSON string or JSON keys passed individually."
        )
298
        if dataclass_cls is not None:
299
300
301
302
303
304
305
306

            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
307
            kwargs[name]["help"] += _maybe_add_docs_url(dataclass_cls)
308
            kwargs[name]["help"] += f"\n\n{json_tip}"
309
        elif contains_type(type_hints, bool):
310
311
312
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
313
            kwargs[name].update(literal_to_kwargs(type_hints))
314
        elif contains_type(type_hints, tuple):
315
            kwargs[name].update(collection_to_kwargs(type_hints, tuple))
316
        elif contains_type(type_hints, list):
317
318
319
            kwargs[name].update(collection_to_kwargs(type_hints, list))
        elif contains_type(type_hints, set):
            kwargs[name].update(collection_to_kwargs(type_hints, set))
320
        elif contains_type(type_hints, int):
321
322
323
324
            if name == "max_model_len":
                kwargs[name]["type"] = human_readable_int_or_auto
                kwargs[name]["help"] += f"\n\n{human_readable_int_or_auto.__doc__}"
            elif name in ("max_num_batched_tokens", "kv_cache_memory_bytes"):
325
                kwargs[name]["type"] = human_readable_int
326
                kwargs[name]["help"] += f"\n\n{human_readable_int.__doc__}"
327
328
            else:
                kwargs[name]["type"] = int
329
330
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
331
332
333
334
        elif contains_type(type_hints, dict) and (
            contains_type(type_hints, str)
            or any(is_not_builtin(th) for th in type_hints)
        ):
335
            kwargs[name]["type"] = union_dict_and_str
336
        elif contains_type(type_hints, dict):
337
            kwargs[name]["type"] = parse_type(json.loads)
338
            kwargs[name]["help"] += f"\n\n{json_tip}"
339
340
341
        elif contains_type(type_hints, str) or any(
            is_not_builtin(th) for th in type_hints
        ):
342
343
            kwargs[name]["type"] = str
        else:
344
            raise ValueError(f"Unsupported type {type_hints} for argument {name}.")
345

346
347
348
349
350
        # 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"]}))

351
352
353
354
355
356
357
        # 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
358
359


360
def get_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
361
362
    """Return argparse kwargs for the given Config dataclass.

363
364
365
    If `--help` or `mkdocs` are not present in the command line command, the
    attribute documentation will not be included in the help output.

366
367
368
369
370
371
372
    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))


373
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
374
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
375
    """Arguments for vLLM engine."""
376

377
    model: str = ModelConfig.model
378
    enable_return_routed_experts: bool = ModelConfig.enable_return_routed_experts
379
    model_weights: str = ModelConfig.model_weights
380
    served_model_name: str | list[str] | None = ModelConfig.served_model_name
381
    tokenizer: str | None = ModelConfig.tokenizer
382
    hf_config_path: str | None = ModelConfig.hf_config_path
383
384
    runner: RunnerOption = ModelConfig.runner
    convert: ConvertOption = ModelConfig.convert
385
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
386
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
387
    tokenizer_mode: TokenizerMode | str = ModelConfig.tokenizer_mode
388
    trust_remote_code: bool = ModelConfig.trust_remote_code
389
390
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
    allowed_media_domains: list[str] | None = ModelConfig.allowed_media_domains
391
    download_dir: str | None = LoadConfig.download_dir
392
    safetensors_load_strategy: str | None = LoadConfig.safetensors_load_strategy
393
    load_format: str | LoadFormats = LoadConfig.load_format
394
395
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
396
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
397
    seed: int = ModelConfig.seed
398
    max_model_len: int = ModelConfig.max_model_len
399
400
401
402
403
404
    cudagraph_capture_sizes: list[int] | None = (
        CompilationConfig.cudagraph_capture_sizes
    )
    max_cudagraph_capture_size: int | None = get_field(
        CompilationConfig, "max_cudagraph_capture_size"
    )
405
    ir_op_priority: IrOpPriorityConfig = get_field(KernelConfig, "ir_op_priority")
406
407
408
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
409
    distributed_executor_backend: (
410
        str | DistributedExecutorBackend | type[Executor] | None
411
    ) = ParallelConfig.distributed_executor_backend
412
    # number of P/D disaggregation (or other disaggregation) workers
413
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
414
415
416
417
    master_addr: str = ParallelConfig.master_addr
    master_port: int = ParallelConfig.master_port
    nnodes: int = ParallelConfig.nnodes
    node_rank: int = ParallelConfig.node_rank
418
    distributed_timeout_seconds: int | None = ParallelConfig.distributed_timeout_seconds
419
420
421
    numa_bind: bool = ParallelConfig.numa_bind
    numa_bind_nodes: list[int] | None = ParallelConfig.numa_bind_nodes
    numa_bind_cpus: list[str] | None = ParallelConfig.numa_bind_cpus
422
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
423
    prefill_context_parallel_size: int = ParallelConfig.prefill_context_parallel_size
424
    decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
425
    dcp_comm_backend: DCPCommBackend = ParallelConfig.dcp_comm_backend
426
    dcp_kv_cache_interleave_size: int = ParallelConfig.dcp_kv_cache_interleave_size
427
    cp_kv_cache_interleave_size: int = ParallelConfig.cp_kv_cache_interleave_size
428
    data_parallel_size: int = ParallelConfig.data_parallel_size
429
430
431
432
433
    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
434
    data_parallel_hybrid_lb: bool = False
435
    data_parallel_external_lb: bool = False
436
    data_parallel_backend: DataParallelBackend = ParallelConfig.data_parallel_backend
437
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
438
    enable_ep_weight_filter: bool = ParallelConfig.enable_ep_weight_filter
439
    moe_backend: MoEBackend = KernelConfig.moe_backend
440
    all2all_backend: All2AllBackend = ParallelConfig.all2all_backend
441
    enable_elastic_ep: bool = ParallelConfig.enable_elastic_ep
442
    enable_dbo: bool = ParallelConfig.enable_dbo
443
    ubatch_size: int = ParallelConfig.ubatch_size
444
445
    dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold
    dbo_prefill_token_threshold: int = ParallelConfig.dbo_prefill_token_threshold
446
    disable_nccl_for_dp_synchronization: bool | None = (
447
448
        ParallelConfig.disable_nccl_for_dp_synchronization
    )
449
    eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config")
450
    enable_eplb: bool = ParallelConfig.enable_eplb
451
    expert_placement_strategy: ExpertPlacementStrategy = (
452
        ParallelConfig.expert_placement_strategy
453
    )
454
455
    _api_process_count: int = ParallelConfig._api_process_count
    _api_process_rank: int = ParallelConfig._api_process_rank
456
    max_parallel_loading_workers: int | None = (
457
458
        ParallelConfig.max_parallel_loading_workers
    )
459
    block_size: int | None = None
460
    enable_prefix_caching: bool | None = None
461
    prefix_caching_hash_algo: PrefixCachingHashAlgo = (
462
        CacheConfig.prefix_caching_hash_algo
463
    )
464
465
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
466
467
468
469
470
471
472
    offload_backend: str = OffloadConfig.offload_backend
    cpu_offload_gb: float = UVAOffloadConfig.cpu_offload_gb
    cpu_offload_params: set[str] = get_field(UVAOffloadConfig, "cpu_offload_params")
    offload_group_size: int = PrefetchOffloadConfig.offload_group_size
    offload_num_in_group: int = PrefetchOffloadConfig.offload_num_in_group
    offload_prefetch_step: int = PrefetchOffloadConfig.offload_prefetch_step
    offload_params: set[str] = get_field(PrefetchOffloadConfig, "offload_params")
473
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
474
    kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
475
    max_num_batched_tokens: int | None = None
476
477
    max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
    max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
478
    long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold
479
    max_num_seqs: int | None = None
480
    max_logprobs: int = ModelConfig.max_logprobs
481
    logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode
482
    disable_log_stats: bool = False
483
    aggregate_engine_logging: bool = False
484
485
486
    revision: str | None = ModelConfig.revision
    code_revision: str | None = ModelConfig.code_revision
    hf_token: bool | str | None = ModelConfig.hf_token
487
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
488
    tokenizer_revision: str | None = ModelConfig.tokenizer_revision
489
    quantization: QuantizationMethods | str | None = ModelConfig.quantization
490
    quantization_config: "dict[str, Any] | OnlineQuantizationConfigArgs | None" = None
491
    allow_deprecated_quantization: bool = ModelConfig.allow_deprecated_quantization
492
    enforce_eager: bool = ModelConfig.enforce_eager
493
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
494
    language_model_only: bool = MultiModalConfig.language_model_only
495
    limit_mm_per_prompt: dict[str, int | dict[str, int]] = get_field(
496
497
        MultiModalConfig, "limit_per_prompt"
    )
498
    enable_mm_embeds: bool = MultiModalConfig.enable_mm_embeds
499
    interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings
500
501
502
    media_io_kwargs: dict[str, dict[str, Any]] = get_field(
        MultiModalConfig, "media_io_kwargs"
    )
503
    mm_processor_kwargs: dict[str, Any] | None = MultiModalConfig.mm_processor_kwargs
504
    mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb
505
    mm_processor_cache_type: MMCacheType | None = (
506
        MultiModalConfig.mm_processor_cache_type
507
508
    )
    mm_shm_cache_max_object_size_mb: int = (
509
        MultiModalConfig.mm_shm_cache_max_object_size_mb
510
    )
511
    mm_encoder_only: bool = MultiModalConfig.mm_encoder_only
512
    mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
513
    mm_encoder_attn_backend: AttentionBackendEnum | str | None = (
514
515
        MultiModalConfig.mm_encoder_attn_backend
    )
516
    io_processor_plugin: str | None = None
517
    renderer_num_workers: int = 1
518
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
519
    video_pruning_rate: float | None = MultiModalConfig.video_pruning_rate
520
    mm_tensor_ipc: MMTensorIPC = MultiModalConfig.mm_tensor_ipc
521
    # LoRA fields
522
    enable_lora: bool = False
523
    max_loras: int = LoRAConfig.max_loras
524
    max_lora_rank: MaxLoRARanks = LoRAConfig.max_lora_rank
525
    default_mm_loras: dict[str, str] | None = LoRAConfig.default_mm_loras
526
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
527
528
    max_cpu_loras: int | None = LoRAConfig.max_cpu_loras
    lora_dtype: str | torch.dtype | None = LoRAConfig.lora_dtype
529
    lora_target_modules: list[str] | None = LoRAConfig.target_modules
530
    enable_tower_connector_lora: bool = LoRAConfig.enable_tower_connector_lora
531
    specialize_active_lora: bool = LoRAConfig.specialize_active_lora
532

533
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
534
    num_gpu_blocks_override: int | None = CacheConfig.num_gpu_blocks_override
535
    model_loader_extra_config: dict = get_field(LoadConfig, "model_loader_extra_config")
536
    ignore_patterns: str | list[str] = get_field(LoadConfig, "ignore_patterns")
537

538
    enable_chunked_prefill: bool | None = None
539
    disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
540

541
542
    scheduler_reserve_full_isl: bool = SchedulerConfig.scheduler_reserve_full_isl

543
    disable_hybrid_kv_cache_manager: bool | None = (
544
545
        SchedulerConfig.disable_hybrid_kv_cache_manager
    )
546

547
    structured_outputs_config: StructuredOutputsConfig = get_field(
548
549
        VllmConfig, "structured_outputs_config"
    )
550
    reasoning_parser: str = StructuredOutputsConfig.reasoning_parser
551
    reasoning_parser_plugin: str | None = None
552

553
    speculative_config: dict[str, Any] | None = None
554

555
    show_hidden_metrics_for_version: str | None = (
556
        ObservabilityConfig.show_hidden_metrics_for_version
557
    )
558
559
    otlp_traces_endpoint: str | None = ObservabilityConfig.otlp_traces_endpoint
    collect_detailed_traces: list[DetailedTraceModules] | None = (
560
        ObservabilityConfig.collect_detailed_traces
561
    )
562
563
564
565
    kv_cache_metrics: bool = ObservabilityConfig.kv_cache_metrics
    kv_cache_metrics_sample: float = get_field(
        ObservabilityConfig, "kv_cache_metrics_sample"
    )
566
    cudagraph_metrics: bool = ObservabilityConfig.cudagraph_metrics
567
568
569
    enable_layerwise_nvtx_tracing: bool = (
        ObservabilityConfig.enable_layerwise_nvtx_tracing
    )
570
    enable_mfu_metrics: bool = ObservabilityConfig.enable_mfu_metrics
571
572
573
    enable_logging_iteration_details: bool = (
        ObservabilityConfig.enable_logging_iteration_details
    )
574
    enable_mm_processor_stats: bool = ObservabilityConfig.enable_mm_processor_stats
575
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
576
    scheduler_cls: str | type[object] | None = SchedulerConfig.scheduler_cls
577

578
    pooler_config: PoolerConfig | None = ModelConfig.pooler_config
579
    compilation_config: CompilationConfig = get_field(VllmConfig, "compilation_config")
580
    attention_config: AttentionConfig = get_field(VllmConfig, "attention_config")
581
582
583
584
    kernel_config: KernelConfig = get_field(VllmConfig, "kernel_config")
    enable_flashinfer_autotune: bool = get_field(
        KernelConfig, "enable_flashinfer_autotune"
    )
585
586
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
587

588
589
    profiler_config: ProfilerConfig = get_field(VllmConfig, "profiler_config")

590
591
    kv_transfer_config: KVTransferConfig | None = None
    kv_events_config: KVEventsConfig | None = None
592

593
    ec_transfer_config: ECTransferConfig | None = None
594
    reasoning_config: ReasoningConfig = get_field(VllmConfig, "reasoning_config")
595

596
597
    generation_config: str = ModelConfig.generation_config
    enable_sleep_mode: bool = ModelConfig.enable_sleep_mode
598
599
600
    override_generation_config: dict[str, Any] = get_field(
        ModelConfig, "override_generation_config"
    )
601
    model_impl: str = ModelConfig.model_impl
602
    override_attention_dtype: str | None = ModelConfig.override_attention_dtype
603
    attention_backend: AttentionBackendEnum | None = AttentionConfig.backend
604

605
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
606
607
608
    kv_cache_dtype_skip_layers: list[str] = get_field(
        CacheConfig, "kv_cache_dtype_skip_layers"
    )
609
610
    mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype
    mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype
611
    mamba_block_size: int | None = get_field(CacheConfig, "mamba_block_size")
612
    mamba_cache_mode: MambaCacheMode = CacheConfig.mamba_cache_mode
613
614
615
616
    enable_mamba_cache_stochastic_rounding: bool = (
        CacheConfig.enable_mamba_cache_stochastic_rounding
    )
    mamba_cache_philox_rounds: int = CacheConfig.mamba_cache_philox_rounds
617

618
    additional_config: dict[str, Any] = get_field(VllmConfig, "additional_config")
619

620
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
621
    pt_load_map_location: str | dict[str, str] = LoadConfig.pt_load_map_location
622

623
    logits_processors: list[str | type[LogitsProcessor]] | None = (
624
625
        ModelConfig.logits_processors
    )
626
627
    """Custom logitproc types"""

628
    async_scheduling: bool | None = SchedulerConfig.async_scheduling
629

630
631
    stream_interval: int = SchedulerConfig.stream_interval

632
    kv_sharing_fast_prefill: bool = CacheConfig.kv_sharing_fast_prefill
633
    optimization_level: OptimizationLevel = VllmConfig.optimization_level
634
    performance_mode: PerformanceMode = VllmConfig.performance_mode
635

636
    kv_offloading_size: float | None = CacheConfig.kv_offloading_size
637
    kv_offloading_backend: KVOffloadingBackend = CacheConfig.kv_offloading_backend
638
    tokens_only: bool = False
639

640
641
    shutdown_timeout: int = 0

642
643
644
645
    weight_transfer_config: WeightTransferConfig | None = get_field(
        VllmConfig,
        "weight_transfer_config",
    )
646

647
    fail_on_environ_validation: bool = False
648
    gdn_prefill_backend: Literal["flashinfer", "triton"] | None = None
649

650
    def __post_init__(self):
651
652
653
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
654
        if isinstance(self.compilation_config, dict):
655
            self.compilation_config = CompilationConfig(**self.compilation_config)
656
657
        if isinstance(self.attention_config, dict):
            self.attention_config = AttentionConfig(**self.attention_config)
658
659
        if isinstance(self.kernel_config, dict):
            self.kernel_config = KernelConfig(**self.kernel_config)
660
        if isinstance(self.eplb_config, dict):
661
            self.eplb_config = EPLBConfig(**self.eplb_config)
662
663
664
665
        if isinstance(self.weight_transfer_config, dict):
            self.weight_transfer_config = WeightTransferConfig(
                **self.weight_transfer_config
            )
666
667
668
        if isinstance(self.ir_op_priority, dict):
            self.ir_op_priority = IrOpPriorityConfig(**self.ir_op_priority)

669
670
671
672
673
674
        from vllm.config.quantization import resolve_online_quant_config

        self.quantization_config = resolve_online_quant_config(
            self.quantization, self.quantization_config
        )

675
        # Setup plugins
676
        from vllm.plugins import load_general_plugins
677

678
        load_general_plugins()
679
        # when use hf offline,replace model and tokenizer id to local model path
680
681
682
        if huggingface_hub.constants.HF_HUB_OFFLINE:
            model_id = self.model
            self.model = get_model_path(self.model, self.revision)
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
            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,
                    )
699
700

    @staticmethod
701
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
702
        """Shared CLI arguments for vLLM engine."""
703

704
        # Model arguments
705
706
707
708
709
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
710
        if not ("serve" in sys.argv[1:] and "--help" in sys.argv[1:]):
711
            model_group.add_argument("--model", **model_kwargs["model"])
712
713
        model_group.add_argument("--runner", **model_kwargs["runner"])
        model_group.add_argument("--convert", **model_kwargs["convert"])
714
715
        model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"])
        model_group.add_argument("--tokenizer-mode", **model_kwargs["tokenizer_mode"])
716
717
718
        model_group.add_argument(
            "--trust-remote-code", **model_kwargs["trust_remote_code"]
        )
719
720
        model_group.add_argument("--dtype", **model_kwargs["dtype"])
        model_group.add_argument("--seed", **model_kwargs["seed"])
721
        model_group.add_argument("--hf-config-path", **model_kwargs["hf_config_path"])
722
723
724
725
726
727
        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"]
        )
728
        model_group.add_argument("--revision", **model_kwargs["revision"])
729
        model_group.add_argument("--code-revision", **model_kwargs["code_revision"])
730
731
732
        model_group.add_argument(
            "--tokenizer-revision", **model_kwargs["tokenizer_revision"]
        )
733
734
        model_group.add_argument("--max-model-len", **model_kwargs["max_model_len"])
        model_group.add_argument("--quantization", "-q", **model_kwargs["quantization"])
735
736
737
738
        model_group.add_argument(
            "--allow-deprecated-quantization",
            **model_kwargs["allow_deprecated_quantization"],
        )
739
        model_group.add_argument("--enforce-eager", **model_kwargs["enforce_eager"])
740
741
742
743
        model_group.add_argument(
            "--enable-return-routed-experts",
            **model_kwargs["enable_return_routed_experts"],
        )
744
745
746
747
748
749
750
751
        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"]
        )
752
753
754
        model_group.add_argument(
            "--skip-tokenizer-init", **model_kwargs["skip_tokenizer_init"]
        )
755
756
757
758
759
760
761
        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"])
762
763
        # This one is a special case because it can bool
        # or str. TODO: Handle this in get_kwargs
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
        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(
            "--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"]
        )
783
        model_group.add_argument("--model-impl", **model_kwargs["model_impl"])
784
785
786
787
788
789
        model_group.add_argument(
            "--override-attention-dtype", **model_kwargs["override_attention_dtype"]
        )
        model_group.add_argument(
            "--logits-processors", **model_kwargs["logits_processors"]
        )
790
791
        model_group.add_argument(
            "--io-processor-plugin", **model_kwargs["io_processor_plugin"]
792
        )
793
794
795
796
        model_group.add_argument(
            "--renderer-num-workers",
            **model_kwargs["renderer_num_workers"],
        )
797

798
799
800
801
802
803
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
804
        load_group.add_argument("--load-format", **load_kwargs["load_format"])
805
806
807
808
809
810
811
812
813
814
815
816
        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"]
        )
817

818
819
820
821
822
823
824
825
826
827
        # Attention arguments
        attention_kwargs = get_kwargs(AttentionConfig)
        attention_group = parser.add_argument_group(
            title="AttentionConfig",
            description=AttentionConfig.__doc__,
        )
        attention_group.add_argument(
            "--attention-backend", **attention_kwargs["backend"]
        )

828
829
830
831
832
        # Structured outputs arguments
        structured_outputs_kwargs = get_kwargs(StructuredOutputsConfig)
        structured_outputs_group = parser.add_argument_group(
            title="StructuredOutputsConfig",
            description=StructuredOutputsConfig.__doc__,
833
        )
834
        structured_outputs_group.add_argument(
835
            "--reasoning-parser",
836
            # Choices need to be validated after parsing to include plugins
837
838
            **structured_outputs_kwargs["reasoning_parser"],
        )
839
840
841
842
        structured_outputs_group.add_argument(
            "--reasoning-parser-plugin",
            **structured_outputs_kwargs["reasoning_parser_plugin"],
        )
843

844
        # Parallel arguments
845
846
847
848
849
850
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
851
            "--distributed-executor-backend",
852
853
            **parallel_kwargs["distributed_executor_backend"],
        )
854
        parallel_group.add_argument(
855
856
857
858
            "--pipeline-parallel-size",
            "-pp",
            **parallel_kwargs["pipeline_parallel_size"],
        )
859
860
861
862
        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"])
863
864
865
866
        parallel_group.add_argument(
            "--distributed-timeout-seconds",
            **parallel_kwargs["distributed_timeout_seconds"],
        )
867
868
869
870
871
872
873
        parallel_group.add_argument("--numa-bind", **parallel_kwargs["numa_bind"])
        parallel_group.add_argument(
            "--numa-bind-nodes", **parallel_kwargs["numa_bind_nodes"]
        )
        parallel_group.add_argument(
            "--numa-bind-cpus", **parallel_kwargs["numa_bind_cpus"]
        )
874
        parallel_group.add_argument(
875
876
            "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
        )
877
        parallel_group.add_argument(
878
879
880
881
            "--decode-context-parallel-size",
            "-dcp",
            **parallel_kwargs["decode_context_parallel_size"],
        )
882
883
884
885
        parallel_group.add_argument(
            "--dcp-comm-backend",
            **parallel_kwargs["dcp_comm_backend"],
        )
886
887
888
889
        parallel_group.add_argument(
            "--dcp-kv-cache-interleave-size",
            **parallel_kwargs["dcp_kv_cache_interleave_size"],
        )
890
891
892
893
894
895
896
897
898
        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"],
        )
899
900
901
902
903
904
        parallel_group.add_argument(
            "--data-parallel-size", "-dp", **parallel_kwargs["data_parallel_size"]
        )
        parallel_group.add_argument(
            "--data-parallel-rank",
            "-dpn",
905
            type=int,
906
907
908
            help="Data parallel rank of this instance. "
            "When set, enables external load balancer mode.",
        )
909
        parallel_group.add_argument(
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
            "--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".',
        )
940
        parallel_group.add_argument(
941
942
943
944
945
946
947
948
            "--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"],
949
950
        )
        parallel_group.add_argument(
951
952
953
            "--enable-expert-parallel",
            "-ep",
            **parallel_kwargs["enable_expert_parallel"],
954
        )
955
956
957
958
        parallel_group.add_argument(
            "--enable-ep-weight-filter",
            **parallel_kwargs["enable_ep_weight_filter"],
        )
959
960
961
        parallel_group.add_argument(
            "--all2all-backend", **parallel_kwargs["all2all_backend"]
        )
962
        parallel_group.add_argument("--enable-dbo", **parallel_kwargs["enable_dbo"])
963
964
965
966
        parallel_group.add_argument(
            "--ubatch-size",
            **parallel_kwargs["ubatch_size"],
        )
967
968
969
        parallel_group.add_argument(
            "--enable-elastic-ep", **parallel_kwargs["enable_elastic_ep"]
        )
970
971
        parallel_group.add_argument(
            "--dbo-decode-token-threshold",
972
973
            **parallel_kwargs["dbo_decode_token_threshold"],
        )
974
975
        parallel_group.add_argument(
            "--dbo-prefill-token-threshold",
976
977
            **parallel_kwargs["dbo_prefill_token_threshold"],
        )
978
979
980
981
        parallel_group.add_argument(
            "--disable-nccl-for-dp-synchronization",
            **parallel_kwargs["disable_nccl_for_dp_synchronization"],
        )
982
983
        parallel_group.add_argument("--enable-eplb", **parallel_kwargs["enable_eplb"])
        parallel_group.add_argument("--eplb-config", **parallel_kwargs["eplb_config"])
984
985
        parallel_group.add_argument(
            "--expert-placement-strategy",
986
987
            **parallel_kwargs["expert_placement_strategy"],
        )
988

989
        parallel_group.add_argument(
990
            "--max-parallel-loading-workers",
991
992
            **parallel_kwargs["max_parallel_loading_workers"],
        )
993
        parallel_group.add_argument(
994
995
            "--ray-workers-use-nsight", **parallel_kwargs["ray_workers_use_nsight"]
        )
996
        parallel_group.add_argument(
997
            "--disable-custom-all-reduce",
998
999
1000
1001
1002
1003
            **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"]
        )
1004

1005
1006
1007
1008
1009
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
1010
        )
1011
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
        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"]
        )
        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(
1023
1024
1025
1026
1027
            "--enable-prefix-caching",
            **{
                **cache_kwargs["enable_prefix_caching"],
                "default": None,
            },
1028
1029
1030
1031
1032
1033
1034
        )
        cache_group.add_argument(
            "--prefix-caching-hash-algo", **cache_kwargs["prefix_caching_hash_algo"]
        )
        cache_group.add_argument(
            "--calculate-kv-scales", **cache_kwargs["calculate_kv_scales"]
        )
1035
1036
1037
        cache_group.add_argument(
            "--kv-cache-dtype-skip-layers", **cache_kwargs["kv_cache_dtype_skip_layers"]
        )
1038
1039
1040
1041
1042
1043
1044
1045
1046
        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"]
        )
1047
1048
1049
        cache_group.add_argument(
            "--mamba-block-size", **cache_kwargs["mamba_block_size"]
        )
1050
1051
1052
        cache_group.add_argument(
            "--mamba-cache-mode", **cache_kwargs["mamba_cache_mode"]
        )
1053
1054
1055
1056
1057
1058
1059
        cache_group.add_argument(
            "--enable-mamba-cache-stochastic-rounding",
            **cache_kwargs["enable_mamba_cache_stochastic_rounding"],
        )
        cache_group.add_argument(
            "--mamba-cache-philox-rounds", **cache_kwargs["mamba_cache_philox_rounds"]
        )
1060
1061
1062
1063
1064
1065
        cache_group.add_argument(
            "--kv-offloading-size", **cache_kwargs["kv_offloading_size"]
        )
        cache_group.add_argument(
            "--kv-offloading-backend", **cache_kwargs["kv_offloading_backend"]
        )
1066

1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
        # Model weight offload related configs
        offload_kwargs = get_kwargs(OffloadConfig)
        uva_kwargs = get_kwargs(UVAOffloadConfig)
        prefetch_kwargs = get_kwargs(PrefetchOffloadConfig)
        offload_group = parser.add_argument_group(
            title="OffloadConfig",
            description=OffloadConfig.__doc__,
        )
        offload_group.add_argument(
            "--offload-backend", **offload_kwargs["offload_backend"]
        )
        offload_group.add_argument("--cpu-offload-gb", **uva_kwargs["cpu_offload_gb"])
        offload_group.add_argument(
            "--cpu-offload-params", **uva_kwargs["cpu_offload_params"]
        )
        offload_group.add_argument(
            "--offload-group-size",
            **prefetch_kwargs["offload_group_size"],
        )
        offload_group.add_argument(
            "--offload-num-in-group",
            **prefetch_kwargs["offload_num_in_group"],
        )
        offload_group.add_argument(
            "--offload-prefetch-step",
            **prefetch_kwargs["offload_prefetch_step"],
        )
        offload_group.add_argument(
            "--offload-params", **prefetch_kwargs["offload_params"]
        )

1098
        # Multimodal related configs
1099
1100
1101
1102
1103
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
1104
1105
1106
        multimodal_group.add_argument(
            "--language-model-only", **multimodal_kwargs["language_model_only"]
        )
1107
        multimodal_group.add_argument(
1108
1109
            "--limit-mm-per-prompt", **multimodal_kwargs["limit_per_prompt"]
        )
1110
1111
1112
        multimodal_group.add_argument(
            "--enable-mm-embeds", **multimodal_kwargs["enable_mm_embeds"]
        )
1113
1114
1115
        multimodal_group.add_argument(
            "--media-io-kwargs", **multimodal_kwargs["media_io_kwargs"]
        )
1116
1117
1118
1119
1120
1121
        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"]
        )
1122
        multimodal_group.add_argument(
1123
1124
            "--mm-processor-cache-type", **multimodal_kwargs["mm_processor_cache_type"]
        )
1125
1126
        multimodal_group.add_argument(
            "--mm-shm-cache-max-object-size-mb",
1127
1128
            **multimodal_kwargs["mm_shm_cache_max_object_size_mb"],
        )
1129
1130
1131
        multimodal_group.add_argument(
            "--mm-encoder-only", **multimodal_kwargs["mm_encoder_only"]
        )
1132
        multimodal_group.add_argument(
1133
1134
            "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"]
        )
1135
1136
1137
1138
        multimodal_group.add_argument(
            "--mm-encoder-attn-backend",
            **multimodal_kwargs["mm_encoder_attn_backend"],
        )
1139
1140
1141
        multimodal_group.add_argument(
            "--interleave-mm-strings", **multimodal_kwargs["interleave_mm_strings"]
        )
1142
        multimodal_group.add_argument(
1143
1144
            "--skip-mm-profiling", **multimodal_kwargs["skip_mm_profiling"]
        )
1145

1146
        multimodal_group.add_argument(
1147
1148
            "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"]
        )
1149
1150
1151
        multimodal_group.add_argument(
            "--mm-tensor-ipc", **multimodal_kwargs["mm_tensor_ipc"]
        )
1152

1153
        # LoRA related configs
1154
1155
1156
1157
1158
1159
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
1160
            "--enable-lora",
1161
            action=argparse.BooleanOptionalAction,
1162
1163
            help="If True, enable handling of LoRA adapters.",
        )
1164
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
1165
        lora_group.add_argument("--max-lora-rank", **lora_kwargs["max_lora_rank"])
1166
        lora_group.add_argument(
1167
            "--lora-dtype",
1168
1169
            **lora_kwargs["lora_dtype"],
        )
1170
1171
1172
1173
        lora_group.add_argument(
            "--enable-tower-connector-lora",
            **lora_kwargs["enable_tower_connector_lora"],
        )
1174
1175
1176
1177
        lora_group.add_argument("--max-cpu-loras", **lora_kwargs["max_cpu_loras"])
        lora_group.add_argument(
            "--fully-sharded-loras", **lora_kwargs["fully_sharded_loras"]
        )
1178
1179
1180
        lora_group.add_argument(
            "--lora-target-modules", **lora_kwargs["target_modules"]
        )
1181
        lora_group.add_argument("--default-mm-loras", **lora_kwargs["default_mm_loras"])
1182
1183
1184
        lora_group.add_argument(
            "--specialize-active-lora", **lora_kwargs["specialize_active_lora"]
        )
1185

1186
1187
1188
1189
1190
1191
1192
1193
        # 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",
1194
1195
            **observability_kwargs["show_hidden_metrics_for_version"],
        )
1196
        observability_group.add_argument(
1197
1198
            "--otlp-traces-endpoint", **observability_kwargs["otlp_traces_endpoint"]
        )
1199
1200
1201
1202
1203
        # 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"] += [
1204
            ",".join(p) for p in permutations(get_args(DetailedTraceModules), r=2)
1205
1206
1207
        ]
        observability_group.add_argument(
            "--collect-detailed-traces",
1208
1209
            **observability_kwargs["collect_detailed_traces"],
        )
1210
1211
1212
1213
1214
1215
1216
        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"],
        )
1217
1218
1219
1220
        observability_group.add_argument(
            "--cudagraph-metrics",
            **observability_kwargs["cudagraph_metrics"],
        )
1221
1222
1223
1224
        observability_group.add_argument(
            "--enable-layerwise-nvtx-tracing",
            **observability_kwargs["enable_layerwise_nvtx_tracing"],
        )
1225
1226
1227
1228
        observability_group.add_argument(
            "--enable-mfu-metrics",
            **observability_kwargs["enable_mfu_metrics"],
        )
1229
1230
1231
1232
        observability_group.add_argument(
            "--enable-logging-iteration-details",
            **observability_kwargs["enable_logging_iteration_details"],
        )
1233

1234
1235
1236
1237
1238
1239
1240
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
1241
1242
1243
1244
1245
            "--max-num-batched-tokens",
            **{
                **scheduler_kwargs["max_num_batched_tokens"],
                "default": None,
            },
1246
        )
1247
        scheduler_group.add_argument(
1248
1249
1250
1251
1252
            "--max-num-seqs",
            **{
                **scheduler_kwargs["max_num_seqs"],
                "default": None,
            },
1253
1254
1255
1256
        )
        scheduler_group.add_argument(
            "--max-num-partial-prefills", **scheduler_kwargs["max_num_partial_prefills"]
        )
1257
1258
        scheduler_group.add_argument(
            "--max-long-partial-prefills",
1259
1260
            **scheduler_kwargs["max_long_partial_prefills"],
        )
1261
1262
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
1263
1264
            **scheduler_kwargs["long_prefill_token_threshold"],
        )
1265
1266
        # multi-step scheduling has been removed; corresponding arguments
        # are no longer supported.
1267
        scheduler_group.add_argument(
1268
1269
            "--scheduling-policy", **scheduler_kwargs["policy"]
        )
1270
        scheduler_group.add_argument(
1271
1272
1273
1274
1275
            "--enable-chunked-prefill",
            **{
                **scheduler_kwargs["enable_chunked_prefill"],
                "default": None,
            },
1276
1277
1278
1279
1280
1281
1282
        )
        scheduler_group.add_argument(
            "--disable-chunked-mm-input", **scheduler_kwargs["disable_chunked_mm_input"]
        )
        scheduler_group.add_argument(
            "--scheduler-cls", **scheduler_kwargs["scheduler_cls"]
        )
1283
1284
1285
1286
        scheduler_group.add_argument(
            "--scheduler-reserve-full-isl",
            **scheduler_kwargs["scheduler_reserve_full_isl"],
        )
1287
1288
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
1289
1290
1291
1292
1293
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"],
        )
        scheduler_group.add_argument(
            "--async-scheduling", **scheduler_kwargs["async_scheduling"]
        )
1294
1295
1296
        scheduler_group.add_argument(
            "--stream-interval", **scheduler_kwargs["stream_interval"]
        )
1297

1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
        # 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_group.add_argument(
            "--max-cudagraph-capture-size",
            **compilation_kwargs["max_cudagraph_capture_size"],
        )

1312
1313
1314
1315
1316
1317
        # Kernel arguments
        kernel_kwargs = get_kwargs(KernelConfig)
        kernel_group = parser.add_argument_group(
            title="KernelConfig",
            description=KernelConfig.__doc__,
        )
1318
        kernel_group.add_argument("--ir-op-priority", **kernel_kwargs["ir_op_priority"])
1319
1320
1321
1322
        kernel_group.add_argument(
            "--enable-flashinfer-autotune",
            **kernel_kwargs["enable_flashinfer_autotune"],
        )
1323
1324
1325
        moe_backend_kwargs = kernel_kwargs["moe_backend"]
        moe_backend_kwargs["type"] = lambda s: s.lower().replace("-", "_")
        kernel_group.add_argument("--moe-backend", **moe_backend_kwargs)
1326

1327
        # vLLM arguments
1328
        vllm_kwargs = get_kwargs(VllmConfig)
1329
1330
1331
1332
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
1333
1334
1335
1336
        # 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)
1337
        vllm_group.add_argument(
1338
            "--speculative-config", "-sc", **vllm_kwargs["speculative_config"]
1339
1340
1341
1342
1343
        )
        vllm_group.add_argument(
            "--kv-transfer-config", **vllm_kwargs["kv_transfer_config"]
        )
        vllm_group.add_argument("--kv-events-config", **vllm_kwargs["kv_events_config"])
1344
1345
1346
        vllm_group.add_argument(
            "--ec-transfer-config", **vllm_kwargs["ec_transfer_config"]
        )
1347
        vllm_group.add_argument(
1348
            "--compilation-config", "-cc", **vllm_kwargs["compilation_config"]
1349
        )
1350
1351
1352
        vllm_group.add_argument(
            "--attention-config", "-ac", **vllm_kwargs["attention_config"]
        )
1353
        vllm_group.add_argument("--reasoning-config", **vllm_kwargs["reasoning_config"])
1354
        vllm_group.add_argument("--kernel-config", **vllm_kwargs["kernel_config"])
1355
1356
1357
1358
1359
1360
        vllm_group.add_argument(
            "--additional-config", **vllm_kwargs["additional_config"]
        )
        vllm_group.add_argument(
            "--structured-outputs-config", **vllm_kwargs["structured_outputs_config"]
        )
1361
        vllm_group.add_argument("--profiler-config", **vllm_kwargs["profiler_config"])
1362
1363
1364
        vllm_group.add_argument(
            "--optimization-level", **vllm_kwargs["optimization_level"]
        )
1365
        vllm_group.add_argument("--performance-mode", **vllm_kwargs["performance_mode"])
1366
1367
1368
        vllm_group.add_argument(
            "--weight-transfer-config", **vllm_kwargs["weight_transfer_config"]
        )
1369

1370
        # Other arguments
1371
1372
1373
1374
1375
        parser.add_argument(
            "--disable-log-stats",
            action="store_true",
            help="Disable logging statistics.",
        )
1376

1377
1378
1379
1380
1381
1382
        parser.add_argument(
            "--aggregate-engine-logging",
            action="store_true",
            help="Log aggregate rather than per-engine statistics "
            "when using data parallelism.",
        )
1383
1384
1385
1386
1387
1388
1389
1390

        parser.add_argument(
            "--fail-on-environ-validation",
            help="If set, the engine will raise an error if "
            "environment validation fails.",
            default=False,
            action=argparse.BooleanOptionalAction,
        )
1391
1392
1393
1394
1395
1396
1397
1398

        parser.add_argument(
            "--shutdown-timeout",
            type=int,
            default=0,
            help="Shutdown timeout in seconds. 0 = abort, >0 = wait.",
        )

1399
1400
1401
1402
1403
1404
1405
        parser.add_argument(
            "--gdn-prefill-backend",
            dest="gdn_prefill_backend",
            choices=["flashinfer", "triton"],
            default=None,
            help="Select GDN prefill backend.",
        )
1406
        return parser
1407
1408

    @classmethod
1409
    def from_cli_args(cls, args: argparse.Namespace):
1410
1411
1412
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
1413
1414
1415
        engine_args = cls(
            **{attr: getattr(args, attr) for attr in attrs if hasattr(args, attr)}
        )
Zhuohan Li's avatar
Zhuohan Li committed
1416
        return engine_args
1417

1418
    def create_model_config(self) -> ModelConfig:
1419
1420
        # gguf file needs a specific model loader
        if is_gguf(self.model):
1421
1422
            self.quantization = self.load_format = "gguf"

1423
1424
1425
1426
1427
1428
1429
        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,
1430
1431
            )

1432
        return ModelConfig(
1433
            model=self.model,
1434
            model_weights=self.model_weights,
1435
            hf_config_path=self.hf_config_path,
1436
1437
            runner=self.runner,
            convert=self.convert,
1438
            tokenizer=self.tokenizer,  # type: ignore[arg-type]
1439
            tokenizer_mode=self.tokenizer_mode,
1440
            trust_remote_code=self.trust_remote_code,
1441
1442
            allowed_local_media_path=self.allowed_local_media_path,
            allowed_media_domains=self.allowed_media_domains,
1443
1444
1445
1446
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
1447
            hf_token=self.hf_token,
1448
            hf_overrides=self.hf_overrides,
1449
            tokenizer_revision=self.tokenizer_revision,
1450
1451
            max_model_len=self.max_model_len,
            quantization=self.quantization,
1452
            quantization_config=self.quantization_config,
1453
            allow_deprecated_quantization=self.allow_deprecated_quantization,
1454
            enforce_eager=self.enforce_eager,
1455
            enable_return_routed_experts=self.enable_return_routed_experts,
1456
            max_logprobs=self.max_logprobs,
1457
            logprobs_mode=self.logprobs_mode,
1458
            disable_sliding_window=self.disable_sliding_window,
1459
            disable_cascade_attn=self.disable_cascade_attn,
1460
            skip_tokenizer_init=self.skip_tokenizer_init,
1461
            enable_prompt_embeds=self.enable_prompt_embeds,
1462
            served_model_name=self.served_model_name,
1463
            language_model_only=self.language_model_only,
1464
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1465
            enable_mm_embeds=self.enable_mm_embeds,
1466
            interleave_mm_strings=self.interleave_mm_strings,
1467
            media_io_kwargs=self.media_io_kwargs,
1468
            skip_mm_profiling=self.skip_mm_profiling,
1469
            config_format=self.config_format,
1470
            mm_processor_kwargs=self.mm_processor_kwargs,
1471
            mm_processor_cache_gb=self.mm_processor_cache_gb,
1472
            mm_processor_cache_type=self.mm_processor_cache_type,
1473
            mm_shm_cache_max_object_size_mb=self.mm_shm_cache_max_object_size_mb,
1474
            mm_encoder_only=self.mm_encoder_only,
1475
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
1476
            mm_encoder_attn_backend=self.mm_encoder_attn_backend,
1477
            pooler_config=self.pooler_config,
1478
            generation_config=self.generation_config,
1479
            override_generation_config=self.override_generation_config,
1480
            enable_sleep_mode=self.enable_sleep_mode,
1481
            model_impl=self.model_impl,
1482
            override_attention_dtype=self.override_attention_dtype,
1483
            logits_processors=self.logits_processors,
1484
            video_pruning_rate=self.video_pruning_rate,
1485
            mm_tensor_ipc=self.mm_tensor_ipc,
1486
            io_processor_plugin=self.io_processor_plugin,
1487
            renderer_num_workers=self.renderer_num_workers,
1488
        )
1489

1490
    def validate_tensorizer_args(self):
1491
1492
        from vllm.model_executor.model_loader.tensorizer import TensorizerConfig

1493
1494
        for key in self.model_loader_extra_config:
            if key in TensorizerConfig._fields:
1495
1496
1497
                self.model_loader_extra_config["tensorizer_config"][key] = (
                    self.model_loader_extra_config[key]
                )
1498

1499
    def create_load_config(self) -> LoadConfig:
1500
1501
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1502

1503
1504
1505
        if self.load_format == "tensorizer":
            if hasattr(self.model_loader_extra_config, "to_serializable"):
                self.model_loader_extra_config = (
1506
1507
                    self.model_loader_extra_config.to_serializable()
                )
1508
            self.model_loader_extra_config["tensorizer_config"] = {}
1509
1510
1511
            self.model_loader_extra_config["tensorizer_config"]["tensorizer_dir"] = (
                self.model
            )
1512
            self.validate_tensorizer_args()
1513

1514
1515
1516
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
1517
            safetensors_load_strategy=self.safetensors_load_strategy,
1518
1519
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1520
            use_tqdm_on_load=self.use_tqdm_on_load,
1521
            pt_load_map_location=self.pt_load_map_location,
1522
        )
1523

1524
1525
1526
1527
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
1528
    ) -> SpeculativeConfig | None:
1529
1530
1531
1532
1533
1534
        """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
1535
        dictionary from the engine.
1536
1537
        """
        if self.speculative_config is None:
1538
            return None
1539

1540
1541
1542
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.
1543
1544
1545
1546
1547
1548
        self.speculative_config.update(
            {
                "target_model_config": target_model_config,
                "target_parallel_config": target_parallel_config,
            }
        )
1549
        return SpeculativeConfig(**self.speculative_config)
1550

1551
1552
    def create_engine_config(
        self,
1553
        usage_context: UsageContext | None = None,
1554
        headless: bool = False,
1555
1556
1557
1558
    ) -> VllmConfig:
        """
        Create the VllmConfig.

1559
        NOTE: If VllmConfig is incompatible, we raise an error.
1560
        """
1561
        current_platform.pre_register_and_update()
1562

1563
        device_config = DeviceConfig(device=cast(Device, current_platform.device_type))
1564

1565
1566
        envs.validate_environ(self.fail_on_environ_validation)

1567
1568
        # Check if the model is a speculator and override model/tokenizer/config
        # BEFORE creating ModelConfig, so the config is created with the target model
1569
1570
1571
1572
        # 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):
1573
1574
1575
1576
1577
1578
1579
            (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,
1580
                    hf_token=self.hf_token,
1581
1582
1583
                )
            )

1584
        model_config = self.create_model_config()
1585
        self.model = model_config.model
1586
        self.model_weights = model_config.model_weights
1587
1588
        self.tokenizer = model_config.tokenizer

1589
        self._check_feature_supported()
1590
        self._set_default_chunked_prefill_and_prefix_caching_args(model_config)
1591
        self._set_default_reasoning_config_args()
1592
        sliding_window: int | None = None
1593
1594
1595
1596
1597
1598
        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()

1599
1600
1601
1602
1603
        # Resolve "auto" kv_cache_dtype to actual value from model config
        resolved_cache_dtype = resolve_kv_cache_dtype_string(
            self.kv_cache_dtype, model_config
        )

1604
1605
1606
1607
        assert self.enable_prefix_caching is not None, (
            "enable_prefix_caching must be set by this point"
        )

1608
        cache_config = CacheConfig(
1609
            block_size=self.block_size,  # type: ignore[arg-type]
1610
            gpu_memory_utilization=self.gpu_memory_utilization,
1611
            kv_cache_memory_bytes=self.kv_cache_memory_bytes,
1612
            cache_dtype=resolved_cache_dtype,  # type: ignore[arg-type]
1613
            is_attention_free=model_config.is_attention_free,
1614
            num_gpu_blocks_override=self.num_gpu_blocks_override,
1615
            sliding_window=sliding_window,
1616
            enable_prefix_caching=self.enable_prefix_caching,
1617
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1618
            calculate_kv_scales=self.calculate_kv_scales,
1619
            kv_cache_dtype_skip_layers=self.kv_cache_dtype_skip_layers,
1620
            kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
1621
1622
            mamba_cache_dtype=self.mamba_cache_dtype,
            mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype,
1623
            mamba_block_size=self.mamba_block_size,
1624
            mamba_cache_mode=self.mamba_cache_mode,
1625
1626
            enable_mamba_cache_stochastic_rounding=self.enable_mamba_cache_stochastic_rounding,
            mamba_cache_philox_rounds=self.mamba_cache_philox_rounds,
1627
1628
            kv_offloading_size=self.kv_offloading_size,
            kv_offloading_backend=self.kv_offloading_backend,
1629
        )
1630

1631
1632
1633
1634
1635
1636
        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
1637

1638
            ray_runtime_env = ray.get_runtime_context().runtime_env
1639
1640
1641
1642
1643
1644
1645
            # 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)
1646

1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
        # 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()

1658
        assert not headless or not self.data_parallel_hybrid_lb, (
1659
1660
            "data_parallel_hybrid_lb is not applicable in headless mode"
        )
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
        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
        )
1702
        # Local DP rank = 1, use pure-external LB.
1703
        if data_parallel_external_lb:
1704
            assert self.data_parallel_rank is not None, (
1705
                "data_parallel_rank or node_rank must be specified if "
1706
1707
                "data_parallel_external_lb is enable."
            )
1708
            assert self.data_parallel_size_local in (1, None), (
1709
1710
                "data_parallel_size_local must be 1 or None when data_parallel_rank "
                "is set"
1711
            )
1712
            data_parallel_size_local = 1
1713
1714
            # Use full external lb if we have local_size of 1.
            self.data_parallel_hybrid_lb = False
1715
1716
        elif self.data_parallel_size_local is not None:
            data_parallel_size_local = self.data_parallel_size_local
1717
1718
1719
1720
1721
1722
1723

            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.
1724
1725
1726
1727
1728
                logger.warning(
                    "data_parallel_hybrid_lb is not eligible when "
                    "data_parallel_size_local = 1, autoswitch to "
                    "data_parallel_external_lb."
                )
1729
1730
1731
1732
1733
1734
1735
                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

1736
1737
1738
1739
1740
1741
1742
1743
1744
            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,
                )
1745
        else:
1746
            assert not self.data_parallel_hybrid_lb, (
1747
1748
                "data_parallel_size_local must be set to use data_parallel_hybrid_lb."
            )
1749

1750
1751
1752
1753
1754
1755
1756
1757
1758
            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
1759
1760
1761

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1762
1763
1764
1765
        if self.data_parallel_address is None:
            if self.data_parallel_backend == "ray":
                host_ip = get_ip()
                logger.info(
1766
1767
                    "Using host IP %s as ray-based data parallel address", host_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1768
1769
1770
1771
                data_parallel_address = host_ip
            else:
                assert self.data_parallel_backend == "mp", (
                    "data_parallel_backend can only be ray or mp, got %s",
1772
1773
                    self.data_parallel_backend,
                )
1774
1775
1776
                data_parallel_address = (
                    self.master_addr or ParallelConfig.data_parallel_master_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1777
1778
        else:
            data_parallel_address = self.data_parallel_address
1779
1780
1781

        # This port is only used when there are remote data parallel engines,
        # otherwise the local IPC transport is used.
1782
        data_parallel_rpc_port = (
1783
            self.data_parallel_rpc_port
1784
1785
1786
            if (self.data_parallel_rpc_port is not None)
            else ParallelConfig.data_parallel_rpc_port
        )
1787

1788
1789
1790
1791
        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.")

1792
        parallel_config = ParallelConfig(
1793
1794
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1795
            prefill_context_parallel_size=self.prefill_context_parallel_size,
1796
            data_parallel_size=self.data_parallel_size,
1797
1798
            data_parallel_rank=self.data_parallel_rank or 0,
            data_parallel_external_lb=data_parallel_external_lb,
1799
            data_parallel_size_local=data_parallel_size_local,
1800
1801
1802
1803
            master_addr=self.master_addr,
            master_port=self.master_port,
            nnodes=self.nnodes,
            node_rank=self.node_rank,
1804
            distributed_timeout_seconds=self.distributed_timeout_seconds,
1805
1806
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
1807
            data_parallel_backend=self.data_parallel_backend,
1808
            data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
1809
            is_moe_model=model_config.is_moe,
1810
            enable_expert_parallel=self.enable_expert_parallel,
1811
            enable_ep_weight_filter=self.enable_ep_weight_filter,
1812
            all2all_backend=self.all2all_backend,
1813
            enable_elastic_ep=self.enable_elastic_ep,
1814
            enable_dbo=self.enable_dbo,
1815
            ubatch_size=self.ubatch_size,
1816
            dbo_decode_token_threshold=self.dbo_decode_token_threshold,
1817
            dbo_prefill_token_threshold=self.dbo_prefill_token_threshold,
1818
            disable_nccl_for_dp_synchronization=self.disable_nccl_for_dp_synchronization,
1819
            enable_eplb=self.enable_eplb,
1820
            eplb_config=self.eplb_config,
1821
            expert_placement_strategy=self.expert_placement_strategy,
1822
1823
1824
            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,
1825
            ray_runtime_env=ray_runtime_env,
1826
            placement_group=placement_group,
1827
1828
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1829
            worker_extension_cls=self.worker_extension_cls,
1830
            decode_context_parallel_size=self.decode_context_parallel_size,
1831
            dcp_comm_backend=self.dcp_comm_backend,
1832
            dcp_kv_cache_interleave_size=self.dcp_kv_cache_interleave_size,
1833
            cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size,
1834
1835
            _api_process_count=self._api_process_count,
            _api_process_rank=self._api_process_rank,
1836
1837
1838
            numa_bind=self.numa_bind,
            numa_bind_nodes=self.numa_bind_nodes,
            numa_bind_cpus=self.numa_bind_cpus,
1839
        )
1840

1841
        speculative_config = self.create_speculative_config(
1842
1843
1844
1845
            target_model_config=model_config,
            target_parallel_config=parallel_config,
        )

1846
1847
1848
1849
1850
1851
        self._set_default_max_num_seqs_and_batched_tokens_args(
            usage_context,
            model_config,
            parallel_config,
        )

1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
        assert self.max_num_batched_tokens is not None, (
            "max_num_batched_tokens must be set by this point"
        )
        assert self.max_num_seqs is not None, "max_num_seqs must be set by this point"
        assert self.enable_chunked_prefill is not None, (
            "enable_chunked_prefill must be set by this point"
        )
        assert model_config.max_model_len is not None, (
            "max_model_len must be set by this point"
        )
1862
        scheduler_config = SchedulerConfig(
1863
            runner_type=model_config.runner_type,
1864
1865
1866
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1867
            enable_chunked_prefill=self.enable_chunked_prefill,
1868
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1869
            is_multimodal_model=model_config.is_multimodal_model,
1870
            is_encoder_decoder=model_config.is_encoder_decoder,
1871
            policy=self.scheduling_policy,
1872
            scheduler_cls=self.scheduler_cls,
1873
1874
1875
            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,
1876
            scheduler_reserve_full_isl=self.scheduler_reserve_full_isl,
1877
            disable_hybrid_kv_cache_manager=self.disable_hybrid_kv_cache_manager,
1878
            async_scheduling=self.async_scheduling,
1879
            stream_interval=self.stream_interval,
1880
        )
1881

1882
1883
1884
        if not model_config.is_multimodal_model and self.default_mm_loras:
            raise ValueError(
                "Default modality-specific LoRA(s) were provided for a "
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
                "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,
1895
                target_modules=self.lora_target_modules,
1896
                enable_tower_connector_lora=self.enable_tower_connector_lora,
1897
                specialize_active_lora=self.specialize_active_lora,
1898
1899
1900
1901
1902
1903
1904
                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
        )
1905

1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
        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"
            )

1920
1921
1922
1923
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1924
1925
1926
1927
1928
1929
1930
1931
        # Attention config overrides
        attention_config = copy.deepcopy(self.attention_config)
        if self.attention_backend is not None:
            if attention_config.backend is not None:
                raise ValueError(
                    "attention_backend and attention_config.backend "
                    "are mutually exclusive"
                )
1932
1933
1934
1935
            # Reuse the validator to handle "auto" and string-to-enum conversion
            attention_config.backend = AttentionConfig.validate_backend_before(
                self.attention_backend
            )
1936

1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
        # Kernel config overrides
        kernel_config = copy.deepcopy(self.kernel_config)
        if self.enable_flashinfer_autotune is not None:
            if kernel_config.enable_flashinfer_autotune is not None:
                raise ValueError(
                    "enable_flashinfer_autotune and "
                    "kernel_config.enable_flashinfer_autotune "
                    "are mutually exclusive"
                )
            kernel_config.enable_flashinfer_autotune = self.enable_flashinfer_autotune
1947
1948
        if self.moe_backend != "auto":
            kernel_config.moe_backend = self.moe_backend
1949

1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
        # Transfer top-level ir_op_priority into KernelConfig.ir_op_priority
        for op_name, op_priority in asdict(self.ir_op_priority).items():
            # Empty means unset
            if not op_priority:
                continue

            # Priority cannot be set 2x for the same op
            if getattr(kernel_config.ir_op_priority, op_name):
                raise ValueError(
                    f"Op priority for {op_name} specified via both ir_op_priority "
                    f"and KernelConfig.ir_op_priority, only one allowed at a time."
                )

            # Set the attribute
            setattr(kernel_config.ir_op_priority, op_name, op_priority)

1966
        load_config = self.create_load_config()
1967

1968
1969
        # Pass reasoning_parser into StructuredOutputsConfig
        if self.reasoning_parser:
1970
            self.structured_outputs_config.reasoning_parser = self.reasoning_parser
1971

1972
1973
1974
1975
1976
        if self.reasoning_parser_plugin:
            self.structured_outputs_config.reasoning_parser_plugin = (
                self.reasoning_parser_plugin
            )

1977
        observability_config = ObservabilityConfig(
1978
            show_hidden_metrics_for_version=self.show_hidden_metrics_for_version,
1979
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1980
            collect_detailed_traces=self.collect_detailed_traces,
1981
1982
            kv_cache_metrics=self.kv_cache_metrics,
            kv_cache_metrics_sample=self.kv_cache_metrics_sample,
1983
            cudagraph_metrics=self.cudagraph_metrics,
1984
            enable_layerwise_nvtx_tracing=self.enable_layerwise_nvtx_tracing,
1985
            enable_mfu_metrics=self.enable_mfu_metrics,
1986
            enable_mm_processor_stats=self.enable_mm_processor_stats,
1987
            enable_logging_iteration_details=self.enable_logging_iteration_details,
1988
        )
1989

1990
        # Compilation config overrides
1991
        compilation_config = copy.deepcopy(self.compilation_config)
1992
        if self.cudagraph_capture_sizes is not None:
1993
            if compilation_config.cudagraph_capture_sizes is not None:
1994
1995
1996
1997
                raise ValueError(
                    "cudagraph_capture_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
1998
            compilation_config.cudagraph_capture_sizes = self.cudagraph_capture_sizes
1999
        if self.max_cudagraph_capture_size is not None:
2000
            if compilation_config.max_cudagraph_capture_size is not None:
2001
2002
2003
2004
                raise ValueError(
                    "max_cudagraph_capture_size and compilation_config."
                    "max_cudagraph_capture_size are mutually exclusive"
                )
2005
            compilation_config.max_cudagraph_capture_size = (
2006
2007
                self.max_cudagraph_capture_size
            )
2008
2009

        offload_config = OffloadConfig(
2010
            offload_backend=self.offload_backend,
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
            uva=UVAOffloadConfig(
                cpu_offload_gb=self.cpu_offload_gb,
                cpu_offload_params=self.cpu_offload_params,
            ),
            prefetch=PrefetchOffloadConfig(
                offload_group_size=self.offload_group_size,
                offload_num_in_group=self.offload_num_in_group,
                offload_prefetch_step=self.offload_prefetch_step,
                offload_params=self.offload_params,
            ),
        )

2023
2024
2025
        if self.gdn_prefill_backend is not None:
            self.additional_config["gdn_prefill_backend"] = self.gdn_prefill_backend

2026
        config = VllmConfig(
2027
2028
2029
2030
2031
            model_config=model_config,
            cache_config=cache_config,
            parallel_config=parallel_config,
            scheduler_config=scheduler_config,
            device_config=device_config,
2032
            load_config=load_config,
2033
            offload_config=offload_config,
2034
            attention_config=attention_config,
2035
            kernel_config=kernel_config,
2036
2037
            lora_config=lora_config,
            speculative_config=speculative_config,
2038
            structured_outputs_config=self.structured_outputs_config,
2039
            observability_config=observability_config,
2040
            compilation_config=compilation_config,
2041
            kv_transfer_config=self.kv_transfer_config,
2042
            kv_events_config=self.kv_events_config,
2043
            ec_transfer_config=self.ec_transfer_config,
2044
            reasoning_config=self.reasoning_config,
2045
            profiler_config=self.profiler_config,
2046
            additional_config=self.additional_config,
2047
            optimization_level=self.optimization_level,
2048
            performance_mode=self.performance_mode,
2049
            weight_transfer_config=self.weight_transfer_config,
2050
            shutdown_timeout=self.shutdown_timeout,
2051
        )
2052

2053
2054
        return config

2055
    def _check_feature_supported(self):
2056
        """Raise an error if the feature is not supported."""
2057
        # No Concurrent Partial Prefills so far.
2058
2059
2060
2061
2062
        if (
            self.max_num_partial_prefills != SchedulerConfig.max_num_partial_prefills
            or self.max_long_partial_prefills
            != SchedulerConfig.max_long_partial_prefills
        ):
2063
            _raise_unsupported_error(feature_name="Concurrent Partial Prefill")
2064

2065
        if self.pipeline_parallel_size > 1:
2066
2067
2068
            supports_pp = getattr(
                self.distributed_executor_backend, "supports_pp", False
            )
2069
            if not supports_pp and self.distributed_executor_backend not in (
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
                ParallelConfig.distributed_executor_backend,
                "ray",
                "mp",
                "external_launcher",
            ):
                name = (
                    "Pipeline Parallelism without Ray distributed "
                    "executor or multiprocessing executor or external "
                    "launcher"
                )
2080
                _raise_unsupported_error(feature_name=name)
2081

2082
2083
2084
2085
2086
2087
    @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
2088

2089
2090
        default_max_num_batched_tokens: dict[UsageContext | None, int]
        default_max_num_seqs: dict[UsageContext | None, int]
2091

2092
2093
        # When no user override, set the default values based on the usage
        # context.
2094
        # Use different default values for different hardware.
2095
2096
2097
2098
2099
2100
2101

        # 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:
2102
            device_memory = current_platform.get_device_total_memory()
2103
            device_name = current_platform.get_device_name().lower()
2104
2105
        except Exception:
            # This is only used to set default_max_num_batched_tokens
2106
            device_memory = 0
2107
            device_name = ""
2108

2109
2110
2111
2112
        # 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:
2113
            # For GPUs like H100 and MI300x, use larger default values.
2114
2115
2116
2117
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
2118
2119
2120
2121
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
2122
2123
2124
2125
2126
2127
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
2128
2129
2130
2131
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
2132

2133
2134
        # tpu specific default values.
        if current_platform.is_tpu():
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
            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,
                }
2152

2153
2154
2155
        # cpu specific default values.
        if current_platform.is_cpu():
            default_max_num_batched_tokens = {
2156
2157
                UsageContext.LLM_CLASS: 4096 * world_size,
                UsageContext.OPENAI_API_SERVER: 2048 * world_size,
2158
2159
            }
            default_max_num_seqs = {
2160
2161
                UsageContext.LLM_CLASS: 256 * world_size,
                UsageContext.OPENAI_API_SERVER: 128 * world_size,
2162
2163
            }

2164
2165
        return default_max_num_batched_tokens, default_max_num_seqs

2166
2167
    def _set_default_chunked_prefill_and_prefix_caching_args(
        self, model_config: ModelConfig
2168
    ) -> None:
2169
2170
        default_chunked_prefill = model_config.is_chunked_prefill_supported
        default_prefix_caching = model_config.is_prefix_caching_supported
2171
2172
2173
2174
2175
2176
2177
2178

        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",
            )
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
        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",
            )
2190
2191
2192
2193
        elif (
            model_config.runner_type == "pooling"
            and self.enable_chunked_prefill
            and not default_chunked_prefill
2194
        ):
2195
            logger.warning_once(
2196
2197
2198
                "This model does not officially support chunked prefill. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
2199
                scope="local",
2200
2201
2202
2203
2204
            )

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

2205
            logger.debug(
2206
2207
2208
2209
2210
2211
2212
2213
                "%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
        ):
2214
            logger.warning_once(
2215
2216
2217
                "This model does not officially support prefix caching. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
2218
                scope="local",
2219
2220
            )

2221
        # Disable chunked prefill and prefix caching for:
2222
        # RISCV CPUs in V1
2223
2224
2225
2226
        if current_platform.is_cpu() and current_platform.get_cpu_architecture() in (
            CpuArchEnum.RISCV,
        ):
            logger.info(
2227
2228
                "Chunked prefill is not supported for"
                "RISC-V CPUs; "
2229
2230
2231
2232
                "disabling it for V1 backend."
            )
            self.enable_chunked_prefill = False
            logger.info(
2233
2234
                "Prefix caching is not supported for "
                "RISC-V CPUs; "
2235
2236
2237
2238
                "disabling it for V1 backend."
            )
            self.enable_prefix_caching = False

2239
2240
2241
2242
2243
2244
2245
    def _set_default_reasoning_config_args(self):
        if not self.reasoning_parser:
            return
        if self.reasoning_config is None:
            self.reasoning_config = ReasoningConfig()
        self.reasoning_config.reasoning_parser = self.reasoning_parser

2246
    def _set_default_max_num_seqs_and_batched_tokens_args(
2247
2248
2249
        self,
        usage_context: UsageContext | None,
        model_config: ModelConfig,
2250
        parallel_config: ParallelConfig,
2251
    ):
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
        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:
2262
2263
2264
2265
2266
2267
2268
2269
2270
            if parallel_config.use_batched_dp_moe:
                self.max_num_batched_tokens = (
                    SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS_FOR_BATCHED_DP
                )
            else:
                self.max_num_batched_tokens = default_max_num_batched_tokens.get(
                    usage_context,
                    SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS,
                )
2271
2272
2273
2274
2275
2276
2277

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

2278
2279
2280
2281
2282
2283
2284
        # If throughput mode is set, double max_num_batched_tokens and max_num_seqs.
        if self.performance_mode == "throughput":
            if orig_max_num_batched_tokens is None:
                self.max_num_batched_tokens *= 2
            if orig_max_num_seqs is None:
                self.max_num_seqs *= 2

2285
        if orig_max_num_batched_tokens is None:
2286
2287
2288
            assert model_config.max_model_len is not None, (
                "max_model_len must be set by this point"
            )
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
            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,
2301
2302
                self.max_num_batched_tokens,
            )
2303

2304
2305
2306
2307
            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,
2308
            )
2309

2310
2311
2312
2313
        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)

2314
            logger.debug(
2315
                "Defaulting max_num_seqs to %d for %s usage context.",
2316
                self.max_num_seqs,
2317
                usage_context.value if usage_context else None,
2318
            )
2319

2320

2321
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
2322
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
2323
    """Arguments for asynchronous vLLM engine."""
2324

2325
2326
    enable_log_requests: bool = False

2327
    @staticmethod
2328
2329
2330
    def add_cli_args(
        parser: FlexibleArgumentParser, async_args_only: bool = False
    ) -> FlexibleArgumentParser:
2331
        # Initialize plugin to update the parser, for example, The plugin may
2332
        # add a new kind of quantization method to --quantization argument or
2333
2334
        # a new device to --device argument.
        load_general_plugins()
2335
2336
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
2337
2338
2339
2340
        parser.add_argument(
            "--enable-log-requests",
            action=argparse.BooleanOptionalAction,
            default=AsyncEngineArgs.enable_log_requests,
2341
            help="Enable logging request information, dependent on log level:\n"
2342
2343
2344
            "- INFO: Request ID, parameters and LoRA request.\n"
            "- DEBUG: Prompt inputs (e.g: text, token IDs).\n"
            "You can set the minimum log level via `VLLM_LOGGING_LEVEL`.",
2345
        )
2346
        current_platform.pre_register_and_update(parser)
2347
        return parser
2348
2349


2350
2351
2352
2353
2354
2355
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)
2356
2357


2358
def human_readable_int(value: str) -> int:
2359
2360
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
2361

2362
2363
2364
2365
2366
2367
    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    """
    value = value.strip()
2368

2369
    match = re.fullmatch(r"(\d+(?:\.\d+)?)([kKmMgGtT])", value)
2370
2371
    if match:
        decimal_multiplier = {
2372
2373
2374
            "k": 10**3,
            "m": 10**6,
            "g": 10**9,
2375
            "t": 10**12,
2376
2377
        }
        binary_multiplier = {
2378
2379
2380
            "K": 2**10,
            "M": 2**20,
            "G": 2**30,
2381
            "T": 2**40,
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
        }

        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:
2394
2395
2396
2397
2398
                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
2399
2400
2401

    # Regular plain number.
    return int(value)
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420


def human_readable_int_or_auto(value: str) -> int:
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
    Also accepts -1 or 'auto' as a special value for auto-detection.

    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    - '-1' or 'auto' -> -1 (special value for auto-detection)
    """
    value = value.strip()

    if value == "-1" or value.lower() == "auto":
        return -1

    return human_readable_int(value)