arg_utils.py 96.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

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

34
import vllm.envs as envs
35
from vllm.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 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
115
if TYPE_CHECKING:
    from vllm.model_executor.layers.quantization import QuantizationMethods
116
    from vllm.model_executor.model_loader import LoadFormats
117
    from vllm.usage.usage_lib import UsageContext
118
    from vllm.v1.executor import Executor
119
else:
120
    Executor = Any
121
    QuantizationMethods = Any
122
    LoadFormats = Any
123
124
    UsageContext = Any

125

126
127
logger = init_logger(__name__)

128
129
# object is used to allow for special typing forms
T = TypeVar("T")
130
131
TypeHint: TypeAlias = type[Any] | object
TypeHintT: TypeAlias = type[T] | object
132

133

134
135
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
    def _parse_type(val: str) -> T:
136
137
138
139
        try:
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
140
141
                f"Value {val} cannot be converted to {return_type}."
            ) from e
142

143
144
145
    return _parse_type


146
147
def optional_type(return_type: Callable[[str], T]) -> Callable[[str], T | None]:
    def _optional_type(val: str) -> T | None:
148
149
150
151
        if val == "" or val == "None":
            return None
        return parse_type(return_type)(val)

152
    return _optional_type
153
154


155
def union_dict_and_str(val: str) -> str | dict[str, str] | None:
156
    if not re.match(r"(?s)^\s*{.*}\s*$", val):
157
        return str(val)
158
    return optional_type(json.loads)(val)
159
160


161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
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)


176
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
177
178
179
180
    """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`.
    """
181
    type_hint = get_type(type_hints, Literal)
182
183
184
    options = get_args(type_hint)
    option_type = type(options[0])
    if not all(isinstance(option, option_type) for option in options):
185
        raise ValueError(
186
            "All options must be of the same type. "
187
188
            f"Got {options} with types {[type(c) for c in options]}"
        )
189
190
    kwarg = "metavar" if contains_type(type_hints, str) else "choices"
    return {"type": option_type, kwarg: sorted(options)}
191
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
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),
    }


218
219
220
221
222
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


223
224
225
226
227
228
229
230
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]))
231
232
    elif origin in {Union, UnionType}:
        # Union for Union[X, Y] and UnionType for X | Y
233
234
235
236
237
238
239
240
        for arg in args:
            type_hints.update(get_type_hints(arg))
    else:
        type_hints.add(type_hint)

    return type_hints


241
NEEDS_HELP = (
242
243
    any("--help" in arg for arg in sys.argv)  # vllm SUBCOMMAND --help
    or (argv0 := sys.argv[0]).endswith("mkdocs")  # mkdocs SUBCOMMAND
244
245
246
247
    or argv0.endswith("mkdocs/__main__.py")  # python -m mkdocs SUBCOMMAND
)


248
249
250
251
252
253
254
255
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__}"


256
@functools.lru_cache(maxsize=30)
257
def _compute_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
258
259
    # Save time only getting attr docs if we're generating help text
    cls_docs = get_attr_docs(cls) if NEEDS_HELP else {}
260
261
    kwargs = {}
    for field in fields(cls):
262
        # Get the set of possible types for the field
263
        type_hints: set[TypeHint] = get_type_hints(field.type)
264
265
266
267
268

        # 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)

269
        # Get the default value of the field
270
271
        if field.default is not MISSING:
            default = field.default
272
273
            # Handle pydantic.Field defaults
            if isinstance(default, FieldInfo):
274
275
276
277
278
279
                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():
280
                        default = default.default_factory()  # type: ignore[call-arg]
281
        elif field.default_factory is not MISSING:
282
            default = field.default_factory()
283
284
285

        # Get the help text for the field
        name = field.name
286
        help = cls_docs.get(name, "").strip()
287
288
289
290
291
292
293
        # 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
294
295
296
        json_tip = (
            "Should either be a valid JSON string or JSON keys passed individually."
        )
297
        if dataclass_cls is not None:
298
299
300
301
302
303
304
305

            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
306
            kwargs[name]["help"] += _maybe_add_docs_url(dataclass_cls)
307
            kwargs[name]["help"] += f"\n\n{json_tip}"
308
        elif contains_type(type_hints, bool):
309
310
311
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
312
            kwargs[name].update(literal_to_kwargs(type_hints))
313
        elif contains_type(type_hints, tuple):
314
            kwargs[name].update(collection_to_kwargs(type_hints, tuple))
315
        elif contains_type(type_hints, list):
316
317
318
            kwargs[name].update(collection_to_kwargs(type_hints, list))
        elif contains_type(type_hints, set):
            kwargs[name].update(collection_to_kwargs(type_hints, set))
319
        elif contains_type(type_hints, int):
320
321
322
323
            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"):
324
                kwargs[name]["type"] = human_readable_int
325
                kwargs[name]["help"] += f"\n\n{human_readable_int.__doc__}"
326
327
            else:
                kwargs[name]["type"] = int
328
329
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
330
331
332
333
        elif contains_type(type_hints, dict) and (
            contains_type(type_hints, str)
            or any(is_not_builtin(th) for th in type_hints)
        ):
334
            kwargs[name]["type"] = union_dict_and_str
335
        elif contains_type(type_hints, dict):
336
            kwargs[name]["type"] = parse_type(json.loads)
337
            kwargs[name]["help"] += f"\n\n{json_tip}"
338
339
340
        elif contains_type(type_hints, str) or any(
            is_not_builtin(th) for th in type_hints
        ):
341
342
            kwargs[name]["type"] = str
        else:
343
            raise ValueError(f"Unsupported type {type_hints} for argument {name}.")
344

345
346
347
348
349
        # 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"]}))

350
351
352
353
354
355
356
        # 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
357
358


359
def get_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]:
360
361
    """Return argparse kwargs for the given Config dataclass.

362
363
364
    If `--help` or `mkdocs` are not present in the command line command, the
    attribute documentation will not be included in the help output.

365
366
367
368
369
370
371
    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))


372
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
373
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
374
    """Arguments for vLLM engine."""
375

376
    model: str = ModelConfig.model
377
    enable_return_routed_experts: bool = ModelConfig.enable_return_routed_experts
378
    model_weights: str = ModelConfig.model_weights
379
    served_model_name: str | list[str] | None = ModelConfig.served_model_name
380
    tokenizer: str | None = ModelConfig.tokenizer
381
    hf_config_path: str | None = ModelConfig.hf_config_path
382
383
    runner: RunnerOption = ModelConfig.runner
    convert: ConvertOption = ModelConfig.convert
384
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
385
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
386
    tokenizer_mode: TokenizerMode | str = ModelConfig.tokenizer_mode
387
    trust_remote_code: bool = ModelConfig.trust_remote_code
388
389
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
    allowed_media_domains: list[str] | None = ModelConfig.allowed_media_domains
390
    download_dir: str | None = LoadConfig.download_dir
391
    safetensors_load_strategy: str | None = LoadConfig.safetensors_load_strategy
392
    load_format: str | LoadFormats = LoadConfig.load_format
393
394
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
395
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
396
    seed: int = ModelConfig.seed
397
    max_model_len: int = ModelConfig.max_model_len
398
399
400
401
402
403
    cudagraph_capture_sizes: list[int] | None = (
        CompilationConfig.cudagraph_capture_sizes
    )
    max_cudagraph_capture_size: int | None = get_field(
        CompilationConfig, "max_cudagraph_capture_size"
    )
404
405
406
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
407
    distributed_executor_backend: (
408
        str | DistributedExecutorBackend | type[Executor] | None
409
    ) = ParallelConfig.distributed_executor_backend
410
    # number of P/D disaggregation (or other disaggregation) workers
411
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
412
413
414
415
    master_addr: str = ParallelConfig.master_addr
    master_port: int = ParallelConfig.master_port
    nnodes: int = ParallelConfig.nnodes
    node_rank: int = ParallelConfig.node_rank
416
    distributed_timeout_seconds: int | None = ParallelConfig.distributed_timeout_seconds
417
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
418
    prefill_context_parallel_size: int = ParallelConfig.prefill_context_parallel_size
419
    decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
420
    dcp_comm_backend: DCPCommBackend = ParallelConfig.dcp_comm_backend
421
    dcp_kv_cache_interleave_size: int = ParallelConfig.dcp_kv_cache_interleave_size
422
    cp_kv_cache_interleave_size: int = ParallelConfig.cp_kv_cache_interleave_size
423
    data_parallel_size: int = ParallelConfig.data_parallel_size
424
425
426
427
428
    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
429
    data_parallel_hybrid_lb: bool = False
430
    data_parallel_external_lb: bool = False
431
    data_parallel_backend: DataParallelBackend = ParallelConfig.data_parallel_backend
432
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
433
    enable_ep_weight_filter: bool = ParallelConfig.enable_ep_weight_filter
434
    moe_backend: MoEBackend = KernelConfig.moe_backend
435
    all2all_backend: All2AllBackend = ParallelConfig.all2all_backend
436
    enable_elastic_ep: bool = ParallelConfig.enable_elastic_ep
437
    enable_dbo: bool = ParallelConfig.enable_dbo
438
    ubatch_size: int = ParallelConfig.ubatch_size
439
440
    dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold
    dbo_prefill_token_threshold: int = ParallelConfig.dbo_prefill_token_threshold
441
    disable_nccl_for_dp_synchronization: bool | None = (
442
443
        ParallelConfig.disable_nccl_for_dp_synchronization
    )
444
    eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config")
445
    enable_eplb: bool = ParallelConfig.enable_eplb
446
    expert_placement_strategy: ExpertPlacementStrategy = (
447
        ParallelConfig.expert_placement_strategy
448
    )
449
450
    _api_process_count: int = ParallelConfig._api_process_count
    _api_process_rank: int = ParallelConfig._api_process_rank
451
    max_parallel_loading_workers: int | None = (
452
453
        ParallelConfig.max_parallel_loading_workers
    )
454
    block_size: int | None = None
455
    enable_prefix_caching: bool | None = None
456
    prefix_caching_hash_algo: PrefixCachingHashAlgo = (
457
        CacheConfig.prefix_caching_hash_algo
458
    )
459
460
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
461
462
463
464
465
466
467
    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")
468
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
469
    kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
470
    max_num_batched_tokens: int | None = None
471
472
    max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
    max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
473
    long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold
474
    max_num_seqs: int | None = None
475
    max_logprobs: int = ModelConfig.max_logprobs
476
    logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode
477
    disable_log_stats: bool = False
478
    aggregate_engine_logging: bool = False
479
480
481
    revision: str | None = ModelConfig.revision
    code_revision: str | None = ModelConfig.code_revision
    hf_token: bool | str | None = ModelConfig.hf_token
482
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
483
    tokenizer_revision: str | None = ModelConfig.tokenizer_revision
484
    quantization: QuantizationMethods | str | None = ModelConfig.quantization
485
    allow_deprecated_quantization: bool = ModelConfig.allow_deprecated_quantization
486
    enforce_eager: bool = ModelConfig.enforce_eager
487
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
488
    language_model_only: bool = MultiModalConfig.language_model_only
489
    limit_mm_per_prompt: dict[str, int | dict[str, int]] = get_field(
490
491
        MultiModalConfig, "limit_per_prompt"
    )
492
    enable_mm_embeds: bool = MultiModalConfig.enable_mm_embeds
493
    interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings
494
495
496
    media_io_kwargs: dict[str, dict[str, Any]] = get_field(
        MultiModalConfig, "media_io_kwargs"
    )
497
    mm_processor_kwargs: dict[str, Any] | None = MultiModalConfig.mm_processor_kwargs
498
    mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb
499
    mm_processor_cache_type: MMCacheType | None = (
500
        MultiModalConfig.mm_processor_cache_type
501
502
    )
    mm_shm_cache_max_object_size_mb: int = (
503
        MultiModalConfig.mm_shm_cache_max_object_size_mb
504
    )
505
    mm_encoder_only: bool = MultiModalConfig.mm_encoder_only
506
    mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
507
    mm_encoder_attn_backend: AttentionBackendEnum | str | None = (
508
509
        MultiModalConfig.mm_encoder_attn_backend
    )
510
    io_processor_plugin: str | None = None
511
    renderer_num_workers: int = 1
512
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
513
    video_pruning_rate: float | None = MultiModalConfig.video_pruning_rate
514
    mm_tensor_ipc: MMTensorIPC = MultiModalConfig.mm_tensor_ipc
515
    # LoRA fields
516
    enable_lora: bool = False
517
    max_loras: int = LoRAConfig.max_loras
518
    max_lora_rank: MaxLoRARanks = LoRAConfig.max_lora_rank
519
    default_mm_loras: dict[str, str] | None = LoRAConfig.default_mm_loras
520
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
521
522
    max_cpu_loras: int | None = LoRAConfig.max_cpu_loras
    lora_dtype: str | torch.dtype | None = LoRAConfig.lora_dtype
523
    lora_target_modules: list[str] | None = LoRAConfig.target_modules
524
    enable_tower_connector_lora: bool = LoRAConfig.enable_tower_connector_lora
525
    specialize_active_lora: bool = LoRAConfig.specialize_active_lora
526

527
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
528
    num_gpu_blocks_override: int | None = CacheConfig.num_gpu_blocks_override
529
    model_loader_extra_config: dict = get_field(LoadConfig, "model_loader_extra_config")
530
    ignore_patterns: str | list[str] = get_field(LoadConfig, "ignore_patterns")
531

532
    enable_chunked_prefill: bool | None = None
533
    disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
534

535
536
    scheduler_reserve_full_isl: bool = SchedulerConfig.scheduler_reserve_full_isl

537
    disable_hybrid_kv_cache_manager: bool | None = (
538
539
        SchedulerConfig.disable_hybrid_kv_cache_manager
    )
540

541
    structured_outputs_config: StructuredOutputsConfig = get_field(
542
543
        VllmConfig, "structured_outputs_config"
    )
544
    reasoning_parser: str = StructuredOutputsConfig.reasoning_parser
545
    reasoning_parser_plugin: str | None = None
546

547
    speculative_config: dict[str, Any] | None = None
548

549
    show_hidden_metrics_for_version: str | None = (
550
        ObservabilityConfig.show_hidden_metrics_for_version
551
    )
552
553
    otlp_traces_endpoint: str | None = ObservabilityConfig.otlp_traces_endpoint
    collect_detailed_traces: list[DetailedTraceModules] | None = (
554
        ObservabilityConfig.collect_detailed_traces
555
    )
556
557
558
559
    kv_cache_metrics: bool = ObservabilityConfig.kv_cache_metrics
    kv_cache_metrics_sample: float = get_field(
        ObservabilityConfig, "kv_cache_metrics_sample"
    )
560
    cudagraph_metrics: bool = ObservabilityConfig.cudagraph_metrics
561
562
563
    enable_layerwise_nvtx_tracing: bool = (
        ObservabilityConfig.enable_layerwise_nvtx_tracing
    )
564
    enable_mfu_metrics: bool = ObservabilityConfig.enable_mfu_metrics
565
566
567
    enable_logging_iteration_details: bool = (
        ObservabilityConfig.enable_logging_iteration_details
    )
568
    enable_mm_processor_stats: bool = ObservabilityConfig.enable_mm_processor_stats
569
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
570
    scheduler_cls: str | type[object] | None = SchedulerConfig.scheduler_cls
571

572
    pooler_config: PoolerConfig | None = ModelConfig.pooler_config
573
    compilation_config: CompilationConfig = get_field(VllmConfig, "compilation_config")
574
    attention_config: AttentionConfig = get_field(VllmConfig, "attention_config")
575
576
577
578
    kernel_config: KernelConfig = get_field(VllmConfig, "kernel_config")
    enable_flashinfer_autotune: bool = get_field(
        KernelConfig, "enable_flashinfer_autotune"
    )
579
580
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
581

582
583
    profiler_config: ProfilerConfig = get_field(VllmConfig, "profiler_config")

584
585
    kv_transfer_config: KVTransferConfig | None = None
    kv_events_config: KVEventsConfig | None = None
586

587
    ec_transfer_config: ECTransferConfig | None = None
588
    reasoning_config: ReasoningConfig = get_field(VllmConfig, "reasoning_config")
589

590
591
    generation_config: str = ModelConfig.generation_config
    enable_sleep_mode: bool = ModelConfig.enable_sleep_mode
592
593
594
    override_generation_config: dict[str, Any] = get_field(
        ModelConfig, "override_generation_config"
    )
595
    model_impl: str = ModelConfig.model_impl
596
    override_attention_dtype: str | None = ModelConfig.override_attention_dtype
597
    attention_backend: AttentionBackendEnum | None = AttentionConfig.backend
598

599
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
600
601
602
    kv_cache_dtype_skip_layers: list[str] = get_field(
        CacheConfig, "kv_cache_dtype_skip_layers"
    )
603
604
    mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype
    mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype
605
    mamba_block_size: int | None = get_field(CacheConfig, "mamba_block_size")
606
    mamba_cache_mode: MambaCacheMode = CacheConfig.mamba_cache_mode
607
608
609
610
    enable_mamba_cache_stochastic_rounding: bool = (
        CacheConfig.enable_mamba_cache_stochastic_rounding
    )
    mamba_cache_philox_rounds: int = CacheConfig.mamba_cache_philox_rounds
611

612
    additional_config: dict[str, Any] = get_field(VllmConfig, "additional_config")
613

614
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
615
    pt_load_map_location: str | dict[str, str] = LoadConfig.pt_load_map_location
616

617
    logits_processors: list[str | type[LogitsProcessor]] | None = (
618
619
        ModelConfig.logits_processors
    )
620
621
    """Custom logitproc types"""

622
    async_scheduling: bool | None = SchedulerConfig.async_scheduling
623

624
625
    stream_interval: int = SchedulerConfig.stream_interval

626
    kv_sharing_fast_prefill: bool = CacheConfig.kv_sharing_fast_prefill
627
    optimization_level: OptimizationLevel = VllmConfig.optimization_level
628
    performance_mode: PerformanceMode = VllmConfig.performance_mode
629

630
    kv_offloading_size: float | None = CacheConfig.kv_offloading_size
631
    kv_offloading_backend: KVOffloadingBackend = CacheConfig.kv_offloading_backend
632
    tokens_only: bool = False
633

634
635
    shutdown_timeout: int = 0

636
637
638
639
    weight_transfer_config: WeightTransferConfig | None = get_field(
        VllmConfig,
        "weight_transfer_config",
    )
640

641
    fail_on_environ_validation: bool = False
642
    gdn_prefill_backend: Literal["flashinfer", "triton"] | None = None
643

644
    def __post_init__(self):
645
646
647
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
648
        if isinstance(self.compilation_config, dict):
649
            self.compilation_config = CompilationConfig(**self.compilation_config)
650
651
        if isinstance(self.attention_config, dict):
            self.attention_config = AttentionConfig(**self.attention_config)
652
653
        if isinstance(self.kernel_config, dict):
            self.kernel_config = KernelConfig(**self.kernel_config)
654
        if isinstance(self.eplb_config, dict):
655
            self.eplb_config = EPLBConfig(**self.eplb_config)
656
657
658
659
        if isinstance(self.weight_transfer_config, dict):
            self.weight_transfer_config = WeightTransferConfig(
                **self.weight_transfer_config
            )
660
        # Setup plugins
661
        from vllm.plugins import load_general_plugins
662

663
        load_general_plugins()
664
        # when use hf offline,replace model and tokenizer id to local model path
665
666
667
        if huggingface_hub.constants.HF_HUB_OFFLINE:
            model_id = self.model
            self.model = get_model_path(self.model, self.revision)
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
            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,
                    )
684
685

    @staticmethod
686
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
687
        """Shared CLI arguments for vLLM engine."""
688

689
        # Model arguments
690
691
692
693
694
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
695
        if not ("serve" in sys.argv[1:] and "--help" in sys.argv[1:]):
696
            model_group.add_argument("--model", **model_kwargs["model"])
697
698
        model_group.add_argument("--runner", **model_kwargs["runner"])
        model_group.add_argument("--convert", **model_kwargs["convert"])
699
700
        model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"])
        model_group.add_argument("--tokenizer-mode", **model_kwargs["tokenizer_mode"])
701
702
703
        model_group.add_argument(
            "--trust-remote-code", **model_kwargs["trust_remote_code"]
        )
704
705
        model_group.add_argument("--dtype", **model_kwargs["dtype"])
        model_group.add_argument("--seed", **model_kwargs["seed"])
706
        model_group.add_argument("--hf-config-path", **model_kwargs["hf_config_path"])
707
708
709
710
711
712
        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"]
        )
713
        model_group.add_argument("--revision", **model_kwargs["revision"])
714
        model_group.add_argument("--code-revision", **model_kwargs["code_revision"])
715
716
717
        model_group.add_argument(
            "--tokenizer-revision", **model_kwargs["tokenizer_revision"]
        )
718
719
        model_group.add_argument("--max-model-len", **model_kwargs["max_model_len"])
        model_group.add_argument("--quantization", "-q", **model_kwargs["quantization"])
720
721
722
723
        model_group.add_argument(
            "--allow-deprecated-quantization",
            **model_kwargs["allow_deprecated_quantization"],
        )
724
        model_group.add_argument("--enforce-eager", **model_kwargs["enforce_eager"])
725
726
727
728
        model_group.add_argument(
            "--enable-return-routed-experts",
            **model_kwargs["enable_return_routed_experts"],
        )
729
730
731
732
733
734
735
736
        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"]
        )
737
738
739
        model_group.add_argument(
            "--skip-tokenizer-init", **model_kwargs["skip_tokenizer_init"]
        )
740
741
742
743
744
745
746
        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"])
747
748
        # This one is a special case because it can bool
        # or str. TODO: Handle this in get_kwargs
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
        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"]
        )
768
        model_group.add_argument("--model-impl", **model_kwargs["model_impl"])
769
770
771
772
773
774
        model_group.add_argument(
            "--override-attention-dtype", **model_kwargs["override_attention_dtype"]
        )
        model_group.add_argument(
            "--logits-processors", **model_kwargs["logits_processors"]
        )
775
776
        model_group.add_argument(
            "--io-processor-plugin", **model_kwargs["io_processor_plugin"]
777
        )
778
779
780
781
        model_group.add_argument(
            "--renderer-num-workers",
            **model_kwargs["renderer_num_workers"],
        )
782

783
784
785
786
787
788
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
789
        load_group.add_argument("--load-format", **load_kwargs["load_format"])
790
791
792
793
794
795
796
797
798
799
800
801
        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"]
        )
802

803
804
805
806
807
808
809
810
811
812
        # 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"]
        )

813
814
815
816
817
        # Structured outputs arguments
        structured_outputs_kwargs = get_kwargs(StructuredOutputsConfig)
        structured_outputs_group = parser.add_argument_group(
            title="StructuredOutputsConfig",
            description=StructuredOutputsConfig.__doc__,
818
        )
819
        structured_outputs_group.add_argument(
820
            "--reasoning-parser",
821
            # Choices need to be validated after parsing to include plugins
822
823
            **structured_outputs_kwargs["reasoning_parser"],
        )
824
825
826
827
        structured_outputs_group.add_argument(
            "--reasoning-parser-plugin",
            **structured_outputs_kwargs["reasoning_parser_plugin"],
        )
828

829
        # Parallel arguments
830
831
832
833
834
835
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
836
            "--distributed-executor-backend",
837
838
            **parallel_kwargs["distributed_executor_backend"],
        )
839
        parallel_group.add_argument(
840
841
842
843
            "--pipeline-parallel-size",
            "-pp",
            **parallel_kwargs["pipeline_parallel_size"],
        )
844
845
846
847
        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"])
848
849
850
851
        parallel_group.add_argument(
            "--distributed-timeout-seconds",
            **parallel_kwargs["distributed_timeout_seconds"],
        )
852
        parallel_group.add_argument(
853
854
            "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
        )
855
        parallel_group.add_argument(
856
857
858
859
            "--decode-context-parallel-size",
            "-dcp",
            **parallel_kwargs["decode_context_parallel_size"],
        )
860
861
862
863
        parallel_group.add_argument(
            "--dcp-comm-backend",
            **parallel_kwargs["dcp_comm_backend"],
        )
864
865
866
867
        parallel_group.add_argument(
            "--dcp-kv-cache-interleave-size",
            **parallel_kwargs["dcp_kv_cache_interleave_size"],
        )
868
869
870
871
872
873
874
875
876
        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"],
        )
877
878
879
880
881
882
        parallel_group.add_argument(
            "--data-parallel-size", "-dp", **parallel_kwargs["data_parallel_size"]
        )
        parallel_group.add_argument(
            "--data-parallel-rank",
            "-dpn",
883
            type=int,
884
885
886
            help="Data parallel rank of this instance. "
            "When set, enables external load balancer mode.",
        )
887
        parallel_group.add_argument(
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
            "--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".',
        )
918
        parallel_group.add_argument(
919
920
921
922
923
924
925
926
            "--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"],
927
928
        )
        parallel_group.add_argument(
929
930
931
            "--enable-expert-parallel",
            "-ep",
            **parallel_kwargs["enable_expert_parallel"],
932
        )
933
934
935
936
        parallel_group.add_argument(
            "--enable-ep-weight-filter",
            **parallel_kwargs["enable_ep_weight_filter"],
        )
937
938
939
        parallel_group.add_argument(
            "--all2all-backend", **parallel_kwargs["all2all_backend"]
        )
940
        parallel_group.add_argument("--enable-dbo", **parallel_kwargs["enable_dbo"])
941
942
943
944
        parallel_group.add_argument(
            "--ubatch-size",
            **parallel_kwargs["ubatch_size"],
        )
945
946
947
        parallel_group.add_argument(
            "--enable-elastic-ep", **parallel_kwargs["enable_elastic_ep"]
        )
948
949
        parallel_group.add_argument(
            "--dbo-decode-token-threshold",
950
951
            **parallel_kwargs["dbo_decode_token_threshold"],
        )
952
953
        parallel_group.add_argument(
            "--dbo-prefill-token-threshold",
954
955
            **parallel_kwargs["dbo_prefill_token_threshold"],
        )
956
957
958
959
        parallel_group.add_argument(
            "--disable-nccl-for-dp-synchronization",
            **parallel_kwargs["disable_nccl_for_dp_synchronization"],
        )
960
961
        parallel_group.add_argument("--enable-eplb", **parallel_kwargs["enable_eplb"])
        parallel_group.add_argument("--eplb-config", **parallel_kwargs["eplb_config"])
962
963
        parallel_group.add_argument(
            "--expert-placement-strategy",
964
965
            **parallel_kwargs["expert_placement_strategy"],
        )
966

967
        parallel_group.add_argument(
968
            "--max-parallel-loading-workers",
969
970
            **parallel_kwargs["max_parallel_loading_workers"],
        )
971
        parallel_group.add_argument(
972
973
            "--ray-workers-use-nsight", **parallel_kwargs["ray_workers_use_nsight"]
        )
974
        parallel_group.add_argument(
975
            "--disable-custom-all-reduce",
976
977
978
979
980
981
            **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"]
        )
982

983
984
985
986
987
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
988
        )
989
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
990
991
992
993
994
995
996
997
998
999
1000
        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(
1001
1002
1003
1004
1005
            "--enable-prefix-caching",
            **{
                **cache_kwargs["enable_prefix_caching"],
                "default": None,
            },
1006
1007
1008
1009
1010
1011
1012
        )
        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"]
        )
1013
1014
1015
        cache_group.add_argument(
            "--kv-cache-dtype-skip-layers", **cache_kwargs["kv_cache_dtype_skip_layers"]
        )
1016
1017
1018
1019
1020
1021
1022
1023
1024
        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"]
        )
1025
1026
1027
        cache_group.add_argument(
            "--mamba-block-size", **cache_kwargs["mamba_block_size"]
        )
1028
1029
1030
        cache_group.add_argument(
            "--mamba-cache-mode", **cache_kwargs["mamba_cache_mode"]
        )
1031
1032
1033
1034
1035
1036
1037
        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"]
        )
1038
1039
1040
1041
1042
1043
        cache_group.add_argument(
            "--kv-offloading-size", **cache_kwargs["kv_offloading_size"]
        )
        cache_group.add_argument(
            "--kv-offloading-backend", **cache_kwargs["kv_offloading_backend"]
        )
1044

1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
        # 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"]
        )

1076
        # Multimodal related configs
1077
1078
1079
1080
1081
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
1082
1083
1084
        multimodal_group.add_argument(
            "--language-model-only", **multimodal_kwargs["language_model_only"]
        )
1085
        multimodal_group.add_argument(
1086
1087
            "--limit-mm-per-prompt", **multimodal_kwargs["limit_per_prompt"]
        )
1088
1089
1090
        multimodal_group.add_argument(
            "--enable-mm-embeds", **multimodal_kwargs["enable_mm_embeds"]
        )
1091
1092
1093
        multimodal_group.add_argument(
            "--media-io-kwargs", **multimodal_kwargs["media_io_kwargs"]
        )
1094
1095
1096
1097
1098
1099
        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"]
        )
1100
        multimodal_group.add_argument(
1101
1102
            "--mm-processor-cache-type", **multimodal_kwargs["mm_processor_cache_type"]
        )
1103
1104
        multimodal_group.add_argument(
            "--mm-shm-cache-max-object-size-mb",
1105
1106
            **multimodal_kwargs["mm_shm_cache_max_object_size_mb"],
        )
1107
1108
1109
        multimodal_group.add_argument(
            "--mm-encoder-only", **multimodal_kwargs["mm_encoder_only"]
        )
1110
        multimodal_group.add_argument(
1111
1112
            "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"]
        )
1113
1114
1115
1116
        multimodal_group.add_argument(
            "--mm-encoder-attn-backend",
            **multimodal_kwargs["mm_encoder_attn_backend"],
        )
1117
1118
1119
        multimodal_group.add_argument(
            "--interleave-mm-strings", **multimodal_kwargs["interleave_mm_strings"]
        )
1120
        multimodal_group.add_argument(
1121
1122
            "--skip-mm-profiling", **multimodal_kwargs["skip_mm_profiling"]
        )
1123

1124
        multimodal_group.add_argument(
1125
1126
            "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"]
        )
1127
1128
1129
        multimodal_group.add_argument(
            "--mm-tensor-ipc", **multimodal_kwargs["mm_tensor_ipc"]
        )
1130

1131
        # LoRA related configs
1132
1133
1134
1135
1136
1137
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
1138
            "--enable-lora",
1139
            action=argparse.BooleanOptionalAction,
1140
1141
            help="If True, enable handling of LoRA adapters.",
        )
1142
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
1143
        lora_group.add_argument("--max-lora-rank", **lora_kwargs["max_lora_rank"])
1144
        lora_group.add_argument(
1145
            "--lora-dtype",
1146
1147
            **lora_kwargs["lora_dtype"],
        )
1148
1149
1150
1151
        lora_group.add_argument(
            "--enable-tower-connector-lora",
            **lora_kwargs["enable_tower_connector_lora"],
        )
1152
1153
1154
1155
        lora_group.add_argument("--max-cpu-loras", **lora_kwargs["max_cpu_loras"])
        lora_group.add_argument(
            "--fully-sharded-loras", **lora_kwargs["fully_sharded_loras"]
        )
1156
1157
1158
        lora_group.add_argument(
            "--lora-target-modules", **lora_kwargs["target_modules"]
        )
1159
        lora_group.add_argument("--default-mm-loras", **lora_kwargs["default_mm_loras"])
1160
1161
1162
        lora_group.add_argument(
            "--specialize-active-lora", **lora_kwargs["specialize_active_lora"]
        )
1163

1164
1165
1166
1167
1168
1169
1170
1171
        # 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",
1172
1173
            **observability_kwargs["show_hidden_metrics_for_version"],
        )
1174
        observability_group.add_argument(
1175
1176
            "--otlp-traces-endpoint", **observability_kwargs["otlp_traces_endpoint"]
        )
1177
1178
1179
1180
1181
        # 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"] += [
1182
            ",".join(p) for p in permutations(get_args(DetailedTraceModules), r=2)
1183
1184
1185
        ]
        observability_group.add_argument(
            "--collect-detailed-traces",
1186
1187
            **observability_kwargs["collect_detailed_traces"],
        )
1188
1189
1190
1191
1192
1193
1194
        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"],
        )
1195
1196
1197
1198
        observability_group.add_argument(
            "--cudagraph-metrics",
            **observability_kwargs["cudagraph_metrics"],
        )
1199
1200
1201
1202
        observability_group.add_argument(
            "--enable-layerwise-nvtx-tracing",
            **observability_kwargs["enable_layerwise_nvtx_tracing"],
        )
1203
1204
1205
1206
        observability_group.add_argument(
            "--enable-mfu-metrics",
            **observability_kwargs["enable_mfu_metrics"],
        )
1207
1208
1209
1210
        observability_group.add_argument(
            "--enable-logging-iteration-details",
            **observability_kwargs["enable_logging_iteration_details"],
        )
1211

1212
1213
1214
1215
1216
1217
1218
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
1219
1220
1221
1222
1223
            "--max-num-batched-tokens",
            **{
                **scheduler_kwargs["max_num_batched_tokens"],
                "default": None,
            },
1224
        )
1225
        scheduler_group.add_argument(
1226
1227
1228
1229
1230
            "--max-num-seqs",
            **{
                **scheduler_kwargs["max_num_seqs"],
                "default": None,
            },
1231
1232
1233
1234
        )
        scheduler_group.add_argument(
            "--max-num-partial-prefills", **scheduler_kwargs["max_num_partial_prefills"]
        )
1235
1236
        scheduler_group.add_argument(
            "--max-long-partial-prefills",
1237
1238
            **scheduler_kwargs["max_long_partial_prefills"],
        )
1239
1240
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
1241
1242
            **scheduler_kwargs["long_prefill_token_threshold"],
        )
1243
1244
        # multi-step scheduling has been removed; corresponding arguments
        # are no longer supported.
1245
        scheduler_group.add_argument(
1246
1247
            "--scheduling-policy", **scheduler_kwargs["policy"]
        )
1248
        scheduler_group.add_argument(
1249
1250
1251
1252
1253
            "--enable-chunked-prefill",
            **{
                **scheduler_kwargs["enable_chunked_prefill"],
                "default": None,
            },
1254
1255
1256
1257
1258
1259
1260
        )
        scheduler_group.add_argument(
            "--disable-chunked-mm-input", **scheduler_kwargs["disable_chunked_mm_input"]
        )
        scheduler_group.add_argument(
            "--scheduler-cls", **scheduler_kwargs["scheduler_cls"]
        )
1261
1262
1263
1264
        scheduler_group.add_argument(
            "--scheduler-reserve-full-isl",
            **scheduler_kwargs["scheduler_reserve_full_isl"],
        )
1265
1266
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
1267
1268
1269
1270
1271
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"],
        )
        scheduler_group.add_argument(
            "--async-scheduling", **scheduler_kwargs["async_scheduling"]
        )
1272
1273
1274
        scheduler_group.add_argument(
            "--stream-interval", **scheduler_kwargs["stream_interval"]
        )
1275

1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
        # 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"],
        )

1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
        # Kernel arguments
        kernel_kwargs = get_kwargs(KernelConfig)
        kernel_group = parser.add_argument_group(
            title="KernelConfig",
            description=KernelConfig.__doc__,
        )
        kernel_group.add_argument(
            "--enable-flashinfer-autotune",
            **kernel_kwargs["enable_flashinfer_autotune"],
        )
1300
1301
1302
        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)
1303

1304
        # vLLM arguments
1305
        vllm_kwargs = get_kwargs(VllmConfig)
1306
1307
1308
1309
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
1310
1311
1312
1313
        # 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)
1314
        vllm_group.add_argument(
1315
            "--speculative-config", "-sc", **vllm_kwargs["speculative_config"]
1316
1317
1318
1319
1320
        )
        vllm_group.add_argument(
            "--kv-transfer-config", **vllm_kwargs["kv_transfer_config"]
        )
        vllm_group.add_argument("--kv-events-config", **vllm_kwargs["kv_events_config"])
1321
1322
1323
        vllm_group.add_argument(
            "--ec-transfer-config", **vllm_kwargs["ec_transfer_config"]
        )
1324
        vllm_group.add_argument(
1325
            "--compilation-config", "-cc", **vllm_kwargs["compilation_config"]
1326
        )
1327
1328
1329
        vllm_group.add_argument(
            "--attention-config", "-ac", **vllm_kwargs["attention_config"]
        )
1330
        vllm_group.add_argument("--reasoning-config", **vllm_kwargs["reasoning_config"])
1331
        vllm_group.add_argument("--kernel-config", **vllm_kwargs["kernel_config"])
1332
1333
1334
1335
1336
1337
        vllm_group.add_argument(
            "--additional-config", **vllm_kwargs["additional_config"]
        )
        vllm_group.add_argument(
            "--structured-outputs-config", **vllm_kwargs["structured_outputs_config"]
        )
1338
        vllm_group.add_argument("--profiler-config", **vllm_kwargs["profiler_config"])
1339
1340
1341
        vllm_group.add_argument(
            "--optimization-level", **vllm_kwargs["optimization_level"]
        )
1342
        vllm_group.add_argument("--performance-mode", **vllm_kwargs["performance_mode"])
1343
1344
1345
        vllm_group.add_argument(
            "--weight-transfer-config", **vllm_kwargs["weight_transfer_config"]
        )
1346

1347
        # Other arguments
1348
1349
1350
1351
1352
        parser.add_argument(
            "--disable-log-stats",
            action="store_true",
            help="Disable logging statistics.",
        )
1353

1354
1355
1356
1357
1358
1359
        parser.add_argument(
            "--aggregate-engine-logging",
            action="store_true",
            help="Log aggregate rather than per-engine statistics "
            "when using data parallelism.",
        )
1360
1361
1362
1363
1364
1365
1366
1367

        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,
        )
1368
1369
1370
1371
1372
1373
1374
1375

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

1376
1377
1378
1379
1380
1381
1382
        parser.add_argument(
            "--gdn-prefill-backend",
            dest="gdn_prefill_backend",
            choices=["flashinfer", "triton"],
            default=None,
            help="Select GDN prefill backend.",
        )
1383
        return parser
1384
1385

    @classmethod
1386
    def from_cli_args(cls, args: argparse.Namespace):
1387
1388
1389
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
1390
1391
1392
        engine_args = cls(
            **{attr: getattr(args, attr) for attr in attrs if hasattr(args, attr)}
        )
Zhuohan Li's avatar
Zhuohan Li committed
1393
        return engine_args
1394

1395
    def create_model_config(self) -> ModelConfig:
1396
1397
        # gguf file needs a specific model loader
        if is_gguf(self.model):
1398
1399
            self.quantization = self.load_format = "gguf"

1400
1401
1402
1403
1404
1405
1406
        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,
1407
1408
            )

1409
        return ModelConfig(
1410
            model=self.model,
1411
            model_weights=self.model_weights,
1412
            hf_config_path=self.hf_config_path,
1413
1414
            runner=self.runner,
            convert=self.convert,
1415
            tokenizer=self.tokenizer,  # type: ignore[arg-type]
1416
            tokenizer_mode=self.tokenizer_mode,
1417
            trust_remote_code=self.trust_remote_code,
1418
1419
            allowed_local_media_path=self.allowed_local_media_path,
            allowed_media_domains=self.allowed_media_domains,
1420
1421
1422
1423
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
1424
            hf_token=self.hf_token,
1425
            hf_overrides=self.hf_overrides,
1426
            tokenizer_revision=self.tokenizer_revision,
1427
1428
            max_model_len=self.max_model_len,
            quantization=self.quantization,
1429
            allow_deprecated_quantization=self.allow_deprecated_quantization,
1430
            enforce_eager=self.enforce_eager,
1431
            enable_return_routed_experts=self.enable_return_routed_experts,
1432
            max_logprobs=self.max_logprobs,
1433
            logprobs_mode=self.logprobs_mode,
1434
            disable_sliding_window=self.disable_sliding_window,
1435
            disable_cascade_attn=self.disable_cascade_attn,
1436
            skip_tokenizer_init=self.skip_tokenizer_init,
1437
            enable_prompt_embeds=self.enable_prompt_embeds,
1438
            served_model_name=self.served_model_name,
1439
            language_model_only=self.language_model_only,
1440
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1441
            enable_mm_embeds=self.enable_mm_embeds,
1442
            interleave_mm_strings=self.interleave_mm_strings,
1443
            media_io_kwargs=self.media_io_kwargs,
1444
            skip_mm_profiling=self.skip_mm_profiling,
1445
            config_format=self.config_format,
1446
            mm_processor_kwargs=self.mm_processor_kwargs,
1447
            mm_processor_cache_gb=self.mm_processor_cache_gb,
1448
            mm_processor_cache_type=self.mm_processor_cache_type,
1449
            mm_shm_cache_max_object_size_mb=self.mm_shm_cache_max_object_size_mb,
1450
            mm_encoder_only=self.mm_encoder_only,
1451
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
1452
            mm_encoder_attn_backend=self.mm_encoder_attn_backend,
1453
            pooler_config=self.pooler_config,
1454
            generation_config=self.generation_config,
1455
            override_generation_config=self.override_generation_config,
1456
            enable_sleep_mode=self.enable_sleep_mode,
1457
            model_impl=self.model_impl,
1458
            override_attention_dtype=self.override_attention_dtype,
1459
            logits_processors=self.logits_processors,
1460
            video_pruning_rate=self.video_pruning_rate,
1461
            mm_tensor_ipc=self.mm_tensor_ipc,
1462
            io_processor_plugin=self.io_processor_plugin,
1463
            renderer_num_workers=self.renderer_num_workers,
1464
        )
1465

1466
    def validate_tensorizer_args(self):
1467
1468
        from vllm.model_executor.model_loader.tensorizer import TensorizerConfig

1469
1470
        for key in self.model_loader_extra_config:
            if key in TensorizerConfig._fields:
1471
1472
1473
                self.model_loader_extra_config["tensorizer_config"][key] = (
                    self.model_loader_extra_config[key]
                )
1474

1475
    def create_load_config(self) -> LoadConfig:
1476
1477
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1478

1479
1480
1481
        if self.load_format == "tensorizer":
            if hasattr(self.model_loader_extra_config, "to_serializable"):
                self.model_loader_extra_config = (
1482
1483
                    self.model_loader_extra_config.to_serializable()
                )
1484
            self.model_loader_extra_config["tensorizer_config"] = {}
1485
1486
1487
            self.model_loader_extra_config["tensorizer_config"]["tensorizer_dir"] = (
                self.model
            )
1488
            self.validate_tensorizer_args()
1489

1490
1491
1492
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
1493
            safetensors_load_strategy=self.safetensors_load_strategy,
1494
1495
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1496
            use_tqdm_on_load=self.use_tqdm_on_load,
1497
            pt_load_map_location=self.pt_load_map_location,
1498
        )
1499

1500
1501
1502
1503
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
1504
    ) -> SpeculativeConfig | None:
1505
1506
1507
1508
1509
1510
        """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
1511
        dictionary from the engine.
1512
1513
        """
        if self.speculative_config is None:
1514
            return None
1515

1516
1517
1518
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.
1519
1520
1521
1522
1523
1524
        self.speculative_config.update(
            {
                "target_model_config": target_model_config,
                "target_parallel_config": target_parallel_config,
            }
        )
1525
        return SpeculativeConfig(**self.speculative_config)
1526

1527
1528
    def create_engine_config(
        self,
1529
        usage_context: UsageContext | None = None,
1530
        headless: bool = False,
1531
1532
1533
1534
    ) -> VllmConfig:
        """
        Create the VllmConfig.

1535
        NOTE: If VllmConfig is incompatible, we raise an error.
1536
        """
1537
        current_platform.pre_register_and_update()
1538

1539
        device_config = DeviceConfig(device=cast(Device, current_platform.device_type))
1540

1541
1542
        envs.validate_environ(self.fail_on_environ_validation)

1543
1544
        # Check if the model is a speculator and override model/tokenizer/config
        # BEFORE creating ModelConfig, so the config is created with the target model
1545
1546
1547
1548
        # 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):
1549
1550
1551
1552
1553
1554
1555
            (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,
1556
                    hf_token=self.hf_token,
1557
1558
1559
                )
            )

1560
        model_config = self.create_model_config()
1561
        self.model = model_config.model
1562
        self.model_weights = model_config.model_weights
1563
1564
        self.tokenizer = model_config.tokenizer

1565
        self._check_feature_supported()
1566
1567
1568
1569
        self._set_default_chunked_prefill_and_prefix_caching_args(model_config)
        self._set_default_max_num_seqs_and_batched_tokens_args(
            usage_context, model_config
        )
1570

1571
        sliding_window: int | None = None
1572
1573
1574
1575
1576
1577
        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()

1578
1579
1580
1581
1582
        # 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
        )

1583
1584
1585
1586
        assert self.enable_prefix_caching is not None, (
            "enable_prefix_caching must be set by this point"
        )

1587
        cache_config = CacheConfig(
1588
            block_size=self.block_size,  # type: ignore[arg-type]
1589
            gpu_memory_utilization=self.gpu_memory_utilization,
1590
            kv_cache_memory_bytes=self.kv_cache_memory_bytes,
1591
            cache_dtype=resolved_cache_dtype,  # type: ignore[arg-type]
1592
            is_attention_free=model_config.is_attention_free,
1593
            num_gpu_blocks_override=self.num_gpu_blocks_override,
1594
            sliding_window=sliding_window,
1595
            enable_prefix_caching=self.enable_prefix_caching,
1596
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1597
            calculate_kv_scales=self.calculate_kv_scales,
1598
            kv_cache_dtype_skip_layers=self.kv_cache_dtype_skip_layers,
1599
            kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
1600
1601
            mamba_cache_dtype=self.mamba_cache_dtype,
            mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype,
1602
            mamba_block_size=self.mamba_block_size,
1603
            mamba_cache_mode=self.mamba_cache_mode,
1604
1605
            enable_mamba_cache_stochastic_rounding=self.enable_mamba_cache_stochastic_rounding,
            mamba_cache_philox_rounds=self.mamba_cache_philox_rounds,
1606
1607
            kv_offloading_size=self.kv_offloading_size,
            kv_offloading_backend=self.kv_offloading_backend,
1608
        )
1609

1610
1611
1612
1613
1614
1615
        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
1616

1617
            ray_runtime_env = ray.get_runtime_context().runtime_env
1618
1619
1620
1621
1622
1623
1624
            # 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)
1625

1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
        # 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()

1637
        assert not headless or not self.data_parallel_hybrid_lb, (
1638
1639
            "data_parallel_hybrid_lb is not applicable in headless mode"
        )
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
        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
        )
1681
        # Local DP rank = 1, use pure-external LB.
1682
        if data_parallel_external_lb:
1683
            assert self.data_parallel_rank is not None, (
1684
                "data_parallel_rank or node_rank must be specified if "
1685
1686
                "data_parallel_external_lb is enable."
            )
1687
            assert self.data_parallel_size_local in (1, None), (
1688
1689
                "data_parallel_size_local must be 1 or None when data_parallel_rank "
                "is set"
1690
            )
1691
            data_parallel_size_local = 1
1692
1693
            # Use full external lb if we have local_size of 1.
            self.data_parallel_hybrid_lb = False
1694
1695
        elif self.data_parallel_size_local is not None:
            data_parallel_size_local = self.data_parallel_size_local
1696
1697
1698
1699
1700
1701
1702

            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.
1703
1704
1705
1706
1707
                logger.warning(
                    "data_parallel_hybrid_lb is not eligible when "
                    "data_parallel_size_local = 1, autoswitch to "
                    "data_parallel_external_lb."
                )
1708
1709
1710
1711
1712
1713
1714
                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

1715
1716
1717
1718
1719
1720
1721
1722
1723
            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,
                )
1724
        else:
1725
            assert not self.data_parallel_hybrid_lb, (
1726
1727
                "data_parallel_size_local must be set to use data_parallel_hybrid_lb."
            )
1728

1729
1730
1731
1732
1733
1734
1735
1736
1737
            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
1738
1739
1740

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1741
1742
1743
1744
        if self.data_parallel_address is None:
            if self.data_parallel_backend == "ray":
                host_ip = get_ip()
                logger.info(
1745
1746
                    "Using host IP %s as ray-based data parallel address", host_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1747
1748
1749
1750
                data_parallel_address = host_ip
            else:
                assert self.data_parallel_backend == "mp", (
                    "data_parallel_backend can only be ray or mp, got %s",
1751
1752
                    self.data_parallel_backend,
                )
1753
1754
1755
                data_parallel_address = (
                    self.master_addr or ParallelConfig.data_parallel_master_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1756
1757
        else:
            data_parallel_address = self.data_parallel_address
1758
1759
1760

        # This port is only used when there are remote data parallel engines,
        # otherwise the local IPC transport is used.
1761
        data_parallel_rpc_port = (
1762
            self.data_parallel_rpc_port
1763
1764
1765
            if (self.data_parallel_rpc_port is not None)
            else ParallelConfig.data_parallel_rpc_port
        )
1766

1767
1768
1769
1770
        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.")

1771
        parallel_config = ParallelConfig(
1772
1773
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1774
            prefill_context_parallel_size=self.prefill_context_parallel_size,
1775
            data_parallel_size=self.data_parallel_size,
1776
1777
            data_parallel_rank=self.data_parallel_rank or 0,
            data_parallel_external_lb=data_parallel_external_lb,
1778
            data_parallel_size_local=data_parallel_size_local,
1779
1780
1781
1782
            master_addr=self.master_addr,
            master_port=self.master_port,
            nnodes=self.nnodes,
            node_rank=self.node_rank,
1783
            distributed_timeout_seconds=self.distributed_timeout_seconds,
1784
1785
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
1786
            data_parallel_backend=self.data_parallel_backend,
1787
            data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
1788
            is_moe_model=model_config.is_moe,
1789
            enable_expert_parallel=self.enable_expert_parallel,
1790
            enable_ep_weight_filter=self.enable_ep_weight_filter,
1791
            all2all_backend=self.all2all_backend,
1792
            enable_elastic_ep=self.enable_elastic_ep,
1793
            enable_dbo=self.enable_dbo,
1794
            ubatch_size=self.ubatch_size,
1795
            dbo_decode_token_threshold=self.dbo_decode_token_threshold,
1796
            dbo_prefill_token_threshold=self.dbo_prefill_token_threshold,
1797
            disable_nccl_for_dp_synchronization=self.disable_nccl_for_dp_synchronization,
1798
            enable_eplb=self.enable_eplb,
1799
            eplb_config=self.eplb_config,
1800
            expert_placement_strategy=self.expert_placement_strategy,
1801
1802
1803
            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,
1804
            ray_runtime_env=ray_runtime_env,
1805
            placement_group=placement_group,
1806
1807
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1808
            worker_extension_cls=self.worker_extension_cls,
1809
            decode_context_parallel_size=self.decode_context_parallel_size,
1810
            dcp_comm_backend=self.dcp_comm_backend,
1811
            dcp_kv_cache_interleave_size=self.dcp_kv_cache_interleave_size,
1812
            cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size,
1813
1814
            _api_process_count=self._api_process_count,
            _api_process_rank=self._api_process_rank,
1815
        )
1816

1817
        speculative_config = self.create_speculative_config(
1818
1819
1820
1821
            target_model_config=model_config,
            target_parallel_config=parallel_config,
        )

1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
        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"
        )
1832
        scheduler_config = SchedulerConfig(
1833
            runner_type=model_config.runner_type,
1834
1835
1836
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1837
            enable_chunked_prefill=self.enable_chunked_prefill,
1838
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1839
            is_multimodal_model=model_config.is_multimodal_model,
1840
            is_encoder_decoder=model_config.is_encoder_decoder,
1841
            policy=self.scheduling_policy,
1842
            scheduler_cls=self.scheduler_cls,
1843
1844
1845
            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,
1846
            scheduler_reserve_full_isl=self.scheduler_reserve_full_isl,
1847
            disable_hybrid_kv_cache_manager=self.disable_hybrid_kv_cache_manager,
1848
            async_scheduling=self.async_scheduling,
1849
            stream_interval=self.stream_interval,
1850
        )
1851

1852
1853
1854
        if not model_config.is_multimodal_model and self.default_mm_loras:
            raise ValueError(
                "Default modality-specific LoRA(s) were provided for a "
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
                "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,
1865
                target_modules=self.lora_target_modules,
1866
                enable_tower_connector_lora=self.enable_tower_connector_lora,
1867
                specialize_active_lora=self.specialize_active_lora,
1868
1869
1870
1871
1872
1873
1874
                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
        )
1875

1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
        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"
            )

1890
1891
1892
1893
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1894
1895
1896
1897
1898
1899
1900
1901
        # 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"
                )
1902
1903
1904
1905
            # Reuse the validator to handle "auto" and string-to-enum conversion
            attention_config.backend = AttentionConfig.validate_backend_before(
                self.attention_backend
            )
1906

1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
        # 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
1917
1918
        if self.moe_backend != "auto":
            kernel_config.moe_backend = self.moe_backend
1919

1920
        load_config = self.create_load_config()
1921

1922
1923
        # Pass reasoning_parser into StructuredOutputsConfig
        if self.reasoning_parser:
1924
            self.structured_outputs_config.reasoning_parser = self.reasoning_parser
1925

1926
1927
1928
1929
1930
        if self.reasoning_parser_plugin:
            self.structured_outputs_config.reasoning_parser_plugin = (
                self.reasoning_parser_plugin
            )

1931
        observability_config = ObservabilityConfig(
1932
            show_hidden_metrics_for_version=self.show_hidden_metrics_for_version,
1933
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1934
            collect_detailed_traces=self.collect_detailed_traces,
1935
1936
            kv_cache_metrics=self.kv_cache_metrics,
            kv_cache_metrics_sample=self.kv_cache_metrics_sample,
1937
            cudagraph_metrics=self.cudagraph_metrics,
1938
            enable_layerwise_nvtx_tracing=self.enable_layerwise_nvtx_tracing,
1939
            enable_mfu_metrics=self.enable_mfu_metrics,
1940
            enable_mm_processor_stats=self.enable_mm_processor_stats,
1941
            enable_logging_iteration_details=self.enable_logging_iteration_details,
1942
        )
1943

1944
        # Compilation config overrides
1945
        compilation_config = copy.deepcopy(self.compilation_config)
1946
        if self.cudagraph_capture_sizes is not None:
1947
            if compilation_config.cudagraph_capture_sizes is not None:
1948
1949
1950
1951
                raise ValueError(
                    "cudagraph_capture_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
1952
            compilation_config.cudagraph_capture_sizes = self.cudagraph_capture_sizes
1953
        if self.max_cudagraph_capture_size is not None:
1954
            if compilation_config.max_cudagraph_capture_size is not None:
1955
1956
1957
1958
                raise ValueError(
                    "max_cudagraph_capture_size and compilation_config."
                    "max_cudagraph_capture_size are mutually exclusive"
                )
1959
            compilation_config.max_cudagraph_capture_size = (
1960
1961
                self.max_cudagraph_capture_size
            )
1962
1963

        offload_config = OffloadConfig(
1964
            offload_backend=self.offload_backend,
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
            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,
            ),
        )

1977
1978
1979
        if self.gdn_prefill_backend is not None:
            self.additional_config["gdn_prefill_backend"] = self.gdn_prefill_backend

1980
        config = VllmConfig(
1981
1982
1983
1984
1985
            model_config=model_config,
            cache_config=cache_config,
            parallel_config=parallel_config,
            scheduler_config=scheduler_config,
            device_config=device_config,
1986
            load_config=load_config,
1987
            offload_config=offload_config,
1988
            attention_config=attention_config,
1989
            kernel_config=kernel_config,
1990
1991
            lora_config=lora_config,
            speculative_config=speculative_config,
1992
            structured_outputs_config=self.structured_outputs_config,
1993
            observability_config=observability_config,
1994
            compilation_config=compilation_config,
1995
            kv_transfer_config=self.kv_transfer_config,
1996
            kv_events_config=self.kv_events_config,
1997
            ec_transfer_config=self.ec_transfer_config,
1998
            reasoning_config=self.reasoning_config,
1999
            profiler_config=self.profiler_config,
2000
            additional_config=self.additional_config,
2001
            optimization_level=self.optimization_level,
2002
            performance_mode=self.performance_mode,
2003
            weight_transfer_config=self.weight_transfer_config,
2004
            shutdown_timeout=self.shutdown_timeout,
2005
        )
2006

2007
2008
        return config

2009
    def _check_feature_supported(self):
2010
        """Raise an error if the feature is not supported."""
2011
        # No Concurrent Partial Prefills so far.
2012
2013
2014
2015
2016
        if (
            self.max_num_partial_prefills != SchedulerConfig.max_num_partial_prefills
            or self.max_long_partial_prefills
            != SchedulerConfig.max_long_partial_prefills
        ):
2017
            _raise_unsupported_error(feature_name="Concurrent Partial Prefill")
2018

2019
        if self.pipeline_parallel_size > 1:
2020
2021
2022
            supports_pp = getattr(
                self.distributed_executor_backend, "supports_pp", False
            )
2023
            if not supports_pp and self.distributed_executor_backend not in (
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
                ParallelConfig.distributed_executor_backend,
                "ray",
                "mp",
                "external_launcher",
            ):
                name = (
                    "Pipeline Parallelism without Ray distributed "
                    "executor or multiprocessing executor or external "
                    "launcher"
                )
2034
                _raise_unsupported_error(feature_name=name)
2035

2036
2037
2038
2039
2040
2041
    @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
2042

2043
2044
        default_max_num_batched_tokens: dict[UsageContext | None, int]
        default_max_num_seqs: dict[UsageContext | None, int]
2045

2046
2047
        # When no user override, set the default values based on the usage
        # context.
2048
        # Use different default values for different hardware.
2049
2050
2051
2052
2053
2054
2055

        # 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:
2056
            device_memory = current_platform.get_device_total_memory()
2057
            device_name = current_platform.get_device_name().lower()
2058
2059
        except Exception:
            # This is only used to set default_max_num_batched_tokens
2060
            device_memory = 0
2061
            device_name = ""
2062

2063
2064
2065
2066
        # 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:
2067
            # For GPUs like H100 and MI300x, use larger default values.
2068
2069
2070
2071
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
2072
2073
2074
2075
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
2076
2077
2078
2079
2080
2081
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
2082
2083
2084
2085
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
2086

2087
2088
        # tpu specific default values.
        if current_platform.is_tpu():
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
            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,
                }
2106

2107
2108
2109
        # cpu specific default values.
        if current_platform.is_cpu():
            default_max_num_batched_tokens = {
2110
2111
                UsageContext.LLM_CLASS: 4096 * world_size,
                UsageContext.OPENAI_API_SERVER: 2048 * world_size,
2112
2113
            }
            default_max_num_seqs = {
2114
2115
                UsageContext.LLM_CLASS: 256 * world_size,
                UsageContext.OPENAI_API_SERVER: 128 * world_size,
2116
2117
            }

2118
2119
        return default_max_num_batched_tokens, default_max_num_seqs

2120
2121
    def _set_default_chunked_prefill_and_prefix_caching_args(
        self, model_config: ModelConfig
2122
    ) -> None:
2123
2124
        default_chunked_prefill = model_config.is_chunked_prefill_supported
        default_prefix_caching = model_config.is_prefix_caching_supported
2125
2126
2127
2128
2129
2130
2131
2132

        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",
            )
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
        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",
            )
2144
2145
2146
2147
        elif (
            model_config.runner_type == "pooling"
            and self.enable_chunked_prefill
            and not default_chunked_prefill
2148
        ):
2149
            logger.warning_once(
2150
2151
2152
                "This model does not officially support chunked prefill. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
2153
                scope="local",
2154
2155
2156
2157
2158
            )

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

2159
            logger.debug(
2160
2161
2162
2163
2164
2165
2166
2167
                "%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
        ):
2168
            logger.warning_once(
2169
2170
2171
                "This model does not officially support prefix caching. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
2172
                scope="local",
2173
2174
            )

2175
        # Disable chunked prefill and prefix caching for:
2176
        # RISCV CPUs in V1
2177
2178
2179
2180
        if current_platform.is_cpu() and current_platform.get_cpu_architecture() in (
            CpuArchEnum.RISCV,
        ):
            logger.info(
2181
2182
                "Chunked prefill is not supported for"
                "RISC-V CPUs; "
2183
2184
2185
2186
                "disabling it for V1 backend."
            )
            self.enable_chunked_prefill = False
            logger.info(
2187
2188
                "Prefix caching is not supported for "
                "RISC-V CPUs; "
2189
2190
2191
2192
2193
                "disabling it for V1 backend."
            )
            self.enable_prefix_caching = False

    def _set_default_max_num_seqs_and_batched_tokens_args(
2194
2195
2196
        self,
        usage_context: UsageContext | None,
        model_config: ModelConfig,
2197
    ):
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
        world_size = self.pipeline_parallel_size * self.tensor_parallel_size
        (
            default_max_num_batched_tokens,
            default_max_num_seqs,
        ) = self.get_batch_defaults(world_size)

        orig_max_num_batched_tokens = self.max_num_batched_tokens
        orig_max_num_seqs = self.max_num_seqs

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

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

2219
2220
2221
2222
2223
2224
2225
        # 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

2226
        if orig_max_num_batched_tokens is None:
2227
2228
2229
            assert model_config.max_model_len is not None, (
                "max_model_len must be set by this point"
            )
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
            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,
2242
2243
                self.max_num_batched_tokens,
            )
2244

2245
2246
2247
2248
            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,
2249
            )
2250

2251
2252
2253
2254
        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)

2255
            logger.debug(
2256
                "Defaulting max_num_seqs to %d for %s usage context.",
2257
                self.max_num_seqs,
2258
                usage_context.value if usage_context else None,
2259
            )
2260

2261

2262
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
2263
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
2264
    """Arguments for asynchronous vLLM engine."""
2265

2266
2267
    enable_log_requests: bool = False

2268
    @staticmethod
2269
2270
2271
    def add_cli_args(
        parser: FlexibleArgumentParser, async_args_only: bool = False
    ) -> FlexibleArgumentParser:
2272
        # Initialize plugin to update the parser, for example, The plugin may
2273
        # add a new kind of quantization method to --quantization argument or
2274
2275
        # a new device to --device argument.
        load_general_plugins()
2276
2277
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
2278
2279
2280
2281
        parser.add_argument(
            "--enable-log-requests",
            action=argparse.BooleanOptionalAction,
            default=AsyncEngineArgs.enable_log_requests,
2282
            help="Enable logging request information, dependent on log level:\n"
2283
2284
2285
            "- 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`.",
2286
        )
2287
        current_platform.pre_register_and_update(parser)
2288
        return parser
2289
2290


2291
2292
2293
2294
2295
2296
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)
2297
2298


2299
def human_readable_int(value: str) -> int:
2300
2301
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
2302

2303
2304
2305
2306
2307
2308
    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    """
    value = value.strip()
2309

2310
    match = re.fullmatch(r"(\d+(?:\.\d+)?)([kKmMgGtT])", value)
2311
2312
    if match:
        decimal_multiplier = {
2313
2314
2315
            "k": 10**3,
            "m": 10**6,
            "g": 10**9,
2316
            "t": 10**12,
2317
2318
        }
        binary_multiplier = {
2319
2320
2321
            "K": 2**10,
            "M": 2**20,
            "G": 2**30,
2322
            "T": 2**40,
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
        }

        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:
2335
2336
2337
2338
2339
                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
2340
2341
2342

    # Regular plain number.
    return int(value)
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361


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)