arg_utils.py 87.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

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

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

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

106
107
logger = init_logger(__name__)

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

113

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

123
124
125
    return _parse_type


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

132
    return _optional_type
133
134


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


141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
def is_type(type_hint: TypeHint, type: TypeHintT) -> TypeIs[TypeHintT]:
    """Check if the type hint is a specific type."""
    return type_hint is type or get_origin(type_hint) is type


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


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


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

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


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

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

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

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


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


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

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

    return type_hints


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


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


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

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

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

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

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

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

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

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

320
321
322
323
324
        # If the type hint was a sequence of literals, use the helper function
        # to update the type and choices
        if get_origin(kwargs[name].get("type")) is Literal:
            kwargs[name].update(literal_to_kwargs({kwargs[name]["type"]}))

325
326
327
328
329
330
331
        # If None is in type_hints, make the argument optional.
        # But not if it's a bool, argparse will handle this better.
        if type(None) in type_hints and not contains_type(type_hints, bool):
            kwargs[name]["type"] = optional_type(kwargs[name]["type"])
            if kwargs[name].get("choices"):
                kwargs[name]["choices"].append("None")
    return kwargs
332
333


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

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

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


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

351
    model: str = ModelConfig.model
352
353
354
    served_model_name: str | list[str] | None = ModelConfig.served_model_name
    tokenizer: str | None = ModelConfig.tokenizer
    hf_config_path: str | None = ModelConfig.hf_config_path
355
356
    runner: RunnerOption = ModelConfig.runner
    convert: ConvertOption = ModelConfig.convert
357
    task: TaskOption | None = ModelConfig.task
358
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
359
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
360
361
362
    tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
363
364
    allowed_media_domains: list[str] | None = ModelConfig.allowed_media_domains
    download_dir: str | None = LoadConfig.download_dir
365
    safetensors_load_strategy: str = LoadConfig.safetensors_load_strategy
366
    load_format: str | LoadFormats = LoadConfig.load_format
367
368
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
369
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
370
371
    seed: int | None = ModelConfig.seed
    max_model_len: int | None = ModelConfig.max_model_len
372
373
374
375
376
377
378
    cuda_graph_sizes: list[int] | None = CompilationConfig.cudagraph_capture_sizes
    cudagraph_capture_sizes: list[int] | None = (
        CompilationConfig.cudagraph_capture_sizes
    )
    max_cudagraph_capture_size: int | None = get_field(
        CompilationConfig, "max_cudagraph_capture_size"
    )
379
380
381
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
382
    distributed_executor_backend: (
383
        str | DistributedExecutorBackend | type[Executor] | None
384
    ) = ParallelConfig.distributed_executor_backend
385
    # number of P/D disaggregation (or other disaggregation) workers
386
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
387
388
389
390
    master_addr: str = ParallelConfig.master_addr
    master_port: int = ParallelConfig.master_port
    nnodes: int = ParallelConfig.nnodes
    node_rank: int = ParallelConfig.node_rank
391
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
392
    decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
393
    dcp_kv_cache_interleave_size: int = ParallelConfig.dcp_kv_cache_interleave_size
394
    data_parallel_size: int = ParallelConfig.data_parallel_size
395
396
397
398
399
    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
400
    data_parallel_hybrid_lb: bool = False
401
    data_parallel_external_lb: bool = False
Rui Qiao's avatar
Rui Qiao committed
402
    data_parallel_backend: str = ParallelConfig.data_parallel_backend
403
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
404
    all2all_backend: str | None = ParallelConfig.all2all_backend
405
    enable_dbo: bool = ParallelConfig.enable_dbo
406
407
    dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold
    dbo_prefill_token_threshold: int = ParallelConfig.dbo_prefill_token_threshold
408
409
410
    disable_nccl_for_dp_synchronization: bool = (
        ParallelConfig.disable_nccl_for_dp_synchronization
    )
411
    eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config")
412
    enable_eplb: bool = ParallelConfig.enable_eplb
413
    expert_placement_strategy: ExpertPlacementStrategy = (
414
        ParallelConfig.expert_placement_strategy
415
    )
416
417
    _api_process_count: int = ParallelConfig._api_process_count
    _api_process_rank: int = ParallelConfig._api_process_rank
418
419
420
421
    num_redundant_experts: int = EPLBConfig.num_redundant_experts
    eplb_window_size: int = EPLBConfig.window_size
    eplb_step_interval: int = EPLBConfig.step_interval
    eplb_log_balancedness: bool = EPLBConfig.log_balancedness
422
    max_parallel_loading_workers: int | None = (
423
424
        ParallelConfig.max_parallel_loading_workers
    )
425
426
    block_size: BlockSize | None = CacheConfig.block_size
    enable_prefix_caching: bool | None = CacheConfig.enable_prefix_caching
427
    prefix_caching_hash_algo: PrefixCachingHashAlgo = (
428
        CacheConfig.prefix_caching_hash_algo
429
    )
430
431
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
432
433
434
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
435
    kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
436
    max_num_batched_tokens: int | None = None
437
438
    max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
    max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
439
    long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold
440
    max_num_seqs: int | None = None
441
    max_logprobs: int = ModelConfig.max_logprobs
442
    logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode
443
    disable_log_stats: bool = False
444
    aggregate_engine_logging: bool = False
445
446
447
    revision: str | None = ModelConfig.revision
    code_revision: str | None = ModelConfig.code_revision
    hf_token: bool | str | None = ModelConfig.hf_token
448
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
449
450
    tokenizer_revision: str | None = ModelConfig.tokenizer_revision
    quantization: QuantizationMethods | None = ModelConfig.quantization
451
    enforce_eager: bool = ModelConfig.enforce_eager
452
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
453
    limit_mm_per_prompt: dict[str, int | dict[str, int]] = get_field(
454
455
        MultiModalConfig, "limit_per_prompt"
    )
456
    enable_mm_embeds: bool = MultiModalConfig.enable_mm_embeds
457
    interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings
458
459
460
    media_io_kwargs: dict[str, dict[str, Any]] = get_field(
        MultiModalConfig, "media_io_kwargs"
    )
461
    mm_processor_kwargs: dict[str, Any] | None = MultiModalConfig.mm_processor_kwargs
462
    disable_mm_preprocessor_cache: bool = False  # DEPRECATED
463
    mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb
464
    mm_processor_cache_type: MMCacheType | None = (
465
        MultiModalConfig.mm_processor_cache_type
466
467
    )
    mm_shm_cache_max_object_size_mb: int = (
468
        MultiModalConfig.mm_shm_cache_max_object_size_mb
469
    )
470
    mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
471
    mm_encoder_attn_backend: AttentionBackendEnum | str | None = (
472
473
        MultiModalConfig.mm_encoder_attn_backend
    )
474
    io_processor_plugin: str | None = None
475
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
476
    video_pruning_rate: float = MultiModalConfig.video_pruning_rate
477
    # LoRA fields
478
    enable_lora: bool = False
479
480
    max_loras: int = LoRAConfig.max_loras
    max_lora_rank: int = LoRAConfig.max_lora_rank
481
    default_mm_loras: dict[str, str] | None = LoRAConfig.default_mm_loras
482
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
483
484
    max_cpu_loras: int | None = LoRAConfig.max_cpu_loras
    lora_dtype: str | torch.dtype | None = LoRAConfig.lora_dtype
485
486
    lora_extra_vocab_size: int = LoRAConfig.lora_extra_vocab_size

487
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
488
    num_gpu_blocks_override: int | None = CacheConfig.num_gpu_blocks_override
489
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
490
    model_loader_extra_config: dict = get_field(LoadConfig, "model_loader_extra_config")
491
    ignore_patterns: str | list[str] = get_field(LoadConfig, "ignore_patterns")
492

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

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

500
    structured_outputs_config: StructuredOutputsConfig = get_field(
501
502
        VllmConfig, "structured_outputs_config"
    )
503
    reasoning_parser: str = StructuredOutputsConfig.reasoning_parser
504
    reasoning_parser_plugin: str | None = None
505
    # Deprecated guided decoding fields
506
507
508
509
    guided_decoding_backend: str | None = None
    guided_decoding_disable_fallback: bool | None = None
    guided_decoding_disable_any_whitespace: bool | None = None
    guided_decoding_disable_additional_properties: bool | None = None
510

511
    logits_processor_pattern: str | None = ModelConfig.logits_processor_pattern
512

513
    speculative_config: dict[str, Any] | None = None
514

515
    show_hidden_metrics_for_version: str | None = (
516
        ObservabilityConfig.show_hidden_metrics_for_version
517
    )
518
519
    otlp_traces_endpoint: str | None = ObservabilityConfig.otlp_traces_endpoint
    collect_detailed_traces: list[DetailedTraceModules] | None = (
520
        ObservabilityConfig.collect_detailed_traces
521
    )
522
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
523
    scheduler_cls: str | type[object] | None = SchedulerConfig.scheduler_cls
524

525
526
    pooler_config: PoolerConfig | None = ModelConfig.pooler_config
    override_pooler_config: dict | PoolerConfig | None = (
527
        ModelConfig.override_pooler_config
528
529
    )
    compilation_config: CompilationConfig = get_field(VllmConfig, "compilation_config")
530
531
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
532

533
534
    kv_transfer_config: KVTransferConfig | None = None
    kv_events_config: KVEventsConfig | None = None
535

536
537
    ec_transfer_config: ECTransferConfig | None = None

538
539
    generation_config: str = ModelConfig.generation_config
    enable_sleep_mode: bool = ModelConfig.enable_sleep_mode
540
541
542
    override_generation_config: dict[str, Any] = get_field(
        ModelConfig, "override_generation_config"
    )
543
    model_impl: str = ModelConfig.model_impl
544
    override_attention_dtype: str = ModelConfig.override_attention_dtype
545

546
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
547
548
    mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype
    mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype
549
    mamba_block_size: int | None = get_field(CacheConfig, "mamba_block_size")
550

551
    additional_config: dict[str, Any] = get_field(VllmConfig, "additional_config")
552

553
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
554
    pt_load_map_location: str = LoadConfig.pt_load_map_location
555

556
557
    # DEPRECATED
    enable_multimodal_encoder_data_parallel: bool = False
558

559
    logits_processors: list[str | type[LogitsProcessor]] | None = (
560
561
        ModelConfig.logits_processors
    )
562
563
    """Custom logitproc types"""

564
    async_scheduling: bool | None = SchedulerConfig.async_scheduling
565

566
567
    stream_interval: int = SchedulerConfig.stream_interval

568
    kv_sharing_fast_prefill: bool = CacheConfig.kv_sharing_fast_prefill
569

570
571
572
573
    kv_offloading_size: float | None = CacheConfig.kv_offloading_size
    kv_offloading_backend: KVOffloadingBackend | None = (
        CacheConfig.kv_offloading_backend
    )
574
    tokens_only: bool = False
575

576
    def __post_init__(self):
577
578
579
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
580
        if isinstance(self.compilation_config, dict):
581
            self.compilation_config = CompilationConfig(**self.compilation_config)
582
        if isinstance(self.eplb_config, dict):
583
            self.eplb_config = EPLBConfig(**self.eplb_config)
584
        # Setup plugins
585
        from vllm.plugins import load_general_plugins
586

587
        load_general_plugins()
588
589
590
591
592
        # when use hf offline,replace model id to local model path
        if huggingface_hub.constants.HF_HUB_OFFLINE:
            model_id = self.model
            self.model = get_model_path(self.model, self.revision)
            logger.info(
593
594
595
596
                "HF_HUB_OFFLINE is True, replace model_id [%s] to model_path [%s]",
                model_id,
                self.model,
            )
597
598

    @staticmethod
599
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
600
        """Shared CLI arguments for vLLM engine."""
601

602
        # Model arguments
603
604
605
606
607
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
608
        if not ("serve" in sys.argv[1:] and "--help" in sys.argv[1:]):
609
            model_group.add_argument("--model", **model_kwargs["model"])
610
611
        model_group.add_argument("--runner", **model_kwargs["runner"])
        model_group.add_argument("--convert", **model_kwargs["convert"])
612
        model_group.add_argument("--task", **model_kwargs["task"], deprecated=True)
613
        model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"])
614
615
616
617
        model_group.add_argument("--tokenizer-mode", **model_kwargs["tokenizer_mode"])
        model_group.add_argument(
            "--trust-remote-code", **model_kwargs["trust_remote_code"]
        )
618
619
        model_group.add_argument("--dtype", **model_kwargs["dtype"])
        model_group.add_argument("--seed", **model_kwargs["seed"])
620
621
622
623
624
625
626
        model_group.add_argument("--hf-config-path", **model_kwargs["hf_config_path"])
        model_group.add_argument(
            "--allowed-local-media-path", **model_kwargs["allowed_local_media_path"]
        )
        model_group.add_argument(
            "--allowed-media-domains", **model_kwargs["allowed_media_domains"]
        )
627
        model_group.add_argument("--revision", **model_kwargs["revision"])
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
        model_group.add_argument("--code-revision", **model_kwargs["code_revision"])
        model_group.add_argument(
            "--tokenizer-revision", **model_kwargs["tokenizer_revision"]
        )
        model_group.add_argument("--max-model-len", **model_kwargs["max_model_len"])
        model_group.add_argument("--quantization", "-q", **model_kwargs["quantization"])
        model_group.add_argument("--enforce-eager", **model_kwargs["enforce_eager"])
        model_group.add_argument("--max-logprobs", **model_kwargs["max_logprobs"])
        model_group.add_argument("--logprobs-mode", **model_kwargs["logprobs_mode"])
        model_group.add_argument(
            "--disable-sliding-window", **model_kwargs["disable_sliding_window"]
        )
        model_group.add_argument(
            "--disable-cascade-attn", **model_kwargs["disable_cascade_attn"]
        )
        model_group.add_argument(
            "--skip-tokenizer-init", **model_kwargs["skip_tokenizer_init"]
        )
        model_group.add_argument(
            "--enable-prompt-embeds", **model_kwargs["enable_prompt_embeds"]
        )
        model_group.add_argument(
            "--served-model-name", **model_kwargs["served_model_name"]
        )
        model_group.add_argument("--config-format", **model_kwargs["config_format"])
653
654
        # This one is a special case because it can bool
        # or str. TODO: Handle this in get_kwargs
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
        model_group.add_argument(
            "--hf-token",
            type=str,
            nargs="?",
            const=True,
            default=model_kwargs["hf_token"]["default"],
            help=model_kwargs["hf_token"]["help"],
        )
        model_group.add_argument("--hf-overrides", **model_kwargs["hf_overrides"])
        model_group.add_argument("--pooler-config", **model_kwargs["pooler_config"])
        model_group.add_argument(
            "--override-pooler-config",
            **model_kwargs["override_pooler_config"],
            deprecated=True,
        )
        model_group.add_argument(
            "--logits-processor-pattern", **model_kwargs["logits_processor_pattern"]
        )
        model_group.add_argument(
            "--generation-config", **model_kwargs["generation_config"]
        )
        model_group.add_argument(
            "--override-generation-config", **model_kwargs["override_generation_config"]
        )
        model_group.add_argument(
            "--enable-sleep-mode", **model_kwargs["enable_sleep_mode"]
        )
682
        model_group.add_argument("--model-impl", **model_kwargs["model_impl"])
683
684
685
686
687
688
689
690
691
        model_group.add_argument(
            "--override-attention-dtype", **model_kwargs["override_attention_dtype"]
        )
        model_group.add_argument(
            "--logits-processors", **model_kwargs["logits_processors"]
        )
        model_group.add_argument(
            "--io-processor-plugin", **model_kwargs["io_processor_plugin"]
        )
692

693
694
695
696
697
698
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
699
        load_group.add_argument("--load-format", **load_kwargs["load_format"])
700
701
702
703
704
705
706
707
708
709
710
711
        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"]
        )
712

713
714
715
716
717
        # Structured outputs arguments
        structured_outputs_kwargs = get_kwargs(StructuredOutputsConfig)
        structured_outputs_group = parser.add_argument_group(
            title="StructuredOutputsConfig",
            description=StructuredOutputsConfig.__doc__,
718
        )
719
        structured_outputs_group.add_argument(
720
            "--reasoning-parser",
721
            # Choices need to be validated after parsing to include plugins
722
723
            **structured_outputs_kwargs["reasoning_parser"],
        )
724
725
726
727
        structured_outputs_group.add_argument(
            "--reasoning-parser-plugin",
            **structured_outputs_kwargs["reasoning_parser_plugin"],
        )
728
729
730
731
732
733
734
735
736
737
738
        # Deprecated guided decoding arguments
        for arg, type in [
            ("--guided-decoding-backend", str),
            ("--guided-decoding-disable-fallback", bool),
            ("--guided-decoding-disable-any-whitespace", bool),
            ("--guided-decoding-disable-additional-properties", bool),
        ]:
            structured_outputs_group.add_argument(
                arg,
                type=type,
                help=(f"[DEPRECATED] {arg} will be removed in v0.12.0."),
739
740
                deprecated=True,
            )
741

742
        # Parallel arguments
743
744
745
746
747
748
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
749
            "--distributed-executor-backend",
750
751
            **parallel_kwargs["distributed_executor_backend"],
        )
752
        parallel_group.add_argument(
753
754
755
756
            "--pipeline-parallel-size",
            "-pp",
            **parallel_kwargs["pipeline_parallel_size"],
        )
757
758
759
760
        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"])
761
        parallel_group.add_argument(
762
763
            "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
        )
764
        parallel_group.add_argument(
765
766
767
768
            "--decode-context-parallel-size",
            "-dcp",
            **parallel_kwargs["decode_context_parallel_size"],
        )
769
770
771
772
        parallel_group.add_argument(
            "--dcp-kv-cache-interleave-size",
            **parallel_kwargs["dcp_kv_cache_interleave_size"],
        )
773
774
775
776
777
778
        parallel_group.add_argument(
            "--data-parallel-size", "-dp", **parallel_kwargs["data_parallel_size"]
        )
        parallel_group.add_argument(
            "--data-parallel-rank",
            "-dpn",
779
            type=int,
780
781
782
            help="Data parallel rank of this instance. "
            "When set, enables external load balancer mode.",
        )
783
        parallel_group.add_argument(
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
            "--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".',
        )
814
        parallel_group.add_argument(
815
816
817
818
819
820
821
822
            "--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"],
823
824
825
826
        )
        parallel_group.add_argument(
            "--enable-expert-parallel", **parallel_kwargs["enable_expert_parallel"]
        )
827
828
829
        parallel_group.add_argument(
            "--all2all-backend", **parallel_kwargs["all2all_backend"]
        )
830
        parallel_group.add_argument("--enable-dbo", **parallel_kwargs["enable_dbo"])
831
832
        parallel_group.add_argument(
            "--dbo-decode-token-threshold",
833
834
            **parallel_kwargs["dbo_decode_token_threshold"],
        )
835
836
        parallel_group.add_argument(
            "--dbo-prefill-token-threshold",
837
838
            **parallel_kwargs["dbo_prefill_token_threshold"],
        )
839
840
841
842
        parallel_group.add_argument(
            "--disable-nccl-for-dp-synchronization",
            **parallel_kwargs["disable_nccl_for_dp_synchronization"],
        )
843
844
        parallel_group.add_argument("--enable-eplb", **parallel_kwargs["enable_eplb"])
        parallel_group.add_argument("--eplb-config", **parallel_kwargs["eplb_config"])
845
846
        parallel_group.add_argument(
            "--expert-placement-strategy",
847
848
            **parallel_kwargs["expert_placement_strategy"],
        )
849
850
851
        parallel_group.add_argument(
            "--num-redundant-experts",
            type=int,
852
853
854
            help="[DEPRECATED] --num-redundant-experts will be removed in v0.12.0.",
            deprecated=True,
        )
855
856
857
858
        parallel_group.add_argument(
            "--eplb-window-size",
            type=int,
            help="[DEPRECATED] --eplb-window-size will be removed in v0.12.0.",
859
860
            deprecated=True,
        )
861
862
863
        parallel_group.add_argument(
            "--eplb-step-interval",
            type=int,
864
865
866
            help="[DEPRECATED] --eplb-step-interval will be removed in v0.12.0.",
            deprecated=True,
        )
867
868
869
        parallel_group.add_argument(
            "--eplb-log-balancedness",
            action=argparse.BooleanOptionalAction,
870
871
872
            help="[DEPRECATED] --eplb-log-balancedness will be removed in v0.12.0.",
            deprecated=True,
        )
873

874
        parallel_group.add_argument(
875
            "--max-parallel-loading-workers",
876
877
            **parallel_kwargs["max_parallel_loading_workers"],
        )
878
        parallel_group.add_argument(
879
880
            "--ray-workers-use-nsight", **parallel_kwargs["ray_workers_use_nsight"]
        )
881
        parallel_group.add_argument(
882
            "--disable-custom-all-reduce",
883
884
885
886
887
888
            **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"]
        )
889
890
        parallel_group.add_argument(
            "--enable-multimodal-encoder-data-parallel",
891
            action="store_true",
892
893
            deprecated=True,
        )
894

895
896
897
898
899
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
900
        )
901
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
902
903
904
905
906
907
        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"]
        )
908
        cache_group.add_argument("--swap-space", **cache_kwargs["swap_space"])
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
        cache_group.add_argument("--kv-cache-dtype", **cache_kwargs["cache_dtype"])
        cache_group.add_argument(
            "--num-gpu-blocks-override", **cache_kwargs["num_gpu_blocks_override"]
        )
        cache_group.add_argument(
            "--enable-prefix-caching", **cache_kwargs["enable_prefix_caching"]
        )
        cache_group.add_argument(
            "--prefix-caching-hash-algo", **cache_kwargs["prefix_caching_hash_algo"]
        )
        cache_group.add_argument("--cpu-offload-gb", **cache_kwargs["cpu_offload_gb"])
        cache_group.add_argument(
            "--calculate-kv-scales", **cache_kwargs["calculate_kv_scales"]
        )
        cache_group.add_argument(
            "--kv-sharing-fast-prefill", **cache_kwargs["kv_sharing_fast_prefill"]
        )
        cache_group.add_argument(
            "--mamba-cache-dtype", **cache_kwargs["mamba_cache_dtype"]
        )
        cache_group.add_argument(
            "--mamba-ssm-cache-dtype", **cache_kwargs["mamba_ssm_cache_dtype"]
        )
932
933
934
        cache_group.add_argument(
            "--mamba-block-size", **cache_kwargs["mamba_block_size"]
        )
935
936
937
938
939
940
        cache_group.add_argument(
            "--kv-offloading-size", **cache_kwargs["kv_offloading_size"]
        )
        cache_group.add_argument(
            "--kv-offloading-backend", **cache_kwargs["kv_offloading_backend"]
        )
941

942
        # Multimodal related configs
943
944
945
946
947
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
948
        multimodal_group.add_argument(
949
950
            "--limit-mm-per-prompt", **multimodal_kwargs["limit_per_prompt"]
        )
951
952
953
        multimodal_group.add_argument(
            "--enable-mm-embeds", **multimodal_kwargs["enable_mm_embeds"]
        )
954
955
956
957
958
959
960
961
962
        multimodal_group.add_argument(
            "--media-io-kwargs", **multimodal_kwargs["media_io_kwargs"]
        )
        multimodal_group.add_argument(
            "--mm-processor-kwargs", **multimodal_kwargs["mm_processor_kwargs"]
        )
        multimodal_group.add_argument(
            "--mm-processor-cache-gb", **multimodal_kwargs["mm_processor_cache_gb"]
        )
963
        multimodal_group.add_argument(
964
965
            "--disable-mm-preprocessor-cache", action="store_true", deprecated=True
        )
966
        multimodal_group.add_argument(
967
968
            "--mm-processor-cache-type", **multimodal_kwargs["mm_processor_cache_type"]
        )
969
970
        multimodal_group.add_argument(
            "--mm-shm-cache-max-object-size-mb",
971
972
            **multimodal_kwargs["mm_shm_cache_max_object_size_mb"],
        )
973
        multimodal_group.add_argument(
974
975
            "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"]
        )
976
977
978
979
        multimodal_group.add_argument(
            "--mm-encoder-attn-backend",
            **multimodal_kwargs["mm_encoder_attn_backend"],
        )
980
981
982
        multimodal_group.add_argument(
            "--interleave-mm-strings", **multimodal_kwargs["interleave_mm_strings"]
        )
983
        multimodal_group.add_argument(
984
985
            "--skip-mm-profiling", **multimodal_kwargs["skip_mm_profiling"]
        )
986

987
        multimodal_group.add_argument(
988
989
            "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"]
        )
990

991
        # LoRA related configs
992
993
994
995
996
997
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
998
            "--enable-lora",
999
            action=argparse.BooleanOptionalAction,
1000
1001
            help="If True, enable handling of LoRA adapters.",
        )
1002
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
1003
1004
1005
1006
        lora_group.add_argument("--max-lora-rank", **lora_kwargs["max_lora_rank"])
        lora_group.add_argument(
            "--lora-extra-vocab-size", **lora_kwargs["lora_extra_vocab_size"]
        )
1007
        lora_group.add_argument(
1008
            "--lora-dtype",
1009
1010
            **lora_kwargs["lora_dtype"],
        )
1011
1012
1013
1014
1015
        lora_group.add_argument("--max-cpu-loras", **lora_kwargs["max_cpu_loras"])
        lora_group.add_argument(
            "--fully-sharded-loras", **lora_kwargs["fully_sharded_loras"]
        )
        lora_group.add_argument("--default-mm-loras", **lora_kwargs["default_mm_loras"])
1016

1017
1018
1019
1020
1021
1022
1023
1024
        # 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",
1025
1026
            **observability_kwargs["show_hidden_metrics_for_version"],
        )
1027
        observability_group.add_argument(
1028
1029
            "--otlp-traces-endpoint", **observability_kwargs["otlp_traces_endpoint"]
        )
1030
1031
1032
1033
1034
        # 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"] += [
1035
            ",".join(p) for p in permutations(get_args(DetailedTraceModules), r=2)
1036
1037
1038
        ]
        observability_group.add_argument(
            "--collect-detailed-traces",
1039
1040
            **observability_kwargs["collect_detailed_traces"],
        )
1041

1042
1043
1044
1045
1046
1047
1048
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
1049
1050
1051
1052
1053
            "--max-num-batched-tokens",
            **{
                **scheduler_kwargs["max_num_batched_tokens"],
                "default": None,
            },
1054
        )
1055
        scheduler_group.add_argument(
1056
1057
1058
1059
1060
            "--max-num-seqs",
            **{
                **scheduler_kwargs["max_num_seqs"],
                "default": None,
            },
1061
1062
1063
1064
        )
        scheduler_group.add_argument(
            "--max-num-partial-prefills", **scheduler_kwargs["max_num_partial_prefills"]
        )
1065
1066
        scheduler_group.add_argument(
            "--max-long-partial-prefills",
1067
1068
            **scheduler_kwargs["max_long_partial_prefills"],
        )
1069
1070
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
1071
1072
1073
1074
1075
            **scheduler_kwargs["long_prefill_token_threshold"],
        )
        scheduler_group.add_argument(
            "--num-lookahead-slots", **scheduler_kwargs["num_lookahead_slots"]
        )
1076
1077
        # multi-step scheduling has been removed; corresponding arguments
        # are no longer supported.
1078
        scheduler_group.add_argument(
1079
1080
            "--scheduling-policy", **scheduler_kwargs["policy"]
        )
1081
        scheduler_group.add_argument(
1082
1083
1084
1085
1086
            "--enable-chunked-prefill",
            **{
                **scheduler_kwargs["enable_chunked_prefill"],
                "default": None,
            },
1087
1088
1089
1090
1091
1092
1093
        )
        scheduler_group.add_argument(
            "--disable-chunked-mm-input", **scheduler_kwargs["disable_chunked_mm_input"]
        )
        scheduler_group.add_argument(
            "--scheduler-cls", **scheduler_kwargs["scheduler_cls"]
        )
1094
1095
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
1096
1097
1098
1099
1100
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"],
        )
        scheduler_group.add_argument(
            "--async-scheduling", **scheduler_kwargs["async_scheduling"]
        )
1101
1102
1103
        scheduler_group.add_argument(
            "--stream-interval", **scheduler_kwargs["stream_interval"]
        )
1104

1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
        # Compilation arguments
        compilation_kwargs = get_kwargs(CompilationConfig)
        compilation_group = parser.add_argument_group(
            title="CompilationConfig",
            description=CompilationConfig.__doc__,
        )
        compilation_group.add_argument(
            "--cudagraph-capture-sizes", **compilation_kwargs["cudagraph_capture_sizes"]
        )
        compilation_kwargs["cudagraph_capture_sizes"]["help"] = (
            "--cuda-graph-sizes is deprecated and will be removed in v0.13.0 or v1.0.0,"
            " whichever is soonest. Please use --cudagraph-capture-sizes instead."
        )
        compilation_group.add_argument(
            "--cuda-graph-sizes",
            **compilation_kwargs["cudagraph_capture_sizes"],
            deprecated=True,
        )
        compilation_group.add_argument(
            "--max-cudagraph-capture-size",
            **compilation_kwargs["max_cudagraph_capture_size"],
        )

1128
        # vLLM arguments
1129
        vllm_kwargs = get_kwargs(VllmConfig)
1130
1131
1132
1133
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
1134
1135
1136
1137
        # 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)
1138
1139
1140
1141
1142
1143
1144
        vllm_group.add_argument(
            "--speculative-config", **vllm_kwargs["speculative_config"]
        )
        vllm_group.add_argument(
            "--kv-transfer-config", **vllm_kwargs["kv_transfer_config"]
        )
        vllm_group.add_argument("--kv-events-config", **vllm_kwargs["kv_events_config"])
1145
1146
1147
        vllm_group.add_argument(
            "--ec-transfer-config", **vllm_kwargs["ec_transfer_config"]
        )
1148
1149
1150
1151
1152
1153
1154
1155
1156
        vllm_group.add_argument(
            "--compilation-config", "-O", **vllm_kwargs["compilation_config"]
        )
        vllm_group.add_argument(
            "--additional-config", **vllm_kwargs["additional_config"]
        )
        vllm_group.add_argument(
            "--structured-outputs-config", **vllm_kwargs["structured_outputs_config"]
        )
1157

1158
        # Other arguments
1159
1160
1161
1162
1163
        parser.add_argument(
            "--disable-log-stats",
            action="store_true",
            help="Disable logging statistics.",
        )
1164

1165
1166
1167
1168
1169
1170
        parser.add_argument(
            "--aggregate-engine-logging",
            action="store_true",
            help="Log aggregate rather than per-engine statistics "
            "when using data parallelism.",
        )
1171
        return parser
1172
1173

    @classmethod
1174
    def from_cli_args(cls, args: argparse.Namespace):
1175
1176
1177
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
1178
1179
1180
        engine_args = cls(
            **{attr: getattr(args, attr) for attr in attrs if hasattr(args, attr)}
        )
Zhuohan Li's avatar
Zhuohan Li committed
1181
        return engine_args
1182

1183
    def create_model_config(self) -> ModelConfig:
1184
1185
1186
1187
        # gguf file needs a specific model loader and doesn't use hf_repo
        if check_gguf_file(self.model):
            self.quantization = self.load_format = "gguf"

1188
1189
1190
1191
        if self.disable_mm_preprocessor_cache:
            logger.warning(
                "`--disable-mm-preprocessor-cache` is deprecated "
                "and will be removed in v0.13. "
1192
1193
                "Please use `--mm-processor-cache-gb 0` instead.",
            )
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205

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

            self.mm_processor_cache_gb = envs.VLLM_MM_INPUT_CACHE_GIB

1206
1207
1208
1209
        if self.enable_multimodal_encoder_data_parallel:
            logger.warning(
                "--enable-multimodal-encoder-data-parallel` is deprecated "
                "and will be removed in v0.13. "
1210
1211
                "Please use `--mm-encoder-tp-mode data` instead."
            )
1212
1213
1214

            self.mm_encoder_tp_mode = "data"

1215
        return ModelConfig(
1216
            model=self.model,
1217
            hf_config_path=self.hf_config_path,
1218
1219
            runner=self.runner,
            convert=self.convert,
1220
            task=self.task,
1221
            tokenizer=self.tokenizer,
1222
1223
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1224
            allowed_local_media_path=self.allowed_local_media_path,
1225
            allowed_media_domains=self.allowed_media_domains,
1226
1227
1228
1229
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
1230
            hf_token=self.hf_token,
1231
            hf_overrides=self.hf_overrides,
1232
1233
1234
1235
1236
            tokenizer_revision=self.tokenizer_revision,
            max_model_len=self.max_model_len,
            quantization=self.quantization,
            enforce_eager=self.enforce_eager,
            max_logprobs=self.max_logprobs,
1237
            logprobs_mode=self.logprobs_mode,
1238
            disable_sliding_window=self.disable_sliding_window,
1239
            disable_cascade_attn=self.disable_cascade_attn,
1240
            skip_tokenizer_init=self.skip_tokenizer_init,
1241
            enable_prompt_embeds=self.enable_prompt_embeds,
1242
            served_model_name=self.served_model_name,
1243
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1244
            enable_mm_embeds=self.enable_mm_embeds,
1245
            interleave_mm_strings=self.interleave_mm_strings,
1246
            media_io_kwargs=self.media_io_kwargs,
1247
            skip_mm_profiling=self.skip_mm_profiling,
1248
            config_format=self.config_format,
1249
            mm_processor_kwargs=self.mm_processor_kwargs,
1250
            mm_processor_cache_gb=self.mm_processor_cache_gb,
1251
            mm_processor_cache_type=self.mm_processor_cache_type,
1252
            mm_shm_cache_max_object_size_mb=self.mm_shm_cache_max_object_size_mb,
1253
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
1254
            mm_encoder_attn_backend=self.mm_encoder_attn_backend,
1255
            pooler_config=self.pooler_config,
1256
            override_pooler_config=self.override_pooler_config,
1257
            logits_processor_pattern=self.logits_processor_pattern,
1258
            generation_config=self.generation_config,
1259
            override_generation_config=self.override_generation_config,
1260
            enable_sleep_mode=self.enable_sleep_mode,
1261
            model_impl=self.model_impl,
1262
            override_attention_dtype=self.override_attention_dtype,
1263
            logits_processors=self.logits_processors,
1264
            video_pruning_rate=self.video_pruning_rate,
1265
            io_processor_plugin=self.io_processor_plugin,
1266
        )
1267

1268
    def validate_tensorizer_args(self):
1269
1270
        from vllm.model_executor.model_loader.tensorizer import TensorizerConfig

1271
1272
        for key in self.model_loader_extra_config:
            if key in TensorizerConfig._fields:
1273
1274
1275
                self.model_loader_extra_config["tensorizer_config"][key] = (
                    self.model_loader_extra_config[key]
                )
1276

1277
    def create_load_config(self) -> LoadConfig:
1278
1279
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1280

1281
1282
1283
        if self.load_format == "tensorizer":
            if hasattr(self.model_loader_extra_config, "to_serializable"):
                self.model_loader_extra_config = (
1284
1285
                    self.model_loader_extra_config.to_serializable()
                )
1286
            self.model_loader_extra_config["tensorizer_config"] = {}
1287
1288
1289
            self.model_loader_extra_config["tensorizer_config"]["tensorizer_dir"] = (
                self.model
            )
1290
            self.validate_tensorizer_args()
1291

1292
1293
1294
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
1295
            safetensors_load_strategy=self.safetensors_load_strategy,
1296
            device="cpu" if is_online_quantization(self.quantization) else None,
1297
1298
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1299
            use_tqdm_on_load=self.use_tqdm_on_load,
1300
            pt_load_map_location=self.pt_load_map_location,
1301
        )
1302

1303
1304
1305
1306
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
1307
    ) -> SpeculativeConfig | None:
1308
1309
1310
1311
1312
1313
        """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
1314
        dictionary from the engine.
1315
1316
        """
        if self.speculative_config is None:
1317
            return None
1318

1319
1320
1321
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.
1322
1323
1324
1325
1326
1327
        self.speculative_config.update(
            {
                "target_model_config": target_model_config,
                "target_parallel_config": target_parallel_config,
            }
        )
1328
        return SpeculativeConfig(**self.speculative_config)
1329

1330
1331
    def create_engine_config(
        self,
1332
        usage_context: UsageContext | None = None,
1333
        headless: bool = False,
1334
1335
1336
1337
    ) -> VllmConfig:
        """
        Create the VllmConfig.

1338
        NOTE: If VllmConfig is incompatible, we raise an error.
1339
        """
1340
        current_platform.pre_register_and_update()
1341

1342
        device_config = DeviceConfig(device=cast(Device, current_platform.device_type))
1343

1344
1345
        # Check if the model is a speculator and override model/tokenizer/config
        # BEFORE creating ModelConfig, so the config is created with the target model
1346
1347
1348
1349
        # 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):
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
            (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,
                )
            )

1360
1361
1362
1363
        model_config = self.create_model_config()
        self.model = model_config.model
        self.tokenizer = model_config.tokenizer

1364
        self._check_feature_supported(model_config)
1365

1366
1367
        # Set default arguments for V1 Engine.
        self._set_default_args(usage_context, model_config)
1368
1369
        # Disable chunked prefill and prefix caching for:
        # POWER (ppc64le)/ARM/s390x/RISCV CPUs in V1
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
        if current_platform.is_cpu() and current_platform.get_cpu_architecture() in (
            CpuArchEnum.POWERPC,
            CpuArchEnum.S390X,
            CpuArchEnum.ARM,
            CpuArchEnum.RISCV,
        ):
            logger.info(
                "Chunked prefill is not supported for ARM and POWER, "
                "S390X and RISC-V CPUs; "
                "disabling it for V1 backend."
            )
1381
            self.enable_chunked_prefill = False
1382
1383
1384
1385
1386
1387
1388
            logger.info(
                "Prefix caching is not supported for ARM and POWER, "
                "S390X and RISC-V CPUs; "
                "disabling it for V1 backend."
            )
            self.enable_prefix_caching = False

1389
1390
        assert self.enable_chunked_prefill is not None

1391
        sliding_window: int | None = None
1392
1393
1394
1395
1396
1397
        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()

1398
1399
1400
        # Note(hc): In the current implementation of decode context
        # parallel(DCP), tp_size needs to be divisible by dcp_size,
        # because the world size does not change by dcp, it simply
1401
        # reuses the GPUs of TP group, and split one TP group into
1402
        # tp_size//dcp_size DCP groups.
1403
        assert self.tensor_parallel_size % self.decode_context_parallel_size == 0, (
1404
1405
1406
1407
            f"tp_size={self.tensor_parallel_size} must be divisible by"
            f"dcp_size={self.decode_context_parallel_size}."
        )

1408
        cache_config = CacheConfig(
1409
            block_size=self.block_size,
1410
            gpu_memory_utilization=self.gpu_memory_utilization,
1411
            kv_cache_memory_bytes=self.kv_cache_memory_bytes,
1412
1413
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1414
            is_attention_free=model_config.is_attention_free,
1415
            num_gpu_blocks_override=self.num_gpu_blocks_override,
1416
            sliding_window=sliding_window,
1417
            enable_prefix_caching=self.enable_prefix_caching,
1418
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1419
            cpu_offload_gb=self.cpu_offload_gb,
1420
            calculate_kv_scales=self.calculate_kv_scales,
1421
            kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
1422
1423
            mamba_cache_dtype=self.mamba_cache_dtype,
            mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype,
1424
            mamba_block_size=self.mamba_block_size,
1425
1426
            kv_offloading_size=self.kv_offloading_size,
            kv_offloading_backend=self.kv_offloading_backend,
1427
        )
1428

1429
1430
1431
1432
1433
1434
        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
1435

1436
            ray_runtime_env = ray.get_runtime_context().runtime_env
1437
1438
1439
1440
1441
1442
1443
            # 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)
1444

1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
        # 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()

1456
        assert not headless or not self.data_parallel_hybrid_lb, (
1457
1458
            "data_parallel_hybrid_lb is not applicable in headless mode"
        )
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
        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
        )
1500
        # Local DP rank = 1, use pure-external LB.
1501
        if data_parallel_external_lb:
1502
1503
1504
1505
            assert self.data_parallel_rank is not None, (
                "data_parallel_rank or node_rank must be spefified if "
                "data_parallel_external_lb is enable."
            )
1506
            assert self.data_parallel_size_local in (1, None), (
1507
1508
                "data_parallel_size_local must be 1 or None when data_parallel_rank "
                "is set"
1509
            )
1510
            data_parallel_size_local = 1
1511
1512
            # Use full external lb if we have local_size of 1.
            self.data_parallel_hybrid_lb = False
1513
1514
        elif self.data_parallel_size_local is not None:
            data_parallel_size_local = self.data_parallel_size_local
1515
1516
1517
1518
1519
1520
1521

            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.
1522
1523
1524
1525
1526
                logger.warning(
                    "data_parallel_hybrid_lb is not eligible when "
                    "data_parallel_size_local = 1, autoswitch to "
                    "data_parallel_external_lb."
                )
1527
1528
1529
1530
1531
1532
1533
                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

1534
1535
1536
1537
1538
1539
1540
1541
1542
            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,
                )
1543
        else:
1544
            assert not self.data_parallel_hybrid_lb, (
1545
1546
                "data_parallel_size_local must be set to use data_parallel_hybrid_lb."
            )
1547

1548
1549
1550
1551
1552
1553
1554
1555
1556
            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
1557
1558
1559

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1560
1561
1562
1563
        if self.data_parallel_address is None:
            if self.data_parallel_backend == "ray":
                host_ip = get_ip()
                logger.info(
1564
1565
                    "Using host IP %s as ray-based data parallel address", host_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1566
1567
1568
1569
                data_parallel_address = host_ip
            else:
                assert self.data_parallel_backend == "mp", (
                    "data_parallel_backend can only be ray or mp, got %s",
1570
1571
                    self.data_parallel_backend,
                )
1572
1573
1574
                data_parallel_address = (
                    self.master_addr or ParallelConfig.data_parallel_master_ip
                )
Rui Qiao's avatar
Rui Qiao committed
1575
1576
        else:
            data_parallel_address = self.data_parallel_address
1577
1578
1579

        # This port is only used when there are remote data parallel engines,
        # otherwise the local IPC transport is used.
1580
        data_parallel_rpc_port = (
1581
            self.data_parallel_rpc_port
1582
1583
1584
            if (self.data_parallel_rpc_port is not None)
            else ParallelConfig.data_parallel_rpc_port
        )
1585

1586
1587
1588
1589
        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.")

1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
        # Forward the deprecated CLI args to the EPLB config.
        if self.num_redundant_experts is not None:
            self.eplb_config.num_redundant_experts = self.num_redundant_experts
        if self.eplb_window_size is not None:
            self.eplb_config.window_size = self.eplb_window_size
        if self.eplb_step_interval is not None:
            self.eplb_config.step_interval = self.eplb_step_interval
        if self.eplb_log_balancedness is not None:
            self.eplb_config.log_balancedness = self.eplb_log_balancedness

1600
        parallel_config = ParallelConfig(
1601
1602
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1603
            data_parallel_size=self.data_parallel_size,
1604
1605
            data_parallel_rank=self.data_parallel_rank or 0,
            data_parallel_external_lb=data_parallel_external_lb,
1606
            data_parallel_size_local=data_parallel_size_local,
1607
1608
1609
1610
            master_addr=self.master_addr,
            master_port=self.master_port,
            nnodes=self.nnodes,
            node_rank=self.node_rank,
1611
1612
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
1613
            data_parallel_backend=self.data_parallel_backend,
1614
            data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
1615
            enable_expert_parallel=self.enable_expert_parallel,
1616
            all2all_backend=self.all2all_backend,
1617
1618
            enable_dbo=self.enable_dbo,
            dbo_decode_token_threshold=self.dbo_decode_token_threshold,
1619
            dbo_prefill_token_threshold=self.dbo_prefill_token_threshold,
1620
            disable_nccl_for_dp_synchronization=self.disable_nccl_for_dp_synchronization,
1621
            enable_eplb=self.enable_eplb,
1622
            eplb_config=self.eplb_config,
1623
            expert_placement_strategy=self.expert_placement_strategy,
1624
1625
1626
            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,
1627
            ray_runtime_env=ray_runtime_env,
1628
            placement_group=placement_group,
1629
1630
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1631
            worker_extension_cls=self.worker_extension_cls,
1632
            decode_context_parallel_size=self.decode_context_parallel_size,
1633
            dcp_kv_cache_interleave_size=self.dcp_kv_cache_interleave_size,
1634
1635
            _api_process_count=self._api_process_count,
            _api_process_rank=self._api_process_rank,
1636
        )
1637

1638
        speculative_config = self.create_speculative_config(
1639
1640
1641
1642
            target_model_config=model_config,
            target_parallel_config=parallel_config,
        )

1643
1644
1645
1646
1647
        # make sure num_lookahead_slots is set appropriately depending on
        # whether speculative decoding is enabled
        num_lookahead_slots = self.num_lookahead_slots
        if speculative_config is not None:
            num_lookahead_slots = speculative_config.num_lookahead_slots
1648

1649
        scheduler_config = SchedulerConfig(
1650
            runner_type=model_config.runner_type,
1651
1652
1653
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1654
            num_lookahead_slots=num_lookahead_slots,
1655
            enable_chunked_prefill=self.enable_chunked_prefill,
1656
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1657
            is_multimodal_model=model_config.is_multimodal_model,
1658
            is_encoder_decoder=model_config.is_encoder_decoder,
1659
            policy=self.scheduling_policy,
1660
            scheduler_cls=self.scheduler_cls,
1661
1662
1663
            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,
1664
            disable_hybrid_kv_cache_manager=self.disable_hybrid_kv_cache_manager,
1665
            async_scheduling=self.async_scheduling,
1666
            stream_interval=self.stream_interval,
1667
        )
1668

1669
1670
1671
        if not model_config.is_multimodal_model and self.default_mm_loras:
            raise ValueError(
                "Default modality-specific LoRA(s) were provided for a "
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
                "non multimodal model"
            )

        lora_config = (
            LoRAConfig(
                max_lora_rank=self.max_lora_rank,
                max_loras=self.max_loras,
                default_mm_loras=self.default_mm_loras,
                fully_sharded_loras=self.fully_sharded_loras,
                lora_extra_vocab_size=self.lora_extra_vocab_size,
                lora_dtype=self.lora_dtype,
                max_cpu_loras=self.max_cpu_loras
                if self.max_cpu_loras and self.max_cpu_loras > 0
                else None,
            )
            if self.enable_lora
            else None
        )
1690

1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
        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"
            )

1705
1706
1707
1708
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1709
        load_config = self.create_load_config()
1710

1711
1712
        # Pass reasoning_parser into StructuredOutputsConfig
        if self.reasoning_parser:
1713
            self.structured_outputs_config.reasoning_parser = self.reasoning_parser
1714

1715
1716
1717
1718
1719
        if self.reasoning_parser_plugin:
            self.structured_outputs_config.reasoning_parser_plugin = (
                self.reasoning_parser_plugin
            )

1720
1721
1722
        # Forward the deprecated CLI args to the StructuredOutputsConfig
        so_config = self.structured_outputs_config
        if self.guided_decoding_backend is not None:
1723
            so_config.guided_decoding_backend = self.guided_decoding_backend
1724
        if self.guided_decoding_disable_fallback is not None:
1725
            so_config.disable_fallback = self.guided_decoding_disable_fallback
1726
        if self.guided_decoding_disable_any_whitespace is not None:
1727
            so_config.disable_any_whitespace = (
1728
1729
                self.guided_decoding_disable_any_whitespace
            )
1730
        if self.guided_decoding_disable_additional_properties is not None:
1731
            so_config.disable_additional_properties = (
1732
1733
                self.guided_decoding_disable_additional_properties
            )
1734

1735
        observability_config = ObservabilityConfig(
1736
            show_hidden_metrics_for_version=self.show_hidden_metrics_for_version,
1737
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1738
            collect_detailed_traces=self.collect_detailed_traces,
1739
        )
1740

1741
        # Compilation config overrides
1742
        compilation_config = copy.deepcopy(self.compilation_config)
1743
1744
1745
1746
1747
1748
        if self.cuda_graph_sizes is not None:
            logger.warning(
                "--cuda-graph-sizes is deprecated and will be removed in v0.13.0 or "
                "v1.0.0, whichever is soonest. Please use --cudagraph-capture-sizes "
                "instead."
            )
1749
            if compilation_config.cudagraph_capture_sizes is not None:
1750
1751
1752
1753
                raise ValueError(
                    "cuda_graph_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
1754
            compilation_config.cudagraph_capture_sizes = self.cuda_graph_sizes
1755
        if self.cudagraph_capture_sizes is not None:
1756
            if compilation_config.cudagraph_capture_sizes is not None:
1757
1758
1759
1760
                raise ValueError(
                    "cudagraph_capture_sizes and compilation_config."
                    "cudagraph_capture_sizes are mutually exclusive"
                )
1761
            compilation_config.cudagraph_capture_sizes = self.cudagraph_capture_sizes
1762
        if self.max_cudagraph_capture_size is not None:
1763
            if compilation_config.max_cudagraph_capture_size is not None:
1764
1765
1766
1767
                raise ValueError(
                    "max_cudagraph_capture_size and compilation_config."
                    "max_cudagraph_capture_size are mutually exclusive"
                )
1768
            compilation_config.max_cudagraph_capture_size = (
1769
1770
1771
                self.max_cudagraph_capture_size
            )

1772
        config = VllmConfig(
1773
1774
1775
1776
1777
1778
1779
1780
            model_config=model_config,
            cache_config=cache_config,
            parallel_config=parallel_config,
            scheduler_config=scheduler_config,
            device_config=device_config,
            lora_config=lora_config,
            speculative_config=speculative_config,
            load_config=load_config,
1781
            structured_outputs_config=self.structured_outputs_config,
1782
            observability_config=observability_config,
1783
            compilation_config=compilation_config,
1784
            kv_transfer_config=self.kv_transfer_config,
1785
            kv_events_config=self.kv_events_config,
1786
            ec_transfer_config=self.ec_transfer_config,
1787
            additional_config=self.additional_config,
1788
        )
1789

1790
1791
        return config

1792
1793
    def _check_feature_supported(self, model_config: ModelConfig):
        """Raise an error if the feature is not supported."""
1794
        if self.logits_processor_pattern != EngineArgs.logits_processor_pattern:
1795
            _raise_unsupported_error(feature_name="--logits-processor-pattern")
1796
1797

        # No Concurrent Partial Prefills so far.
1798
1799
1800
1801
1802
        if (
            self.max_num_partial_prefills != SchedulerConfig.max_num_partial_prefills
            or self.max_long_partial_prefills
            != SchedulerConfig.max_long_partial_prefills
        ):
1803
            _raise_unsupported_error(feature_name="Concurrent Partial Prefill")
1804

1805
        # N-gram, Medusa, and Eagle are supported for speculative decoding.
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
        if self.speculative_config is not None:
            # speculative_config could still be a dict at this point
            if isinstance(self.speculative_config, dict):
                method = self.speculative_config.get("method", None)
            else:
                method = self.speculative_config.method

            if method == "draft_model":
                raise NotImplementedError(
                    "Draft model speculative decoding is not supported yet. "
                    "Please consider using other speculative decoding methods "
1817
1818
                    "such as ngram, medusa, eagle, or mtp."
                )
1819

1820
        if self.pipeline_parallel_size > 1:
1821
1822
1823
            supports_pp = getattr(
                self.distributed_executor_backend, "supports_pp", False
            )
1824
            if not supports_pp and self.distributed_executor_backend not in (
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
                ParallelConfig.distributed_executor_backend,
                "ray",
                "mp",
                "external_launcher",
            ):
                name = (
                    "Pipeline Parallelism without Ray distributed "
                    "executor or multiprocessing executor or external "
                    "launcher"
                )
1835
                _raise_unsupported_error(feature_name=name)
1836

1837
1838
1839
1840
1841
    @classmethod
    def get_chunked_prefill_prefix_caching_defaults(
        cls,
        model_config: ModelConfig,
    ) -> tuple[bool, bool]:
1842
        if model_config.runner_type != "pooling":
1843
1844
1845
1846
1847
            default_chunked_prefill = True

            # Disable prefix caching default for hybrid models
            # since the feature is still experimental.
            default_prefix_caching = not model_config.is_hybrid
1848
        else:
1849
1850
            assert model_config.pooler_config is not None

1851
            pooling_type = model_config.pooler_config.pooling_type
1852
1853
1854
            incremental_prefill_supported = (
                pooling_type is not None
                and pooling_type.lower() == "last"
1855
                and getattr(model_config.hf_config, "is_causal", True)
1856
            )
1857

1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
            default_chunked_prefill = incremental_prefill_supported
            default_prefix_caching = incremental_prefill_supported

        return default_chunked_prefill, default_prefix_caching

    @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
1869

1870
1871
        default_max_num_batched_tokens: dict[UsageContext | None, int]
        default_max_num_seqs: dict[UsageContext | None, int]
1872

1873
1874
        # When no user override, set the default values based on the usage
        # context.
1875
        # Use different default values for different hardware.
1876
1877
1878
1879
1880
1881
1882

        # 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:
1883
            device_memory = current_platform.get_device_total_memory()
1884
            device_name = current_platform.get_device_name().lower()
1885
1886
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1887
            device_memory = 0
1888

1889
1890
1891
1892
        # 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:
1893
            # For GPUs like H100 and MI300x, use larger default values.
1894
1895
1896
1897
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1898
1899
1900
1901
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
1902
1903
1904
1905
1906
1907
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1908
1909
1910
1911
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
1912

1913
1914
        # tpu specific default values.
        if current_platform.is_tpu():
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
            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,
                }
1932

1933
1934
1935
        # cpu specific default values.
        if current_platform.is_cpu():
            default_max_num_batched_tokens = {
1936
1937
                UsageContext.LLM_CLASS: 4096 * world_size,
                UsageContext.OPENAI_API_SERVER: 2048 * world_size,
1938
1939
            }
            default_max_num_seqs = {
1940
1941
                UsageContext.LLM_CLASS: 256 * world_size,
                UsageContext.OPENAI_API_SERVER: 128 * world_size,
1942
1943
            }

1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
        return default_max_num_batched_tokens, default_max_num_seqs

    def _set_default_args(
        self, usage_context: UsageContext, model_config: ModelConfig
    ) -> None:
        """Set Default Arguments for V1 Engine."""
        (
            default_chunked_prefill,
            default_prefix_caching,
        ) = self.get_chunked_prefill_prefix_caching_defaults(model_config)

        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",
            )
        elif (
            model_config.runner_type == "pooling"
            and self.enable_chunked_prefill
            and not default_chunked_prefill
1966
        ):
1967
1968
1969
1970
1971
1972
1973
1974
1975
            logger.warning(
                "This model does not officially support chunked prefill. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
            )

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

1976
            logger.debug(
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
                "%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
        ):
            logger.warning(
                "This model does not officially support prefix caching. "
                "Enabling this manually may cause the engine to crash "
                "or produce incorrect outputs.",
            )

        world_size = self.pipeline_parallel_size * self.tensor_parallel_size
        (
            default_max_num_batched_tokens,
            default_max_num_seqs,
        ) = self.get_batch_defaults(world_size)

        orig_max_num_batched_tokens = self.max_num_batched_tokens
        orig_max_num_seqs = self.max_num_seqs

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

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

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

            # When using default settings,
            # Ensure max_num_batched_tokens does not exceed model limit.
            # Some models (e.g., Whisper) have embeddings tied to max length.
            self.max_num_batched_tokens = min(
                self.max_num_seqs * model_config.max_model_len,
2025
2026
                self.max_num_batched_tokens,
            )
2027

2028
2029
2030
2031
            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,
2032
            )
2033

2034
2035
2036
2037
        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)

2038
            logger.debug(
2039
                "Defaulting max_num_seqs to %d for %s usage context.",
2040
                self.max_num_seqs,
2041
                usage_context.value if usage_context else None,
2042
            )
2043

2044

2045
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
2046
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
2047
    """Arguments for asynchronous vLLM engine."""
2048

2049
2050
2051
2052
2053
2054
    enable_log_requests: bool = False

    @property
    @deprecated(
        "`disable_log_requests` is deprecated and has been replaced with "
        "`enable_log_requests`. This will be removed in v0.12.0. Please use "
2055
2056
        "`enable_log_requests` instead."
    )
2057
2058
2059
2060
2061
2062
2063
    def disable_log_requests(self) -> bool:
        return not self.enable_log_requests

    @disable_log_requests.setter
    @deprecated(
        "`disable_log_requests` is deprecated and has been replaced with "
        "`enable_log_requests`. This will be removed in v0.12.0. Please use "
2064
2065
        "`enable_log_requests` instead."
    )
2066
2067
    def disable_log_requests(self, value: bool):
        self.enable_log_requests = not value
2068
2069

    @staticmethod
2070
2071
2072
    def add_cli_args(
        parser: FlexibleArgumentParser, async_args_only: bool = False
    ) -> FlexibleArgumentParser:
2073
        # Initialize plugin to update the parser, for example, The plugin may
2074
        # add a new kind of quantization method to --quantization argument or
2075
2076
        # a new device to --device argument.
        load_general_plugins()
2077
2078
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
        parser.add_argument(
            "--enable-log-requests",
            action=argparse.BooleanOptionalAction,
            default=AsyncEngineArgs.enable_log_requests,
            help="Enable logging requests.",
        )
        parser.add_argument(
            "--disable-log-requests",
            action=argparse.BooleanOptionalAction,
            default=not AsyncEngineArgs.enable_log_requests,
            help="[DEPRECATED] Disable logging requests.",
            deprecated=True,
        )
2092
        current_platform.pre_register_and_update(parser)
2093
        return parser
2094
2095


2096
2097
2098
2099
2100
2101
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)
2102
2103


2104
2105
2106
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
2107

2108
2109
2110
2111
2112
2113
    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    """
    value = value.strip()
2114
    match = re.fullmatch(r"(\d+(?:\.\d+)?)([kKmMgGtT])", value)
2115
2116
    if match:
        decimal_multiplier = {
2117
2118
2119
            "k": 10**3,
            "m": 10**6,
            "g": 10**9,
2120
2121
        }
        binary_multiplier = {
2122
2123
2124
            "K": 2**10,
            "M": 2**20,
            "G": 2**30,
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
        }

        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:
2137
2138
2139
2140
2141
                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
2142
2143
2144

    # Regular plain number.
    return int(value)