arg_utils.py 81.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
# yapf: disable
5
import os
6
import argparse
7
import copy
8
import dataclasses
9
import functools
10
import json
11
import sys
12
from dataclasses import MISSING, dataclass, fields, is_dataclass
13
from itertools import permutations
14
15
16
from typing import (TYPE_CHECKING, Annotated, Any, Callable, Dict, List,
                    Literal, Optional, Type, TypeVar, Union, cast, get_args,
                    get_origin)
17

18
import huggingface_hub
19
import regex as re
20
import torch
21
from pydantic import TypeAdapter, ValidationError
22
from typing_extensions import TypeIs, deprecated
23

24
import vllm.envs as envs
25
from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
26
27
28
                         ConfigType, ConvertOption, DetailedTraceModules,
                         Device, DeviceConfig, DistributedExecutorBackend,
                         EPLBConfig, HfOverrides, KVEventsConfig,
29
                         KVTransferConfig, LoadConfig, LogprobsMode,
30
                         LoRAConfig, MambaDType, MMEncoderTPMode, ModelConfig,
31
32
33
34
35
                         ModelDType, ObservabilityConfig, ParallelConfig,
                         PoolerConfig, PrefixCachingHashAlgo, RunnerOption,
                         SchedulerConfig, SchedulerPolicy, SpeculativeConfig,
                         StructuredOutputsConfig, TaskOption, TokenizerMode,
                         VllmConfig, get_attr_docs)
36
from vllm.config.multimodal import MMCacheType, MultiModalConfig
37
from vllm.config.parallel import ExpertPlacementStrategy
38
from vllm.config.utils import get_field
39
from vllm.logger import init_logger
40
from vllm.platforms import CpuArchEnum, current_platform
41
from vllm.plugins import load_general_plugins
42
from vllm.ray.lazy_utils import is_ray_initialized
43
from vllm.reasoning import ReasoningParserManager
44
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
45
46
from vllm.transformers_utils.config import (get_model_path, is_interleaved,
                                            maybe_override_with_speculators)
47
from vllm.transformers_utils.utils import check_gguf_file
48
49
from vllm.utils import (FlexibleArgumentParser, GiB_bytes, get_ip,
                        is_in_ray_actor)
50
from vllm.v1.sample.logits_processor import LogitsProcessor
51
52

# yapf: enable
53

54
55
56
if TYPE_CHECKING:
    from vllm.executor.executor_base import ExecutorBase
    from vllm.model_executor.layers.quantization import QuantizationMethods
57
    from vllm.model_executor.model_loader import LoadFormats
58
59
60
61
    from vllm.usage.usage_lib import UsageContext
else:
    ExecutorBase = Any
    QuantizationMethods = Any
62
    LoadFormats = Any
63
64
    UsageContext = Any

65
66
logger = init_logger(__name__)

67
68
69
70
# object is used to allow for special typing forms
T = TypeVar("T")
TypeHint = Union[type[Any], object]
TypeHintT = Union[type[T], object]
71

72

73
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
74

75
    def _parse_type(val: str) -> T:
76
77
78
79
80
        try:
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
                f"Value {val} cannot be converted to {return_type}.") from e
81

82
83
84
85
86
87
88
89
90
91
92
    return _parse_type


def optional_type(
        return_type: Callable[[str], T]) -> Callable[[str], Optional[T]]:

    def _optional_type(val: str) -> Optional[T]:
        if val == "" or val == "None":
            return None
        return parse_type(return_type)(val)

93
    return _optional_type
94
95


96
def union_dict_and_str(val: str) -> Optional[Union[str, dict[str, str]]]:
97
    if not re.match(r"(?s)^\s*{.*}\s*$", val):
98
        return str(val)
99
    return optional_type(json.loads)(val)
100
101


102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
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)


117
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
118
119
120
121
    """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`.
    """
122
    type_hint = get_type(type_hints, Literal)
123
124
125
    options = get_args(type_hint)
    option_type = type(options[0])
    if not all(isinstance(option, option_type) for option in options):
126
        raise ValueError(
127
128
129
130
            "All options must be of the same type. "
            f"Got {options} with types {[type(c) for c in options]}")
    kwarg = "metavar" if contains_type(type_hints, str) else "choices"
    return {"type": option_type, kwarg: sorted(options)}
131
132


133
134
135
136
137
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
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]))
    elif origin is Union:
        for arg in args:
            type_hints.update(get_type_hints(arg))
    else:
        type_hints.add(type_hint)

    return type_hints


155
156
157
158
def is_online_quantization(quantization: Any) -> bool:
    return quantization in ["inc"]


159
NEEDS_HELP = (
160
161
    any("--help" in arg for arg in sys.argv)  # vllm SUBCOMMAND --help
    or (argv0 := sys.argv[0]).endswith("mkdocs")  # mkdocs SUBCOMMAND
162
163
164
165
    or argv0.endswith("mkdocs/__main__.py")  # python -m mkdocs SUBCOMMAND
)


166
167
@functools.lru_cache(maxsize=30)
def _compute_kwargs(cls: ConfigType) -> dict[str, Any]:
168
169
    # Save time only getting attr docs if we're generating help text
    cls_docs = get_attr_docs(cls) if NEEDS_HELP else {}
170
171
    kwargs = {}
    for field in fields(cls):
172
        # Get the set of possible types for the field
173
        type_hints: set[TypeHint] = get_type_hints(field.type)
174
175
176
177
178

        # 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)

179
        # Get the default value of the field
180
181
182
        if field.default is not MISSING:
            default = field.default
        elif field.default_factory is not MISSING:
183
184
185
186
            default = field.default_factory()

        # Get the help text for the field
        name = field.name
187
        help = cls_docs.get(name, "").strip()
188
189
190
191
192
193
194
        # 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
195
196
        json_tip = ("Should either be a valid JSON string or JSON keys passed "
                    "individually.")
197
        if dataclass_cls is not None:
198
199
200
201
202
203
204
205

            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
206
            kwargs[name]["help"] += f"\n\n{json_tip}"
207
        elif contains_type(type_hints, bool):
208
209
210
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
211
            kwargs[name].update(literal_to_kwargs(type_hints))
212
213
214
215
216
217
218
219
220
221
222
223
        elif contains_type(type_hints, tuple):
            type_hint = get_type(type_hints, tuple)
            types = get_args(type_hint)
            tuple_type = types[0]
            assert all(t is tuple_type for t in types if t is not Ellipsis), (
                "All non-Ellipsis tuple elements must be of the same "
                f"type. Got {types}.")
            kwargs[name]["type"] = tuple_type
            kwargs[name]["nargs"] = "+" if Ellipsis in types else len(types)
        elif contains_type(type_hints, list):
            type_hint = get_type(type_hints, list)
            types = get_args(type_hint)
224
225
226
227
228
229
            list_type = types[0]
            if get_origin(list_type) is Union:
                msg = "List type must contain str if it is a Union."
                assert str in get_args(list_type), msg
                list_type = str
            kwargs[name]["type"] = list_type
230
231
232
            kwargs[name]["nargs"] = "+"
        elif contains_type(type_hints, int):
            kwargs[name]["type"] = int
233
            # Special case for large integers
234
235
236
237
238
239
            human_readable_ints = {
                "max_model_len",
                "max_num_batched_tokens",
                "kv_cache_memory_bytes",
            }
            if name in human_readable_ints:
240
                kwargs[name]["type"] = human_readable_int
241
                kwargs[name]["help"] += f"\n\n{human_readable_int.__doc__}"
242
243
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
244
245
246
        elif (contains_type(type_hints, dict)
              and (contains_type(type_hints, str)
                   or any(is_not_builtin(th) for th in type_hints))):
247
            kwargs[name]["type"] = union_dict_and_str
248
        elif contains_type(type_hints, dict):
249
            kwargs[name]["type"] = parse_type(json.loads)
250
            kwargs[name]["help"] += f"\n\n{json_tip}"
251
252
253
254
255
256
257
        elif (contains_type(type_hints, str)
              or any(is_not_builtin(th) for th in type_hints)):
            kwargs[name]["type"] = str
        else:
            raise ValueError(
                f"Unsupported type {type_hints} for argument {name}.")

258
259
260
261
262
        # 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"]}))

263
264
265
266
267
268
269
        # 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
270
271


272
273
274
def get_kwargs(cls: ConfigType) -> dict[str, Any]:
    """Return argparse kwargs for the given Config dataclass.

275
276
277
    If `--help` or `mkdocs` are not present in the command line command, the
    attribute documentation will not be included in the help output.

278
279
280
281
282
283
284
    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))


285
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
286
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
287
    """Arguments for vLLM engine."""
288
289
290
291
292
    model: str = ModelConfig.model
    served_model_name: Optional[Union[
        str, List[str]]] = ModelConfig.served_model_name
    tokenizer: Optional[str] = ModelConfig.tokenizer
    hf_config_path: Optional[str] = ModelConfig.hf_config_path
293
294
295
    runner: RunnerOption = ModelConfig.runner
    convert: ConvertOption = ModelConfig.convert
    task: Optional[TaskOption] = ModelConfig.task
296
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
297
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
298
299
300
    tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
301
302
    allowed_media_domains: Optional[
        list[str]] = ModelConfig.allowed_media_domains
303
    download_dir: Optional[str] = LoadConfig.download_dir
304
    safetensors_load_strategy: str = LoadConfig.safetensors_load_strategy
305
    load_format: Union[str, LoadFormats] = LoadConfig.load_format
306
307
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
308
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
309
310
    seed: Optional[int] = ModelConfig.seed
    max_model_len: Optional[int] = ModelConfig.max_model_len
311
312
    cuda_graph_sizes: list[int] = get_field(SchedulerConfig,
                                            "cuda_graph_sizes")
313
314
315
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
316
    distributed_executor_backend: Optional[Union[
317
        str, DistributedExecutorBackend,
318
        Type[ExecutorBase]]] = ParallelConfig.distributed_executor_backend
319
    # number of P/D disaggregation (or other disaggregation) workers
320
321
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
322
323
    decode_context_parallel_size: int = \
        ParallelConfig.decode_context_parallel_size
324
    data_parallel_size: int = ParallelConfig.data_parallel_size
325
    data_parallel_rank: Optional[int] = None
326
    data_parallel_start_rank: Optional[int] = None
327
328
329
    data_parallel_size_local: Optional[int] = None
    data_parallel_address: Optional[str] = None
    data_parallel_rpc_port: Optional[int] = None
330
    data_parallel_hybrid_lb: bool = False
Rui Qiao's avatar
Rui Qiao committed
331
    data_parallel_backend: str = ParallelConfig.data_parallel_backend
332
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
333
334
335
    enable_dbo: bool = ParallelConfig.enable_dbo
    dbo_decode_token_threshold: int = \
        ParallelConfig.dbo_decode_token_threshold
336
337
    dbo_prefill_token_threshold: int = \
        ParallelConfig.dbo_prefill_token_threshold
338
    eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config")
339
    enable_eplb: bool = ParallelConfig.enable_eplb
340
341
    expert_placement_strategy: ExpertPlacementStrategy = \
        ParallelConfig.expert_placement_strategy
342
343
    _api_process_count: int = ParallelConfig._api_process_count
    _api_process_rank: int = ParallelConfig._api_process_rank
344
345
346
347
    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
348
349
    max_parallel_loading_workers: Optional[
        int] = ParallelConfig.max_parallel_loading_workers
350
351
352
353
    block_size: Optional[BlockSize] = CacheConfig.block_size
    enable_prefix_caching: Optional[bool] = CacheConfig.enable_prefix_caching
    prefix_caching_hash_algo: PrefixCachingHashAlgo = \
        CacheConfig.prefix_caching_hash_algo
354
355
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
356
357
358
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
359
    kv_cache_memory_bytes: Optional[int] = CacheConfig.kv_cache_memory_bytes
360
361
362
363
364
365
366
    max_num_batched_tokens: Optional[
        int] = SchedulerConfig.max_num_batched_tokens
    max_num_partial_prefills: int = SchedulerConfig.max_num_partial_prefills
    max_long_partial_prefills: int = SchedulerConfig.max_long_partial_prefills
    long_prefill_token_threshold: int = \
        SchedulerConfig.long_prefill_token_threshold
    max_num_seqs: Optional[int] = SchedulerConfig.max_num_seqs
367
    max_logprobs: int = ModelConfig.max_logprobs
368
    logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode
369
    disable_log_stats: bool = False
370
371
372
373
374
    revision: Optional[str] = ModelConfig.revision
    code_revision: Optional[str] = ModelConfig.code_revision
    rope_scaling: dict[str, Any] = get_field(ModelConfig, "rope_scaling")
    rope_theta: Optional[float] = ModelConfig.rope_theta
    hf_token: Optional[Union[bool, str]] = ModelConfig.hf_token
375
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
376
377
378
    tokenizer_revision: Optional[str] = ModelConfig.tokenizer_revision
    quantization: Optional[QuantizationMethods] = ModelConfig.quantization
    enforce_eager: bool = ModelConfig.enforce_eager
379
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
380
    limit_mm_per_prompt: dict[str, int] = \
381
        get_field(MultiModalConfig, "limit_per_prompt")
382
    enable_mm_embeds: bool = MultiModalConfig.enable_mm_embeds
383
    interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings
384
385
386
    media_io_kwargs: dict[str, dict[str,
                                    Any]] = get_field(MultiModalConfig,
                                                      "media_io_kwargs")
387
388
    mm_processor_kwargs: Optional[Dict[str, Any]] = \
        MultiModalConfig.mm_processor_kwargs
389
    disable_mm_preprocessor_cache: bool = False  # DEPRECATED
390
    mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb
391
392
393
394
    mm_processor_cache_type: Optional[MMCacheType] = \
        MultiModalConfig.mm_processor_cache_type
    mm_shm_cache_max_object_size_mb: int = \
        MultiModalConfig.mm_shm_cache_max_object_size_mb
395
    mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
396
    io_processor_plugin: Optional[str] = None
397
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
398
    video_pruning_rate: float = MultiModalConfig.video_pruning_rate
399
    # LoRA fields
400
    enable_lora: bool = False
401
402
403
    enable_lora_bias: bool = LoRAConfig.bias_enabled
    max_loras: int = LoRAConfig.max_loras
    max_lora_rank: int = LoRAConfig.max_lora_rank
404
405
    default_mm_loras: Optional[Dict[str, str]] = \
        LoRAConfig.default_mm_loras
406
407
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
    max_cpu_loras: Optional[int] = LoRAConfig.max_cpu_loras
zhuwenwen's avatar
zhuwenwen committed
408
    lora_target_modules: Optional[List[str]] = LoRAConfig.lora_target_modules
409
410
411
    lora_dtype: Optional[Union[str, torch.dtype]] = LoRAConfig.lora_dtype
    lora_extra_vocab_size: int = LoRAConfig.lora_extra_vocab_size

412
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
413
414
    num_gpu_blocks_override: Optional[
        int] = CacheConfig.num_gpu_blocks_override
415
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
416
417
    model_loader_extra_config: dict = \
        get_field(LoadConfig, "model_loader_extra_config")
418
419
    ignore_patterns: Optional[Union[str,
                                    List[str]]] = LoadConfig.ignore_patterns
420

421
422
423
    enable_chunked_prefill: Optional[
        bool] = SchedulerConfig.enable_chunked_prefill
    disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
424

425
426
427
    disable_hybrid_kv_cache_manager: bool = (
        SchedulerConfig.disable_hybrid_kv_cache_manager)

428
429
430
431
432
433
434
435
436
    structured_outputs_config: StructuredOutputsConfig = get_field(
        VllmConfig, "structured_outputs_config")
    reasoning_parser: str = StructuredOutputsConfig.reasoning_parser
    # Deprecated guided decoding fields
    guided_decoding_backend: Optional[str] = None
    guided_decoding_disable_fallback: Optional[bool] = None
    guided_decoding_disable_any_whitespace: Optional[bool] = None
    guided_decoding_disable_additional_properties: Optional[bool] = None

437
438
    logits_processor_pattern: Optional[
        str] = ModelConfig.logits_processor_pattern
439

440
    speculative_config: Optional[Dict[str, Any]] = None
zhuwenwen's avatar
zhuwenwen committed
441
    num_speculative_heads: Optional[int] = None
442

443
444
445
446
447
448
    show_hidden_metrics_for_version: Optional[str] = \
        ObservabilityConfig.show_hidden_metrics_for_version
    otlp_traces_endpoint: Optional[str] = \
        ObservabilityConfig.otlp_traces_endpoint
    collect_detailed_traces: Optional[list[DetailedTraceModules]] = \
        ObservabilityConfig.collect_detailed_traces
449
450
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
    scheduler_cls: Union[str, Type[object]] = SchedulerConfig.scheduler_cls
451

452
    pooler_config: Optional[PoolerConfig] = ModelConfig.pooler_config
453
454
    override_pooler_config: Optional[Union[dict, PoolerConfig]] = \
        ModelConfig.override_pooler_config
455
456
    compilation_config: CompilationConfig = \
        get_field(VllmConfig, "compilation_config")
457
458
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
459

460
    kv_transfer_config: Optional[KVTransferConfig] = None
461
    kv_events_config: Optional[KVEventsConfig] = None
462

463
464
465
466
467
    generation_config: str = ModelConfig.generation_config
    enable_sleep_mode: bool = ModelConfig.enable_sleep_mode
    override_generation_config: dict[str, Any] = \
        get_field(ModelConfig, "override_generation_config")
    model_impl: str = ModelConfig.model_impl
468
    override_attention_dtype: str = ModelConfig.override_attention_dtype
469

470
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
471
472
    mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype
    mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype
473

474
475
    additional_config: dict[str, Any] = \
        get_field(VllmConfig, "additional_config")
476

477
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
478
    pt_load_map_location: str = LoadConfig.pt_load_map_location
王敏's avatar
王敏 committed
479

480
481
    # DEPRECATED
    enable_multimodal_encoder_data_parallel: bool = False
王敏's avatar
王敏 committed
482

483
484
485
486
    logits_processors: Optional[list[Union[
        str, type[LogitsProcessor]]]] = ModelConfig.logits_processors
    """Custom logitproc types"""

487
488
    async_scheduling: bool = SchedulerConfig.async_scheduling

489
490
491
    kv_sharing_fast_prefill: bool = \
        CacheConfig.kv_sharing_fast_prefill

492
    def __post_init__(self):
493
494
495
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
496
497
498
        if isinstance(self.compilation_config, dict):
            self.compilation_config = CompilationConfig(
                **self.compilation_config)
499
        if isinstance(self.eplb_config, dict):
500
            self.eplb_config = EPLBConfig(**self.eplb_config)
501
        # Setup plugins
502
503
        from vllm.plugins import load_general_plugins
        load_general_plugins()
504
505
506
507
508
509
510
        # 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(
                "HF_HUB_OFFLINE is True, replace model_id [%s] " \
                "to model_path [%s]",model_id, self.model)
511
512

    @staticmethod
513
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
514
        """Shared CLI arguments for vLLM engine."""
515

516
        # Model arguments
517
518
519
520
521
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
Reid's avatar
Reid committed
522
        if not ('serve' in sys.argv[1:] and '--help' in sys.argv[1:]):
523
            model_group.add_argument("--model", **model_kwargs["model"])
524
525
526
527
528
        model_group.add_argument("--runner", **model_kwargs["runner"])
        model_group.add_argument("--convert", **model_kwargs["convert"])
        model_group.add_argument("--task",
                                 **model_kwargs["task"],
                                 deprecated=True)
529
530
531
532
533
534
535
536
537
538
539
        model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"])
        model_group.add_argument("--tokenizer-mode",
                                 **model_kwargs["tokenizer_mode"])
        model_group.add_argument("--trust-remote-code",
                                 **model_kwargs["trust_remote_code"])
        model_group.add_argument("--dtype", **model_kwargs["dtype"])
        model_group.add_argument("--seed", **model_kwargs["seed"])
        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"])
540
541
        model_group.add_argument("--allowed-media-domains",
                                 **model_kwargs["allowed_media_domains"])
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
        model_group.add_argument("--revision", **model_kwargs["revision"])
        model_group.add_argument("--code-revision",
                                 **model_kwargs["code_revision"])
        model_group.add_argument("--rope-scaling",
                                 **model_kwargs["rope_scaling"])
        model_group.add_argument("--rope-theta", **model_kwargs["rope_theta"])
        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"])
558
559
        model_group.add_argument("--logprobs-mode",
                                 **model_kwargs["logprobs_mode"])
560
561
562
563
564
565
        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"])
566
567
        model_group.add_argument("--enable-prompt-embeds",
                                 **model_kwargs["enable_prompt_embeds"])
568
569
570
571
572
573
574
575
576
577
578
579
580
581
        model_group.add_argument("--served-model-name",
                                 **model_kwargs["served_model_name"])
        model_group.add_argument("--config-format",
                                 **model_kwargs["config_format"])
        # This one is a special case because it can bool
        # or str. TODO: Handle this in get_kwargs
        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"])
582
583
        model_group.add_argument("--pooler-config",
                                 **model_kwargs["pooler_config"])
584
        model_group.add_argument("--override-pooler-config",
585
586
                                 **model_kwargs["override_pooler_config"],
                                 deprecated=True)
587
588
589
590
591
592
593
594
        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"])
595
        model_group.add_argument("--model-impl", **model_kwargs["model_impl"])
596
597
        model_group.add_argument("--override-attention-dtype",
                                 **model_kwargs["override_attention_dtype"])
598
599
        model_group.add_argument("--logits-processors",
                                 **model_kwargs["logits_processors"])
600
601
        model_group.add_argument("--io-processor-plugin",
                                 **model_kwargs["io_processor_plugin"])
602

603
604
605
606
607
608
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
609
        load_group.add_argument("--load-format", **load_kwargs["load_format"])
610
        load_group.add_argument("--download-dir",
611
                                **load_kwargs["download_dir"])
612
613
        load_group.add_argument("--safetensors-load-strategy",
                                **load_kwargs["safetensors_load_strategy"])
614
        load_group.add_argument("--model-loader-extra-config",
615
                                **load_kwargs["model_loader_extra_config"])
616
617
618
        load_group.add_argument("--ignore-patterns",
                                **load_kwargs["ignore_patterns"])
        load_group.add_argument("--use-tqdm-on-load",
619
                                **load_kwargs["use_tqdm_on_load"])
620
621
        load_group.add_argument('--pt-load-map-location',
                                **load_kwargs["pt_load_map_location"])
622

623
624
625
626
627
        # Structured outputs arguments
        structured_outputs_kwargs = get_kwargs(StructuredOutputsConfig)
        structured_outputs_group = parser.add_argument_group(
            title="StructuredOutputsConfig",
            description=StructuredOutputsConfig.__doc__,
628
        )
629
        structured_outputs_group.add_argument(
630
            "--reasoning-parser",
631
            # This choice is a special case because it's not static
632
            choices=list(ReasoningParserManager.reasoning_parsers),
633
634
635
636
637
638
639
640
641
642
643
644
645
            **structured_outputs_kwargs["reasoning_parser"])
        # 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."),
                deprecated=True)
646

647
        # Parallel arguments
648
649
650
651
652
653
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
654
            "--distributed-executor-backend",
655
656
            **parallel_kwargs["distributed_executor_backend"])
        parallel_group.add_argument(
657
            "--pipeline-parallel-size", "-pp",
658
            **parallel_kwargs["pipeline_parallel_size"])
659
        parallel_group.add_argument("--tensor-parallel-size", "-tp",
660
                                    **parallel_kwargs["tensor_parallel_size"])
661
662
663
        parallel_group.add_argument(
            "--decode-context-parallel-size", "-dcp",
            **parallel_kwargs["decode_context_parallel_size"])
664
        parallel_group.add_argument("--data-parallel-size", "-dp",
665
                                    **parallel_kwargs["data_parallel_size"])
666
667
668
669
670
671
        parallel_group.add_argument(
            '--data-parallel-rank',
            '-dpn',
            type=int,
            help='Data parallel rank of this instance. '
            'When set, enables external load balancer mode.')
672
673
674
675
676
        parallel_group.add_argument('--data-parallel-start-rank',
                                    '-dpr',
                                    type=int,
                                    help='Starting data parallel rank '
                                    'for secondary nodes.')
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
        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.')
Rui Qiao's avatar
Rui Qiao committed
692
693
694
695
696
697
        parallel_group.add_argument('--data-parallel-backend',
                                    '-dpb',
                                    type=str,
                                    default='mp',
                                    help='Backend for data parallel, either '
                                    '"mp" or "ray".')
698
699
700
        parallel_group.add_argument(
            "--data-parallel-hybrid-lb",
            **parallel_kwargs["data_parallel_hybrid_lb"])
701
        parallel_group.add_argument(
702
            "--enable-expert-parallel",
703
            **parallel_kwargs["enable_expert_parallel"])
704
705
706
707
708
        parallel_group.add_argument("--enable-dbo",
                                    **parallel_kwargs["enable_dbo"])
        parallel_group.add_argument(
            "--dbo-decode-token-threshold",
            **parallel_kwargs["dbo_decode_token_threshold"])
709
710
711
        parallel_group.add_argument(
            "--dbo-prefill-token-threshold",
            **parallel_kwargs["dbo_prefill_token_threshold"])
712
713
        parallel_group.add_argument("--enable-eplb",
                                    **parallel_kwargs["enable_eplb"])
714
715
        parallel_group.add_argument("--eplb-config",
                                    **parallel_kwargs["eplb_config"])
716
717
718
        parallel_group.add_argument(
            "--expert-placement-strategy",
            **parallel_kwargs["expert_placement_strategy"])
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
        parallel_group.add_argument(
            "--num-redundant-experts",
            type=int,
            help=
            "[DEPRECATED] --num-redundant-experts will be removed in v0.12.0.",
            deprecated=True)
        parallel_group.add_argument(
            "--eplb-window-size",
            type=int,
            help="[DEPRECATED] --eplb-window-size will be removed in v0.12.0.",
            deprecated=True)
        parallel_group.add_argument(
            "--eplb-step-interval",
            type=int,
            help=
            "[DEPRECATED] --eplb-step-interval will be removed in v0.12.0.",
            deprecated=True)
        parallel_group.add_argument(
            "--eplb-log-balancedness",
            action=argparse.BooleanOptionalAction,
            help=
            "[DEPRECATED] --eplb-log-balancedness will be removed in v0.12.0.",
            deprecated=True)

743
        parallel_group.add_argument(
744
            "--max-parallel-loading-workers",
745
746
            **parallel_kwargs["max_parallel_loading_workers"])
        parallel_group.add_argument(
747
            "--ray-workers-use-nsight",
748
749
            **parallel_kwargs["ray_workers_use_nsight"])
        parallel_group.add_argument(
750
            "--disable-custom-all-reduce",
751
            **parallel_kwargs["disable_custom_all_reduce"])
752
753
754
755
        parallel_group.add_argument("--worker-cls",
                                    **parallel_kwargs["worker_cls"])
        parallel_group.add_argument("--worker-extension-cls",
                                    **parallel_kwargs["worker_extension_cls"])
756
757
        parallel_group.add_argument(
            "--enable-multimodal-encoder-data-parallel",
758
759
            action="store_true",
            deprecated=True)
760

761
762
763
764
765
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
766
        )
767
768
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
        cache_group.add_argument("--gpu-memory-utilization",
769
                                 **cache_kwargs["gpu_memory_utilization"])
770
771
        cache_group.add_argument("--kv-cache-memory-bytes",
                                 **cache_kwargs["kv_cache_memory_bytes"])
772
773
        cache_group.add_argument("--swap-space", **cache_kwargs["swap_space"])
        cache_group.add_argument("--kv-cache-dtype",
774
                                 **cache_kwargs["cache_dtype"])
775
        cache_group.add_argument("--num-gpu-blocks-override",
776
777
778
779
780
                                 **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"])
781
        cache_group.add_argument("--cpu-offload-gb",
782
                                 **cache_kwargs["cpu_offload_gb"])
783
        cache_group.add_argument("--calculate-kv-scales",
784
                                 **cache_kwargs["calculate_kv_scales"])
785
786
        cache_group.add_argument("--kv-sharing-fast-prefill",
                                 **cache_kwargs["kv_sharing_fast_prefill"])
787
788
789
790
        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"])
791

792
        # Multimodal related configs
793
794
795
796
797
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
798
        multimodal_group.add_argument("--limit-mm-per-prompt",
799
                                      **multimodal_kwargs["limit_per_prompt"])
800
801
802
        multimodal_group.add_argument(
            "--enable-mm-embeds", **multimodal_kwargs["enable_mm_embeds"]
        )
803
804
        multimodal_group.add_argument("--media-io-kwargs",
                                      **multimodal_kwargs["media_io_kwargs"])
805
        multimodal_group.add_argument(
806
            "--mm-processor-kwargs",
807
808
            **multimodal_kwargs["mm_processor_kwargs"])
        multimodal_group.add_argument(
809
810
811
            "--mm-processor-cache-gb",
            **multimodal_kwargs["mm_processor_cache_gb"])
        multimodal_group.add_argument("--disable-mm-preprocessor-cache",
812
                                      action="store_true",
813
                                      deprecated=True)
814
815
816
817
818
819
        multimodal_group.add_argument(
            "--mm-processor-cache-type",
            **multimodal_kwargs["mm_processor_cache_type"])
        multimodal_group.add_argument(
            "--mm-shm-cache-max-object-size-mb",
            **multimodal_kwargs["mm_shm_cache_max_object_size_mb"])
820
821
        multimodal_group.add_argument(
            "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"])
822
823
824
        multimodal_group.add_argument(
            "--interleave-mm-strings",
            **multimodal_kwargs["interleave_mm_strings"])
825
826
        multimodal_group.add_argument("--skip-mm-profiling",
                                      **multimodal_kwargs["skip_mm_profiling"])
827

828
829
830
        multimodal_group.add_argument(
            "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"])

831
        # LoRA related configs
832
833
834
835
836
837
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
838
            "--enable-lora",
839
            action=argparse.BooleanOptionalAction,
840
841
            help="If True, enable handling of LoRA adapters.")
        lora_group.add_argument("--enable-lora-bias",
842
                                **lora_kwargs["bias_enabled"])
843
844
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
        lora_group.add_argument("--max-lora-rank",
845
                                **lora_kwargs["max_lora_rank"])
zhuwenwen's avatar
zhuwenwen committed
846
847
        lora_group.add_argument('--lora-target-modules',
                            **lora_kwargs["lora_target_modules"])
848
        lora_group.add_argument("--lora-extra-vocab-size",
849
850
                                **lora_kwargs["lora_extra_vocab_size"])
        lora_group.add_argument(
851
            "--lora-dtype",
852
853
            **lora_kwargs["lora_dtype"],
        )
854
        lora_group.add_argument("--max-cpu-loras",
855
                                **lora_kwargs["max_cpu_loras"])
856
        lora_group.add_argument("--fully-sharded-loras",
857
                                **lora_kwargs["fully_sharded_loras"])
858
859
        lora_group.add_argument("--default-mm-loras",
                                **lora_kwargs["default_mm_loras"])
860

861

862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
        # 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",
            **observability_kwargs["show_hidden_metrics_for_version"])
        observability_group.add_argument(
            "--otlp-traces-endpoint",
            **observability_kwargs["otlp_traces_endpoint"])
        # 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"] += [
            ",".join(p)
            for p in permutations(get_args(DetailedTraceModules), r=2)
        ]
        observability_group.add_argument(
            "--collect-detailed-traces",
            **observability_kwargs["collect_detailed_traces"])
885

886
887
888
889
890
891
892
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
893
            "--max-num-batched-tokens",
894
            **scheduler_kwargs["max_num_batched_tokens"])
895
        scheduler_group.add_argument("--max-num-seqs",
896
897
898
899
900
901
902
                                     **scheduler_kwargs["max_num_seqs"])
        scheduler_group.add_argument(
            "--max-num-partial-prefills",
            **scheduler_kwargs["max_num_partial_prefills"])
        scheduler_group.add_argument(
            "--max-long-partial-prefills",
            **scheduler_kwargs["max_long_partial_prefills"])
903
904
        scheduler_group.add_argument('--cuda-graph-sizes',
                                     **scheduler_kwargs["cuda_graph_sizes"])
905
906
907
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
            **scheduler_kwargs["long_prefill_token_threshold"])
908
        scheduler_group.add_argument("--num-lookahead-slots",
909
                                     **scheduler_kwargs["num_lookahead_slots"])
910
911
        # multi-step scheduling has been removed; corresponding arguments
        # are no longer supported.
912
        scheduler_group.add_argument("--scheduling-policy",
913
                                     **scheduler_kwargs["policy"])
914
        scheduler_group.add_argument(
915
            "--enable-chunked-prefill",
916
            **scheduler_kwargs["enable_chunked_prefill"])
917
918
919
        scheduler_group.add_argument(
            "--disable-chunked-mm-input",
            **scheduler_kwargs["disable_chunked_mm_input"])
920
921
        scheduler_group.add_argument("--scheduler-cls",
                                     **scheduler_kwargs["scheduler_cls"])
922
923
924
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"])
925
926
        scheduler_group.add_argument("--async-scheduling",
                                     **scheduler_kwargs["async_scheduling"])
927
928

        # vLLM arguments
929
        vllm_kwargs = get_kwargs(VllmConfig)
930
931
932
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
933
        )
934
935
936
937
        # 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)
938
939
        vllm_group.add_argument("--speculative-config",
                                **vllm_kwargs["speculative_config"])
940
941
942
943
944
945
946
947
        vllm_group.add_argument("--kv-transfer-config",
                                **vllm_kwargs["kv_transfer_config"])
        vllm_group.add_argument('--kv-events-config',
                                **vllm_kwargs["kv_events_config"])
        vllm_group.add_argument("--compilation-config", "-O",
                                **vllm_kwargs["compilation_config"])
        vllm_group.add_argument("--additional-config",
                                **vllm_kwargs["additional_config"])
948
949
        vllm_group.add_argument('--structured-outputs-config',
                                **vllm_kwargs["structured_outputs_config"])
950

951
952
953
954
        # Other arguments
        parser.add_argument('--disable-log-stats',
                            action='store_true',
                            help='Disable logging statistics.')
955
956
957
958
959
960
        
        parser.add_argument('--num-speculative-heads',
                            type=int,
                            default=EngineArgs.num_speculative_heads,
                            help='The number of speculative heads to sample from '
                                'the draft model in speculative decoding.')
961

962
        return parser
963
964

    @classmethod
965
    def from_cli_args(cls, args: argparse.Namespace):
966
967
968
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
969
970
971
972
        engine_args = cls(**{
            attr: getattr(args, attr)
            for attr in attrs if hasattr(args, attr)
        })
Zhuohan Li's avatar
Zhuohan Li committed
973
        return engine_args
974

975
    def create_model_config(self) -> ModelConfig:
976
977
978
979
980
981
        # 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"

        # NOTE: This is to allow model loading from S3 in CI
        if (not isinstance(self, AsyncEngineArgs) and envs.VLLM_CI_USE_S3
982
                and self.model in MODELS_ON_S3 and self.load_format == "auto"):
983
984
            self.model = f"{MODEL_WEIGHTS_S3_BUCKET}/{self.model}"

985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
        if self.disable_mm_preprocessor_cache:
            logger.warning(
                "`--disable-mm-preprocessor-cache` is deprecated "
                "and will be removed in v0.13. "
                "Please use `--mm-processor-cache-gb 0` instead.", )

            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
1001

1002
1003
1004
1005
1006
1007
1008
1009
        if self.enable_multimodal_encoder_data_parallel:
            logger.warning(
                "--enable-multimodal-encoder-data-parallel` is deprecated "
                "and will be removed in v0.13. "
                "Please use `--mm-encoder-tp-mode data` instead.")

            self.mm_encoder_tp_mode = "data"

1010
        return ModelConfig(
1011
            model=self.model,
1012
            hf_config_path=self.hf_config_path,
1013
1014
            runner=self.runner,
            convert=self.convert,
1015
            task=self.task,
1016
            tokenizer=self.tokenizer,
1017
1018
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1019
            allowed_local_media_path=self.allowed_local_media_path,
1020
            allowed_media_domains=self.allowed_media_domains,
1021
1022
1023
1024
1025
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
1026
            rope_theta=self.rope_theta,
1027
            hf_token=self.hf_token,
1028
            hf_overrides=self.hf_overrides,
1029
1030
1031
1032
1033
            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,
1034
            logprobs_mode=self.logprobs_mode,
1035
            disable_sliding_window=self.disable_sliding_window,
1036
            disable_cascade_attn=self.disable_cascade_attn,
1037
            skip_tokenizer_init=self.skip_tokenizer_init,
1038
            enable_prompt_embeds=self.enable_prompt_embeds,
1039
            served_model_name=self.served_model_name,
1040
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1041
            enable_mm_embeds=self.enable_mm_embeds,
1042
            interleave_mm_strings=self.interleave_mm_strings,
1043
            media_io_kwargs=self.media_io_kwargs,
1044
            skip_mm_profiling=self.skip_mm_profiling,
1045
            config_format=self.config_format,
1046
            mm_processor_kwargs=self.mm_processor_kwargs,
1047
            mm_processor_cache_gb=self.mm_processor_cache_gb,
1048
1049
1050
            mm_processor_cache_type=self.mm_processor_cache_type,
            mm_shm_cache_max_object_size_mb=self.
            mm_shm_cache_max_object_size_mb,
1051
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
1052
            pooler_config=self.pooler_config,
1053
            override_pooler_config=self.override_pooler_config,
1054
            logits_processor_pattern=self.logits_processor_pattern,
1055
            generation_config=self.generation_config,
1056
            override_generation_config=self.override_generation_config,
1057
            enable_sleep_mode=self.enable_sleep_mode,
1058
            model_impl=self.model_impl,
1059
            override_attention_dtype=self.override_attention_dtype,
1060
            logits_processors=self.logits_processors,
1061
            video_pruning_rate=self.video_pruning_rate,
1062
            io_processor_plugin=self.io_processor_plugin,
1063
            enable_chunked_prefill=self.enable_chunked_prefill,
1064
        )
1065

1066
1067
1068
1069
1070
1071
1072
    def validate_tensorizer_args(self):
        from vllm.model_executor.model_loader.tensorizer import (
            TensorizerConfig)
        for key in self.model_loader_extra_config:
            if key in TensorizerConfig._fields:
                self.model_loader_extra_config["tensorizer_config"][
                    key] = self.model_loader_extra_config[key]
1073

1074
1075
    def create_load_config(self) -> LoadConfig:

1076
1077
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1078

1079
1080
1081
1082
1083
1084
1085
1086
        if self.load_format == "tensorizer":
            if hasattr(self.model_loader_extra_config, "to_serializable"):
                self.model_loader_extra_config = (
                    self.model_loader_extra_config.to_serializable())
            self.model_loader_extra_config["tensorizer_config"] = {}
            self.model_loader_extra_config["tensorizer_config"][
                "tensorizer_dir"] = self.model
            self.validate_tensorizer_args()
1087

1088
1089
1090
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
1091
            safetensors_load_strategy=self.safetensors_load_strategy,
1092
1093
            device="cpu"
            if is_online_quantization(self.quantization) else None,
1094
1095
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1096
            use_tqdm_on_load=self.use_tqdm_on_load,
1097
            pt_load_map_location=self.pt_load_map_location,
1098
1099
        )

1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        enable_chunked_prefill: bool,
        disable_log_stats: bool,
    ) -> Optional["SpeculativeConfig"]:
        """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
1113
        dictionary from the engine.
1114
1115
        """
        if self.speculative_config is None:
1116
            return None
1117

1118
1119
1120
1121
1122
1123
1124
1125
1126
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.
        self.speculative_config.update({
            "target_model_config": target_model_config,
            "target_parallel_config": target_parallel_config,
            "enable_chunked_prefill": enable_chunked_prefill,
            "disable_log_stats": disable_log_stats,
        })
1127
        return SpeculativeConfig(**self.speculative_config)
1128

1129
1130
1131
    def create_engine_config(
        self,
        usage_context: Optional[UsageContext] = None,
1132
        headless: bool = False,
1133
1134
1135
    ) -> VllmConfig:
        """
        Create the VllmConfig.
1136

1137
1138
1139
        NOTE: for autoselection of V0 vs V1 engine, we need to
        create the ModelConfig first, since ModelConfig's attrs
        (e.g. the model arch) are needed to make the decision.
1140

1141
1142
        This function set VLLM_USE_V1=X if VLLM_USE_V1 is
        unspecified by the user.
1143

1144
1145
1146
        If VLLM_USE_V1 is specified by the user but the VllmConfig
        is incompatible, we raise an error.
        """
1147
        current_platform.pre_register_and_update()
1148

1149
1150
        device_config = DeviceConfig(
            device=cast(Device, current_platform.device_type))
1151
1152
1153
1154
1155
1156
1157
1158
1159

        (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,
         )
1160
1161
        model_config = self.create_model_config()

1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
        # * If VLLM_USE_V1 is unset, we enable V1 for "supported features"
        #   and fall back to V0 for experimental or unsupported features.
        # * If VLLM_USE_V1=1, we enable V1 for supported + experimental
        #   features and raise error for unsupported features.
        # * If VLLM_USE_V1=0, we disable V1.
        use_v1 = False
        try_v1 = envs.VLLM_USE_V1 or not envs.is_set("VLLM_USE_V1")
        if try_v1 and self._is_v1_supported_oracle(model_config):
            use_v1 = True

        # If user explicitly set VLLM_USE_V1, sanity check we respect it.
        if envs.is_set("VLLM_USE_V1"):
            assert use_v1 == envs.VLLM_USE_V1
        # Otherwise, set the VLLM_USE_V1 variable globally.
        else:
            envs.set_vllm_use_v1(use_v1)

1179
1180
1181
1182
1183
1184
1185
1186
1187
        # Set default arguments for V1 Engine.
        self._set_default_args(usage_context, model_config)
        # Disable chunked prefill for POWER (ppc64le)/ARM/s390x CPUs in V1
        if current_platform.is_cpu() and current_platform.get_cpu_architecture(
        ) in (CpuArchEnum.POWERPC, CpuArchEnum.S390X, CpuArchEnum.ARM):
            logger.info("Chunked prefill is not supported for ARM and POWER "
                        "and S390X CPUs; "
                        "disabling it for V1 backend.")
            self.enable_chunked_prefill = False
1188
        assert self.enable_chunked_prefill is not None
1189

1190
1191
1192
1193
1194
1195
1196
        sliding_window: Optional[int] = None
        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()

1197
1198
1199
        # 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
1200
        # reuses the GPUs of TP group, and split one TP group into
1201
1202
1203
1204
1205
1206
1207
        # tp_size//dcp_size DCP groups.
        assert self.tensor_parallel_size % self.decode_context_parallel_size \
            == 0, (
            f"tp_size={self.tensor_parallel_size} must be divisible by"
            f"dcp_size={self.decode_context_parallel_size}."
        )

1208
        cache_config = CacheConfig(
1209
            block_size=self.block_size,
1210
            gpu_memory_utilization=self.gpu_memory_utilization,
1211
            kv_cache_memory_bytes=self.kv_cache_memory_bytes,
1212
1213
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1214
            is_attention_free=model_config.is_attention_free,
1215
            num_gpu_blocks_override=self.num_gpu_blocks_override,
1216
            sliding_window=sliding_window,
1217
            enable_prefix_caching=self.enable_prefix_caching,
1218
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1219
            cpu_offload_gb=self.cpu_offload_gb,
1220
            calculate_kv_scales=self.calculate_kv_scales,
1221
            kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
1222
1223
            mamba_cache_dtype=self.mamba_cache_dtype,
            mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype,
1224
        )
1225

1226
1227
1228
1229
1230
1231
1232
1233
1234
        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
            ray_runtime_env = ray.get_runtime_context().runtime_env
            logger.info("Using ray runtime env: %s", ray_runtime_env)

1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
        # 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()

1246
1247
1248
1249
        assert not headless or not self.data_parallel_hybrid_lb, (
            "data_parallel_hybrid_lb is not applicable in "
            "headless mode")

1250
        data_parallel_external_lb = self.data_parallel_rank is not None
1251
        # Local DP rank = 1, use pure-external LB.
1252
1253
1254
1255
1256
        if data_parallel_external_lb:
            assert self.data_parallel_size_local in (1, None), (
                "data_parallel_size_local must be 1 when data_parallel_rank "
                "is set")
            data_parallel_size_local = 1
1257
1258
            # Use full external lb if we have local_size of 1.
            self.data_parallel_hybrid_lb = False
1259
1260
        elif self.data_parallel_size_local is not None:
            data_parallel_size_local = self.data_parallel_size_local
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275

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

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

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

            self.data_parallel_rank = self.data_parallel_start_rank or 0
1276
        else:
1277
1278
1279
1280
            assert not self.data_parallel_hybrid_lb, (
                "data_parallel_size_local must be set to use "
                "data_parallel_hybrid_lb.")

1281
1282
            # Local DP size defaults to global DP size if not set.
            data_parallel_size_local = self.data_parallel_size
1283
1284
1285

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
        if self.data_parallel_address is None:
            if self.data_parallel_backend == "ray":
                host_ip = get_ip()
                logger.info(
                    "Using host IP %s as ray-based data parallel address",
                    host_ip)
                data_parallel_address = host_ip
            else:
                assert self.data_parallel_backend == "mp", (
                    "data_parallel_backend can only be ray or mp, got %s",
                    self.data_parallel_backend)
                data_parallel_address = ParallelConfig.data_parallel_master_ip
        else:
            data_parallel_address = self.data_parallel_address
1300
1301
1302
1303
1304
1305
1306

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

1307
1308
1309
1310
        if self.async_scheduling:
            # Async scheduling does not work with the uniprocess backend.
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "mp"
1311
1312
                logger.info("Defaulting to mp-based distributed executor "
                            "backend for async scheduling.")
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
            if self.pipeline_parallel_size > 1:
                raise ValueError("Async scheduling is not supported with "
                                 "pipeline-parallel-size > 1.")

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

1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
        # 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

1334
        parallel_config = ParallelConfig(
1335
1336
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1337
            data_parallel_size=self.data_parallel_size,
1338
1339
            data_parallel_rank=self.data_parallel_rank or 0,
            data_parallel_external_lb=data_parallel_external_lb,
1340
1341
1342
            data_parallel_size_local=data_parallel_size_local,
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
1343
            data_parallel_backend=self.data_parallel_backend,
1344
            data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
1345
            enable_expert_parallel=self.enable_expert_parallel,
1346
1347
            enable_dbo=self.enable_dbo,
            dbo_decode_token_threshold=self.dbo_decode_token_threshold,
1348
            dbo_prefill_token_threshold=self.dbo_prefill_token_threshold,
1349
            enable_eplb=self.enable_eplb,
1350
            eplb_config=self.eplb_config,
1351
            expert_placement_strategy=self.expert_placement_strategy,
1352
1353
1354
            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,
1355
            ray_runtime_env=ray_runtime_env,
1356
            placement_group=placement_group,
1357
1358
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1359
            worker_extension_cls=self.worker_extension_cls,
1360
            decode_context_parallel_size=self.decode_context_parallel_size,
1361
1362
            _api_process_count=self._api_process_count,
            _api_process_rank=self._api_process_rank,
1363
        )
1364

1365
        speculative_config = self.create_speculative_config(
1366
1367
            target_model_config=model_config,
            target_parallel_config=parallel_config,
1368
            enable_chunked_prefill=self.enable_chunked_prefill,
王敏's avatar
王敏 committed
1369
            disable_log_stats=self.disable_log_stats,
1370
1371
        )

1372
1373
1374
1375
1376
        # 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
1377

1378
        scheduler_config = SchedulerConfig(
1379
            runner_type=model_config.runner_type,
1380
1381
1382
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1383
            cuda_graph_sizes=self.cuda_graph_sizes,
1384
            num_lookahead_slots=num_lookahead_slots,
1385
            enable_chunked_prefill=self.enable_chunked_prefill,
1386
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1387
            is_multimodal_model=model_config.is_multimodal_model,
1388
1389
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1390
            policy=self.scheduling_policy,
1391
            scheduler_cls=self.scheduler_cls,
1392
1393
1394
            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,
1395
1396
            disable_hybrid_kv_cache_manager=self.
            disable_hybrid_kv_cache_manager,
1397
            async_scheduling=self.async_scheduling,
1398
        )
1399

1400
1401
1402
1403
1404
        if not model_config.is_multimodal_model and self.default_mm_loras:
            raise ValueError(
                "Default modality-specific LoRA(s) were provided for a "
                "non multimodal model")

1405
        lora_config = LoRAConfig(
1406
            bias_enabled=self.enable_lora_bias,
1407
1408
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1409
            default_mm_loras=self.default_mm_loras,
1410
            fully_sharded_loras=self.fully_sharded_loras,
1411
1412
1413
            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
1414
1415
            and self.max_cpu_loras > 0 else None,
            lora_target_modules=self.lora_target_modules) if self.enable_lora else None
1416

1417
1418
1419
1420
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1421
        load_config = self.create_load_config()
1422

1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
        # Pass reasoning_parser into StructuredOutputsConfig
        if self.reasoning_parser:
            self.structured_outputs_config.reasoning_parser = \
                self.reasoning_parser

        # Forward the deprecated CLI args to the StructuredOutputsConfig
        so_config = self.structured_outputs_config
        if self.guided_decoding_backend is not None:
            so_config.guided_decoding_backend = \
            self.guided_decoding_backend
        if self.guided_decoding_disable_fallback is not None:
            so_config.guided_decoding_disable_fallback = \
            self.guided_decoding_disable_fallback
        if self.guided_decoding_disable_any_whitespace is not None:
            so_config.guided_decoding_disable_any_whitespace = \
            self.guided_decoding_disable_any_whitespace
        if self.guided_decoding_disable_additional_properties is not None:
            so_config.guided_decoding_disable_additional_properties = \
            self.guided_decoding_disable_additional_properties
1442

1443
        observability_config = ObservabilityConfig(
1444
1445
            show_hidden_metrics_for_version=(
                self.show_hidden_metrics_for_version),
1446
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1447
            collect_detailed_traces=self.collect_detailed_traces,
1448
        )
1449

1450
        config = VllmConfig(
1451
1452
1453
1454
1455
1456
1457
1458
            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,
1459
            structured_outputs_config=self.structured_outputs_config,
1460
            observability_config=observability_config,
1461
            compilation_config=self.compilation_config,
1462
            kv_transfer_config=self.kv_transfer_config,
1463
            kv_events_config=self.kv_events_config,
1464
            additional_config=self.additional_config,
1465
        )
1466

1467
1468
        return config

1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
    def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
        """Oracle for whether to use V0 or V1 Engine by default."""

        #############################################################
        # Unsupported Feature Flags on V1.

        if (self.logits_processor_pattern
                != EngineArgs.logits_processor_pattern):
            _raise_or_fallback(feature_name="--logits-processor-pattern",
                               recommend_to_remove=False)
            return False

1481

1482
1483
1484
1485
1486
1487
1488
1489
        # No Mamba or Encoder-Decoder so far.
        if not model_config.is_v1_compatible:
            _raise_or_fallback(feature_name=model_config.architectures,
                               recommend_to_remove=False)
            return False

        # No Concurrent Partial Prefills so far.
        if (self.max_num_partial_prefills
1490
                != SchedulerConfig.max_num_partial_prefills
1491
                or self.max_long_partial_prefills
1492
                != SchedulerConfig.max_long_partial_prefills):
1493
1494
1495
1496
            _raise_or_fallback(feature_name="Concurrent Partial Prefill",
                               recommend_to_remove=False)
            return False

1497
        # V1 supports N-gram, Medusa, and Eagle speculative decoding.
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
        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 "
1509
                    "such as ngram, medusa, eagle, or mtp.")
1510
1511

        V1_BACKENDS = [
1512
1513
            "FLASH_ATTN",
            "PALLAS",
1514
            "TRITON_ATTN",
1515
            "TRITON_MLA",
1516
            "CUTLASS_MLA",
1517
            "FLASHMLA",
1518
            "FLASH_ATTN_MLA",
1519
            "FLASHINFER",
1520
            "FLASHINFER_MLA",
1521
            "ROCM_AITER_MLA",
1522
            "TORCH_SDPA",
1523
            "FLEX_ATTENTION",
1524
            "TREE_ATTN",
1525
1526
            "XFORMERS",
            "ROCM_ATTN",
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
        ]
        if (envs.is_set("VLLM_ATTENTION_BACKEND")
                and envs.VLLM_ATTENTION_BACKEND not in V1_BACKENDS):
            name = f"VLLM_ATTENTION_BACKEND={envs.VLLM_ATTENTION_BACKEND}"
            _raise_or_fallback(feature_name=name, recommend_to_remove=True)
            return False

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

1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
        if self.pipeline_parallel_size > 1:
            supports_pp = getattr(self.distributed_executor_backend,
                                  'supports_pp', False)
            if not supports_pp and self.distributed_executor_backend not in (
                    ParallelConfig.distributed_executor_backend, "ray", "mp",
                    "external_launcher"):
                name = "Pipeline Parallelism without Ray distributed " \
                        "executor or multiprocessing executor or external " \
                        "launcher"
                _raise_or_fallback(feature_name=name,
                                   recommend_to_remove=False)
                return False
1549

1550
1551
1552
1553
1554
1555
        if (current_platform.is_cpu()
                and model_config.get_sliding_window() is not None):
            _raise_or_fallback(feature_name="sliding window (CPU backend)",
                               recommend_to_remove=False)
            return False

1556
1557
1558
1559
        #############################################################

        return True

1560
1561
    def _set_default_args(self, usage_context: UsageContext,
                          model_config: ModelConfig) -> None:
1562
        """Set Default Arguments for V1 Engine."""
1563

1564
1565
1566
1567
1568
        # V1 always uses chunked prefills and prefix caching
        # for non-pooling tasks.
        # For pooling tasks the default is False
        if model_config.runner_type != "pooling":
            self.enable_chunked_prefill = True
1569
1570
1571
1572
            
            if model_config.enable_chunked_prefill is not None and \
                model_config.enable_chunked_prefill is False:
                self.enable_chunked_prefill = False
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583

            # TODO: When prefix caching supports prompt embeds inputs, this
            # check can be removed.
            if (self.enable_prompt_embeds
                    and self.enable_prefix_caching is not False):
                logger.warning(
                    "--enable-prompt-embeds and --enable-prefix-caching "
                    "are not supported together in V1. Prefix caching has "
                    "been disabled.")
                self.enable_prefix_caching = False

1584
1585
1586
1587
1588
            if self.enable_prefix_caching is None:
                self.enable_prefix_caching = True
        else:

            pooling_type = model_config.pooler_config.pooling_type
1589
1590
1591
1592
            is_causal = getattr(model_config.hf_config, "is_causal", True)
            incremental_prefill_supported = (pooling_type is not None
                                             and pooling_type.lower() == "last"
                                             and is_causal)
1593

1594
1595
            action = "Enabling" if \
                incremental_prefill_supported else "Disabling"
1596
1597
1598
1599
                
            if model_config.enable_chunked_prefill is not None and \
                model_config.enable_chunked_prefill is False:
                self.enable_chunked_prefill = False
1600

1601
1602
1603
1604
1605
1606
            if self.enable_chunked_prefill is None:
                self.enable_chunked_prefill = incremental_prefill_supported
                logger.info("(%s) chunked prefill by default", action)
            if self.enable_prefix_caching is None:
                self.enable_prefix_caching = incremental_prefill_supported
                logger.info("(%s) prefix caching by default", action)
1607

1608
1609
1610
        # V1 should use the new scheduler by default.
        # Swap it only if this arg is set to the original V0 default
        if self.scheduler_cls == EngineArgs.scheduler_cls:
1611
            self.scheduler_cls = "vllm.v1.core.sched.scheduler.Scheduler"
1612

1613
1614
        # When no user override, set the default values based on the usage
        # context.
1615
        # Use different default values for different hardware.
1616
1617
1618
1619
1620
1621
1622

        # 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:
1623
            device_memory = current_platform.get_device_total_memory()
1624
            device_name = current_platform.get_device_name().lower()
1625
1626
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1627
            device_memory = 0
1628

1629
1630
1631
        # 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.
1632
        from vllm.usage.usage_lib import UsageContext
1633
        if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
1634
            # For GPUs like H100 and MI300x, use larger default values.
1635
1636
1637
1638
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1639
1640
1641
1642
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
1643
1644
1645
1646
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
1647
                UsageContext.OPENAI_API_SERVER: 10240,
1648
            }
1649
1650
1651
1652
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
1653

1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
        # tpu specific default values.
        if current_platform.is_tpu():
            default_max_num_batched_tokens_tpu = {
                UsageContext.LLM_CLASS: {
                    'V6E': 2048,
                    'V5E': 1024,
                    'V5P': 512,
                },
                UsageContext.OPENAI_API_SERVER: {
                    'V6E': 1024,
                    'V5E': 512,
                    'V5P': 256,
                }
            }

1669
1670
        # cpu specific default values.
        if current_platform.is_cpu():
1671
            world_size = self.pipeline_parallel_size * self.tensor_parallel_size
1672
            default_max_num_batched_tokens = {
1673
1674
                UsageContext.LLM_CLASS: 4096 * world_size,
                UsageContext.OPENAI_API_SERVER: 2048 * world_size,
1675
1676
            }
            default_max_num_seqs = {
1677
1678
                UsageContext.LLM_CLASS: 256 * world_size,
                UsageContext.OPENAI_API_SERVER: 128 * world_size,
1679
1680
            }

1681
        use_context_value = usage_context.value if usage_context else None
1682
1683
        if (self.max_num_batched_tokens is None
                and usage_context in default_max_num_batched_tokens):
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
            if current_platform.is_tpu():
                chip_name = current_platform.get_device_name()
                if chip_name in default_max_num_batched_tokens_tpu[
                        usage_context]:
                    self.max_num_batched_tokens = \
                        default_max_num_batched_tokens_tpu[
                            usage_context][chip_name]
                else:
                    self.max_num_batched_tokens = \
                        default_max_num_batched_tokens[usage_context]
            else:
1695
1696
1697
1698
1699
                if not self.enable_chunked_prefill:
                    self.max_num_batched_tokens = model_config.max_model_len
                else:
                    self.max_num_batched_tokens = \
                        default_max_num_batched_tokens[usage_context]
1700
            logger.debug(
1701
                "Setting max_num_batched_tokens to %d for %s usage context.",
1702
                self.max_num_batched_tokens, use_context_value)
1703

1704
1705
        if (self.max_num_seqs is None
                and usage_context in default_max_num_seqs):
1706
1707
            self.max_num_seqs = min(default_max_num_seqs[usage_context],
                                    self.max_num_batched_tokens or sys.maxsize)
1708
1709
1710

            logger.debug("Setting max_num_seqs to %d for %s usage context.",
                         self.max_num_seqs, use_context_value)
1711

1712

1713
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1714
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1715
    """Arguments for asynchronous vLLM engine."""
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
    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 "
        "`enable_log_requests` instead.")
    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 "
        "`enable_log_requests` instead.")
    def disable_log_requests(self, value: bool):
        self.enable_log_requests = not value
1733
1734

    @staticmethod
1735
1736
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1737
        # Initialize plugin to update the parser, for example, The plugin may
1738
        # add a new kind of quantization method to --quantization argument or
1739
1740
        # a new device to --device argument.
        load_general_plugins()
1741
1742
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1743
1744
1745
1746
        parser.add_argument('--enable-log-requests',
                            action=argparse.BooleanOptionalAction,
                            default=AsyncEngineArgs.enable_log_requests,
                            help='Enable logging requests.')
1747
        parser.add_argument('--disable-log-requests',
1748
1749
1750
1751
                            action=argparse.BooleanOptionalAction,
                            default=not AsyncEngineArgs.enable_log_requests,
                            help='[DEPRECATED] Disable logging requests.',
                            deprecated=True)
1752
        current_platform.pre_register_and_update(parser)
1753
        return parser
1754
1755


1756
def _raise_or_fallback(feature_name: str, recommend_to_remove: bool):
zhuwenwen's avatar
zhuwenwen committed
1757
1758
    # if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
    if envs.VLLM_USE_V1:
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
        raise NotImplementedError(
            f"VLLM_USE_V1=1 is not supported with {feature_name}.")
    msg = f"{feature_name} is not supported by the V1 Engine. "
    msg += "Falling back to V0. "
    if recommend_to_remove:
        msg += f"We recommend to remove {feature_name} from your config "
        msg += "in favor of the V1 Engine."
    logger.warning(msg)



1770
1771
1772
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
1773

1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
    Examples:
    - '1k' -> 1,000
    - '1K' -> 1,024
    - '25.6k' -> 25,600
    """
    value = value.strip()
    match = re.fullmatch(r'(\d+(?:\.\d+)?)([kKmMgGtT])', value)
    if match:
        decimal_multiplier = {
            'k': 10**3,
            'm': 10**6,
            'g': 10**9,
        }
        binary_multiplier = {
            'K': 2**10,
            'M': 2**20,
            'G': 2**30,
        }

        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:
                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

    # Regular plain number.
    return int(value)