arg_utils.py 84.2 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 argparse
6
import copy
7
import dataclasses
8
import functools
9
import json
10
import sys
11
from dataclasses import MISSING, dataclass, fields, is_dataclass
12
from itertools import permutations
13
14
15
from typing import (TYPE_CHECKING, Annotated, Any, Callable, Dict, List,
                    Literal, Optional, Type, TypeVar, Union, cast, get_args,
                    get_origin)
16

17
import huggingface_hub
18
import regex as re
19
import torch
20
from pydantic import TypeAdapter, ValidationError
21
from typing_extensions import TypeIs, deprecated
22

23
import vllm.envs as envs
24
from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
25
26
27
                         ConfigType, ConvertOption, DecodingConfig,
                         DetailedTraceModules, Device, DeviceConfig,
                         DistributedExecutorBackend, EPLBConfig,
28
29
                         GuidedDecodingBackend, HfOverrides, KVEventsConfig,
                         KVTransferConfig, LoadConfig, LogprobsMode,
30
31
32
33
34
35
                         LoRAConfig, MambaDType, MMEncoderTPMode, ModelConfig,
                         ModelDType, ModelImpl, MultiModalConfig,
                         ObservabilityConfig, ParallelConfig, PoolerConfig,
                         PrefixCachingHashAlgo, RunnerOption, SchedulerConfig,
                         SchedulerPolicy, SpeculativeConfig, TaskOption,
                         TokenizerMode, VllmConfig, get_attr_docs, get_field)
36
from vllm.logger import init_logger
37
from vllm.platforms import CpuArchEnum, current_platform
38
from vllm.plugins import load_general_plugins
39
from vllm.ray.lazy_utils import is_ray_initialized
40
from vllm.reasoning import ReasoningParserManager
41
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
42
from vllm.transformers_utils.config import get_model_path, is_interleaved
43
from vllm.transformers_utils.utils import check_gguf_file
44
from vllm.utils import (STR_DUAL_CHUNK_FLASH_ATTN_VAL, FlexibleArgumentParser,
Rui Qiao's avatar
Rui Qiao committed
45
                        GiB_bytes, get_ip, is_in_ray_actor)
46
from vllm.v1.sample.logits_processor import LogitsProcessor
47
48

# yapf: enable
49

50
51
52
if TYPE_CHECKING:
    from vllm.executor.executor_base import ExecutorBase
    from vllm.model_executor.layers.quantization import QuantizationMethods
53
    from vllm.model_executor.model_loader import LoadFormats
54
55
56
57
    from vllm.usage.usage_lib import UsageContext
else:
    ExecutorBase = Any
    QuantizationMethods = Any
58
    LoadFormats = Any
59
60
    UsageContext = Any

61
62
logger = init_logger(__name__)

63
64
65
66
67
# object is used to allow for special typing forms
T = TypeVar("T")
TypeHint = Union[type[Any], object]
TypeHintT = Union[type[T], object]

68

69
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
70

71
    def _parse_type(val: str) -> T:
72
73
74
75
76
        try:
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
                f"Value {val} cannot be converted to {return_type}.") from e
77

78
79
80
81
82
83
84
85
86
87
88
    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)

89
    return _optional_type
90
91


92
def union_dict_and_str(val: str) -> Optional[Union[str, dict[str, str]]]:
93
    if not re.match(r"(?s)^\s*{.*}\s*$", val):
94
        return str(val)
95
    return optional_type(json.loads)(val)
96
97


98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
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)


113
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
114
115
116
117
    """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`.
    """
118
    type_hint = get_type(type_hints, Literal)
119
120
121
    options = get_args(type_hint)
    option_type = type(options[0])
    if not all(isinstance(option, option_type) for option in options):
122
        raise ValueError(
123
124
125
126
            "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)}
127
128


129
130
131
132
133
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
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


151
152
153
154
def is_online_quantization(quantization: Any) -> bool:
    return quantization in ["inc"]


155
156
157
158
159
160
161
NEEDS_HELP = (
    "--help" in (argv := sys.argv)  # vllm SUBCOMMAND --help
    or (argv0 := argv[0]).endswith("mkdocs")  # mkdocs SUBCOMMAND
    or argv0.endswith("mkdocs/__main__.py")  # python -m mkdocs SUBCOMMAND
)


162
163
@functools.lru_cache(maxsize=30)
def _compute_kwargs(cls: ConfigType) -> dict[str, Any]:
164
165
    # Save time only getting attr docs if we're generating help text
    cls_docs = get_attr_docs(cls) if NEEDS_HELP else {}
166
167
    kwargs = {}
    for field in fields(cls):
168
        # Get the set of possible types for the field
169
        type_hints: set[TypeHint] = get_type_hints(field.type)
170
171
172
173
174

        # 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)

175
        # Get the default value of the field
176
177
178
        if field.default is not MISSING:
            default = field.default
        elif field.default_factory is not MISSING:
179
            default = field.default_factory()
180
181
182

        # Get the help text for the field
        name = field.name
183
        help = cls_docs.get(name, "").strip()
184
185
186
187
188
189
190
        # 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
191
192
        json_tip = ("Should either be a valid JSON string or JSON keys passed "
                    "individually.")
193
        if dataclass_cls is not None:
194
195
196
197
198
199
200
201

            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
202
            kwargs[name]["help"] += f"\n\n{json_tip}"
203
        elif contains_type(type_hints, bool):
204
205
206
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
207
            kwargs[name].update(literal_to_kwargs(type_hints))
208
209
210
211
212
213
214
215
216
217
218
219
        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)
220
221
222
223
224
225
            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
226
227
228
            kwargs[name]["nargs"] = "+"
        elif contains_type(type_hints, int):
            kwargs[name]["type"] = int
229
            # Special case for large integers
230
231
232
233
234
235
            human_readable_ints = {
                "max_model_len",
                "max_num_batched_tokens",
                "kv_cache_memory_bytes",
            }
            if name in human_readable_ints:
236
                kwargs[name]["type"] = human_readable_int
237
                kwargs[name]["help"] += f"\n\n{human_readable_int.__doc__}"
238
239
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
240
241
242
        elif (contains_type(type_hints, dict)
              and (contains_type(type_hints, str)
                   or any(is_not_builtin(th) for th in type_hints))):
243
            kwargs[name]["type"] = union_dict_and_str
244
        elif contains_type(type_hints, dict):
245
            kwargs[name]["type"] = parse_type(json.loads)
246
            kwargs[name]["help"] += f"\n\n{json_tip}"
247
248
249
250
251
252
253
        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}.")

254
255
256
257
258
        # 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"]}))

259
260
261
262
263
264
265
        # 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
266
267


268
269
270
def get_kwargs(cls: ConfigType) -> dict[str, Any]:
    """Return argparse kwargs for the given Config dataclass.

271
272
273
    If `--help` or `mkdocs` are not present in the command line command, the
    attribute documentation will not be included in the help output.

274
275
276
277
278
279
280
    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))


281
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
282
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
283
    """Arguments for vLLM engine."""
284
285
286
287
288
    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
289
290
291
    runner: RunnerOption = ModelConfig.runner
    convert: ConvertOption = ModelConfig.convert
    task: Optional[TaskOption] = ModelConfig.task
292
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
293
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
294
295
296
    tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
297
    download_dir: Optional[str] = LoadConfig.download_dir
298
    safetensors_load_strategy: str = LoadConfig.safetensors_load_strategy
299
    load_format: Union[str, LoadFormats] = LoadConfig.load_format
300
301
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
302
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
303
304
    seed: Optional[int] = ModelConfig.seed
    max_model_len: Optional[int] = ModelConfig.max_model_len
305
306
    cuda_graph_sizes: list[int] = get_field(SchedulerConfig,
                                            "cuda_graph_sizes")
307
308
309
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
310
    distributed_executor_backend: Optional[Union[
311
        str, DistributedExecutorBackend,
312
        Type[ExecutorBase]]] = ParallelConfig.distributed_executor_backend
313
    # number of P/D disaggregation (or other disaggregation) workers
314
315
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
316
317
    decode_context_parallel_size: int = \
        ParallelConfig.decode_context_parallel_size
318
    data_parallel_size: int = ParallelConfig.data_parallel_size
319
    data_parallel_rank: Optional[int] = None
320
    data_parallel_start_rank: Optional[int] = None
321
322
323
    data_parallel_size_local: Optional[int] = None
    data_parallel_address: Optional[str] = None
    data_parallel_rpc_port: Optional[int] = None
324
    data_parallel_hybrid_lb: bool = False
Rui Qiao's avatar
Rui Qiao committed
325
    data_parallel_backend: str = ParallelConfig.data_parallel_backend
326
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
327
    eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config")
328
    enable_eplb: bool = ParallelConfig.enable_eplb
329
330
331
332
    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
333
334
    max_parallel_loading_workers: Optional[
        int] = ParallelConfig.max_parallel_loading_workers
335
336
337
338
    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
339
340
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
341
342
343
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
344
    kv_cache_memory_bytes: Optional[int] = CacheConfig.kv_cache_memory_bytes
345
346
347
348
349
350
351
    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
352
    max_logprobs: int = ModelConfig.max_logprobs
353
    logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode
354
    disable_log_stats: bool = False
355
356
357
358
359
    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
360
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
361
362
363
364
    tokenizer_revision: Optional[str] = ModelConfig.tokenizer_revision
    quantization: Optional[QuantizationMethods] = ModelConfig.quantization
    enforce_eager: bool = ModelConfig.enforce_eager
    max_seq_len_to_capture: int = ModelConfig.max_seq_len_to_capture
365
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
366
    limit_mm_per_prompt: dict[str, int] = \
367
        get_field(MultiModalConfig, "limit_per_prompt")
368
    interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings
369
370
371
    media_io_kwargs: dict[str, dict[str,
                                    Any]] = get_field(MultiModalConfig,
                                                      "media_io_kwargs")
372
373
    mm_processor_kwargs: Optional[Dict[str, Any]] = \
        MultiModalConfig.mm_processor_kwargs
374
    disable_mm_preprocessor_cache: bool = False  # DEPRECATED
375
    mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb
376
    mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode
377
    io_processor_plugin: Optional[str] = None
378
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
379
    # LoRA fields
380
    enable_lora: bool = False
381
382
383
    enable_lora_bias: bool = LoRAConfig.bias_enabled
    max_loras: int = LoRAConfig.max_loras
    max_lora_rank: int = LoRAConfig.max_lora_rank
384
385
    default_mm_loras: Optional[Dict[str, str]] = \
        LoRAConfig.default_mm_loras
386
387
388
389
390
    fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
    max_cpu_loras: Optional[int] = LoRAConfig.max_cpu_loras
    lora_dtype: Optional[Union[str, torch.dtype]] = LoRAConfig.lora_dtype
    lora_extra_vocab_size: int = LoRAConfig.lora_extra_vocab_size

391
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
392
393
    num_gpu_blocks_override: Optional[
        int] = CacheConfig.num_gpu_blocks_override
394
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
395
396
    model_loader_extra_config: dict = \
        get_field(LoadConfig, "model_loader_extra_config")
397
398
    ignore_patterns: Optional[Union[str,
                                    List[str]]] = LoadConfig.ignore_patterns
399
    preemption_mode: Optional[str] = SchedulerConfig.preemption_mode
400

401
402
403
404
    scheduler_delay_factor: float = SchedulerConfig.delay_factor
    enable_chunked_prefill: Optional[
        bool] = SchedulerConfig.enable_chunked_prefill
    disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
405

406
407
408
    disable_hybrid_kv_cache_manager: bool = (
        SchedulerConfig.disable_hybrid_kv_cache_manager)

409
410
411
412
413
414
    guided_decoding_backend: GuidedDecodingBackend = DecodingConfig.backend
    guided_decoding_disable_fallback: bool = DecodingConfig.disable_fallback
    guided_decoding_disable_any_whitespace: bool = \
        DecodingConfig.disable_any_whitespace
    guided_decoding_disable_additional_properties: bool = \
        DecodingConfig.disable_additional_properties
415
416
    logits_processor_pattern: Optional[
        str] = ModelConfig.logits_processor_pattern
417

418
    speculative_config: Optional[Dict[str, Any]] = None
419

420
421
422
423
424
425
    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
426
    disable_async_output_proc: bool = not ModelConfig.use_async_output_proc
427
428
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
    scheduler_cls: Union[str, Type[object]] = SchedulerConfig.scheduler_cls
429

430
431
    override_pooler_config: Optional[Union[dict, PoolerConfig]] = \
        ModelConfig.override_pooler_config
432
433
    compilation_config: CompilationConfig = \
        get_field(VllmConfig, "compilation_config")
434
435
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
436

437
    kv_transfer_config: Optional[KVTransferConfig] = None
438
    kv_events_config: Optional[KVEventsConfig] = None
439

440
441
442
443
444
    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
445
    override_attention_dtype: str = ModelConfig.override_attention_dtype
446

447
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
448
449
    mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype
    mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype
450

451
452
    additional_config: dict[str, Any] = \
        get_field(VllmConfig, "additional_config")
453
454
    reasoning_parser: str = DecodingConfig.reasoning_backend

455
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
456
    pt_load_map_location: str = LoadConfig.pt_load_map_location
457

458
459
    # DEPRECATED
    enable_multimodal_encoder_data_parallel: bool = False
460

461
462
463
464
    logits_processors: Optional[list[Union[
        str, type[LogitsProcessor]]]] = ModelConfig.logits_processors
    """Custom logitproc types"""

465
466
    async_scheduling: bool = SchedulerConfig.async_scheduling

467
468
469
    kv_sharing_fast_prefill: bool = \
        CacheConfig.kv_sharing_fast_prefill

470
    def __post_init__(self):
471
472
473
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
474
475
476
        if isinstance(self.compilation_config, dict):
            self.compilation_config = CompilationConfig(
                **self.compilation_config)
477
        if isinstance(self.eplb_config, dict):
478
            self.eplb_config = EPLBConfig(**self.eplb_config)
479
        # Setup plugins
480
481
        from vllm.plugins import load_general_plugins
        load_general_plugins()
482
483
484
485
486
487
488
        # 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)
489
490

    @staticmethod
491
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
492
        """Shared CLI arguments for vLLM engine."""
493

494
        # Model arguments
495
496
497
498
499
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
Reid's avatar
Reid committed
500
        if not ('serve' in sys.argv[1:] and '--help' in sys.argv[1:]):
501
            model_group.add_argument("--model", **model_kwargs["model"])
502
503
504
505
506
        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)
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
        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"])
        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-seq-len-to-capture",
                                 **model_kwargs["max_seq_len_to_capture"])
        model_group.add_argument("--max-logprobs",
                                 **model_kwargs["max_logprobs"])
536
        model_group.add_argument("--logprobs-mode",
537
                                 choices=[f.value for f in LogprobsMode],
538
                                 **model_kwargs["logprobs_mode"])
539
540
541
542
543
544
        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"])
545
546
        model_group.add_argument("--enable-prompt-embeds",
                                 **model_kwargs["enable_prompt_embeds"])
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
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"])
        # This one is a special case because it is the
        # opposite of ModelConfig.use_async_output_proc
        model_group.add_argument(
            "--disable-async-output-proc",
            action="store_true",
            default=EngineArgs.disable_async_output_proc,
            help="Disable async output processing. This may result in "
            "lower performance.")
        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"])
        model_group.add_argument("--override-pooler-config",
                                 **model_kwargs["override_pooler_config"])
        model_group.add_argument("--logits-processor-pattern",
                                 **model_kwargs["logits_processor_pattern"])
        model_group.add_argument("--generation-config",
                                 **model_kwargs["generation_config"])
        model_group.add_argument("--override-generation-config",
                                 **model_kwargs["override_generation_config"])
        model_group.add_argument("--enable-sleep-mode",
                                 **model_kwargs["enable_sleep_mode"])
        model_group.add_argument("--model-impl",
                                 choices=[f.value for f in ModelImpl],
                                 **model_kwargs["model_impl"])
582
583
        model_group.add_argument("--override-attention-dtype",
                                 **model_kwargs["override_attention_dtype"])
584
585
        model_group.add_argument("--logits-processors",
                                 **model_kwargs["logits_processors"])
586
587
        model_group.add_argument("--io-processor-plugin",
                                 **model_kwargs["io_processor_plugin"])
588

589
590
591
592
593
594
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
595
        load_group.add_argument("--load-format", **load_kwargs["load_format"])
596
        load_group.add_argument("--download-dir",
597
                                **load_kwargs["download_dir"])
598
599
        load_group.add_argument("--safetensors-load-strategy",
                                **load_kwargs["safetensors_load_strategy"])
600
        load_group.add_argument("--model-loader-extra-config",
601
                                **load_kwargs["model_loader_extra_config"])
602
603
604
        load_group.add_argument("--ignore-patterns",
                                **load_kwargs["ignore_patterns"])
        load_group.add_argument("--use-tqdm-on-load",
605
                                **load_kwargs["use_tqdm_on_load"])
606
607
        load_group.add_argument('--pt-load-map-location',
                                **load_kwargs["pt_load_map_location"])
608

609
610
611
612
613
614
        # Guided decoding arguments
        guided_decoding_kwargs = get_kwargs(DecodingConfig)
        guided_decoding_group = parser.add_argument_group(
            title="DecodingConfig",
            description=DecodingConfig.__doc__,
        )
615
616
        guided_decoding_group.add_argument("--guided-decoding-backend",
                                           **guided_decoding_kwargs["backend"])
617
        guided_decoding_group.add_argument(
618
619
620
621
622
623
624
625
            "--guided-decoding-disable-fallback",
            **guided_decoding_kwargs["disable_fallback"])
        guided_decoding_group.add_argument(
            "--guided-decoding-disable-any-whitespace",
            **guided_decoding_kwargs["disable_any_whitespace"])
        guided_decoding_group.add_argument(
            "--guided-decoding-disable-additional-properties",
            **guided_decoding_kwargs["disable_additional_properties"])
626
627
        guided_decoding_group.add_argument(
            "--reasoning-parser",
628
            # This choice is a special case because it's not static
629
630
631
            choices=list(ReasoningParserManager.reasoning_parsers),
            **guided_decoding_kwargs["reasoning_backend"])

632
        # Parallel arguments
633
634
635
636
637
638
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
639
            "--distributed-executor-backend",
640
641
            **parallel_kwargs["distributed_executor_backend"])
        parallel_group.add_argument(
642
            "--pipeline-parallel-size", "-pp",
643
            **parallel_kwargs["pipeline_parallel_size"])
644
        parallel_group.add_argument("--tensor-parallel-size", "-tp",
645
                                    **parallel_kwargs["tensor_parallel_size"])
646
647
648
        parallel_group.add_argument(
            "--decode-context-parallel-size", "-dcp",
            **parallel_kwargs["decode_context_parallel_size"])
649
        parallel_group.add_argument("--data-parallel-size", "-dp",
650
                                    **parallel_kwargs["data_parallel_size"])
651
652
653
654
655
656
        parallel_group.add_argument(
            '--data-parallel-rank',
            '-dpn',
            type=int,
            help='Data parallel rank of this instance. '
            'When set, enables external load balancer mode.')
657
658
659
660
661
        parallel_group.add_argument('--data-parallel-start-rank',
                                    '-dpr',
                                    type=int,
                                    help='Starting data parallel rank '
                                    'for secondary nodes.')
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
        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
677
678
679
680
681
682
        parallel_group.add_argument('--data-parallel-backend',
                                    '-dpb',
                                    type=str,
                                    default='mp',
                                    help='Backend for data parallel, either '
                                    '"mp" or "ray".')
683
684
685
        parallel_group.add_argument(
            "--data-parallel-hybrid-lb",
            **parallel_kwargs["data_parallel_hybrid_lb"])
686
        parallel_group.add_argument(
687
            "--enable-expert-parallel",
688
            **parallel_kwargs["enable_expert_parallel"])
689
690
        parallel_group.add_argument("--enable-eplb",
                                    **parallel_kwargs["enable_eplb"])
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
        parallel_group.add_argument("--eplb-config",
                                    **parallel_kwargs["eplb_config"])
        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)

717
        parallel_group.add_argument(
718
            "--max-parallel-loading-workers",
719
720
            **parallel_kwargs["max_parallel_loading_workers"])
        parallel_group.add_argument(
721
            "--ray-workers-use-nsight",
722
723
            **parallel_kwargs["ray_workers_use_nsight"])
        parallel_group.add_argument(
724
            "--disable-custom-all-reduce",
725
            **parallel_kwargs["disable_custom_all_reduce"])
726
727
728
729
        parallel_group.add_argument("--worker-cls",
                                    **parallel_kwargs["worker_cls"])
        parallel_group.add_argument("--worker-extension-cls",
                                    **parallel_kwargs["worker_extension_cls"])
730
731
        parallel_group.add_argument(
            "--enable-multimodal-encoder-data-parallel",
732
733
            action="store_true",
            deprecated=True)
734

735
736
737
738
739
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
740
        )
741
742
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
        cache_group.add_argument("--gpu-memory-utilization",
743
                                 **cache_kwargs["gpu_memory_utilization"])
744
745
        cache_group.add_argument("--kv-cache-memory-bytes",
                                 **cache_kwargs["kv_cache_memory_bytes"])
746
747
        cache_group.add_argument("--swap-space", **cache_kwargs["swap_space"])
        cache_group.add_argument("--kv-cache-dtype",
748
                                 **cache_kwargs["cache_dtype"])
749
        cache_group.add_argument("--num-gpu-blocks-override",
750
751
752
753
754
                                 **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"])
755
        cache_group.add_argument("--cpu-offload-gb",
756
                                 **cache_kwargs["cpu_offload_gb"])
757
        cache_group.add_argument("--calculate-kv-scales",
758
                                 **cache_kwargs["calculate_kv_scales"])
759
760
        cache_group.add_argument("--kv-sharing-fast-prefill",
                                 **cache_kwargs["kv_sharing_fast_prefill"])
761
762
763
764
        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"])
765

766
        # Multimodal related configs
767
768
769
770
771
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
772
        multimodal_group.add_argument("--limit-mm-per-prompt",
773
                                      **multimodal_kwargs["limit_per_prompt"])
774
775
        multimodal_group.add_argument("--media-io-kwargs",
                                      **multimodal_kwargs["media_io_kwargs"])
776
        multimodal_group.add_argument(
777
            "--mm-processor-kwargs",
778
779
            **multimodal_kwargs["mm_processor_kwargs"])
        multimodal_group.add_argument(
780
781
782
            "--mm-processor-cache-gb",
            **multimodal_kwargs["mm_processor_cache_gb"])
        multimodal_group.add_argument("--disable-mm-preprocessor-cache",
783
                                      action="store_true",
784
                                      deprecated=True)
785
786
        multimodal_group.add_argument(
            "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"])
787
788
789
        multimodal_group.add_argument(
            "--interleave-mm-strings",
            **multimodal_kwargs["interleave_mm_strings"])
790
791
        multimodal_group.add_argument("--skip-mm-profiling",
                                      **multimodal_kwargs["skip_mm_profiling"])
792

793
        # LoRA related configs
794
795
796
797
798
799
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
800
            "--enable-lora",
801
            action=argparse.BooleanOptionalAction,
802
803
            help="If True, enable handling of LoRA adapters.")
        lora_group.add_argument("--enable-lora-bias",
804
                                **lora_kwargs["bias_enabled"])
805
806
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
        lora_group.add_argument("--max-lora-rank",
807
                                **lora_kwargs["max_lora_rank"])
808
        lora_group.add_argument("--lora-extra-vocab-size",
809
810
                                **lora_kwargs["lora_extra_vocab_size"])
        lora_group.add_argument(
811
            "--lora-dtype",
812
813
            **lora_kwargs["lora_dtype"],
        )
814
        lora_group.add_argument("--max-cpu-loras",
815
                                **lora_kwargs["max_cpu_loras"])
816
        lora_group.add_argument("--fully-sharded-loras",
817
                                **lora_kwargs["fully_sharded_loras"])
818
819
        lora_group.add_argument("--default-mm-loras",
                                **lora_kwargs["default_mm_loras"])
820

821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
        # 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"])
844

845
846
847
848
849
850
851
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
852
            "--max-num-batched-tokens",
853
            **scheduler_kwargs["max_num_batched_tokens"])
854
        scheduler_group.add_argument("--max-num-seqs",
855
856
857
858
859
860
861
                                     **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"])
862
863
        scheduler_group.add_argument('--cuda-graph-sizes',
                                     **scheduler_kwargs["cuda_graph_sizes"])
864
865
866
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
            **scheduler_kwargs["long_prefill_token_threshold"])
867
        scheduler_group.add_argument("--num-lookahead-slots",
868
                                     **scheduler_kwargs["num_lookahead_slots"])
869
        scheduler_group.add_argument("--scheduler-delay-factor",
870
                                     **scheduler_kwargs["delay_factor"])
871
        scheduler_group.add_argument("--preemption-mode",
872
                                     **scheduler_kwargs["preemption_mode"])
873
874
        # multi-step scheduling has been removed; corresponding arguments
        # are no longer supported.
875
        scheduler_group.add_argument("--scheduling-policy",
876
                                     **scheduler_kwargs["policy"])
877
        scheduler_group.add_argument(
878
            "--enable-chunked-prefill",
879
            **scheduler_kwargs["enable_chunked_prefill"])
880
881
882
        scheduler_group.add_argument(
            "--disable-chunked-mm-input",
            **scheduler_kwargs["disable_chunked_mm_input"])
883
884
        scheduler_group.add_argument("--scheduler-cls",
                                     **scheduler_kwargs["scheduler_cls"])
885
886
887
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"])
888
889
        scheduler_group.add_argument("--async-scheduling",
                                     **scheduler_kwargs["async_scheduling"])
890
891

        # vLLM arguments
892
        vllm_kwargs = get_kwargs(VllmConfig)
893
894
895
896
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
897
898
899
900
        # 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)
901
902
        vllm_group.add_argument("--speculative-config",
                                **vllm_kwargs["speculative_config"])
903
904
905
906
907
908
909
910
        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"])
911

912
913
914
915
        # Other arguments
        parser.add_argument('--disable-log-stats',
                            action='store_true',
                            help='Disable logging statistics.')
916

917
        return parser
918
919

    @classmethod
920
    def from_cli_args(cls, args: argparse.Namespace):
921
922
923
        # Get the list of attributes of this dataclass.
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        # Set the attributes from the parsed arguments.
Zhuohan Li's avatar
Zhuohan Li committed
924
925
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
926

927
    def create_model_config(self) -> ModelConfig:
928
929
930
931
932
933
        # 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
934
                and self.model in MODELS_ON_S3 and self.load_format == "auto"):
935
            self.model = f"{MODEL_WEIGHTS_S3_BUCKET}/{self.model}"
936
            self.load_format = "runai_streamer"
937

938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
        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

955
956
957
958
959
960
961
962
        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"

963
        return ModelConfig(
964
            model=self.model,
965
            hf_config_path=self.hf_config_path,
966
967
            runner=self.runner,
            convert=self.convert,
968
            task=self.task,
969
            tokenizer=self.tokenizer,
970
971
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
972
            allowed_local_media_path=self.allowed_local_media_path,
973
974
975
976
977
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
978
            rope_theta=self.rope_theta,
979
            hf_token=self.hf_token,
980
            hf_overrides=self.hf_overrides,
981
982
983
984
985
986
            tokenizer_revision=self.tokenizer_revision,
            max_model_len=self.max_model_len,
            quantization=self.quantization,
            enforce_eager=self.enforce_eager,
            max_seq_len_to_capture=self.max_seq_len_to_capture,
            max_logprobs=self.max_logprobs,
987
            logprobs_mode=self.logprobs_mode,
988
            disable_sliding_window=self.disable_sliding_window,
989
            disable_cascade_attn=self.disable_cascade_attn,
990
            skip_tokenizer_init=self.skip_tokenizer_init,
991
            enable_prompt_embeds=self.enable_prompt_embeds,
992
            served_model_name=self.served_model_name,
993
            limit_mm_per_prompt=self.limit_mm_per_prompt,
994
            interleave_mm_strings=self.interleave_mm_strings,
995
            media_io_kwargs=self.media_io_kwargs,
996
            skip_mm_profiling=self.skip_mm_profiling,
997
            use_async_output_proc=not self.disable_async_output_proc,
998
            config_format=self.config_format,
999
            mm_processor_kwargs=self.mm_processor_kwargs,
1000
            mm_processor_cache_gb=self.mm_processor_cache_gb,
1001
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
1002
            override_pooler_config=self.override_pooler_config,
1003
            logits_processor_pattern=self.logits_processor_pattern,
1004
            generation_config=self.generation_config,
1005
            override_generation_config=self.override_generation_config,
1006
            enable_sleep_mode=self.enable_sleep_mode,
1007
            model_impl=self.model_impl,
1008
            override_attention_dtype=self.override_attention_dtype,
1009
            logits_processors=self.logits_processors,
1010
            io_processor_plugin=self.io_processor_plugin,
1011
        )
1012

1013
1014
1015
1016
1017
1018
1019
    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]
1020

1021
1022
    def create_load_config(self) -> LoadConfig:

1023
1024
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1025

1026
1027
1028
1029
1030
1031
1032
1033
        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()
1034

1035
1036
1037
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
1038
            safetensors_load_strategy=self.safetensors_load_strategy,
1039
1040
            device="cpu"
            if is_online_quantization(self.quantization) else None,
1041
1042
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
1043
            use_tqdm_on_load=self.use_tqdm_on_load,
1044
            pt_load_map_location=self.pt_load_map_location,
1045
        )
1046

1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
    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
1060
        dictionary from the engine.
1061
        """
1062
1063
1064
1065
1066

        from vllm.transformers_utils.config import get_config
        from vllm.transformers_utils.configs.speculators.base import (
            SpeculatorsConfig)

1067
        if self.speculative_config is None:
1068
1069
1070
1071
            hf_config = get_config(
                self.hf_config_path or target_model_config.model,
                self.trust_remote_code, self.revision, self.code_revision,
                self.config_format)
1072

1073
            # if loading a SpeculatorsConfig, load the speculative_config
1074
1075
1076
            # details from the config directly
            # no user input required / expected
            if isinstance(hf_config, SpeculatorsConfig):
1077
                # We create one since we don't create one
1078
1079
1080
                self.speculative_config = {}
                self.speculative_config[
                    "num_speculative_tokens"] = hf_config.num_lookahead_tokens
1081
                self.speculative_config["model"] = target_model_config.model
1082
1083
1084
                self.speculative_config["method"] = hf_config.method
            else:
                return None
1085

1086
1087
1088
1089
1090
1091
1092
1093
1094
        # 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,
        })
1095
        return SpeculativeConfig(**self.speculative_config)
1096

1097
1098
1099
    def create_engine_config(
        self,
        usage_context: Optional[UsageContext] = None,
1100
        headless: bool = False,
1101
1102
1103
1104
1105
1106
1107
    ) -> VllmConfig:
        """
        Create the VllmConfig.

        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.
Simon Mo's avatar
Simon Mo committed
1108

1109
1110
1111
1112
1113
1114
        This function set VLLM_USE_V1=X if VLLM_USE_V1 is
        unspecified by the user.

        If VLLM_USE_V1 is specified by the user but the VllmConfig
        is incompatible, we raise an error.
        """
1115
        current_platform.pre_register_and_update()
1116

1117
1118
        device_config = DeviceConfig(
            device=cast(Device, current_platform.device_type))
1119
1120
        model_config = self.create_model_config()

1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
        # * 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)

        # Set default arguments for V0 or V1 Engine.
        if use_v1:
1140
            self._set_default_args_v1(usage_context, model_config)
1141
            # Disable chunked prefill for POWER (ppc64le)/ARM/s390x CPUs in V1
1142
1143
            if current_platform.is_cpu(
            ) and current_platform.get_cpu_architecture() in (
1144
                    CpuArchEnum.POWERPC, CpuArchEnum.S390X, CpuArchEnum.ARM):
1145
                logger.info(
1146
1147
                    "Chunked prefill is not supported for ARM and POWER "
                    "and S390X CPUs; "
1148
1149
                    "disabling it for V1 backend.")
                self.enable_chunked_prefill = False
1150
1151
        else:
            self._set_default_args_v0(model_config)
1152
1153
        assert self.enable_chunked_prefill is not None

1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
        if envs.VLLM_ATTENTION_BACKEND in [STR_DUAL_CHUNK_FLASH_ATTN_VAL]:
            assert self.enforce_eager, (
                "Cuda graph is not supported with DualChunkFlashAttention. "
                "To run the model in eager mode, set 'enforce_eager=True' "
                "or use '--enforce-eager' in the CLI.")
            assert current_platform.is_cuda(), (
                "DualChunkFlashAttention is only supported on CUDA platform.")
            assert not use_v1, (
                "DualChunkFlashAttention is not supported on V1 engine. "
                "To run the model in V0 engine, try set 'VLLM_USE_V1=0'")

1165
1166
1167
1168
1169
1170
1171
        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()

1172
1173
1174
        # 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
1175
        # reuses the GPUs of TP group, and split one TP group into
1176
1177
1178
1179
1180
1181
1182
        # 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}."
        )

1183
        cache_config = CacheConfig(
1184
            block_size=self.block_size,
1185
            gpu_memory_utilization=self.gpu_memory_utilization,
1186
            kv_cache_memory_bytes=self.kv_cache_memory_bytes,
1187
1188
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1189
            is_attention_free=model_config.is_attention_free,
1190
            num_gpu_blocks_override=self.num_gpu_blocks_override,
1191
            sliding_window=sliding_window,
1192
            enable_prefix_caching=self.enable_prefix_caching,
1193
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1194
            cpu_offload_gb=self.cpu_offload_gb,
1195
            calculate_kv_scales=self.calculate_kv_scales,
1196
            kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
1197
1198
            mamba_cache_dtype=self.mamba_cache_dtype,
            mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype,
1199
        )
1200

1201
1202
1203
1204
1205
1206
1207
1208
1209
        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)

1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
        # 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()

1221
1222
1223
1224
        assert not headless or not self.data_parallel_hybrid_lb, (
            "data_parallel_hybrid_lb is not applicable in "
            "headless mode")

1225
        data_parallel_external_lb = self.data_parallel_rank is not None
1226
        # Local DP rank = 1, use pure-external LB.
1227
1228
1229
1230
1231
        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
1232
1233
            # Use full external lb if we have local_size of 1.
            self.data_parallel_hybrid_lb = False
1234
1235
        elif self.data_parallel_size_local is not None:
            data_parallel_size_local = self.data_parallel_size_local
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250

            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
1251
        else:
1252
1253
1254
1255
            assert not self.data_parallel_hybrid_lb, (
                "data_parallel_size_local must be set to use "
                "data_parallel_hybrid_lb.")

1256
1257
            # Local DP size defaults to global DP size if not set.
            data_parallel_size_local = self.data_parallel_size
1258
1259
1260

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
        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
1275
1276
1277
1278
1279
1280
1281

        # 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

1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
        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"
                logger.info("Using mp-based distributed executor backend "
                            "for async scheduling.")
            if self.distributed_executor_backend == "uni":
                raise ValueError("Async scheduling is not supported with "
                                 "uni-process backend.")
            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.")

1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
        # 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

1312
        parallel_config = ParallelConfig(
1313
1314
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1315
            data_parallel_size=self.data_parallel_size,
1316
1317
            data_parallel_rank=self.data_parallel_rank or 0,
            data_parallel_external_lb=data_parallel_external_lb,
1318
1319
1320
            data_parallel_size_local=data_parallel_size_local,
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
1321
            data_parallel_backend=self.data_parallel_backend,
1322
            data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
1323
            enable_expert_parallel=self.enable_expert_parallel,
1324
            enable_eplb=self.enable_eplb,
1325
            eplb_config=self.eplb_config,
1326
1327
1328
            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,
1329
            ray_runtime_env=ray_runtime_env,
1330
            placement_group=placement_group,
1331
1332
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1333
            worker_extension_cls=self.worker_extension_cls,
1334
            decode_context_parallel_size=self.decode_context_parallel_size,
1335
        )
1336

1337
        speculative_config = self.create_speculative_config(
1338
1339
            target_model_config=model_config,
            target_parallel_config=parallel_config,
1340
            enable_chunked_prefill=self.enable_chunked_prefill,
1341
            disable_log_stats=self.disable_log_stats,
1342
1343
        )

1344
1345
1346
1347
1348
        # 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
1349

1350
        scheduler_config = SchedulerConfig(
1351
            runner_type=model_config.runner_type,
1352
1353
1354
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1355
            cuda_graph_sizes=self.cuda_graph_sizes,
1356
            num_lookahead_slots=num_lookahead_slots,
1357
1358
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1359
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1360
            is_multimodal_model=model_config.is_multimodal_model,
1361
            preemption_mode=self.preemption_mode,
1362
1363
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1364
            policy=self.scheduling_policy,
1365
            scheduler_cls=self.scheduler_cls,
1366
1367
1368
            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,
1369
1370
            disable_hybrid_kv_cache_manager=self.
            disable_hybrid_kv_cache_manager,
1371
            async_scheduling=self.async_scheduling,
1372
        )
1373

1374
1375
1376
1377
1378
        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")

1379
        lora_config = LoRAConfig(
1380
            bias_enabled=self.enable_lora_bias,
1381
1382
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1383
            default_mm_loras=self.default_mm_loras,
1384
            fully_sharded_loras=self.fully_sharded_loras,
1385
1386
1387
1388
            lora_extra_vocab_size=self.lora_extra_vocab_size,
            lora_dtype=self.lora_dtype,
            max_cpu_loras=self.max_cpu_loras if self.max_cpu_loras
            and self.max_cpu_loras > 0 else None) if self.enable_lora else None
1389

1390
1391
1392
1393
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1394
        load_config = self.create_load_config()
1395

1396
        decoding_config = DecodingConfig(
1397
1398
1399
1400
1401
            backend=self.guided_decoding_backend,
            disable_fallback=self.guided_decoding_disable_fallback,
            disable_any_whitespace=self.guided_decoding_disable_any_whitespace,
            disable_additional_properties=\
                self.guided_decoding_disable_additional_properties,
1402
1403
            reasoning_backend=self.reasoning_parser
        )
1404

1405
        observability_config = ObservabilityConfig(
1406
1407
            show_hidden_metrics_for_version=(
                self.show_hidden_metrics_for_version),
1408
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1409
            collect_detailed_traces=self.collect_detailed_traces,
1410
        )
1411

1412
        config = VllmConfig(
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
            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,
            decoding_config=decoding_config,
            observability_config=observability_config,
1423
            compilation_config=self.compilation_config,
1424
            kv_transfer_config=self.kv_transfer_config,
1425
            kv_events_config=self.kv_events_config,
1426
            additional_config=self.additional_config,
1427
        )
1428

1429
1430
        return config

1431
1432
1433
1434
1435
1436
    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.

1437
        if self.load_format == "sharded_state":
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
            _raise_or_fallback(
                feature_name=f"--load_format {self.load_format}",
                recommend_to_remove=False)
            return False

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

1449
        if self.preemption_mode != SchedulerConfig.preemption_mode:
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
            _raise_or_fallback(feature_name="--preemption-mode",
                               recommend_to_remove=True)
            return False

        if (self.disable_async_output_proc
                != EngineArgs.disable_async_output_proc):
            _raise_or_fallback(feature_name="--disable-async-output-proc",
                               recommend_to_remove=True)
            return False

1460
        if self.scheduler_delay_factor != SchedulerConfig.delay_factor:
1461
1462
1463
1464
1465
            _raise_or_fallback(feature_name="--scheduler-delay-factor",
                               recommend_to_remove=True)
            return False

        if self.kv_cache_dtype != "auto":
1466
            supported = current_platform.is_kv_cache_dtype_supported(
1467
                self.kv_cache_dtype, model_config)
1468
1469
1470
1471
            if not supported:
                _raise_or_fallback(feature_name="--kv-cache-dtype",
                                   recommend_to_remove=False)
                return False
1472

1473
1474
1475
1476
1477
1478
        # No text embedding inputs so far.
        if self.enable_prompt_embeds:
            _raise_or_fallback(feature_name="--enable-prompt-embeds",
                               recommend_to_remove=False)
            return False

1479
        # No Mamba or Encoder-Decoder so far.
1480
1481
1482
1483
1484
1485
1486
        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
1487
                != SchedulerConfig.max_num_partial_prefills
1488
                or self.max_long_partial_prefills
1489
                != SchedulerConfig.max_long_partial_prefills):
1490
1491
1492
1493
            _raise_or_fallback(feature_name="Concurrent Partial Prefill",
                               recommend_to_remove=False)
            return False

1494
        # V1 supports N-gram, Medusa, and Eagle speculative decoding.
1495
1496
1497
1498
1499
1500
        if (self.speculative_config is not None
                and self.speculative_config.get("method") == "draft_model"):
            raise NotImplementedError(
                "Speculative decoding with draft model is not supported yet. "
                "Please consider using other speculative decoding methods "
                "such as ngram, medusa, eagle, or deepseek_mtp.")
1501
1502

        V1_BACKENDS = [
1503
1504
1505
1506
1507
1508
            "FLASH_ATTN_VLLM_V1",
            "FLASH_ATTN",
            "PALLAS",
            "PALLAS_VLLM_V1",
            "TRITON_ATTN_VLLM_V1",
            "TRITON_MLA",
1509
            "CUTLASS_MLA",
1510
            "FLASHMLA",
1511
1512
            "FLASHMLA_VLLM_V1",
            "FLASH_ATTN_MLA",
1513
1514
            "FLASHINFER",
            "FLASHINFER_VLLM_V1",
1515
            "FLASHINFER_MLA",
1516
            "ROCM_AITER_MLA",
1517
            "TORCH_SDPA_VLLM_V1",
1518
            "FLEX_ATTENTION",
1519
            "TREE_ATTN",
1520
            "XFORMERS_VLLM_V1",
1521
1522
1523
1524
1525
1526
1527
        ]
        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

1528
1529
        # Platforms must decide if they can support v1 for this model
        if not current_platform.supports_v1(model_config=model_config):
1530
1531
1532
1533
            _raise_or_fallback(
                feature_name=f"device type={current_platform.device_type}",
                recommend_to_remove=False)
            return False
1534
1535
1536
        #############################################################
        # 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
        # The platform may be supported on V1, but off by default for now.
        if not current_platform.default_v1(  # noqa: SIM103
                model_config=model_config) and _warn_or_fallback(
                    current_platform.device_name):
1554
            return False
1555
1556
1557
1558
1559
1560
1561

        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

1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
        #############################################################

        return True

    def _set_default_args_v0(self, model_config: ModelConfig) -> None:
        """Set Default Arguments for V0 Engine."""

        max_model_len = model_config.max_model_len
        use_long_context = max_model_len > 32768
        if self.enable_chunked_prefill is None:
            # Chunked prefill not supported for Multimodal or MLA in V0.
            if model_config.is_multimodal_model or model_config.use_mla:
                self.enable_chunked_prefill = False

            # Enable chunked prefill by default for long context (> 32K)
            # models to avoid OOM errors in initial memory profiling phase.
            elif use_long_context:
                is_gpu = current_platform.is_cuda()
                use_sliding_window = (model_config.get_sliding_window()
                                      is not None)
1582
                use_spec_decode = self.speculative_config is not None
1583
1584

                if (is_gpu and not use_sliding_window and not use_spec_decode
1585
                        and not self.enable_lora):
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
                    self.enable_chunked_prefill = True
                    logger.warning(
                        "Chunked prefill is enabled by default for models "
                        "with max_model_len > 32K. Chunked prefill might "
                        "not work with some features or models. If you "
                        "encounter any issues, please disable by launching "
                        "with --enable-chunked-prefill=False.")

            if self.enable_chunked_prefill is None:
                self.enable_chunked_prefill = False

        if not self.enable_chunked_prefill and use_long_context:
            logger.warning(
                "The model has a long context length (%s). This may cause"
                "OOM during the initial memory profiling phase, or result "
                "in low performance due to small KV cache size. Consider "
                "setting --max-model-len to a smaller value.", max_model_len)

1604
1605
1606
1607
1608
1609
        # Disable prefix caching for multimodal models for VLLM_V0.
        if self.enable_prefix_caching and model_config.is_multimodal_model:
            logger.warning(
                "--enable-prefix-caching is not supported for multimodal "
                "models in V0 and has been disabled.")
            self.enable_prefix_caching = False
1610
1611
1612
1613
1614

        # Set max_num_seqs to 256 for VLLM_V0.
        if self.max_num_seqs is None:
            self.max_num_seqs = 256

1615
1616
    def _set_default_args_v1(self, usage_context: UsageContext,
                             model_config: ModelConfig) -> None:
1617
        """Set Default Arguments for V1 Engine."""
1618

1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
        # 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
            if self.enable_prefix_caching is None:
                self.enable_prefix_caching = True
        else:

            pooling_type = model_config.pooler_config.pooling_type
1629
1630
1631
1632
            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)
1633

1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
            action = "Enabling" if \
                incremental_prefill_supported else "Disabling"

            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)

1644
1645
1646
        # 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:
1647
            self.scheduler_cls = "vllm.v1.core.sched.scheduler.Scheduler"
1648

1649
1650
        # When no user override, set the default values based on the usage
        # context.
1651
        # Use different default values for different hardware.
1652
1653
1654
1655
1656
1657
1658

        # 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:
1659
            device_memory = current_platform.get_device_total_memory()
1660
            device_name = current_platform.get_device_name().lower()
1661
1662
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1663
            device_memory = 0
1664

1665
1666
1667
        # 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.
1668
        from vllm.usage.usage_lib import UsageContext
1669
        if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
1670
            # For GPUs like H100 and MI300x, use larger default values.
1671
1672
1673
1674
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1675
1676
1677
1678
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
1679
1680
1681
1682
1683
1684
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1685
1686
1687
1688
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
1689

1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
        # 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,
                }
            }

1705
1706
        # cpu specific default values.
        if current_platform.is_cpu():
1707
            world_size = self.pipeline_parallel_size * self.tensor_parallel_size
1708
            default_max_num_batched_tokens = {
1709
1710
                UsageContext.LLM_CLASS: 4096 * world_size,
                UsageContext.OPENAI_API_SERVER: 2048 * world_size,
1711
1712
            }
            default_max_num_seqs = {
1713
1714
                UsageContext.LLM_CLASS: 256 * world_size,
                UsageContext.OPENAI_API_SERVER: 128 * world_size,
1715
1716
            }

1717
        use_context_value = usage_context.value if usage_context else None
1718
1719
        if (self.max_num_batched_tokens is None
                and usage_context in default_max_num_batched_tokens):
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
            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:
1731
1732
1733
1734
1735
                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]
1736
            logger.debug(
1737
                "Setting max_num_batched_tokens to %d for %s usage context.",
1738
                self.max_num_batched_tokens, use_context_value)
1739

1740
1741
        if (self.max_num_seqs is None
                and usage_context in default_max_num_seqs):
1742
1743
            self.max_num_seqs = min(default_max_num_seqs[usage_context],
                                    self.max_num_batched_tokens or sys.maxsize)
1744
1745
1746

            logger.debug("Setting max_num_seqs to %d for %s usage context.",
                         self.max_num_seqs, use_context_value)
1747

1748

1749
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1750
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1751
    """Arguments for asynchronous vLLM engine."""
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
    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
1769
1770

    @staticmethod
1771
1772
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1773
        # Initialize plugin to update the parser, for example, The plugin may
1774
        # add a new kind of quantization method to --quantization argument or
1775
1776
        # a new device to --device argument.
        load_general_plugins()
1777
1778
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1779
1780
1781
1782
        parser.add_argument('--enable-log-requests',
                            action=argparse.BooleanOptionalAction,
                            default=AsyncEngineArgs.enable_log_requests,
                            help='Enable logging requests.')
1783
        parser.add_argument('--disable-log-requests',
1784
1785
1786
1787
                            action=argparse.BooleanOptionalAction,
                            default=not AsyncEngineArgs.enable_log_requests,
                            help='[DEPRECATED] Disable logging requests.',
                            deprecated=True)
1788
        current_platform.pre_register_and_update(parser)
1789
        return parser
1790
1791


1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
def _raise_or_fallback(feature_name: str, recommend_to_remove: bool):
    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
        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)


def _warn_or_fallback(feature_name: str) -> bool:
    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
        logger.warning(
            "Detected VLLM_USE_V1=1 with %s. Usage should "
            "be considered experimental. Please report any "
            "issues on Github.", feature_name)
        should_exit = False
    else:
        logger.info(
            "%s is experimental on VLLM_USE_V1=1. "
            "Falling back to V0 Engine.", feature_name)
        should_exit = True
    return should_exit


1819
1820
1821
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
1822

1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
    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)