arg_utils.py 75.7 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 dataclasses
7
import json
8
import sys
9
import threading
10
import warnings
11
from dataclasses import MISSING, dataclass, fields, is_dataclass
12
from itertools import permutations
13
14
from typing import (Annotated, Any, Callable, Dict, List, Literal, Optional,
                    Type, TypeVar, Union, cast, get_args, get_origin)
15

16
import regex as re
17
import torch
18
from pydantic import SkipValidation, TypeAdapter, ValidationError
19
from typing_extensions import TypeIs, deprecated
20

21
import vllm.envs as envs
22
from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
23
24
25
26
27
28
                         ConfigFormat, ConfigType, DecodingConfig,
                         DetailedTraceModules, Device, DeviceConfig,
                         DistributedExecutorBackend, GuidedDecodingBackend,
                         GuidedDecodingBackendV1, HfOverrides, KVEventsConfig,
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
                         ModelConfig, ModelDType, ModelImpl, MultiModalConfig,
29
30
31
32
33
                         ObservabilityConfig, ParallelConfig, PoolerConfig,
                         PrefixCachingHashAlgo, PromptAdapterConfig,
                         SchedulerConfig, SchedulerPolicy, SpeculativeConfig,
                         TaskOption, TokenizerMode, TokenizerPoolConfig,
                         VllmConfig, get_attr_docs, get_field)
34
from vllm.executor.executor_base import ExecutorBase
35
from vllm.logger import init_logger
36
from vllm.model_executor.layers.quantization import QuantizationMethods
37
from vllm.plugins import load_general_plugins
38
from vllm.reasoning import ReasoningParserManager
39
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
40
from vllm.transformers_utils.utils import check_gguf_file
41
from vllm.usage.usage_lib import UsageContext
42
from vllm.utils import (STR_DUAL_CHUNK_FLASH_ATTN_VAL, FlexibleArgumentParser,
Rui Qiao's avatar
Rui Qiao committed
43
                        GiB_bytes, get_ip, is_in_ray_actor)
44
45

# yapf: enable
46

47
48
logger = init_logger(__name__)

49
50
51
52
53
# object is used to allow for special typing forms
T = TypeVar("T")
TypeHint = Union[type[Any], object]
TypeHintT = Union[type[T], object]

54

55
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
56

57
    def _parse_type(val: str) -> T:
58
59
60
61
62
63
64
        try:
            if return_type is json.loads and not re.match("^{.*}$", val):
                return cast(T, nullable_kvs(val))
            return return_type(val)
        except ValueError as e:
            raise argparse.ArgumentTypeError(
                f"Value {val} cannot be converted to {return_type}.") from e
65

66
67
68
69
70
71
72
73
74
75
76
    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)

77
    return _optional_type
78
79


80
81
82
def union_dict_and_str(val: str) -> Optional[Union[str, dict[str, str]]]:
    if not re.match("^{.*}$", val):
        return str(val)
83
    return optional_type(json.loads)(val)
84
85


86
87
88
89
90
91
@deprecated(
    "Passing a JSON argument as a string containing comma separated key=value "
    "pairs is deprecated. This will be removed in v0.10.0. Please use a JSON "
    "string instead.")
def nullable_kvs(val: str) -> dict[str, int]:
    """Parses a string containing comma separate key [str] to value [int]
92
93
94
95
96
97
98
99
    pairs into a dictionary.

    Args:
        val: String value to be parsed.

    Returns:
        Dictionary with parsed values.
    """
100
    out_dict: dict[str, int] = {}
101
    for item in val.split(","):
102
103
104
105
106
        kv_parts = [part.lower().strip() for part in item.split("=")]
        if len(kv_parts) != 2:
            raise argparse.ArgumentTypeError(
                "Each item should be in the form KEY=VALUE")
        key, value = kv_parts
107
108

        try:
109
            parsed_value = int(value)
110
111
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
112
113
114
115
116
117
            raise argparse.ArgumentTypeError(msg) from exc

        if key in out_dict and out_dict[key] != parsed_value:
            raise argparse.ArgumentTypeError(
                f"Conflicting values specified for key: {key}")
        out_dict[key] = parsed_value
118
119
120
121

    return out_dict


122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
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)


137
138
139
140
141
142
143
144
145
146
147
148
def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]:
    """Convert Literal type hints to argparse kwargs."""
    type_hint = get_type(type_hints, Literal)
    choices = get_args(type_hint)
    choice_type = type(choices[0])
    if not all(isinstance(choice, choice_type) for choice in choices):
        raise ValueError(
            "All choices must be of the same type. "
            f"Got {choices} with types {[type(c) for c in choices]}")
    return {"type": choice_type, "choices": sorted(choices)}


149
150
151
152
153
154
155
156
157
def is_not_builtin(type_hint: TypeHint) -> bool:
    """Check if the class is not a built-in type."""
    return type_hint.__module__ != "builtins"


def get_kwargs(cls: ConfigType) -> dict[str, Any]:
    cls_docs = get_attr_docs(cls)
    kwargs = {}
    for field in fields(cls):
158
159
160
        # Get the set of possible types for the field
        type_hints: set[TypeHint] = set()
        if get_origin(field.type) in {Union, Annotated}:
161
162
            predicate = lambda arg: not isinstance(arg, SkipValidation)
            type_hints.update(filter(predicate, get_args(field.type)))
163
164
165
166
167
168
169
        else:
            type_hints.add(field.type)

        # 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)

170
        # Get the default value of the field
171
172
173
        if field.default is not MISSING:
            default = field.default
        elif field.default_factory is not MISSING:
174
            default = field.default_factory()
175
176
177

        # Get the help text for the field
        name = field.name
178
        help = cls_docs[name].strip()
179
180
181
182
183
184
185
        # 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
186
187
188
189
190
        json_tip = """\n\nShould either be a valid JSON string or JSON keys
        passed individually. For example, the following sets of arguments are
        equivalent:\n\n
        - `--json-arg '{"key1": "value1", "key2": {"key3": "value2"}}'`\n
        - `--json-arg.key1 value1 --json-arg.key2.key3 value2`\n\n"""
191
        if dataclass_cls is not None:
192
193
194
195
196
197
198
199
200
201

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

            kwargs[name]["type"] = parse_dataclass
202
203
            kwargs[name]["help"] += json_tip
        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
220
221
222
223
224
225
226
        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)
            assert len(types) == 1, (
                "List type must have exactly one type. Got "
                f"{type_hint} with types {types}")
            kwargs[name]["type"] = types[0]
            kwargs[name]["nargs"] = "+"
        elif contains_type(type_hints, int):
            kwargs[name]["type"] = int
227
            # Special case for large integers
228
            if name in {"max_model_len", "max_num_batched_tokens"}:
229
                kwargs[name]["type"] = human_readable_int
230
231
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
232
233
234
        elif (contains_type(type_hints, dict)
              and (contains_type(type_hints, str)
                   or any(is_not_builtin(th) for th in type_hints))):
235
            kwargs[name]["type"] = union_dict_and_str
236
        elif contains_type(type_hints, dict):
237
            kwargs[name]["type"] = parse_type(json.loads)
238
            kwargs[name]["help"] += json_tip
239
240
241
242
243
244
245
        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}.")

246
247
248
249
250
        # 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"]}))

251
252
253
254
255
256
257
        # 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
258
259


260
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
261
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
262
    """Arguments for vLLM engine."""
263
264
265
266
267
268
269
    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
    task: TaskOption = ModelConfig.task
    skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init
270
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
271
272
273
    tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
274
275
    download_dir: Optional[str] = LoadConfig.download_dir
    load_format: str = LoadConfig.load_format
276
277
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
278
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
279
280
    seed: Optional[int] = ModelConfig.seed
    max_model_len: Optional[int] = ModelConfig.max_model_len
281
282
    cuda_graph_sizes: list[int] = get_field(SchedulerConfig,
                                            "cuda_graph_sizes")
283
284
285
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
286
    distributed_executor_backend: Optional[Union[
287
288
        DistributedExecutorBackend,
        Type[ExecutorBase]]] = ParallelConfig.distributed_executor_backend
289
    # number of P/D disaggregation (or other disaggregation) workers
290
291
292
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
    data_parallel_size: int = ParallelConfig.data_parallel_size
293
294
295
    data_parallel_size_local: Optional[int] = None
    data_parallel_address: Optional[str] = None
    data_parallel_rpc_port: Optional[int] = None
Rui Qiao's avatar
Rui Qiao committed
296
    data_parallel_backend: str = ParallelConfig.data_parallel_backend
297
298
299
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
    max_parallel_loading_workers: Optional[
        int] = ParallelConfig.max_parallel_loading_workers
300
301
302
303
    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
304
305
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
306
    use_v2_block_manager: bool = True
307
308
309
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
310
311
312
313
314
315
316
    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
317
    max_logprobs: int = ModelConfig.max_logprobs
318
    disable_log_stats: bool = False
319
320
321
322
323
    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
324
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
325
326
327
328
    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
329
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
330
331
332
    # The following three fields are deprecated and will be removed in a future
    # release. Setting them will have no effect. Please remove them from your
    # configurations.
333
    tokenizer_pool_size: int = TokenizerPoolConfig.pool_size
334
335
    tokenizer_pool_type: str = TokenizerPoolConfig.pool_type
    tokenizer_pool_extra_config: dict = \
336
        get_field(TokenizerPoolConfig, "extra_config")
337
    limit_mm_per_prompt: dict[str, int] = \
338
        get_field(MultiModalConfig, "limit_per_prompt")
339
340
341
342
    mm_processor_kwargs: Optional[Dict[str, Any]] = \
        MultiModalConfig.mm_processor_kwargs
    disable_mm_preprocessor_cache: bool = \
        MultiModalConfig.disable_mm_preprocessor_cache
343
    # LoRA fields
344
    enable_lora: bool = False
345
346
347
348
349
350
351
352
353
354
    enable_lora_bias: bool = LoRAConfig.bias_enabled
    max_loras: int = LoRAConfig.max_loras
    max_lora_rank: int = LoRAConfig.max_lora_rank
    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
    long_lora_scaling_factors: Optional[tuple[float, ...]] = \
        LoRAConfig.long_lora_scaling_factors
    # PromptAdapter fields
355
    enable_prompt_adapter: bool = False
356
357
358
359
    max_prompt_adapters: int = PromptAdapterConfig.max_prompt_adapters
    max_prompt_adapter_token: int = \
        PromptAdapterConfig.max_prompt_adapter_token

360
    device: Device = DeviceConfig.device
361
362
    num_scheduler_steps: int = SchedulerConfig.num_scheduler_steps
    multi_step_stream_outputs: bool = SchedulerConfig.multi_step_stream_outputs
363
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
364
365
    num_gpu_blocks_override: Optional[
        int] = CacheConfig.num_gpu_blocks_override
366
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
367
368
    model_loader_extra_config: dict = \
        get_field(LoadConfig, "model_loader_extra_config")
369
370
    ignore_patterns: Optional[Union[str,
                                    List[str]]] = LoadConfig.ignore_patterns
371
    preemption_mode: Optional[str] = SchedulerConfig.preemption_mode
372

373
374
375
376
    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
377

378
379
380
381
382
383
    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
384
385
    logits_processor_pattern: Optional[
        str] = ModelConfig.logits_processor_pattern
386

387
    speculative_config: Optional[Dict[str, Any]] = None
388

389
    qlora_adapter_name_or_path: Optional[str] = None
390
391
392
393
394
395
    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
396
    disable_async_output_proc: bool = not ModelConfig.use_async_output_proc
397
398
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
    scheduler_cls: Union[str, Type[object]] = SchedulerConfig.scheduler_cls
399

400
401
402
403
    override_neuron_config: dict[str, Any] = \
        get_field(ModelConfig, "override_neuron_config")
    override_pooler_config: Optional[Union[dict, PoolerConfig]] = \
        ModelConfig.override_pooler_config
404
405
    compilation_config: CompilationConfig = \
        get_field(VllmConfig, "compilation_config")
406
407
    worker_cls: str = ParallelConfig.worker_cls
    worker_extension_cls: str = ParallelConfig.worker_extension_cls
408

409
    kv_transfer_config: Optional[KVTransferConfig] = None
410
    kv_events_config: Optional[KVEventsConfig] = None
411

412
413
414
415
416
    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
417

418
    calculate_kv_scales: bool = CacheConfig.calculate_kv_scales
419

420
421
    additional_config: dict[str, Any] = \
        get_field(VllmConfig, "additional_config")
422
423
424
    enable_reasoning: Optional[bool] = None  # DEPRECATED
    reasoning_parser: str = DecodingConfig.reasoning_backend

425
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
426
    pt_load_map_location: str = LoadConfig.pt_load_map_location
427

428
429
430
    enable_multimodal_encoder_data_parallel: bool = \
        ParallelConfig.enable_multimodal_encoder_data_parallel

431
    def __post_init__(self):
432
433
434
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
435
        if isinstance(self.compilation_config, (int, dict)):
436
437
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
438
439
440
441
442
443
444
        if self.qlora_adapter_name_or_path is not None:
            warnings.warn(
                "The `qlora_adapter_name_or_path` is deprecated "
                "and will be removed in v0.10.0. ",
                DeprecationWarning,
                stacklevel=2,
            )
445
        # Setup plugins
446
447
        from vllm.plugins import load_general_plugins
        load_general_plugins()
448
449

    @staticmethod
450
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
451
        """Shared CLI arguments for vLLM engine."""
452

453
        # Model arguments
454
455
456
457
458
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
Reid's avatar
Reid committed
459
        if not ('serve' in sys.argv[1:] and '--help' in sys.argv[1:]):
460
            model_group.add_argument("--model", **model_kwargs["model"])
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
        model_group.add_argument("--task", **model_kwargs["task"])
        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"])
        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"])
497
498
        model_group.add_argument("--enable-prompt-embeds",
                                 **model_kwargs["enable_prompt_embeds"])
499
500
501
502
503
504
505
506
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
536
537
        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",
                                 choices=[f.value for f in ConfigFormat],
                                 **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-neuron-config",
                                 **model_kwargs["override_neuron_config"])
        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"])

538
539
540
541
542
543
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
544
        load_group.add_argument("--load-format",
545
546
                                choices=[f.value for f in LoadFormat],
                                **load_kwargs["load_format"])
547
        load_group.add_argument("--download-dir",
548
                                **load_kwargs["download_dir"])
549
        load_group.add_argument("--model-loader-extra-config",
550
                                **load_kwargs["model_loader_extra_config"])
551
552
553
        load_group.add_argument("--ignore-patterns",
                                **load_kwargs["ignore_patterns"])
        load_group.add_argument("--use-tqdm-on-load",
554
                                **load_kwargs["use_tqdm_on_load"])
555
556
557
558
559
560
561
562
        load_group.add_argument(
            "--qlora-adapter-name-or-path",
            type=str,
            default=None,
            help="The `--qlora-adapter-name-or-path` has no effect, do not set"
            " it, and it  will be removed in v0.10.0.",
            deprecated=True,
        )
563
564
        load_group.add_argument('--pt-load-map-location',
                                **load_kwargs["pt_load_map_location"])
565

566
567
568
569
570
571
        # Guided decoding arguments
        guided_decoding_kwargs = get_kwargs(DecodingConfig)
        guided_decoding_group = parser.add_argument_group(
            title="DecodingConfig",
            description=DecodingConfig.__doc__,
        )
572
573
        guided_decoding_group.add_argument("--guided-decoding-backend",
                                           **guided_decoding_kwargs["backend"])
574
        guided_decoding_group.add_argument(
575
576
577
578
579
580
581
582
            "--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"])
583
584
585
        guided_decoding_group.add_argument(
            "--enable-reasoning",
            action=argparse.BooleanOptionalAction,
586
            deprecated=True,
587
            help="[DEPRECATED] The `--enable-reasoning` flag is deprecated as "
588
            "of v0.9.0. Use `--reasoning-parser` to specify the reasoning "
589
            "parser backend instead. This flag (`--enable-reasoning`) will be "
590
591
            "removed in v0.10.0. When `--reasoning-parser` is specified, "
            "reasoning mode is automatically enabled.")
592
593
594
595
596
597
        guided_decoding_group.add_argument(
            "--reasoning-parser",
            # This choices is a special case because it's not static
            choices=list(ReasoningParserManager.reasoning_parsers),
            **guided_decoding_kwargs["reasoning_backend"])

598
        # Parallel arguments
599
600
601
602
603
604
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
605
            "--distributed-executor-backend",
606
607
            **parallel_kwargs["distributed_executor_backend"])
        parallel_group.add_argument(
608
            "--pipeline-parallel-size", "-pp",
609
            **parallel_kwargs["pipeline_parallel_size"])
610
        parallel_group.add_argument("--tensor-parallel-size", "-tp",
611
                                    **parallel_kwargs["tensor_parallel_size"])
612
        parallel_group.add_argument("--data-parallel-size", "-dp",
613
                                    **parallel_kwargs["data_parallel_size"])
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
        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
629
630
631
632
633
634
        parallel_group.add_argument('--data-parallel-backend',
                                    '-dpb',
                                    type=str,
                                    default='mp',
                                    help='Backend for data parallel, either '
                                    '"mp" or "ray".')
635
        parallel_group.add_argument(
636
            "--enable-expert-parallel",
637
638
            **parallel_kwargs["enable_expert_parallel"])
        parallel_group.add_argument(
639
            "--max-parallel-loading-workers",
640
641
            **parallel_kwargs["max_parallel_loading_workers"])
        parallel_group.add_argument(
642
            "--ray-workers-use-nsight",
643
644
            **parallel_kwargs["ray_workers_use_nsight"])
        parallel_group.add_argument(
645
            "--disable-custom-all-reduce",
646
            **parallel_kwargs["disable_custom_all_reduce"])
647
648
649
650
        parallel_group.add_argument("--worker-cls",
                                    **parallel_kwargs["worker_cls"])
        parallel_group.add_argument("--worker-extension-cls",
                                    **parallel_kwargs["worker_extension_cls"])
651
652
653
        parallel_group.add_argument(
            "--enable-multimodal-encoder-data-parallel",
            **parallel_kwargs["enable_multimodal_encoder_data_parallel"])
654

655
656
657
658
659
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
660
        )
661
662
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
        cache_group.add_argument("--gpu-memory-utilization",
663
                                 **cache_kwargs["gpu_memory_utilization"])
664
665
        cache_group.add_argument("--swap-space", **cache_kwargs["swap_space"])
        cache_group.add_argument("--kv-cache-dtype",
666
                                 **cache_kwargs["cache_dtype"])
667
        cache_group.add_argument("--num-gpu-blocks-override",
668
669
670
671
672
                                 **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"])
673
        cache_group.add_argument("--cpu-offload-gb",
674
                                 **cache_kwargs["cpu_offload_gb"])
675
        cache_group.add_argument("--calculate-kv-scales",
676
677
                                 **cache_kwargs["calculate_kv_scales"])

678
679
680
681
682
683
        # Tokenizer arguments
        tokenizer_kwargs = get_kwargs(TokenizerPoolConfig)
        tokenizer_group = parser.add_argument_group(
            title="TokenizerPoolConfig",
            description=TokenizerPoolConfig.__doc__,
        )
684
        tokenizer_group.add_argument("--tokenizer-pool-size",
685
                                     **tokenizer_kwargs["pool_size"])
686
        tokenizer_group.add_argument("--tokenizer-pool-type",
687
                                     **tokenizer_kwargs["pool_type"])
688
        tokenizer_group.add_argument("--tokenizer-pool-extra-config",
689
                                     **tokenizer_kwargs["extra_config"])
690
691

        # Multimodal related configs
692
693
694
695
696
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
697
        multimodal_group.add_argument("--limit-mm-per-prompt",
698
                                      **multimodal_kwargs["limit_per_prompt"])
699
        multimodal_group.add_argument(
700
            "--mm-processor-kwargs",
701
702
            **multimodal_kwargs["mm_processor_kwargs"])
        multimodal_group.add_argument(
703
            "--disable-mm-preprocessor-cache",
704
            **multimodal_kwargs["disable_mm_preprocessor_cache"])
705

706
        # LoRA related configs
707
708
709
710
711
712
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
713
            "--enable-lora",
714
            action=argparse.BooleanOptionalAction,
715
716
            help="If True, enable handling of LoRA adapters.")
        lora_group.add_argument("--enable-lora-bias",
717
                                **lora_kwargs["bias_enabled"])
718
719
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
        lora_group.add_argument("--max-lora-rank",
720
                                **lora_kwargs["max_lora_rank"])
721
        lora_group.add_argument("--lora-extra-vocab-size",
722
723
                                **lora_kwargs["lora_extra_vocab_size"])
        lora_group.add_argument(
724
            "--lora-dtype",
725
726
            **lora_kwargs["lora_dtype"],
        )
727
        lora_group.add_argument("--long-lora-scaling-factors",
728
                                **lora_kwargs["long_lora_scaling_factors"])
729
        lora_group.add_argument("--max-cpu-loras",
730
                                **lora_kwargs["max_cpu_loras"])
731
        lora_group.add_argument("--fully-sharded-loras",
732
733
734
735
736
737
738
739
740
                                **lora_kwargs["fully_sharded_loras"])

        # PromptAdapter related configs
        prompt_adapter_kwargs = get_kwargs(PromptAdapterConfig)
        prompt_adapter_group = parser.add_argument_group(
            title="PromptAdapterConfig",
            description=PromptAdapterConfig.__doc__,
        )
        prompt_adapter_group.add_argument(
741
            "--enable-prompt-adapter",
742
            action=argparse.BooleanOptionalAction,
743
            help="If True, enable handling of PromptAdapters.")
744
        prompt_adapter_group.add_argument(
745
            "--max-prompt-adapters",
746
747
            **prompt_adapter_kwargs["max_prompt_adapters"])
        prompt_adapter_group.add_argument(
748
            "--max-prompt-adapter-token",
749
            **prompt_adapter_kwargs["max_prompt_adapter_token"])
750
751
752
753
754
755
756

        # Device arguments
        device_kwargs = get_kwargs(DeviceConfig)
        device_group = parser.add_argument_group(
            title="DeviceConfig",
            description=DeviceConfig.__doc__,
        )
757
758
759
        device_group.add_argument("--device",
                                  **device_kwargs["device"],
                                  deprecated=True)
760

761
762
763
764
765
766
        # Speculative arguments
        speculative_group = parser.add_argument_group(
            title="SpeculativeConfig",
            description=SpeculativeConfig.__doc__,
        )
        speculative_group.add_argument(
767
            "--speculative-config",
768
769
            type=json.loads,
            default=None,
770
771
            help="The configurations for speculative decoding. Should be a "
            "JSON string.")
772

773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
        # 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"])
796

797
798
799
800
801
802
803
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
804
            "--max-num-batched-tokens",
805
            **scheduler_kwargs["max_num_batched_tokens"])
806
        scheduler_group.add_argument("--max-num-seqs",
807
808
809
810
811
812
813
                                     **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"])
814
815
        scheduler_group.add_argument('--cuda-graph-sizes',
                                     **scheduler_kwargs["cuda_graph_sizes"])
816
817
818
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
            **scheduler_kwargs["long_prefill_token_threshold"])
819
        scheduler_group.add_argument("--num-lookahead-slots",
820
                                     **scheduler_kwargs["num_lookahead_slots"])
821
        scheduler_group.add_argument("--scheduler-delay-factor",
822
                                     **scheduler_kwargs["delay_factor"])
823
        scheduler_group.add_argument("--preemption-mode",
824
                                     **scheduler_kwargs["preemption_mode"])
825
        scheduler_group.add_argument("--num-scheduler-steps",
826
                                     **scheduler_kwargs["num_scheduler_steps"])
827
        scheduler_group.add_argument(
828
            "--multi-step-stream-outputs",
829
            **scheduler_kwargs["multi_step_stream_outputs"])
830
        scheduler_group.add_argument("--scheduling-policy",
831
                                     **scheduler_kwargs["policy"])
832
        scheduler_group.add_argument(
833
            "--enable-chunked-prefill",
834
            **scheduler_kwargs["enable_chunked_prefill"])
835
836
837
        scheduler_group.add_argument(
            "--disable-chunked-mm-input",
            **scheduler_kwargs["disable_chunked_mm_input"])
838
839
840
841
        scheduler_group.add_argument("--scheduler-cls",
                                     **scheduler_kwargs["scheduler_cls"])

        # vLLM arguments
842
        vllm_kwargs = get_kwargs(VllmConfig)
843
844
845
846
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
847
848
849
850
851
852
853
854
        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"])
855

856
857
858
859
        # Other arguments
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
                            default=True,
860
                            deprecated=True,
861
862
863
864
865
866
867
868
                            help='[DEPRECATED] block manager v1 has been '
                            'removed and SelfAttnBlockSpaceManager (i.e. '
                            'block manager v2) is now the default. '
                            'Setting this flag to True or False'
                            ' has no effect on vLLM behavior.')
        parser.add_argument('--disable-log-stats',
                            action='store_true',
                            help='Disable logging statistics.')
869

870
        return parser
871
872

    @classmethod
873
    def from_cli_args(cls, args: argparse.Namespace):
874
875
876
        # 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
877
878
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
879

880
    def create_model_config(self) -> ModelConfig:
881
882
883
884
885
886
887
888
889
890
891
        # 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
                and self.model in MODELS_ON_S3
                and self.load_format == LoadFormat.AUTO):  # noqa: E501
            self.model = f"{MODEL_WEIGHTS_S3_BUCKET}/{self.model}"
            self.load_format = LoadFormat.RUNAI_STREAMER

892
        return ModelConfig(
893
            model=self.model,
894
            hf_config_path=self.hf_config_path,
895
            task=self.task,
896
            tokenizer=self.tokenizer,
897
898
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
899
            allowed_local_media_path=self.allowed_local_media_path,
900
901
902
903
904
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
905
            rope_theta=self.rope_theta,
906
            hf_token=self.hf_token,
907
            hf_overrides=self.hf_overrides,
908
909
910
911
912
913
914
            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,
            disable_sliding_window=self.disable_sliding_window,
915
            disable_cascade_attn=self.disable_cascade_attn,
916
            skip_tokenizer_init=self.skip_tokenizer_init,
917
            enable_prompt_embeds=self.enable_prompt_embeds,
918
            served_model_name=self.served_model_name,
919
            limit_mm_per_prompt=self.limit_mm_per_prompt,
920
            use_async_output_proc=not self.disable_async_output_proc,
921
            config_format=self.config_format,
922
            mm_processor_kwargs=self.mm_processor_kwargs,
923
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
924
925
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
926
            logits_processor_pattern=self.logits_processor_pattern,
927
            generation_config=self.generation_config,
928
            override_generation_config=self.override_generation_config,
929
            enable_sleep_mode=self.enable_sleep_mode,
930
            model_impl=self.model_impl,
931
        )
932

933
934
    def create_load_config(self) -> LoadConfig:

935
936
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
937

938
939
940
941
942
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
943
            use_tqdm_on_load=self.use_tqdm_on_load,
944
            pt_load_map_location=self.pt_load_map_location,
945
        )
946

947
948
949
950
951
952
953
954
955
956
957
958
959
    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
960
        dictionary from the engine.
961
962
        """
        if self.speculative_config is None:
963
964
            return None

965
966
967
968
969
970
971
972
973
974
975
976
977
978
        # 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,
        })
        speculative_config = SpeculativeConfig.from_dict(
            self.speculative_config)

        return speculative_config

979
980
981
982
983
984
985
986
987
988
    def create_engine_config(
        self,
        usage_context: Optional[UsageContext] = None,
    ) -> 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
989

990
991
992
993
994
995
        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.
        """
996
997
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update()
998

999
        device_config = DeviceConfig(device=current_platform.device_type)
1000
1001
        model_config = self.create_model_config()

1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
        # * 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:
            self._set_default_args_v1(usage_context)
        else:
            self._set_default_args_v0(model_config)
1024

1025
1026
        assert self.enable_chunked_prefill is not None

1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
        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'")

1038
        cache_config = CacheConfig(
1039
            block_size=self.block_size,
1040
1041
1042
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1043
            is_attention_free=model_config.is_attention_free,
1044
1045
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1046
            enable_prefix_caching=self.enable_prefix_caching,
1047
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1048
            cpu_offload_gb=self.cpu_offload_gb,
1049
            calculate_kv_scales=self.calculate_kv_scales,
1050
        )
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062

        # 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()

1063
1064
1065
1066
1067
1068
1069
        # Local DP size defaults to global DP size if not set.
        data_parallel_size_local = self.data_parallel_size if (
            self.data_parallel_size_local
            is None) else self.data_parallel_size_local

        # DP address, used in multi-node case for torch distributed group
        # and ZMQ sockets.
Rui Qiao's avatar
Rui Qiao committed
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
        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
1084
1085
1086
1087
1088
1089
1090

        # 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

Rui Qiao's avatar
Rui Qiao committed
1091
1092
        data_parallel_backend = self.data_parallel_backend

1093
        parallel_config = ParallelConfig(
1094
1095
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1096
            data_parallel_size=self.data_parallel_size,
1097
1098
1099
            data_parallel_size_local=data_parallel_size_local,
            data_parallel_master_ip=data_parallel_address,
            data_parallel_rpc_port=data_parallel_rpc_port,
Rui Qiao's avatar
Rui Qiao committed
1100
            data_parallel_backend=data_parallel_backend,
1101
            enable_expert_parallel=self.enable_expert_parallel,
1102
1103
1104
            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,
1105
            placement_group=placement_group,
1106
1107
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1108
            worker_extension_cls=self.worker_extension_cls,
1109
1110
            enable_multimodal_encoder_data_parallel=self.
            enable_multimodal_encoder_data_parallel,
1111
        )
1112

1113
        speculative_config = self.create_speculative_config(
1114
1115
            target_model_config=model_config,
            target_parallel_config=parallel_config,
1116
            enable_chunked_prefill=self.enable_chunked_prefill,
1117
            disable_log_stats=self.disable_log_stats,
1118
1119
        )

1120
        # Reminder: Please update docs/features/compatibility_matrix.md
1121
        # If the feature combo become valid
1122
1123
1124
1125
        if self.num_scheduler_steps > 1:
            if speculative_config is not None:
                raise ValueError("Speculative decoding is not supported with "
                                 "multi-step (--num-scheduler-steps > 1)")
1126
1127
1128
            if self.enable_chunked_prefill and self.pipeline_parallel_size > 1:
                raise ValueError("Multi-Step Chunked-Prefill is not supported "
                                 "for pipeline-parallel-size > 1")
1129
1130
1131
1132
1133
1134
            from vllm.platforms import current_platform
            if current_platform.is_cpu():
                logger.warning("Multi-Step (--num-scheduler-steps > 1) is "
                               "currently not supported for CPUs and has been "
                               "disabled.")
                self.num_scheduler_steps = 1
1135
1136
1137
1138
1139
1140
1141
1142
1143

        # make sure num_lookahead_slots is set the higher value depending on
        # if we are using speculative decoding or multi-step
        num_lookahead_slots = max(self.num_lookahead_slots,
                                  self.num_scheduler_steps - 1)
        num_lookahead_slots = num_lookahead_slots \
            if speculative_config is None \
            else speculative_config.num_lookahead_slots

1144
        scheduler_config = SchedulerConfig(
1145
            runner_type=model_config.runner_type,
1146
1147
1148
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1149
            cuda_graph_sizes=self.cuda_graph_sizes,
1150
            num_lookahead_slots=num_lookahead_slots,
1151
1152
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1153
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1154
            is_multimodal_model=model_config.is_multimodal_model,
1155
            preemption_mode=self.preemption_mode,
1156
            num_scheduler_steps=self.num_scheduler_steps,
1157
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1158
1159
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1160
            policy=self.scheduling_policy,
1161
            scheduler_cls=self.scheduler_cls,
1162
1163
1164
1165
            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,
        )
1166

1167
        lora_config = LoRAConfig(
1168
            bias_enabled=self.enable_lora_bias,
1169
1170
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1171
            fully_sharded_loras=self.fully_sharded_loras,
1172
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1173
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1174
1175
1176
            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
1177

1178
1179
1180
1181
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1182
        load_config = self.create_load_config()
1183

1184
1185
1186
1187
1188
        prompt_adapter_config = PromptAdapterConfig(
            max_prompt_adapters=self.max_prompt_adapters,
            max_prompt_adapter_token=self.max_prompt_adapter_token) \
                                        if self.enable_prompt_adapter else None

1189
        decoding_config = DecodingConfig(
1190
1191
1192
1193
1194
            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,
1195
1196
            reasoning_backend=self.reasoning_parser
        )
1197

1198
        observability_config = ObservabilityConfig(
1199
1200
            show_hidden_metrics_for_version=self.
            show_hidden_metrics_for_version,
1201
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1202
            collect_detailed_traces=self.collect_detailed_traces,
1203
        )
1204

1205
        config = VllmConfig(
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
            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,
1216
            prompt_adapter_config=prompt_adapter_config,
1217
            compilation_config=self.compilation_config,
1218
            kv_transfer_config=self.kv_transfer_config,
1219
            kv_events_config=self.kv_events_config,
1220
            additional_config=self.additional_config,
1221
        )
1222

1223
1224
        return config

1225
1226
1227
1228
1229
1230
    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.

1231
        if self.load_format == LoadFormat.SHARDED_STATE.value:
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
            _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

1243
        if self.preemption_mode != SchedulerConfig.preemption_mode:
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
            _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

1254
        if self.scheduling_policy != SchedulerConfig.policy:
1255
1256
1257
1258
            _raise_or_fallback(feature_name="--scheduling-policy",
                               recommend_to_remove=False)
            return False

1259
        if self.num_scheduler_steps != SchedulerConfig.num_scheduler_steps:
1260
1261
1262
1263
            _raise_or_fallback(feature_name="--num-scheduler-steps",
                               recommend_to_remove=True)
            return False

1264
        if self.scheduler_delay_factor != SchedulerConfig.delay_factor:
1265
1266
1267
1268
            _raise_or_fallback(feature_name="--scheduler-delay-factor",
                               recommend_to_remove=True)
            return False

1269
1270
        if self.guided_decoding_backend not in get_args(
                GuidedDecodingBackendV1):
1271
1272
1273
1274
            _raise_or_fallback(
                feature_name=
                f"--guided-decoding-backend={self.guided_decoding_backend}",
                recommend_to_remove=False)
1275
1276
1277
            return False

        # Need at least Ampere for now (FA support required).
1278
1279
1280
        # Skip this check if we are running on a non-GPU platform,
        # or if the device capability is not available
        # (e.g. in a Ray actor without GPUs).
1281
1282
        from vllm.platforms import current_platform
        if (current_platform.is_cuda()
1283
                and current_platform.get_device_capability()
1284
1285
1286
1287
1288
1289
1290
                and current_platform.get_device_capability().major < 8):
            _raise_or_fallback(feature_name="Compute Capability < 8.0",
                               recommend_to_remove=False)
            return False

        # No Fp8 KV cache so far.
        if self.kv_cache_dtype != "auto":
1291
1292
1293
1294
1295
1296
            fp8_attention = self.kv_cache_dtype.startswith("fp8")
            will_use_fa = (
                current_platform.is_cuda()
                and not envs.is_set("VLLM_ATTENTION_BACKEND")
            ) or envs.VLLM_ATTENTION_BACKEND == "FLASH_ATTN_VLLM_V1"
            supported = False
1297
1298
1299
            if current_platform.is_rocm():
                supported = True
            elif fp8_attention and will_use_fa:
1300
                from vllm.attention.utils.fa_utils import (
1301
1302
1303
1304
1305
1306
                    flash_attn_supports_fp8)
                supported = flash_attn_supports_fp8()
            if not supported:
                _raise_or_fallback(feature_name="--kv-cache-dtype",
                                   recommend_to_remove=False)
                return False
1307
1308
1309
1310
1311
1312
1313

        # No Prompt Adapter so far.
        if self.enable_prompt_adapter:
            _raise_or_fallback(feature_name="--enable-prompt-adapter",
                               recommend_to_remove=False)
            return False

1314
1315
1316
1317
1318
1319
        # 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

1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
        # Only Fp16 and Bf16 dtypes since we only support FA.
        V1_SUPPORTED_DTYPES = [torch.bfloat16, torch.float16]
        if model_config.dtype not in V1_SUPPORTED_DTYPES:
            _raise_or_fallback(feature_name=f"--dtype {model_config.dtype}",
                               recommend_to_remove=False)
            return False

        # No Embedding Models so far.
        if model_config.task not in ["generate"]:
            _raise_or_fallback(feature_name=f"--task {model_config.task}",
                               recommend_to_remove=False)
            return False

        # No Mamba or Encoder-Decoder so far.
        if not model_config.is_v1_compatible:
            _raise_or_fallback(feature_name=model_config.architectures,
                               recommend_to_remove=False)
            return False

        # No Concurrent Partial Prefills so far.
        if (self.max_num_partial_prefills
1341
                != SchedulerConfig.max_num_partial_prefills
1342
                or self.max_long_partial_prefills
1343
                != SchedulerConfig.max_long_partial_prefills):
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
            _raise_or_fallback(feature_name="Concurrent Partial Prefill",
                               recommend_to_remove=False)
            return False

        # No OTLP observability so far.
        if (self.otlp_traces_endpoint or self.collect_detailed_traces):
            _raise_or_fallback(feature_name="--otlp-traces-endpoint",
                               recommend_to_remove=False)
            return False

1354
        # V1 supports N-gram, Medusa, and Eagle speculative decoding.
1355
        is_ngram_enabled = False
1356
        is_eagle_enabled = False
1357
        is_medusa_enabled = False
1358
        if self.speculative_config is not None:
1359
            # This is supported but experimental (handled below).
1360
1361
1362
1363
            speculative_method = self.speculative_config.get("method")
            if speculative_method:
                if speculative_method in ("ngram", "[ngram]"):
                    is_ngram_enabled = True
1364
1365
                elif speculative_method == "medusa":
                    is_medusa_enabled = True
Jiayi Yao's avatar
Jiayi Yao committed
1366
                elif speculative_method in ("eagle", "eagle3", "deepseek_mtp"):
1367
                    is_eagle_enabled = True
1368
            else:
1369
1370
1371
                speculative_model = self.speculative_config.get("model")
                if speculative_model in ("ngram", "[ngram]"):
                    is_ngram_enabled = True
1372
            if not (is_ngram_enabled or is_eagle_enabled or is_medusa_enabled):
1373
                # Other speculative decoding methods are not supported yet.
1374
1375
1376
1377
                _raise_or_fallback(feature_name="Speculative Decoding",
                                   recommend_to_remove=False)
                return False

1378
        # No XFormers so far.
1379
        V1_BACKENDS = [
1380
1381
1382
1383
1384
1385
1386
1387
1388
            "FLASH_ATTN_VLLM_V1",
            "FLASH_ATTN",
            "PALLAS",
            "PALLAS_VLLM_V1",
            "TRITON_ATTN_VLLM_V1",
            "TRITON_MLA",
            "FLASHMLA",
            "FLASHINFER",
            "FLASHINFER_VLLM_V1",
1389
            "ROCM_AITER_MLA",
1390
1391
1392
1393
1394
1395
1396
        ]
        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

1397
1398
        # Platforms must decide if they can support v1 for this model
        if not current_platform.supports_v1(model_config=model_config):
1399
1400
1401
1402
            _raise_or_fallback(
                feature_name=f"device type={current_platform.device_type}",
                recommend_to_remove=False)
            return False
1403
1404
1405
        #############################################################
        # Experimental Features - allow users to opt in.

1406
1407
1408
1409
1410
        # Signal Handlers requires running in main thread.
        if (threading.current_thread() != threading.main_thread()
                and _warn_or_fallback("Engine in background thread")):
            return False

1411
        if (self.pipeline_parallel_size > 1
1412
                and self.distributed_executor_backend
1413
1414
                not in (ParallelConfig.distributed_executor_backend, "ray",
                        "mp", "external_launcher")):
1415
            name = "Pipeline Parallelism without Ray distributed executor " \
1416
                    "or multiprocessing executor or external launcher"
1417
            _raise_or_fallback(feature_name=name, recommend_to_remove=False)
1418
1419
            return False

1420
1421
1422
1423
        # Non-[CUDA, TPU] may be supported on V1, but off by default for now.
        v0_hardware = not any(
            (current_platform.is_cuda(), current_platform.is_tpu()))
        if v0_hardware and _warn_or_fallback(  # noqa: SIM103
1424
                current_platform.device_name):
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
            return False
        #############################################################

        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:
                from vllm.platforms import current_platform
                is_gpu = current_platform.is_cuda()
                use_sliding_window = (model_config.get_sliding_window()
                                      is not None)
1447
                use_spec_decode = self.speculative_config is not None
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474

                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
                        and not self.enable_prompt_adapter
                        and model_config.runner_type != "pooling"):
                    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)
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
            msg = "Chunked prefill is not supported for pooling models"
            raise ValueError(msg)

1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
        # if using prefix caching, we must set a hash algo
        if self.enable_prefix_caching:
            # Disable prefix caching for multimodal models for VLLM_V0.
            if 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

            # VLLM_V0 only supports builtin hash algo for prefix caching.
1485
            if self.prefix_caching_hash_algo == "sha256":
1486
1487
1488
                raise ValueError(
                    "sha256 is not supported for prefix caching in V0 engine. "
                    "Please use 'builtin'.")
1489
1490
1491
1492
1493
1494
1495

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

    def _set_default_args_v1(self, usage_context: UsageContext) -> None:
        """Set Default Arguments for V1 Engine."""
1496

1497
1498
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
1499
1500
1501
1502
1503

        # V1 enables prefix caching by default.
        if self.enable_prefix_caching is None:
            self.enable_prefix_caching = True

1504
1505
1506
        # 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:
1507
            self.scheduler_cls = "vllm.v1.core.sched.scheduler.Scheduler"
1508

1509
1510
        # When no user override, set the default values based on the usage
        # context.
1511
        # Use different default values for different hardware.
1512
1513
1514
1515
1516
1517

        # 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.
1518
        from vllm.platforms import current_platform
1519
        try:
1520
            device_memory = current_platform.get_device_total_memory()
1521
            device_name = current_platform.get_device_name().lower()
1522
1523
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1524
            device_memory = 0
1525

1526
1527
1528
1529
        # NOTE(Kuntai): Setting large `max_num_batched_tokens` for A100 reduces
        # throughput, see PR #17885 for more details.
        # So here we do an extra device name check to prevent such regression.
        if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
1530
            # For GPUs like H100 and MI300x, use larger default values.
1531
1532
1533
1534
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1535
            default_max_num_seqs = 1024
1536
1537
1538
1539
1540
1541
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1542
            default_max_num_seqs = 256
1543

1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
        # 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,
                }
            }

1559
        use_context_value = usage_context.value if usage_context else None
1560
1561
        if (self.max_num_batched_tokens is None
                and usage_context in default_max_num_batched_tokens):
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
            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:
                self.max_num_batched_tokens = default_max_num_batched_tokens[
                    usage_context]
1575
            logger.debug(
1576
                "Setting max_num_batched_tokens to %d for %s usage context.",
1577
                self.max_num_batched_tokens, use_context_value)
1578

1579
1580
1581
1582
1583
        if self.max_num_seqs is None:
            self.max_num_seqs = default_max_num_seqs

            logger.debug("Setting max_num_seqs to %d for %s usage context.",
                         self.max_num_seqs, use_context_value)
1584

1585

1586
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1587
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1588
    """Arguments for asynchronous vLLM engine."""
1589
    disable_log_requests: bool = False
1590
1591

    @staticmethod
1592
1593
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1594
1595
1596
1597
        # Initialize plugin to update the parser, for example, The plugin may
        # adding a new kind of quantization method to --quantization argument or
        # a new device to --device argument.
        load_general_plugins()
1598
1599
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1600
1601
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1602
                            help='Disable logging requests.')
1603
1604
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update(parser)
1605
        return parser
1606
1607


1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
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


1635
1636
1637
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
1638

1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
    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)


1676
1677
# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1678
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1679
1680
1681


def _async_engine_args_parser():
1682
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1683
                                        async_args_only=True)