arg_utils.py 78.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

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
28
29
30
                         ConfigFormat, ConfigType, DecodingConfig,
                         DetailedTraceModules, Device, DeviceConfig,
                         DistributedExecutorBackend, GuidedDecodingBackend,
                         GuidedDecodingBackendV1, HfOverrides, KVEventsConfig,
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
                         ModelConfig, ModelDType, ModelImpl, MultiModalConfig,
31
32
33
34
35
                         ObservabilityConfig, ParallelConfig, PoolerConfig,
                         PrefixCachingHashAlgo, PromptAdapterConfig,
                         SchedulerConfig, SchedulerPolicy, SpeculativeConfig,
                         TaskOption, TokenizerMode, TokenizerPoolConfig,
                         VllmConfig, get_attr_docs, get_field)
36
from vllm.executor.executor_base import ExecutorBase
37
from vllm.logger import init_logger
38
from vllm.model_executor.layers.quantization import QuantizationMethods
39
from vllm.plugins import load_general_plugins
40
from vllm.reasoning import ReasoningParserManager
41
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
42
from vllm.transformers_utils.utils import check_gguf_file
43
from vllm.usage.usage_lib import UsageContext
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
47

# yapf: enable
48

49
50
logger = init_logger(__name__)

51
52
53
54
55
# object is used to allow for special typing forms
T = TypeVar("T")
TypeHint = Union[type[Any], object]
TypeHintT = Union[type[T], object]

56

57
def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]:
58

59
    def _parse_type(val: str) -> T:
60
61
62
63
64
65
66
        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
67

68
69
70
71
72
73
74
75
76
77
78
    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)

79
    return _optional_type
80
81


82
83
84
def union_dict_and_str(val: str) -> Optional[Union[str, dict[str, str]]]:
    if not re.match("^{.*}$", val):
        return str(val)
85
    return optional_type(json.loads)(val)
86
87


88
89
90
91
92
93
@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]
94
95
96
97
98
99
100
101
    pairs into a dictionary.

    Args:
        val: String value to be parsed.

    Returns:
        Dictionary with parsed values.
    """
102
    out_dict: dict[str, int] = {}
103
    for item in val.split(","):
104
105
106
107
108
        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
109
110

        try:
111
            parsed_value = int(value)
112
113
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
114
115
116
117
118
119
            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
120
121
122
123

    return out_dict


124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
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)


139
140
141
142
143
144
145
146
147
148
149
150
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)}


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


156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
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


173
174
@functools.lru_cache(maxsize=30)
def _compute_kwargs(cls: ConfigType) -> dict[str, Any]:
175
176
177
    cls_docs = get_attr_docs(cls)
    kwargs = {}
    for field in fields(cls):
178
        # Get the set of possible types for the field
179
        type_hints: set[TypeHint] = get_type_hints(field.type)
180
181
182
183
184

        # 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)

185
        # Get the default value of the field
186
187
188
        if field.default is not MISSING:
            default = field.default
        elif field.default_factory is not MISSING:
189
            default = field.default_factory()
190
191
192

        # Get the help text for the field
        name = field.name
193
        help = cls_docs[name].strip()
194
195
196
197
198
199
200
        # 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
201
202
203
204
205
        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"""
206
        if dataclass_cls is not None:
207
208
209
210
211
212
213
214
215
216

            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
217
218
            kwargs[name]["help"] += json_tip
        elif contains_type(type_hints, bool):
219
220
221
            # Creates --no-<name> and --<name> flags
            kwargs[name]["action"] = argparse.BooleanOptionalAction
        elif contains_type(type_hints, Literal):
222
            kwargs[name].update(literal_to_kwargs(type_hints))
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
        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
242
            # Special case for large integers
243
            if name in {"max_model_len", "max_num_batched_tokens"}:
244
                kwargs[name]["type"] = human_readable_int
245
246
        elif contains_type(type_hints, float):
            kwargs[name]["type"] = float
247
248
249
        elif (contains_type(type_hints, dict)
              and (contains_type(type_hints, str)
                   or any(is_not_builtin(th) for th in type_hints))):
250
            kwargs[name]["type"] = union_dict_and_str
251
        elif contains_type(type_hints, dict):
252
            kwargs[name]["type"] = parse_type(json.loads)
253
            kwargs[name]["help"] += json_tip
254
255
256
257
258
259
260
        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}.")

261
262
263
264
265
        # 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"]}))

266
267
268
269
270
271
272
        # 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
273
274


275
276
277
278
279
280
281
282
283
284
def get_kwargs(cls: ConfigType) -> dict[str, Any]:
    """Return argparse kwargs for the given Config dataclass.

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


285
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
286
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
287
    """Arguments for vLLM engine."""
288
289
290
291
292
293
294
    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
295
    enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds
296
297
298
    tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
    trust_remote_code: bool = ModelConfig.trust_remote_code
    allowed_local_media_path: str = ModelConfig.allowed_local_media_path
299
300
    download_dir: Optional[str] = LoadConfig.download_dir
    load_format: str = LoadConfig.load_format
301
302
    config_format: str = ModelConfig.config_format
    dtype: ModelDType = ModelConfig.dtype
303
    kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
304
305
    seed: Optional[int] = ModelConfig.seed
    max_model_len: Optional[int] = ModelConfig.max_model_len
306
307
    cuda_graph_sizes: list[int] = get_field(SchedulerConfig,
                                            "cuda_graph_sizes")
308
309
310
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
311
    distributed_executor_backend: Optional[Union[
312
313
        DistributedExecutorBackend,
        Type[ExecutorBase]]] = ParallelConfig.distributed_executor_backend
314
    # number of P/D disaggregation (or other disaggregation) workers
315
316
317
    pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
    data_parallel_size: int = ParallelConfig.data_parallel_size
318
319
320
    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
321
    data_parallel_backend: str = ParallelConfig.data_parallel_backend
322
323
324
    enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
    max_parallel_loading_workers: Optional[
        int] = ParallelConfig.max_parallel_loading_workers
325
326
327
328
    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
329
330
    disable_sliding_window: bool = ModelConfig.disable_sliding_window
    disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
331
    use_v2_block_manager: bool = True
332
333
334
    swap_space: float = CacheConfig.swap_space
    cpu_offload_gb: float = CacheConfig.cpu_offload_gb
    gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
335
336
337
338
339
340
341
    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
342
    max_logprobs: int = ModelConfig.max_logprobs
343
    disable_log_stats: bool = False
344
345
346
347
348
    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
349
    hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
350
351
352
353
    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
354
    disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
355
356
357
    # 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.
358
    tokenizer_pool_size: int = TokenizerPoolConfig.pool_size
359
360
    tokenizer_pool_type: str = TokenizerPoolConfig.pool_type
    tokenizer_pool_extra_config: dict = \
361
        get_field(TokenizerPoolConfig, "extra_config")
362
    limit_mm_per_prompt: dict[str, int] = \
363
        get_field(MultiModalConfig, "limit_per_prompt")
364
365
366
367
    mm_processor_kwargs: Optional[Dict[str, Any]] = \
        MultiModalConfig.mm_processor_kwargs
    disable_mm_preprocessor_cache: bool = \
        MultiModalConfig.disable_mm_preprocessor_cache
368
    # LoRA fields
369
    enable_lora: bool = False
370
371
372
373
374
375
376
377
378
379
    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
380
    enable_prompt_adapter: bool = False
381
382
383
384
    max_prompt_adapters: int = PromptAdapterConfig.max_prompt_adapters
    max_prompt_adapter_token: int = \
        PromptAdapterConfig.max_prompt_adapter_token

385
    device: Device = DeviceConfig.device
386
387
    num_scheduler_steps: int = SchedulerConfig.num_scheduler_steps
    multi_step_stream_outputs: bool = SchedulerConfig.multi_step_stream_outputs
388
    ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight
389
390
    num_gpu_blocks_override: Optional[
        int] = CacheConfig.num_gpu_blocks_override
391
    num_lookahead_slots: int = SchedulerConfig.num_lookahead_slots
392
393
    model_loader_extra_config: dict = \
        get_field(LoadConfig, "model_loader_extra_config")
394
395
    ignore_patterns: Optional[Union[str,
                                    List[str]]] = LoadConfig.ignore_patterns
396
    preemption_mode: Optional[str] = SchedulerConfig.preemption_mode
397

398
399
400
401
    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
402

403
404
405
    disable_hybrid_kv_cache_manager: bool = (
        SchedulerConfig.disable_hybrid_kv_cache_manager)

406
407
408
409
410
411
    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
412
413
    logits_processor_pattern: Optional[
        str] = ModelConfig.logits_processor_pattern
414

415
    speculative_config: Optional[Dict[str, Any]] = None
416

417
    qlora_adapter_name_or_path: Optional[str] = None
418
419
420
421
422
423
    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
424
    disable_async_output_proc: bool = not ModelConfig.use_async_output_proc
425
426
    scheduling_policy: SchedulerPolicy = SchedulerConfig.policy
    scheduler_cls: Union[str, Type[object]] = SchedulerConfig.scheduler_cls
427

428
429
430
431
    override_neuron_config: dict[str, Any] = \
        get_field(ModelConfig, "override_neuron_config")
    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
450
    additional_config: dict[str, Any] = \
        get_field(VllmConfig, "additional_config")
451
452
453
    enable_reasoning: Optional[bool] = None  # DEPRECATED
    reasoning_parser: str = DecodingConfig.reasoning_backend

454
    use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
455
    pt_load_map_location: str = LoadConfig.pt_load_map_location
456

457
458
459
    enable_multimodal_encoder_data_parallel: bool = \
        ParallelConfig.enable_multimodal_encoder_data_parallel

460
    def __post_init__(self):
461
462
463
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
464
        if isinstance(self.compilation_config, (int, dict)):
465
466
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
467
468
469
470
471
472
473
        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,
            )
474
        # Setup plugins
475
476
        from vllm.plugins import load_general_plugins
        load_general_plugins()
477
478

    @staticmethod
479
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
480
        """Shared CLI arguments for vLLM engine."""
481

482
        # Model arguments
483
484
485
486
487
        model_kwargs = get_kwargs(ModelConfig)
        model_group = parser.add_argument_group(
            title="ModelConfig",
            description=ModelConfig.__doc__,
        )
Reid's avatar
Reid committed
488
        if not ('serve' in sys.argv[1:] and '--help' in sys.argv[1:]):
489
            model_group.add_argument("--model", **model_kwargs["model"])
490
491
492
493
494
495
496
497
498
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
        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"])
526
527
        model_group.add_argument("--enable-prompt-embeds",
                                 **model_kwargs["enable_prompt_embeds"])
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
        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"])
566
567
        model_group.add_argument("--override-attention-dtype",
                                 **model_kwargs["override_attention_dtype"])
568

569
570
571
572
573
574
        # Model loading arguments
        load_kwargs = get_kwargs(LoadConfig)
        load_group = parser.add_argument_group(
            title="LoadConfig",
            description=LoadConfig.__doc__,
        )
575
        load_group.add_argument("--load-format",
576
577
                                choices=[f.value for f in LoadFormat],
                                **load_kwargs["load_format"])
578
        load_group.add_argument("--download-dir",
579
                                **load_kwargs["download_dir"])
580
        load_group.add_argument("--model-loader-extra-config",
581
                                **load_kwargs["model_loader_extra_config"])
582
583
584
        load_group.add_argument("--ignore-patterns",
                                **load_kwargs["ignore_patterns"])
        load_group.add_argument("--use-tqdm-on-load",
585
                                **load_kwargs["use_tqdm_on_load"])
586
587
588
589
590
591
592
593
        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,
        )
594
595
        load_group.add_argument('--pt-load-map-location',
                                **load_kwargs["pt_load_map_location"])
596

597
598
599
600
601
602
        # Guided decoding arguments
        guided_decoding_kwargs = get_kwargs(DecodingConfig)
        guided_decoding_group = parser.add_argument_group(
            title="DecodingConfig",
            description=DecodingConfig.__doc__,
        )
603
604
        guided_decoding_group.add_argument("--guided-decoding-backend",
                                           **guided_decoding_kwargs["backend"])
605
        guided_decoding_group.add_argument(
606
607
608
609
610
611
612
613
            "--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"])
614
615
616
        guided_decoding_group.add_argument(
            "--enable-reasoning",
            action=argparse.BooleanOptionalAction,
617
            deprecated=True,
618
            help="[DEPRECATED] The `--enable-reasoning` flag is deprecated as "
619
            "of v0.9.0. Use `--reasoning-parser` to specify the reasoning "
620
            "parser backend instead. This flag (`--enable-reasoning`) will be "
621
622
            "removed in v0.10.0. When `--reasoning-parser` is specified, "
            "reasoning mode is automatically enabled.")
623
624
625
626
627
628
        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"])

629
        # Parallel arguments
630
631
632
633
634
635
        parallel_kwargs = get_kwargs(ParallelConfig)
        parallel_group = parser.add_argument_group(
            title="ParallelConfig",
            description=ParallelConfig.__doc__,
        )
        parallel_group.add_argument(
636
            "--distributed-executor-backend",
637
638
            **parallel_kwargs["distributed_executor_backend"])
        parallel_group.add_argument(
639
            "--pipeline-parallel-size", "-pp",
640
            **parallel_kwargs["pipeline_parallel_size"])
641
        parallel_group.add_argument("--tensor-parallel-size", "-tp",
642
                                    **parallel_kwargs["tensor_parallel_size"])
643
        parallel_group.add_argument("--data-parallel-size", "-dp",
644
                                    **parallel_kwargs["data_parallel_size"])
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
        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
660
661
662
663
664
665
        parallel_group.add_argument('--data-parallel-backend',
                                    '-dpb',
                                    type=str,
                                    default='mp',
                                    help='Backend for data parallel, either '
                                    '"mp" or "ray".')
666
        parallel_group.add_argument(
667
            "--enable-expert-parallel",
668
669
            **parallel_kwargs["enable_expert_parallel"])
        parallel_group.add_argument(
670
            "--max-parallel-loading-workers",
671
672
            **parallel_kwargs["max_parallel_loading_workers"])
        parallel_group.add_argument(
673
            "--ray-workers-use-nsight",
674
675
            **parallel_kwargs["ray_workers_use_nsight"])
        parallel_group.add_argument(
676
            "--disable-custom-all-reduce",
677
            **parallel_kwargs["disable_custom_all_reduce"])
678
679
680
681
        parallel_group.add_argument("--worker-cls",
                                    **parallel_kwargs["worker_cls"])
        parallel_group.add_argument("--worker-extension-cls",
                                    **parallel_kwargs["worker_extension_cls"])
682
683
684
        parallel_group.add_argument(
            "--enable-multimodal-encoder-data-parallel",
            **parallel_kwargs["enable_multimodal_encoder_data_parallel"])
685

686
687
688
689
690
        # KV cache arguments
        cache_kwargs = get_kwargs(CacheConfig)
        cache_group = parser.add_argument_group(
            title="CacheConfig",
            description=CacheConfig.__doc__,
691
        )
692
693
        cache_group.add_argument("--block-size", **cache_kwargs["block_size"])
        cache_group.add_argument("--gpu-memory-utilization",
694
                                 **cache_kwargs["gpu_memory_utilization"])
695
696
        cache_group.add_argument("--swap-space", **cache_kwargs["swap_space"])
        cache_group.add_argument("--kv-cache-dtype",
697
                                 **cache_kwargs["cache_dtype"])
698
        cache_group.add_argument("--num-gpu-blocks-override",
699
700
701
702
703
                                 **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"])
704
        cache_group.add_argument("--cpu-offload-gb",
705
                                 **cache_kwargs["cpu_offload_gb"])
706
        cache_group.add_argument("--calculate-kv-scales",
707
708
                                 **cache_kwargs["calculate_kv_scales"])

709
710
711
712
713
714
        # Tokenizer arguments
        tokenizer_kwargs = get_kwargs(TokenizerPoolConfig)
        tokenizer_group = parser.add_argument_group(
            title="TokenizerPoolConfig",
            description=TokenizerPoolConfig.__doc__,
        )
715
        tokenizer_group.add_argument("--tokenizer-pool-size",
716
                                     **tokenizer_kwargs["pool_size"])
717
        tokenizer_group.add_argument("--tokenizer-pool-type",
718
                                     **tokenizer_kwargs["pool_type"])
719
        tokenizer_group.add_argument("--tokenizer-pool-extra-config",
720
                                     **tokenizer_kwargs["extra_config"])
721
722

        # Multimodal related configs
723
724
725
726
727
        multimodal_kwargs = get_kwargs(MultiModalConfig)
        multimodal_group = parser.add_argument_group(
            title="MultiModalConfig",
            description=MultiModalConfig.__doc__,
        )
728
        multimodal_group.add_argument("--limit-mm-per-prompt",
729
                                      **multimodal_kwargs["limit_per_prompt"])
730
        multimodal_group.add_argument(
731
            "--mm-processor-kwargs",
732
733
            **multimodal_kwargs["mm_processor_kwargs"])
        multimodal_group.add_argument(
734
            "--disable-mm-preprocessor-cache",
735
            **multimodal_kwargs["disable_mm_preprocessor_cache"])
736

737
        # LoRA related configs
738
739
740
741
742
743
        lora_kwargs = get_kwargs(LoRAConfig)
        lora_group = parser.add_argument_group(
            title="LoRAConfig",
            description=LoRAConfig.__doc__,
        )
        lora_group.add_argument(
744
            "--enable-lora",
745
            action=argparse.BooleanOptionalAction,
746
747
            help="If True, enable handling of LoRA adapters.")
        lora_group.add_argument("--enable-lora-bias",
748
                                **lora_kwargs["bias_enabled"])
749
750
        lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"])
        lora_group.add_argument("--max-lora-rank",
751
                                **lora_kwargs["max_lora_rank"])
752
        lora_group.add_argument("--lora-extra-vocab-size",
753
754
                                **lora_kwargs["lora_extra_vocab_size"])
        lora_group.add_argument(
755
            "--lora-dtype",
756
757
            **lora_kwargs["lora_dtype"],
        )
758
        lora_group.add_argument("--long-lora-scaling-factors",
759
                                **lora_kwargs["long_lora_scaling_factors"])
760
        lora_group.add_argument("--max-cpu-loras",
761
                                **lora_kwargs["max_cpu_loras"])
762
        lora_group.add_argument("--fully-sharded-loras",
763
764
765
766
767
768
769
770
771
                                **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(
772
            "--enable-prompt-adapter",
773
            action=argparse.BooleanOptionalAction,
774
            help="If True, enable handling of PromptAdapters.")
775
        prompt_adapter_group.add_argument(
776
            "--max-prompt-adapters",
777
778
            **prompt_adapter_kwargs["max_prompt_adapters"])
        prompt_adapter_group.add_argument(
779
            "--max-prompt-adapter-token",
780
            **prompt_adapter_kwargs["max_prompt_adapter_token"])
781
782
783
784
785
786
787

        # Device arguments
        device_kwargs = get_kwargs(DeviceConfig)
        device_group = parser.add_argument_group(
            title="DeviceConfig",
            description=DeviceConfig.__doc__,
        )
788
789
790
        device_group.add_argument("--device",
                                  **device_kwargs["device"],
                                  deprecated=True)
791

792
793
794
795
796
797
        # Speculative arguments
        speculative_group = parser.add_argument_group(
            title="SpeculativeConfig",
            description=SpeculativeConfig.__doc__,
        )
        speculative_group.add_argument(
798
            "--speculative-config",
799
800
            type=json.loads,
            default=None,
801
802
            help="The configurations for speculative decoding. Should be a "
            "JSON string.")
803

804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
        # 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"])
827

828
829
830
831
832
833
834
        # Scheduler arguments
        scheduler_kwargs = get_kwargs(SchedulerConfig)
        scheduler_group = parser.add_argument_group(
            title="SchedulerConfig",
            description=SchedulerConfig.__doc__,
        )
        scheduler_group.add_argument(
835
            "--max-num-batched-tokens",
836
            **scheduler_kwargs["max_num_batched_tokens"])
837
        scheduler_group.add_argument("--max-num-seqs",
838
839
840
841
842
843
844
                                     **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"])
845
846
        scheduler_group.add_argument('--cuda-graph-sizes',
                                     **scheduler_kwargs["cuda_graph_sizes"])
847
848
849
        scheduler_group.add_argument(
            "--long-prefill-token-threshold",
            **scheduler_kwargs["long_prefill_token_threshold"])
850
        scheduler_group.add_argument("--num-lookahead-slots",
851
                                     **scheduler_kwargs["num_lookahead_slots"])
852
        scheduler_group.add_argument("--scheduler-delay-factor",
853
                                     **scheduler_kwargs["delay_factor"])
854
        scheduler_group.add_argument("--preemption-mode",
855
                                     **scheduler_kwargs["preemption_mode"])
856
        scheduler_group.add_argument("--num-scheduler-steps",
857
                                     **scheduler_kwargs["num_scheduler_steps"])
858
        scheduler_group.add_argument(
859
            "--multi-step-stream-outputs",
860
            **scheduler_kwargs["multi_step_stream_outputs"])
861
        scheduler_group.add_argument("--scheduling-policy",
862
                                     **scheduler_kwargs["policy"])
863
        scheduler_group.add_argument(
864
            "--enable-chunked-prefill",
865
            **scheduler_kwargs["enable_chunked_prefill"])
866
867
868
        scheduler_group.add_argument(
            "--disable-chunked-mm-input",
            **scheduler_kwargs["disable_chunked_mm_input"])
869
870
        scheduler_group.add_argument("--scheduler-cls",
                                     **scheduler_kwargs["scheduler_cls"])
871
872
873
        scheduler_group.add_argument(
            "--disable-hybrid-kv-cache-manager",
            **scheduler_kwargs["disable_hybrid_kv_cache_manager"])
874
875

        # vLLM arguments
876
        vllm_kwargs = get_kwargs(VllmConfig)
877
878
879
880
        vllm_group = parser.add_argument_group(
            title="VllmConfig",
            description=VllmConfig.__doc__,
        )
881
882
883
884
885
886
887
888
        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"])
889

890
891
892
893
        # Other arguments
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
                            default=True,
894
                            deprecated=True,
895
896
897
898
899
900
901
902
                            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.')
903

904
        return parser
905
906

    @classmethod
907
    def from_cli_args(cls, args: argparse.Namespace):
908
909
910
        # 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
911
912
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
913

914
    def create_model_config(self) -> ModelConfig:
915
916
917
918
919
920
921
922
923
924
925
        # 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

926
        return ModelConfig(
927
            model=self.model,
928
            hf_config_path=self.hf_config_path,
929
            task=self.task,
930
            tokenizer=self.tokenizer,
931
932
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
933
            allowed_local_media_path=self.allowed_local_media_path,
934
935
936
937
938
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
939
            rope_theta=self.rope_theta,
940
            hf_token=self.hf_token,
941
            hf_overrides=self.hf_overrides,
942
943
944
945
946
947
948
            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,
949
            disable_cascade_attn=self.disable_cascade_attn,
950
            skip_tokenizer_init=self.skip_tokenizer_init,
951
            enable_prompt_embeds=self.enable_prompt_embeds,
952
            served_model_name=self.served_model_name,
953
            limit_mm_per_prompt=self.limit_mm_per_prompt,
954
            use_async_output_proc=not self.disable_async_output_proc,
955
            config_format=self.config_format,
956
            mm_processor_kwargs=self.mm_processor_kwargs,
957
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
958
959
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
960
            logits_processor_pattern=self.logits_processor_pattern,
961
            generation_config=self.generation_config,
962
            override_generation_config=self.override_generation_config,
963
            enable_sleep_mode=self.enable_sleep_mode,
964
            model_impl=self.model_impl,
965
            override_attention_dtype=self.override_attention_dtype,
966
        )
967

968
969
    def create_load_config(self) -> LoadConfig:

970
971
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
972

973
974
975
976
977
        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,
978
            use_tqdm_on_load=self.use_tqdm_on_load,
979
            pt_load_map_location=self.pt_load_map_location,
980
        )
981

982
983
984
985
986
987
988
989
990
991
992
993
994
    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
995
        dictionary from the engine.
996
997
        """
        if self.speculative_config is None:
998
999
            return None

1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
        # 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

1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
    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
1024

1025
1026
1027
1028
1029
1030
        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.
        """
1031
1032
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update()
1033

1034
1035
        device_config = DeviceConfig(
            device=cast(Device, current_platform.device_type))
1036
1037
        model_config = self.create_model_config()

1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
        # * 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:
1057
            self._set_default_args_v1(usage_context, model_config)
1058
1059
        else:
            self._set_default_args_v0(model_config)
1060

1061
1062
        assert self.enable_chunked_prefill is not None

1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
        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'")

1074
        cache_config = CacheConfig(
1075
            block_size=self.block_size,
1076
1077
1078
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1079
            is_attention_free=model_config.is_attention_free,
1080
1081
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1082
            enable_prefix_caching=self.enable_prefix_caching,
1083
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1084
            cpu_offload_gb=self.cpu_offload_gb,
1085
            calculate_kv_scales=self.calculate_kv_scales,
1086
        )
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098

        # 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()

1099
1100
1101
1102
1103
1104
1105
        # 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
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
        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
1120
1121
1122
1123
1124
1125
1126

        # 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
1127
1128
        data_parallel_backend = self.data_parallel_backend

1129
        parallel_config = ParallelConfig(
1130
1131
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1132
            data_parallel_size=self.data_parallel_size,
1133
1134
1135
            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
1136
            data_parallel_backend=data_parallel_backend,
1137
            enable_expert_parallel=self.enable_expert_parallel,
1138
1139
1140
            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,
1141
            placement_group=placement_group,
1142
1143
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1144
            worker_extension_cls=self.worker_extension_cls,
1145
1146
            enable_multimodal_encoder_data_parallel=self.
            enable_multimodal_encoder_data_parallel,
1147
        )
1148

1149
        speculative_config = self.create_speculative_config(
1150
1151
            target_model_config=model_config,
            target_parallel_config=parallel_config,
1152
            enable_chunked_prefill=self.enable_chunked_prefill,
1153
            disable_log_stats=self.disable_log_stats,
1154
1155
        )

1156
        # Reminder: Please update docs/features/compatibility_matrix.md
1157
        # If the feature combo become valid
1158
1159
1160
1161
        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)")
1162
1163
1164
            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")
1165
1166
1167
1168
1169
1170
            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
1171
1172
1173
1174
1175
1176
1177
1178
1179

        # 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

1180
        scheduler_config = SchedulerConfig(
1181
            runner_type=model_config.runner_type,
1182
1183
1184
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1185
            cuda_graph_sizes=self.cuda_graph_sizes,
1186
            num_lookahead_slots=num_lookahead_slots,
1187
1188
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1189
            disable_chunked_mm_input=self.disable_chunked_mm_input,
1190
            is_multimodal_model=model_config.is_multimodal_model,
1191
            preemption_mode=self.preemption_mode,
1192
            num_scheduler_steps=self.num_scheduler_steps,
1193
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1194
1195
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1196
            policy=self.scheduling_policy,
1197
            scheduler_cls=self.scheduler_cls,
1198
1199
1200
            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,
1201
1202
            disable_hybrid_kv_cache_manager=self.
            disable_hybrid_kv_cache_manager,
1203
        )
1204

1205
        lora_config = LoRAConfig(
1206
            bias_enabled=self.enable_lora_bias,
1207
1208
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1209
            fully_sharded_loras=self.fully_sharded_loras,
1210
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1211
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1212
1213
1214
            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
1215

1216
1217
1218
1219
        # bitsandbytes pre-quantized model need a specific model loader
        if model_config.quantization == "bitsandbytes":
            self.quantization = self.load_format = "bitsandbytes"

1220
        load_config = self.create_load_config()
1221

1222
1223
1224
1225
1226
        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

1227
        decoding_config = DecodingConfig(
1228
1229
1230
1231
1232
            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,
1233
1234
            reasoning_backend=self.reasoning_parser
        )
1235

1236
        observability_config = ObservabilityConfig(
1237
1238
            show_hidden_metrics_for_version=self.
            show_hidden_metrics_for_version,
1239
            otlp_traces_endpoint=self.otlp_traces_endpoint,
1240
            collect_detailed_traces=self.collect_detailed_traces,
1241
        )
1242

1243
        config = VllmConfig(
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
            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,
1254
            prompt_adapter_config=prompt_adapter_config,
1255
            compilation_config=self.compilation_config,
1256
            kv_transfer_config=self.kv_transfer_config,
1257
            kv_events_config=self.kv_events_config,
1258
            additional_config=self.additional_config,
1259
        )
1260

1261
1262
        return config

1263
1264
1265
1266
1267
1268
    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.

1269
        if self.load_format == LoadFormat.SHARDED_STATE.value:
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
            _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

1281
        if self.preemption_mode != SchedulerConfig.preemption_mode:
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
            _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

1292
        if self.num_scheduler_steps != SchedulerConfig.num_scheduler_steps:
1293
1294
1295
1296
            _raise_or_fallback(feature_name="--num-scheduler-steps",
                               recommend_to_remove=True)
            return False

1297
        if self.scheduler_delay_factor != SchedulerConfig.delay_factor:
1298
1299
1300
1301
            _raise_or_fallback(feature_name="--scheduler-delay-factor",
                               recommend_to_remove=True)
            return False

1302
1303
        if self.guided_decoding_backend not in get_args(
                GuidedDecodingBackendV1):
1304
1305
1306
1307
            _raise_or_fallback(
                feature_name=
                f"--guided-decoding-backend={self.guided_decoding_backend}",
                recommend_to_remove=False)
1308
1309
1310
            return False

        # Need at least Ampere for now (FA support required).
1311
1312
1313
        # 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).
1314
        from vllm.platforms import current_platform
1315
        if (current_platform.is_cuda()
1316
                and current_platform.get_device_capability()
1317
1318
1319
1320
1321
1322
1323
                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":
1324
1325
1326
1327
1328
1329
            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
1330
1331
1332
            if current_platform.is_rocm():
                supported = True
            elif fp8_attention and will_use_fa:
1333
                from vllm.attention.utils.fa_utils import (
1334
1335
1336
1337
1338
1339
                    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
1340
1341
1342
1343
1344
1345
1346

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

1347
1348
1349
1350
1351
1352
        # 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

1353
1354
1355
1356
1357
1358
1359
        # 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

1360
        # No Mamba or Encoder-Decoder so far.
1361
1362
1363
1364
1365
        if not model_config.is_v1_compatible:
            _raise_or_fallback(feature_name=model_config.architectures,
                               recommend_to_remove=False)
            return False

Chen Zhang's avatar
Chen Zhang committed
1366
1367
1368
1369
1370
        # V1 mamba models are unoptimized.
        if model_config.has_inner_state and _warn_or_fallback(
                feature_name="Mamba"):
            return False

1371
1372
        # No Concurrent Partial Prefills so far.
        if (self.max_num_partial_prefills
1373
                != SchedulerConfig.max_num_partial_prefills
1374
                or self.max_long_partial_prefills
1375
                != SchedulerConfig.max_long_partial_prefills):
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
            _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

1386
        # V1 supports N-gram, Medusa, and Eagle speculative decoding.
1387
        is_ngram_enabled = False
1388
        is_eagle_enabled = False
1389
        is_medusa_enabled = False
1390
        if self.speculative_config is not None:
1391
            # This is supported but experimental (handled below).
1392
1393
1394
1395
            speculative_method = self.speculative_config.get("method")
            if speculative_method:
                if speculative_method in ("ngram", "[ngram]"):
                    is_ngram_enabled = True
1396
1397
                elif speculative_method == "medusa":
                    is_medusa_enabled = True
Jiayi Yao's avatar
Jiayi Yao committed
1398
                elif speculative_method in ("eagle", "eagle3", "deepseek_mtp"):
1399
                    is_eagle_enabled = True
1400
            else:
1401
1402
1403
                speculative_model = self.speculative_config.get("model")
                if speculative_model in ("ngram", "[ngram]"):
                    is_ngram_enabled = True
1404
            if not (is_ngram_enabled or is_eagle_enabled or is_medusa_enabled):
1405
                # Other speculative decoding methods are not supported yet.
1406
1407
1408
1409
                _raise_or_fallback(feature_name="Speculative Decoding",
                                   recommend_to_remove=False)
                return False

1410
        # No XFormers so far.
1411
        V1_BACKENDS = [
1412
1413
1414
1415
1416
1417
            "FLASH_ATTN_VLLM_V1",
            "FLASH_ATTN",
            "PALLAS",
            "PALLAS_VLLM_V1",
            "TRITON_ATTN_VLLM_V1",
            "TRITON_MLA",
1418
            "CUTLASS_MLA_VLLM_V1",
1419
1420
1421
            "FLASHMLA",
            "FLASHINFER",
            "FLASHINFER_VLLM_V1",
1422
            "ROCM_AITER_MLA",
1423
            "TORCH_SDPA_VLLM_V1",
1424
            "FLEX_ATTENTION",
1425
1426
1427
1428
1429
1430
1431
        ]
        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

1432
1433
        # Platforms must decide if they can support v1 for this model
        if not current_platform.supports_v1(model_config=model_config):
1434
1435
1436
1437
            _raise_or_fallback(
                feature_name=f"device type={current_platform.device_type}",
                recommend_to_remove=False)
            return False
1438
1439
1440
        #############################################################
        # Experimental Features - allow users to opt in.

1441
1442
1443
1444
1445
        # Signal Handlers requires running in main thread.
        if (threading.current_thread() != threading.main_thread()
                and _warn_or_fallback("Engine in background thread")):
            return False

1446
        if (self.pipeline_parallel_size > 1
1447
                and self.distributed_executor_backend
1448
1449
                not in (ParallelConfig.distributed_executor_backend, "ray",
                        "mp", "external_launcher")):
1450
            name = "Pipeline Parallelism without Ray distributed executor " \
1451
                    "or multiprocessing executor or external launcher"
1452
            _raise_or_fallback(feature_name=name, recommend_to_remove=False)
1453
1454
            return False

1455
1456
1457
1458
        # 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):
1459
            return False
1460
1461
1462
1463
1464
1465
1466

        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

1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
        #############################################################

        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)
1488
                use_spec_decode = self.speculative_config is not None
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515

                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)

1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
        # 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.
1526
            if self.prefix_caching_hash_algo == "sha256":
1527
1528
1529
                raise ValueError(
                    "sha256 is not supported for prefix caching in V0 engine. "
                    "Please use 'builtin'.")
1530
1531
1532
1533
1534

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

1535
1536
    def _set_default_args_v1(self, usage_context: UsageContext,
                             model_config: ModelConfig) -> None:
1537
        """Set Default Arguments for V1 Engine."""
1538

1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
        # 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

            # TODO: when encoder models are supported we'll have to
            # check for causal attention here.
            incremental_prefill_supported = (pooling_type is not None and
                                             pooling_type.lower() == "last")
1554

1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
            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)

        if not self.enable_chunked_prefill:
            self.max_num_batched_tokens = model_config.max_model_len
1567

1568
1569
1570
        # 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:
1571
            self.scheduler_cls = "vllm.v1.core.sched.scheduler.Scheduler"
1572

1573
1574
        # When no user override, set the default values based on the usage
        # context.
1575
        # Use different default values for different hardware.
1576
1577
1578
1579
1580
1581

        # 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.
1582
        from vllm.platforms import current_platform
1583
        try:
1584
            device_memory = current_platform.get_device_total_memory()
1585
            device_name = current_platform.get_device_name().lower()
1586
1587
        except Exception:
            # This is only used to set default_max_num_batched_tokens
1588
            device_memory = 0
1589

1590
1591
1592
1593
        # 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:
1594
            # For GPUs like H100 and MI300x, use larger default values.
1595
1596
1597
1598
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
1599
1600
1601
1602
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 1024,
                UsageContext.OPENAI_API_SERVER: 1024,
            }
1603
1604
1605
1606
1607
1608
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
1609
1610
1611
1612
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 256,
                UsageContext.OPENAI_API_SERVER: 256,
            }
1613

1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
        # 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,
                }
            }

1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
        # cpu specific default values.
        if current_platform.is_cpu():
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 4096,
                UsageContext.OPENAI_API_SERVER: 2048,
            }
            default_max_num_seqs = {
                UsageContext.LLM_CLASS: 128,
                UsageContext.OPENAI_API_SERVER: 32,
            }

1640
        use_context_value = usage_context.value if usage_context else None
1641
1642
        if (self.max_num_batched_tokens is None
                and usage_context in default_max_num_batched_tokens):
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
            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]
1656
            logger.debug(
1657
                "Setting max_num_batched_tokens to %d for %s usage context.",
1658
                self.max_num_batched_tokens, use_context_value)
1659

1660
1661
1662
        if (self.max_num_seqs is None
                and usage_context in default_max_num_seqs):
            self.max_num_seqs = default_max_num_seqs[usage_context]
1663
1664
1665

            logger.debug("Setting max_num_seqs to %d for %s usage context.",
                         self.max_num_seqs, use_context_value)
1666

1667

1668
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1669
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1670
    """Arguments for asynchronous vLLM engine."""
1671
    disable_log_requests: bool = False
1672
1673

    @staticmethod
1674
1675
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1676
1677
1678
1679
        # 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()
1680
1681
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1682
1683
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1684
                            help='Disable logging requests.')
1685
1686
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update(parser)
1687
        return parser
1688
1689


1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
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


1717
1718
1719
def human_readable_int(value):
    """Parse human-readable integers like '1k', '2M', etc.
    Including decimal values with decimal multipliers.
1720

1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
    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)


1758
1759
# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1760
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1761
1762
1763


def _async_engine_args_parser():
1764
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1765
                                        async_args_only=True)