arg_utils.py 68.4 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import argparse
4
import dataclasses
5
import json
6
from dataclasses import dataclass
7
from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
8
                    Tuple, Type, Union, cast, get_args)
9

10
11
import torch

12
import vllm.envs as envs
13
from vllm import version
14
from vllm.config import (CacheConfig, CompilationConfig, ConfigFormat,
15
16
                         DecodingConfig, DeviceConfig, HfOverrides,
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
17
18
19
20
                         ModelConfig, ModelImpl, ObservabilityConfig,
                         ParallelConfig, PoolerConfig, PromptAdapterConfig,
                         SchedulerConfig, SpeculativeConfig, TaskOption,
                         TokenizerPoolConfig, VllmConfig)
21
from vllm.executor.executor_base import ExecutorBase
22
from vllm.logger import init_logger
23
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
24
from vllm.plugins import load_general_plugins
25
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
26
from vllm.transformers_utils.utils import check_gguf_file
27
from vllm.usage.usage_lib import UsageContext
28
from vllm.utils import FlexibleArgumentParser, StoreBoolean
29

30
if TYPE_CHECKING:
31
    from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
32

33
34
logger = init_logger(__name__)

35
36
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]

37
38
39
40
41
42
43
44
DEVICE_OPTIONS = [
    "auto",
    "cuda",
    "neuron",
    "cpu",
    "openvino",
    "tpu",
    "xpu",
45
    "hpu",
46
47
]

48

49
50
51
52
53
54
def nullable_str(val: str):
    if not val or val == "None":
        return None
    return val


55
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
56
57
58
59
60
61
62
63
64
    """Parses a string containing comma separate key [str] to value [int]
    pairs into a dictionary.

    Args:
        val: String value to be parsed.

    Returns:
        Dictionary with parsed values.
    """
65
66
67
68
69
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
70
71
72
73
74
        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
75
76

        try:
77
            parsed_value = int(value)
78
79
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
80
81
82
83
84
85
            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
86
87
88
89

    return out_dict


90
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
91
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
92
    """Arguments for vLLM engine."""
93
    model: str = 'facebook/opt-125m'
94
    served_model_name: Optional[Union[str, List[str]]] = None
95
    tokenizer: Optional[str] = None
96
    hf_config_path: Optional[str] = None
97
    task: TaskOption = "auto"
98
    skip_tokenizer_init: bool = False
99
    tokenizer_mode: str = 'auto'
100
    trust_remote_code: bool = False
101
    allowed_local_media_path: str = ""
102
    download_dir: Optional[str] = None
103
    load_format: str = 'auto'
104
    config_format: ConfigFormat = ConfigFormat.AUTO
105
    dtype: str = 'auto'
106
    kv_cache_dtype: str = 'auto'
107
    seed: int = 0
108
    max_model_len: Optional[int] = None
109
110
111
112
113
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
    distributed_executor_backend: Optional[Union[str,
                                                 Type[ExecutorBase]]] = None
114
    # number of P/D disaggregation (or other disaggregation) workers
115
116
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
117
    max_parallel_loading_workers: Optional[int] = None
118
    block_size: Optional[int] = None
119
    enable_prefix_caching: Optional[bool] = None
120
    disable_sliding_window: bool = False
121
    use_v2_block_manager: bool = True
122
123
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
124
    gpu_memory_utilization: float = 0.90
125
    max_num_batched_tokens: Optional[int] = None
126
127
128
    max_num_partial_prefills: Optional[int] = 1
    max_long_partial_prefills: Optional[int] = 1
    long_prefill_token_threshold: Optional[int] = 0
129
    max_num_seqs: Optional[int] = None
130
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
131
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
132
    revision: Optional[str] = None
133
    code_revision: Optional[str] = None
134
    rope_scaling: Optional[Dict[str, Any]] = None
135
    rope_theta: Optional[float] = None
136
    hf_overrides: Optional[HfOverrides] = None
137
    tokenizer_revision: Optional[str] = None
138
    quantization: Optional[str] = None
139
    enforce_eager: Optional[bool] = None
140
    max_seq_len_to_capture: int = 8192
141
    disable_custom_all_reduce: bool = False
142
    tokenizer_pool_size: int = 0
143
144
145
146
    # Note: Specifying a tokenizer pool by passing a class
    # is intended for expert use only. The API may change without
    # notice.
    tokenizer_pool_type: Union[str, Type["BaseTokenizerGroup"]] = "ray"
147
    tokenizer_pool_extra_config: Optional[Dict[str, Any]] = None
148
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
149
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
150
    disable_mm_preprocessor_cache: bool = False
151
    enable_lora: bool = False
152
    enable_lora_bias: bool = False
153
154
    max_loras: int = 1
    max_lora_rank: int = 16
155
156
157
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
158
    fully_sharded_loras: bool = False
159
    lora_extra_vocab_size: int = 256
160
    long_lora_scaling_factors: Optional[Tuple[float]] = None
161
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
162
    max_cpu_loras: Optional[int] = None
163
    device: str = 'auto'
164
    num_scheduler_steps: int = 1
165
    multi_step_stream_outputs: bool = True
166
    ray_workers_use_nsight: bool = False
167
    num_gpu_blocks_override: Optional[int] = None
168
    num_lookahead_slots: int = 0
169
    model_loader_extra_config: Optional[dict] = None
170
    ignore_patterns: Optional[Union[str, List[str]]] = None
171
    preemption_mode: Optional[str] = None
172

173
    scheduler_delay_factor: float = 0.0
174
    enable_chunked_prefill: Optional[bool] = None
175

176
    guided_decoding_backend: str = 'xgrammar'
177
    logits_processor_pattern: Optional[str] = None
178
179
    # Speculative decoding configuration.
    speculative_model: Optional[str] = None
180
    speculative_model_quantization: Optional[str] = None
181
    speculative_draft_tensor_parallel_size: Optional[int] = None
182
    num_speculative_tokens: Optional[int] = None
183
    speculative_disable_mqa_scorer: Optional[bool] = False
184
    speculative_max_model_len: Optional[int] = None
185
    speculative_disable_by_batch_size: Optional[int] = None
186
187
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
188
189
190
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
191
    qlora_adapter_name_or_path: Optional[str] = None
192
    disable_logprobs_during_spec_decoding: Optional[bool] = None
193

194
    show_hidden_metrics_for_version: Optional[str] = None
195
    otlp_traces_endpoint: Optional[str] = None
196
    collect_detailed_traces: Optional[str] = None
197
    disable_async_output_proc: bool = False
198
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
199
    scheduler_cls: Union[str, Type[object]] = "vllm.core.scheduler.Scheduler"
200

201
202
    override_neuron_config: Optional[Dict[str, Any]] = None
    override_pooler_config: Optional[PoolerConfig] = None
203
    compilation_config: Optional[CompilationConfig] = None
204
    worker_cls: str = "auto"
205
    worker_extension_cls: str = ""
206

207
208
    kv_transfer_config: Optional[KVTransferConfig] = None

209
    generation_config: Optional[str] = None
210
    override_generation_config: Optional[Dict[str, Any]] = None
211
    enable_sleep_mode: bool = False
212
    model_impl: str = "auto"
213

214
215
    calculate_kv_scales: Optional[bool] = None

216
    additional_config: Optional[Dict[str, Any]] = None
217
218
    enable_reasoning: Optional[bool] = None
    reasoning_parser: Optional[str] = None
219

220
    def __post_init__(self):
221
        if not self.tokenizer:
222
            self.tokenizer = self.model
223

224
225
226
227
        # Override the default value of enable_prefix_caching if it's not set
        # by user.
        if self.enable_prefix_caching is None:
            self.enable_prefix_caching = bool(envs.VLLM_USE_V1)
228

229
230
231
        # Override max_num_seqs if it's not set by user.
        if self.max_num_seqs is None:
            self.max_num_seqs = 256 if not envs.VLLM_USE_V1 else 1024
232

233
234
235
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
236
        if isinstance(self.compilation_config, (int, dict)):
237
238
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
239

240
        # Setup plugins
241
242
        from vllm.plugins import load_general_plugins
        load_general_plugins()
243
244

    @staticmethod
245
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
246
        """Shared CLI arguments for vLLM engine."""
247

248
        # Model arguments
249
250
251
        parser.add_argument(
            '--model',
            type=str,
252
            default=EngineArgs.model,
253
            help='Name or path of the huggingface model to use.')
254
255
256
257
258
259
        parser.add_argument(
            '--task',
            default=EngineArgs.task,
            choices=get_args(TaskOption),
            help='The task to use the model for. Each vLLM instance only '
            'supports one task, even if the same model can be used for '
260
            'multiple tasks. When the model only supports one task, ``"auto"`` '
261
262
            'can be used to select it; otherwise, you must specify explicitly '
            'which task to use.')
263
264
        parser.add_argument(
            '--tokenizer',
265
            type=nullable_str,
266
            default=EngineArgs.tokenizer,
267
268
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
269
270
271
272
273
274
        parser.add_argument(
            "--hf-config-path",
            type=nullable_str,
            default=EngineArgs.hf_config_path,
            help='Name or path of the huggingface config to use. '
            'If unspecified, model name or path will be used.')
275
276
277
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
278
            help='Skip initialization of tokenizer and detokenizer.')
Jasmond L's avatar
Jasmond L committed
279
280
        parser.add_argument(
            '--revision',
281
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
282
            default=None,
283
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
284
285
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
286
287
        parser.add_argument(
            '--code-revision',
288
            type=nullable_str,
289
            default=None,
290
            help='The specific revision to use for the model code on '
291
292
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
293
294
        parser.add_argument(
            '--tokenizer-revision',
295
            type=nullable_str,
296
            default=None,
297
298
299
            help='Revision of the huggingface tokenizer to use. '
            'It can be a branch name, a tag name, or a commit id. '
            'If unspecified, will use the default version.')
300
301
302
303
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
304
            choices=['auto', 'slow', 'mistral', 'custom'],
305
306
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
307
            'always use the slow tokenizer. \n* '
308
309
310
            '"mistral" will always use the `mistral_common` tokenizer. \n* '
            '"custom" will use --tokenizer to select the '
            'preregistered tokenizer.')
311
312
        parser.add_argument('--trust-remote-code',
                            action='store_true',
313
                            help='Trust remote code from huggingface.')
314
315
316
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
317
318
319
320
            help="Allowing API requests to read local images or videos "
            "from directories specified by the server file system. "
            "This is a security risk. "
            "Should only be enabled in trusted environments.")
321
        parser.add_argument('--download-dir',
322
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
323
                            default=EngineArgs.download_dir,
324
                            help='Directory to download and load the weights, '
325
                            'default to the default cache dir of '
326
                            'huggingface.')
327
328
329
330
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
331
            choices=[f.value for f in LoadFormat],
332
333
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
334
            'and fall back to the pytorch bin format if safetensors format '
335
336
337
338
339
340
341
342
            'is not available.\n'
            '* "pt" will load the weights in the pytorch bin format.\n'
            '* "safetensors" will load the weights in the safetensors format.\n'
            '* "npcache" will load the weights in pytorch format and store '
            'a numpy cache to speed up the loading.\n'
            '* "dummy" will initialize the weights with random values, '
            'which is mainly for profiling.\n'
            '* "tensorizer" will load the weights using tensorizer from '
343
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
344
            'section for more information.\n'
345
346
            '* "runai_streamer" will load the Safetensors weights using Run:ai'
            'Model Streamer \n'
347
348
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
349
350
351
352
353
354
355
        parser.add_argument(
            '--config-format',
            default=EngineArgs.config_format,
            choices=[f.value for f in ConfigFormat],
            help='The format of the model config to load.\n\n'
            '* "auto" will try to load the config in hf format '
            'if available else it will try to load in mistral format ')
356
357
358
359
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
360
361
362
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
363
364
365
366
367
368
369
370
            help='Data type for model weights and activations.\n\n'
            '* "auto" will use FP16 precision for FP32 and FP16 models, and '
            'BF16 precision for BF16 models.\n'
            '* "half" for FP16. Recommended for AWQ quantization.\n'
            '* "float16" is the same as "half".\n'
            '* "bfloat16" for a balance between precision and range.\n'
            '* "float" is shorthand for FP32 precision.\n'
            '* "float32" for FP32 precision.')
371
372
373
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
374
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
375
            default=EngineArgs.kv_cache_dtype,
376
            help='Data type for kv cache storage. If "auto", will use model '
377
378
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
379
380
        parser.add_argument('--max-model-len',
                            type=int,
381
                            default=EngineArgs.max_model_len,
382
383
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
384
385
386
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
387
            default='xgrammar',
388
            help='Which engine will be used for guided decoding'
389
            ' (JSON schema / regex etc) by default. Currently support '
390
            'https://github.com/outlines-dev/outlines, '
391
            'https://github.com/mlc-ai/xgrammar, and '
392
393
            'https://github.com/noamgat/lm-format-enforcer.'
            ' Can be overridden per request via guided_decoding_backend'
394
            ' parameter.\n'
395
            'Backend-specific options can be supplied in a comma-separated '
396
397
            'list following a colon after the backend name. Valid backends and '
            'all available options are: [xgrammar:no-fallback, '
398
            'xgrammar:disable-any-whitespace, '
399
            'outlines:no-fallback, lm-format-enforcer:no-fallback]')
400
401
402
403
404
405
406
407
        parser.add_argument(
            '--logits-processor-pattern',
            type=nullable_str,
            default=None,
            help='Optional regex pattern specifying valid logits processor '
            'qualified names that can be passed with the `logits_processors` '
            'extra completion argument. Defaults to None, which allows no '
            'processors.')
408
409
410
411
412
413
414
415
416
417
418
419
        parser.add_argument(
            '--model-impl',
            type=str,
            default=EngineArgs.model_impl,
            choices=[f.value for f in ModelImpl],
            help='Which implementation of the model to use.\n\n'
            '* "auto" will try to use the vLLM implementation if it exists '
            'and fall back to the Transformers implementation if no vLLM '
            'implementation is available.\n'
            '* "vllm" will use the vLLM model implementation.\n'
            '* "transformers" will use the Transformers model '
            'implementation.\n')
420
        # Parallel arguments
421
422
        parser.add_argument(
            '--distributed-executor-backend',
423
            choices=['ray', 'mp', 'uni', 'external_launcher'],
424
            default=EngineArgs.distributed_executor_backend,
425
426
427
428
429
430
            help='Backend to use for distributed model '
            'workers, either "ray" or "mp" (multiprocessing). If the product '
            'of pipeline_parallel_size and tensor_parallel_size is less than '
            'or equal to the number of GPUs available, "mp" will be used to '
            'keep processing on a single host. Otherwise, this will default '
            'to "ray" if Ray is installed and fail otherwise. Note that tpu '
431
            'only supports Ray for distributed inference.')
432

433
434
435
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
436
                            default=EngineArgs.pipeline_parallel_size,
437
                            help='Number of pipeline stages.')
438
439
440
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
441
                            default=EngineArgs.tensor_parallel_size,
442
                            help='Number of tensor parallel replicas.')
443
444
445
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
446
            default=EngineArgs.max_parallel_loading_workers,
447
            help='Load model sequentially in multiple batches, '
448
            'to avoid RAM OOM when using tensor '
449
            'parallel and large models.')
450
451
452
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
453
            help='If specified, use nsight to profile Ray workers.')
454
        # KV cache arguments
455
456
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
457
                            default=EngineArgs.block_size,
458
                            choices=[8, 16, 32, 64, 128],
459
                            help='Token block size for contiguous chunks of '
460
                            'tokens. This is ignored on neuron devices and '
461
                            'set to ``--max-model-len``. On CUDA devices, '
462
463
                            'only block sizes up to 32 are supported. '
                            'On HPU devices, block size defaults to 128.')
464

465
466
467
468
469
        parser.add_argument(
            "--enable-prefix-caching",
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.enable_prefix_caching,
            help="Enables automatic prefix caching. "
470
            "Use ``--no-enable-prefix-caching`` to disable explicitly.",
471
        )
472
473
474
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
475
                            'capping to sliding window size.')
476
477
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
478
                            default=True,
479
480
481
482
483
                            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.')
484
485
486
487
488
489
490
491
        parser.add_argument(
            '--num-lookahead-slots',
            type=int,
            default=EngineArgs.num_lookahead_slots,
            help='Experimental scheduling config necessary for '
            'speculative decoding. This will be replaced by '
            'speculative config in the future; it is present '
            'to enable correctness tests until then.')
492

493
494
495
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
496
                            help='Random seed for operations.')
497
        parser.add_argument('--swap-space',
498
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
499
                            default=EngineArgs.swap_space,
500
                            help='CPU swap space size (GiB) per GPU.')
501
502
503
504
505
506
507
508
509
        parser.add_argument(
            '--cpu-offload-gb',
            type=float,
            default=0,
            help='The space in GiB to offload to CPU, per GPU. '
            'Default is 0, which means no offloading. Intuitively, '
            'this argument can be seen as a virtual way to increase '
            'the GPU memory size. For example, if you have one 24 GB '
            'GPU and set this to 10, virtually you can think of it as '
510
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
511
            'which requires at least 26GB GPU memory. Note that this '
512
            'requires fast CPU-GPU interconnect, as part of the model is '
513
514
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
515
516
517
518
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
519
520
521
            help='The fraction of GPU memory to be used for the model '
            'executor, which can range from 0 to 1. For example, a value of '
            '0.5 would imply 50%% GPU memory utilization. If unspecified, '
522
523
524
525
526
527
            'will use the default value of 0.9. This is a per-instance '
            'limit, and only applies to the current vLLM instance.'
            'It does not matter if you have another vLLM instance running '
            'on the same GPU. For example, if you have two vLLM instances '
            'running on the same GPU, you can set the GPU memory utilization '
            'to 0.5 for each instance.')
528
        parser.add_argument(
529
            '--num-gpu-blocks-override',
530
531
532
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
533
            ' of GPU blocks. Used for testing preemption.')
534
535
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
536
                            default=EngineArgs.max_num_batched_tokens,
537
538
                            help='Maximum number of batched tokens per '
                            'iteration.')
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
        parser.add_argument(
            "--max-num-partial-prefills",
            type=int,
            default=EngineArgs.max_num_partial_prefills,
            help="For chunked prefill, the max number of concurrent \
            partial prefills."
            "Defaults to 1",
        )
        parser.add_argument(
            "--max-long-partial-prefills",
            type=int,
            default=EngineArgs.max_long_partial_prefills,
            help="For chunked prefill, the maximum number of prompts longer "
            "than --long-prefill-token-threshold that will be prefilled "
            "concurrently. Setting this less than --max-num-partial-prefills "
            "will allow shorter prompts to jump the queue in front of longer "
            "prompts in some cases, improving latency. Defaults to 1.")
        parser.add_argument(
            "--long-prefill-token-threshold",
            type=float,
            default=EngineArgs.long_prefill_token_threshold,
            help="For chunked prefill, a request is considered long if the "
            "prompt is longer than this number of tokens. Defaults to 4%% of "
            "the model's context length.",
        )
564
565
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
566
                            default=EngineArgs.max_num_seqs,
567
                            help='Maximum number of sequences per iteration.')
568
569
570
571
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
572
573
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
574
575
        parser.add_argument('--disable-log-stats',
                            action='store_true',
576
                            help='Disable logging statistics.')
577
578
579
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
580
                            type=nullable_str,
581
                            choices=[*QUANTIZATION_METHODS, None],
582
                            default=EngineArgs.quantization,
583
584
585
586
587
588
                            help='Method used to quantize the weights. If '
                            'None, we first check the `quantization_config` '
                            'attribute in the model config file. If that is '
                            'None, we assume the model weights are not '
                            'quantized and use `dtype` to determine the data '
                            'type of the weights.')
589
590
591
592
593
        parser.add_argument(
            '--rope-scaling',
            default=None,
            type=json.loads,
            help='RoPE scaling configuration in JSON format. '
594
            'For example, ``{"rope_type":"dynamic","factor":2.0}``')
595
596
597
598
599
600
        parser.add_argument('--rope-theta',
                            default=None,
                            type=float,
                            help='RoPE theta. Use with `rope_scaling`. In '
                            'some cases, changing the RoPE theta improves the '
                            'performance of the scaled model.')
601
602
603
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
604
                            help='Extra arguments for the HuggingFace config. '
605
606
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
607
608
609
610
611
        parser.add_argument('--enforce-eager',
                            action='store_true',
                            help='Always use eager-mode PyTorch. If False, '
                            'will use eager mode and CUDA graph in hybrid '
                            'for maximal performance and flexibility.')
612
        parser.add_argument('--max-seq-len-to-capture',
613
614
615
616
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
617
618
619
620
                            'larger than this, we fall back to eager mode. '
                            'Additionally for encoder-decoder models, if the '
                            'sequence length of the encoder input is larger '
                            'than this, we fall back to the eager mode.')
621
622
623
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
624
                            help='See ParallelConfig.')
625
626
627
628
629
630
631
632
633
634
635
636
637
        parser.add_argument('--tokenizer-pool-size',
                            type=int,
                            default=EngineArgs.tokenizer_pool_size,
                            help='Size of tokenizer pool to use for '
                            'asynchronous tokenization. If 0, will '
                            'use synchronous tokenization.')
        parser.add_argument('--tokenizer-pool-type',
                            type=str,
                            default=EngineArgs.tokenizer_pool_type,
                            help='Type of tokenizer pool to use for '
                            'asynchronous tokenization. Ignored '
                            'if tokenizer_pool_size is 0.')
        parser.add_argument('--tokenizer-pool-extra-config',
638
                            type=nullable_str,
639
640
641
642
643
                            default=EngineArgs.tokenizer_pool_extra_config,
                            help='Extra config for tokenizer pool. '
                            'This should be a JSON string that will be '
                            'parsed into a dictionary. Ignored if '
                            'tokenizer_pool_size is 0.')
644
645
646
647
648
649
650
651
652
653
654
655
656
657

        # Multimodal related configs
        parser.add_argument(
            '--limit-mm-per-prompt',
            type=nullable_kvs,
            default=EngineArgs.limit_mm_per_prompt,
            # The default value is given in
            # MultiModalRegistry.init_mm_limits_per_prompt
            help=('For each multimodal plugin, limit how many '
                  'input instances to allow for each prompt. '
                  'Expects a comma-separated list of items, '
                  'e.g.: `image=16,video=2` allows a maximum of 16 '
                  'images and 2 videos per prompt. Defaults to 1 for '
                  'each modality.'))
658
659
660
661
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
662
            help=('Overrides for the multimodal input mapping/processing, '
663
                  'e.g., image processor. For example: ``{"num_crops": 4}``.'))
664
        parser.add_argument(
665
            '--disable-mm-preprocessor-cache',
666
            action='store_true',
667
668
            help='If true, then disables caching of the multi-modal '
            'preprocessor/mapper. (not recommended)')
669

670
671
672
673
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
674
675
676
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
        parser.add_argument('--max-loras',
                            type=int,
                            default=EngineArgs.max_loras,
                            help='Max number of LoRAs in a single batch.')
        parser.add_argument('--max-lora-rank',
                            type=int,
                            default=EngineArgs.max_lora_rank,
                            help='Max LoRA rank.')
        parser.add_argument(
            '--lora-extra-vocab-size',
            type=int,
            default=EngineArgs.lora_extra_vocab_size,
            help=('Maximum size of extra vocabulary that can be '
                  'present in a LoRA adapter (added to the base '
                  'model vocabulary).'))
        parser.add_argument(
            '--lora-dtype',
            type=str,
            default=EngineArgs.lora_dtype,
696
            choices=['auto', 'float16', 'bfloat16'],
697
698
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
699
700
701
702
703
704
705
706
707
708
709
        parser.add_argument(
            '--long-lora-scaling-factors',
            type=nullable_str,
            default=EngineArgs.long_lora_scaling_factors,
            help=('Specify multiple scaling factors (which can '
                  'be different from base model scaling factor '
                  '- see eg. Long LoRA) to allow for multiple '
                  'LoRA adapters trained with those scaling '
                  'factors to be used at the same time. If not '
                  'specified, only adapters trained with the '
                  'base model scaling factor are allowed.'))
710
711
712
713
714
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
715
716
                  'Must be >= than max_loras. '
                  'Defaults to max_loras.'))
717
718
719
720
721
722
723
724
        parser.add_argument(
            '--fully-sharded-loras',
            action='store_true',
            help=('By default, only half of the LoRA computation is '
                  'sharded with tensor parallelism. '
                  'Enabling this will use the fully sharded layers. '
                  'At high sequence length, max rank or '
                  'tensor parallel size, this is likely faster.'))
725
726
727
728
729
730
731
732
733
734
735
        parser.add_argument('--enable-prompt-adapter',
                            action='store_true',
                            help='If True, enable handling of PromptAdapters.')
        parser.add_argument('--max-prompt-adapters',
                            type=int,
                            default=EngineArgs.max_prompt_adapters,
                            help='Max number of PromptAdapters in a batch.')
        parser.add_argument('--max-prompt-adapter-token',
                            type=int,
                            default=EngineArgs.max_prompt_adapter_token,
                            help='Max number of PromptAdapters tokens')
736
737
738
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
739
                            choices=DEVICE_OPTIONS,
740
                            help='Device type for vLLM execution.')
741
742
743
744
745
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
746

747
748
        parser.add_argument(
            '--multi-step-stream-outputs',
749
750
751
752
753
754
            action=StoreBoolean,
            default=EngineArgs.multi_step_stream_outputs,
            nargs="?",
            const="True",
            help='If False, then multi-step will stream outputs at the end '
            'of all steps')
755
756
757
758
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
759
            help='Apply a delay (of delay factor multiplied by previous '
760
            'prompt latency) before scheduling next prompt.')
761
762
        parser.add_argument(
            '--enable-chunked-prefill',
763
764
765
766
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
767
            help='If set, the prefill requests can be chunked based on the '
768
            'max_num_batched_tokens.')
769
770
771

        parser.add_argument(
            '--speculative-model',
772
            type=nullable_str,
773
            default=EngineArgs.speculative_model,
774
775
            help=
            'The name of the draft model to be used in speculative decoding.')
776
777
778
779
780
781
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
782
            help='Method used to quantize the weights of speculative model. '
783
784
785
786
787
            'If None, we first check the `quantization_config` '
            'attribute in the model config file. If that is '
            'None, we assume the model weights are not '
            'quantized and use `dtype` to determine the data '
            'type of the weights.')
788
789
790
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
791
            default=EngineArgs.num_speculative_tokens,
792
            help='The number of speculative tokens to sample from '
793
            'the draft model in speculative decoding.')
794
795
796
797
798
799
        parser.add_argument(
            '--speculative-disable-mqa-scorer',
            action='store_true',
            help=
            'If set to True, the MQA scorer will be disabled in speculative '
            ' and fall back to batch expansion')
800
801
802
803
804
805
806
        parser.add_argument(
            '--speculative-draft-tensor-parallel-size',
            '-spec-draft-tp',
            type=int,
            default=EngineArgs.speculative_draft_tensor_parallel_size,
            help='Number of tensor parallel replicas for '
            'the draft model in speculative decoding.')
807

808
809
        parser.add_argument(
            '--speculative-max-model-len',
810
            type=int,
811
812
813
814
815
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

816
817
818
819
820
821
822
        parser.add_argument(
            '--speculative-disable-by-batch-size',
            type=int,
            default=EngineArgs.speculative_disable_by_batch_size,
            help='Disable speculative decoding for new incoming requests '
            'if the number of enqueue requests is larger than this value.')

823
824
825
826
827
828
829
830
831
832
833
834
835
836
        parser.add_argument(
            '--ngram-prompt-lookup-max',
            type=int,
            default=EngineArgs.ngram_prompt_lookup_max,
            help='Max size of window for ngram prompt lookup in speculative '
            'decoding.')

        parser.add_argument(
            '--ngram-prompt-lookup-min',
            type=int,
            default=EngineArgs.ngram_prompt_lookup_min,
            help='Min size of window for ngram prompt lookup in speculative '
            'decoding.')

837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
        parser.add_argument(
            '--spec-decoding-acceptance-method',
            type=str,
            default=EngineArgs.spec_decoding_acceptance_method,
            choices=['rejection_sampler', 'typical_acceptance_sampler'],
            help='Specify the acceptance method to use during draft token '
            'verification in speculative decoding. Two types of acceptance '
            'routines are supported: '
            '1) RejectionSampler which does not allow changing the '
            'acceptance rate of draft tokens, '
            '2) TypicalAcceptanceSampler which is configurable, allowing for '
            'a higher acceptance rate at the cost of lower quality, '
            'and vice versa.')

        parser.add_argument(
            '--typical-acceptance-sampler-posterior-threshold',
            type=float,
            default=EngineArgs.typical_acceptance_sampler_posterior_threshold,
            help='Set the lower bound threshold for the posterior '
            'probability of a token to be accepted. This threshold is '
            'used by the TypicalAcceptanceSampler to make sampling decisions '
            'during speculative decoding. Defaults to 0.09')

        parser.add_argument(
            '--typical-acceptance-sampler-posterior-alpha',
            type=float,
            default=EngineArgs.typical_acceptance_sampler_posterior_alpha,
            help='A scaling factor for the entropy-based threshold for token '
            'acceptance in the TypicalAcceptanceSampler. Typically defaults '
            'to sqrt of --typical-acceptance-sampler-posterior-threshold '
            'i.e. 0.3')

869
870
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
871
            action=StoreBoolean,
872
            default=EngineArgs.disable_logprobs_during_spec_decoding,
873
874
            nargs="?",
            const="True",
875
876
877
878
879
880
881
882
            help='If set to True, token log probabilities are not returned '
            'during speculative decoding. If set to False, log probabilities '
            'are returned according to the settings in SamplingParams. If '
            'not specified, it defaults to True. Disabling log probabilities '
            'during speculative decoding reduces latency by skipping logprob '
            'calculation in proposal sampling, target sampling, and after '
            'accepted tokens are determined.')

883
        parser.add_argument('--model-loader-extra-config',
884
                            type=nullable_str,
885
886
887
888
889
890
                            default=EngineArgs.model_loader_extra_config,
                            help='Extra config for model loader. '
                            'This will be passed to the model loader '
                            'corresponding to the chosen load_format. '
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
891
892
893
894
895
896
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
897
            "Default to `original/**/*` to avoid repeated loading of llama's "
898
            "checkpoints.")
899
        parser.add_argument(
900
            '--preemption-mode',
901
902
            type=str,
            default=None,
903
904
905
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
906

907
908
909
910
911
912
913
914
915
916
        parser.add_argument(
            "--served-model-name",
            nargs="+",
            type=str,
            default=None,
            help="The model name(s) used in the API. If multiple "
            "names are provided, the server will respond to any "
            "of the provided names. The model name in the model "
            "field of a response will be the first name in this "
            "list. If not specified, the model name will be the "
917
            "same as the ``--model`` argument. Noted that this name(s) "
918
            "will also be used in `model_name` tag content of "
919
            "prometheus metrics, if multiple names provided, metrics "
920
            "tag will take the first one.")
921
922
923
924
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
925

926
927
928
929
930
931
932
933
934
935
936
937
        parser.add_argument('--show-hidden-metrics-for-version',
                            type=str,
                            default=None,
                            help='Enable deprecated Prometheus metrics that '
                            'have been hidden since the specified version. '
                            'For example, if a previously deprecated metric '
                            'has been hidden since the v0.7.0 release, you '
                            'use --show-hidden-metrics-for-version=0.7 as a '
                            'temporary escape hatch while you migrate to new '
                            'metrics. The metric is likely to be removed '
                            'completely in an upcoming release.')

938
939
940
941
942
        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
943
944
945
946
947
948
        parser.add_argument(
            '--collect-detailed-traces',
            type=str,
            default=None,
            help="Valid choices are " +
            ",".join(ALLOWED_DETAILED_TRACE_MODULES) +
949
            ". It makes sense to set this only if ``--otlp-traces-endpoint`` is"
950
951
952
            " set. If set, it will collect detailed traces for the specified "
            "modules. This involves use of possibly costly and or blocking "
            "operations and hence might have a performance impact.")
953

954
955
956
957
958
959
        parser.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.")
960

961
962
963
964
965
966
967
968
969
970
        parser.add_argument(
            '--scheduling-policy',
            choices=['fcfs', 'priority'],
            default="fcfs",
            help='The scheduling policy to use. "fcfs" (first come first served'
            ', i.e. requests are handled in order of arrival; default) '
            'or "priority" (requests are handled based on given '
            'priority (lower value means earlier handling) and time of '
            'arrival deciding any ties).')

971
972
973
974
975
976
977
        parser.add_argument(
            '--scheduler-cls',
            default=EngineArgs.scheduler_cls,
            help='The scheduler class to use. "vllm.core.scheduler.Scheduler" '
            'is the default scheduler. Can be a class directly or the path to '
            'a class of form "mod.custom_class".')

978
        parser.add_argument(
979
980
            '--override-neuron-config',
            type=json.loads,
981
            default=None,
982
            help="Override or set neuron device configuration. "
983
            "e.g. ``{\"cast_logits_dtype\": \"bloat16\"}``.")
984
        parser.add_argument(
985
986
            '--override-pooler-config',
            type=PoolerConfig.from_json,
987
            default=None,
988
            help="Override or set the pooling method for pooling models. "
989
            "e.g. ``{\"pooling_type\": \"mean\", \"normalize\": false}``.")
990

991
992
993
994
995
996
997
998
999
1000
1001
1002
        parser.add_argument('--compilation-config',
                            '-O',
                            type=CompilationConfig.from_cli,
                            default=None,
                            help='torch.compile configuration for the model.'
                            'When it is a number (0, 1, 2, 3), it will be '
                            'interpreted as the optimization level.\n'
                            'NOTE: level 0 is the default level without '
                            'any optimization. level 1 and 2 are for internal '
                            'testing only. level 3 is the recommended level '
                            'for production.\n'
                            'To specify the full compilation config, '
1003
1004
1005
1006
                            'use a JSON string.\n'
                            'Following the convention of traditional '
                            'compilers, using -O without space is also '
                            'supported. -O3 is equivalent to -O 3.')
1007

1008
1009
1010
1011
1012
1013
        parser.add_argument('--kv-transfer-config',
                            type=KVTransferConfig.from_cli,
                            default=None,
                            help='The configurations for distributed KV cache '
                            'transfer. Should be a JSON string.')

1014
1015
1016
1017
1018
        parser.add_argument(
            '--worker-cls',
            type=str,
            default="auto",
            help='The worker class to use for distributed execution.')
1019
1020
1021
1022
1023
1024
1025
        parser.add_argument(
            '--worker-extension-cls',
            type=str,
            default="",
            help='The worker extension class on top of the worker cls, '
            'it is useful if you just want to add new functions to the worker '
            'class without changing the existing functions.')
1026
1027
1028
1029
1030
        parser.add_argument(
            "--generation-config",
            type=nullable_str,
            default=None,
            help="The folder path to the generation config. "
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
            "Defaults to None, no generation config is loaded, vLLM defaults "
            "will be used. If set to 'auto', the generation config will be "
            "loaded from model path. If set to a folder path, the generation "
            "config will be loaded from the specified folder path. If "
            "`max_new_tokens` is specified in generation config, then "
            "it sets a server-wide limit on the number of output tokens "
            "for all requests.")

        parser.add_argument(
            "--override-generation-config",
            type=json.loads,
            default=None,
            help="Overrides or sets generation config in JSON format. "
            "e.g. ``{\"temperature\": 0.5}``. If used with "
            "--generation-config=auto, the override parameters will be merged "
            "with the default config from the model. If generation-config is "
            "None, only the override parameters are used.")
1048

1049
1050
1051
1052
1053
1054
        parser.add_argument("--enable-sleep-mode",
                            action="store_true",
                            default=False,
                            help="Enable sleep mode for the engine. "
                            "(only cuda platform is supported)")

1055
1056
1057
1058
1059
1060
1061
1062
1063
        parser.add_argument(
            '--calculate-kv-scales',
            action='store_true',
            help='This enables dynamic calculation of '
            'k_scale and v_scale when kv-cache-dtype is fp8. '
            'If calculate-kv-scales is false, the scales will '
            'be loaded from the model checkpoint if available. '
            'Otherwise, the scales will default to 1.0.')

1064
1065
1066
1067
1068
1069
1070
1071
        parser.add_argument(
            "--additional-config",
            type=json.loads,
            default=None,
            help="Additional config for specified platform in JSON format. "
            "Different platforms may support different configs. Make sure the "
            "configs are valid for the platform you are using. The input format"
            " is like '{\"config_key\":\"config_value\"}'")
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090

        parser.add_argument(
            "--enable-reasoning",
            action="store_true",
            default=False,
            help="Whether to enable reasoning_content for the model. "
            "If enabled, the model will be able to generate reasoning content."
        )

        parser.add_argument(
            "--reasoning-parser",
            type=str,
            choices=["deepseek_r1"],
            default=None,
            help=
            "Select the reasoning parser depending on the model that you're "
            "using. This is used to parse the reasoning content into OpenAI "
            "API format. Required for ``--enable-reasoning``.")

1091
        return parser
1092
1093

    @classmethod
1094
    def from_cli_args(cls, args: argparse.Namespace):
1095
1096
1097
        # 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
1098
1099
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
1100

1101
    def create_model_config(self) -> ModelConfig:
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
        # 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

1113
        return ModelConfig(
1114
            model=self.model,
1115
            hf_config_path=self.hf_config_path,
1116
            task=self.task,
1117
1118
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
1119
1120
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1121
            allowed_local_media_path=self.allowed_local_media_path,
1122
1123
1124
1125
1126
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
1127
            rope_theta=self.rope_theta,
1128
            hf_overrides=self.hf_overrides,
1129
1130
1131
1132
1133
1134
1135
1136
            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,
            skip_tokenizer_init=self.skip_tokenizer_init,
1137
            served_model_name=self.served_model_name,
1138
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1139
            use_async_output_proc=not self.disable_async_output_proc,
1140
            config_format=self.config_format,
1141
            mm_processor_kwargs=self.mm_processor_kwargs,
1142
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
1143
1144
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
1145
            logits_processor_pattern=self.logits_processor_pattern,
1146
            generation_config=self.generation_config,
1147
            override_generation_config=self.override_generation_config,
1148
            enable_sleep_mode=self.enable_sleep_mode,
1149
            model_impl=self.model_impl,
1150
        )
1151

1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
    def create_load_config(self) -> LoadConfig:
        # bitsandbytes quantization needs a specific model loader
        # so we make sure the quant method and the load format are consistent
        if (self.quantization == "bitsandbytes" or
           self.qlora_adapter_name_or_path is not None) and \
           self.load_format != "bitsandbytes":
            raise ValueError(
                "BitsAndBytes quantization and QLoRA adapter only support "
                f"'bitsandbytes' load format, but got {self.load_format}")

        if (self.load_format == "bitsandbytes" or
            self.qlora_adapter_name_or_path is not None) and \
            self.quantization != "bitsandbytes":
            raise ValueError(
                "BitsAndBytes load format and QLoRA adapter only support "
                f"'bitsandbytes' quantization, but got {self.quantization}")

1169
1170
1171
1172
1173
1174
        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,
        )
1175

1176
1177
1178
1179
1180
    def create_engine_config(self,
                             usage_context: Optional[UsageContext] = None
                             ) -> VllmConfig:
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update()
1181

1182
1183
        if envs.VLLM_USE_V1:
            self._override_v1_engine_args(usage_context)
1184

1185
        device_config = DeviceConfig(device=self.device)
1186
1187
        model_config = self.create_model_config()

1188
1189
1190
1191
1192
        if (model_config.is_multimodal_model and not envs.VLLM_USE_V1
                and self.enable_prefix_caching):
            logger.warning("--enable-prefix-caching is currently not "
                           "supported for multimodal models in v0 and "
                           "has been disabled.")
1193
1194
            self.enable_prefix_caching = False

1195
        cache_config = CacheConfig(
1196
            block_size=self.block_size,
1197
1198
1199
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1200
            is_attention_free=model_config.is_attention_free,
1201
1202
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1203
1204
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
1205
            calculate_kv_scales=self.calculate_kv_scales,
1206
        )
1207
        parallel_config = ParallelConfig(
1208
1209
1210
1211
1212
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
            max_parallel_loading_workers=self.max_parallel_loading_workers,
            disable_custom_all_reduce=self.disable_custom_all_reduce,
            tokenizer_pool_config=TokenizerPoolConfig.create_config(
1213
1214
1215
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1216
            ),
1217
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1218
1219
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1220
            worker_extension_cls=self.worker_extension_cls,
1221
        )
1222

1223
1224
1225
1226
1227
1228
        max_model_len = model_config.max_model_len
        use_long_context = max_model_len > 32768
        if self.enable_chunked_prefill is None:
            # If not explicitly set, enable chunked prefill by default for
            # long context (> 32K) models. This is to avoid OOM errors in the
            # initial memory profiling phase.
1229

1230
1231
1232
            # For multimodal models and models with MLA, chunked prefill is
            # disabled by default in V0, but enabled by design in V1
            if model_config.is_multimodal_model or model_config.use_mla:
1233
1234
1235
                self.enable_chunked_prefill = bool(envs.VLLM_USE_V1)

            elif use_long_context:
1236
1237
1238
1239
                is_gpu = device_config.device_type == "cuda"
                use_sliding_window = (model_config.get_sliding_window()
                                      is not None)
                use_spec_decode = self.speculative_model is not None
1240
                from vllm.platforms import current_platform
1241
1242
                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
1243
                        and not self.enable_prompt_adapter
1244
1245
                        and model_config.runner_type != "pooling"
                        and not current_platform.is_rocm()):
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
                    self.enable_chunked_prefill = True
                    logger.warning(
                        "Chunked prefill is enabled by default for models with "
                        "max_model_len > 32K. Currently, chunked prefill might "
                        "not work with some features or models. If you "
                        "encounter any issues, please disable chunked prefill "
                        "by setting --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 "
                "errors during the initial memory profiling phase, or result "
                "in low performance due to small KV cache space. Consider "
                "setting --max-model-len to a smaller value.", max_model_len)
1262
1263
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
1264
            msg = "Chunked prefill is not supported for pooling models"
1265
            raise ValueError(msg)
1266

1267
1268
1269
1270
1271
        speculative_config = SpeculativeConfig.maybe_create_spec_config(
            target_model_config=model_config,
            target_parallel_config=parallel_config,
            target_dtype=self.dtype,
            speculative_model=self.speculative_model,
1272
1273
            speculative_model_quantization = \
                self.speculative_model_quantization,
1274
1275
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
1276
            num_speculative_tokens=self.num_speculative_tokens,
1277
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
1278
1279
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
1280
1281
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
1282
            disable_log_stats=self.disable_log_stats,
1283
1284
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
1285
1286
1287
1288
1289
1290
            draft_token_acceptance_method=\
                self.spec_decoding_acceptance_method,
            typical_acceptance_sampler_posterior_threshold=self.
            typical_acceptance_sampler_posterior_threshold,
            typical_acceptance_sampler_posterior_alpha=self.
            typical_acceptance_sampler_posterior_alpha,
1291
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
1292
1293
        )

1294
        # Reminder: Please update docs/source/features/compatibility_matrix.md
1295
        # If the feature combo become valid
1296
1297
1298
1299
        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)")
1300
1301
1302
            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")
1303
1304
1305
1306
1307
1308
            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
1309
1310
1311
1312
1313
1314
1315
1316
1317

        # 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

1318
        scheduler_config = SchedulerConfig(
1319
            runner_type=model_config.runner_type,
1320
1321
1322
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1323
            num_lookahead_slots=num_lookahead_slots,
1324
1325
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1326
            is_multimodal_model=model_config.is_multimodal_model,
1327
            preemption_mode=self.preemption_mode,
1328
            num_scheduler_steps=self.num_scheduler_steps,
1329
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1330
1331
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1332
            policy=self.scheduling_policy,
1333
            scheduler_cls=self.scheduler_cls,
1334
1335
1336
1337
            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,
        )
1338

1339
        lora_config = LoRAConfig(
1340
            bias_enabled=self.enable_lora_bias,
1341
1342
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1343
            fully_sharded_loras=self.fully_sharded_loras,
1344
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1345
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1346
1347
1348
            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
1349

1350
1351
1352
1353
1354
1355
1356
        if self.qlora_adapter_name_or_path is not None and \
            self.qlora_adapter_name_or_path != "":
            if self.model_loader_extra_config is None:
                self.model_loader_extra_config = {}
            self.model_loader_extra_config[
                "qlora_adapter_name_or_path"] = self.qlora_adapter_name_or_path

1357
        load_config = self.create_load_config()
1358

1359
1360
1361
1362
1363
        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

1364
        decoding_config = DecodingConfig(
1365
1366
1367
1368
            guided_decoding_backend=self.guided_decoding_backend,
            reasoning_backend=self.reasoning_parser
            if self.enable_reasoning else None,
        )
1369

1370
1371
1372
1373
1374
        show_hidden_metrics = False
        if self.show_hidden_metrics_for_version is not None:
            show_hidden_metrics = version._prev_minor_version_was(
                self.show_hidden_metrics_for_version)

1375
1376
1377
1378
1379
1380
1381
1382
        detailed_trace_modules = []
        if self.collect_detailed_traces is not None:
            detailed_trace_modules = self.collect_detailed_traces.split(",")
        for m in detailed_trace_modules:
            if m not in ALLOWED_DETAILED_TRACE_MODULES:
                raise ValueError(
                    f"Invalid module {m} in collect_detailed_traces. "
                    f"Valid modules are {ALLOWED_DETAILED_TRACE_MODULES}")
1383
        observability_config = ObservabilityConfig(
1384
            show_hidden_metrics=show_hidden_metrics,
1385
1386
1387
1388
1389
1390
            otlp_traces_endpoint=self.otlp_traces_endpoint,
            collect_model_forward_time="model" in detailed_trace_modules
            or "all" in detailed_trace_modules,
            collect_model_execute_time="worker" in detailed_trace_modules
            or "all" in detailed_trace_modules,
        )
1391

1392
        config = VllmConfig(
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
            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,
1403
            prompt_adapter_config=prompt_adapter_config,
1404
            compilation_config=self.compilation_config,
1405
            kv_transfer_config=self.kv_transfer_config,
1406
            additional_config=self.additional_config,
1407
        )
1408

1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
        if envs.VLLM_USE_V1:
            self._override_v1_engine_config(config)
        return config

    def _override_v1_engine_args(self, usage_context: UsageContext) -> None:
        """
        Override the EngineArgs's args based on the usage context for V1.
        """
        assert envs.VLLM_USE_V1, "V1 is not enabled"

1419
1420
1421
1422
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
        # When no user override, set the default values based on the usage
        # context.
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
        # Use different default values for different hardware.
        from vllm.platforms import current_platform
        device_name = current_platform.get_device_name().lower()
        if "h100" in device_name or "h200" in device_name:
            # For H100 and H200, we use larger default values.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }

1439
1440
1441
1442
        if (self.max_num_batched_tokens is None
                and usage_context in default_max_num_batched_tokens):
            self.max_num_batched_tokens = default_max_num_batched_tokens[
                usage_context]
1443
1444
1445
            logger.warning(
                "Setting max_num_batched_tokens to %d for %s usage context.",
                self.max_num_batched_tokens, usage_context.value)
1446
1447
1448
1449
1450
1451
1452

    def _override_v1_engine_config(self, engine_config: VllmConfig) -> None:
        """
        Override the EngineConfig's configs based on the usage context for V1.
        """
        assert envs.VLLM_USE_V1, "V1 is not enabled"

1453

1454
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1455
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1456
    """Arguments for asynchronous vLLM engine."""
1457
    disable_log_requests: bool = False
1458
1459

    @staticmethod
1460
1461
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1462
1463
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1464
1465
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1466
                            help='Disable logging requests.')
1467
1468
1469
1470
1471
1472
        # 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()
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update(parser)
1473
        return parser
1474
1475
1476
1477


# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1478
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1479
1480
1481


def _async_engine_args_parser():
1482
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1483
                                        async_args_only=True)