arg_utils.py 51.9 KB
Newer Older
1
import argparse
2
import dataclasses
3
import json
4
from dataclasses import dataclass
5
6
from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
                    Tuple, Type, Union)
7

8
9
import torch

10
import vllm.envs as envs
11
12
13
14
from vllm.config import (CacheConfig, ConfigFormat, DecodingConfig,
                         DeviceConfig, EngineConfig, LoadConfig, LoadFormat,
                         LoRAConfig, ModelConfig, ObservabilityConfig,
                         ParallelConfig, PromptAdapterConfig, SchedulerConfig,
15
                         SpeculativeConfig, TokenizerPoolConfig)
16
from vllm.executor.executor_base import ExecutorBase
17
from vllm.logger import init_logger
18
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
19
from vllm.transformers_utils.utils import check_gguf_file
20
from vllm.utils import FlexibleArgumentParser
21

22
if TYPE_CHECKING:
23
    from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
24

25
26
logger = init_logger(__name__)

27
28
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]

29
30
31
32
33
34
35
36
37
38
DEVICE_OPTIONS = [
    "auto",
    "cuda",
    "neuron",
    "cpu",
    "openvino",
    "tpu",
    "xpu",
]

39

40
41
42
43
44
45
def nullable_str(val: str):
    if not val or val == "None":
        return None
    return val


46
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
47
48
49
50
51
52
53
54
55
    """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.
    """
56
57
58
59
60
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
61
62
63
64
65
        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
66
67

        try:
68
            parsed_value = int(value)
69
70
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
71
72
73
74
75
76
            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
77
78
79
80

    return out_dict


81
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
82
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
83
    """Arguments for vLLM engine."""
84
    model: str = 'facebook/opt-125m'
85
    served_model_name: Optional[Union[str, List[str]]] = None
86
    tokenizer: Optional[str] = None
87
    skip_tokenizer_init: bool = False
88
    tokenizer_mode: str = 'auto'
89
    trust_remote_code: bool = False
90
    download_dir: Optional[str] = None
91
    load_format: str = 'auto'
92
    config_format: str = 'auto'
93
    dtype: str = 'auto'
94
    kv_cache_dtype: str = 'auto'
95
    quantization_param_path: Optional[str] = None
96
    seed: int = 0
97
    max_model_len: Optional[int] = None
98
    worker_use_ray: bool = False
99
100
101
102
103
    # 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
104
105
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
106
    max_parallel_loading_workers: Optional[int] = None
107
    block_size: int = 16
108
    enable_prefix_caching: bool = False
109
    disable_sliding_window: bool = False
110
    use_v2_block_manager: bool = True
111
112
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
113
    gpu_memory_utilization: float = 0.90
114
    max_num_batched_tokens: Optional[int] = None
115
    max_num_seqs: int = 256
116
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
117
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
118
    revision: Optional[str] = None
119
    code_revision: Optional[str] = None
120
    rope_scaling: Optional[dict] = None
121
    rope_theta: Optional[float] = None
122
    tokenizer_revision: Optional[str] = None
123
    quantization: Optional[str] = None
124
    enforce_eager: Optional[bool] = None
125
126
    max_context_len_to_capture: Optional[int] = None
    max_seq_len_to_capture: int = 8192
127
    disable_custom_all_reduce: bool = False
128
    tokenizer_pool_size: int = 0
129
130
131
132
    # 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"
133
    tokenizer_pool_extra_config: Optional[dict] = None
134
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
135
136
137
    enable_lora: bool = False
    max_loras: int = 1
    max_lora_rank: int = 16
138
139
140
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
141
    fully_sharded_loras: bool = False
142
    lora_extra_vocab_size: int = 256
143
    long_lora_scaling_factors: Optional[Tuple[float]] = None
144
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
145
    max_cpu_loras: Optional[int] = None
146
    device: str = 'auto'
147
    num_scheduler_steps: int = 1
148
    multi_step_stream_outputs: bool = False
149
    ray_workers_use_nsight: bool = False
150
    num_gpu_blocks_override: Optional[int] = None
151
    num_lookahead_slots: int = 0
152
    model_loader_extra_config: Optional[dict] = None
153
    ignore_patterns: Optional[Union[str, List[str]]] = None
154
    preemption_mode: Optional[str] = None
155

156
    scheduler_delay_factor: float = 0.0
157
    enable_chunked_prefill: Optional[bool] = None
158

159
    guided_decoding_backend: str = 'outlines'
160
161
    # Speculative decoding configuration.
    speculative_model: Optional[str] = None
162
    speculative_model_quantization: Optional[str] = None
163
    speculative_draft_tensor_parallel_size: Optional[int] = None
164
    num_speculative_tokens: Optional[int] = None
165
    speculative_disable_mqa_scorer: Optional[bool] = False
166
    speculative_max_model_len: Optional[int] = None
167
    speculative_disable_by_batch_size: Optional[int] = None
168
169
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
170
171
172
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
173
    qlora_adapter_name_or_path: Optional[str] = None
174
    disable_logprobs_during_spec_decoding: Optional[bool] = None
175

176
    otlp_traces_endpoint: Optional[str] = None
177
    collect_detailed_traces: Optional[str] = None
178
    disable_async_output_proc: bool = False
179
    override_neuron_config: Optional[Dict[str, Any]] = None
180
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
181
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
182

183
    def __post_init__(self):
184
185
        if self.tokenizer is None:
            self.tokenizer = self.model
186
187

    @staticmethod
188
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
189
        """Shared CLI arguments for vLLM engine."""
190

191
        # Model arguments
192
193
194
        parser.add_argument(
            '--model',
            type=str,
195
            default=EngineArgs.model,
196
            help='Name or path of the huggingface model to use.')
197
198
        parser.add_argument(
            '--tokenizer',
199
            type=nullable_str,
200
            default=EngineArgs.tokenizer,
201
202
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
203
204
205
206
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
            help='Skip initialization of tokenizer and detokenizer')
Jasmond L's avatar
Jasmond L committed
207
208
        parser.add_argument(
            '--revision',
209
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
210
            default=None,
211
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
212
213
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
214
215
        parser.add_argument(
            '--code-revision',
216
            type=nullable_str,
217
            default=None,
218
            help='The specific revision to use for the model code on '
219
220
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
221
222
        parser.add_argument(
            '--tokenizer-revision',
223
            type=nullable_str,
224
            default=None,
225
226
227
            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.')
228
229
230
231
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
232
            choices=['auto', 'slow', 'mistral'],
233
234
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
235
236
            'always use the slow tokenizer. \n* '
            '"mistral" will always use the `mistral_common` tokenizer.')
237
238
        parser.add_argument('--trust-remote-code',
                            action='store_true',
239
                            help='Trust remote code from huggingface.')
240
        parser.add_argument('--download-dir',
241
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
242
                            default=EngineArgs.download_dir,
243
                            help='Directory to download and load the weights, '
244
                            'default to the default cache dir of '
245
                            'huggingface.')
246
247
248
249
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
250
            choices=[f.value for f in LoadFormat],
251
252
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
253
            'and fall back to the pytorch bin format if safetensors format '
254
255
256
257
258
259
260
261
            '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 '
262
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
263
264
265
            'section for more information.\n'
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
266
267
268
269
270
271
272
        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 ')
273
274
275
276
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
277
278
279
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
280
281
282
283
284
285
286
287
            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.')
288
289
290
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
291
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
292
            default=EngineArgs.kv_cache_dtype,
293
            help='Data type for kv cache storage. If "auto", will use model '
294
295
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
296
297
        parser.add_argument(
            '--quantization-param-path',
298
            type=nullable_str,
299
300
301
302
303
304
305
            default=None,
            help='Path to the JSON file containing the KV cache '
            'scaling factors. This should generally be supplied, when '
            'KV cache dtype is FP8. Otherwise, KV cache scaling factors '
            'default to 1.0, which may cause accuracy issues. '
            'FP8_E5M2 (without scaling) is only supported on cuda version'
            'greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead '
306
            'supported for common inference criteria.')
307
308
        parser.add_argument('--max-model-len',
                            type=int,
309
                            default=EngineArgs.max_model_len,
310
311
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
312
313
314
315
316
317
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
            default='outlines',
            choices=['outlines', 'lm-format-enforcer'],
            help='Which engine will be used for guided decoding'
318
319
320
321
322
            ' (JSON schema / regex etc) by default. Currently support '
            'https://github.com/outlines-dev/outlines and '
            'https://github.com/noamgat/lm-format-enforcer.'
            ' Can be overridden per request via guided_decoding_backend'
            ' parameter.')
323
        # Parallel arguments
324
325
326
327
328
329
330
331
332
333
334
        parser.add_argument(
            '--distributed-executor-backend',
            choices=['ray', 'mp'],
            default=EngineArgs.distributed_executor_backend,
            help='Backend to use for distributed serving. When more than 1 GPU '
            'is used, will be automatically set to "ray" if installed '
            'or "mp" (multiprocessing) otherwise.')
        parser.add_argument(
            '--worker-use-ray',
            action='store_true',
            help='Deprecated, use --distributed-executor-backend=ray.')
335
336
337
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
338
                            default=EngineArgs.pipeline_parallel_size,
339
                            help='Number of pipeline stages.')
340
341
342
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
343
                            default=EngineArgs.tensor_parallel_size,
344
                            help='Number of tensor parallel replicas.')
345
346
347
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
348
            default=EngineArgs.max_parallel_loading_workers,
349
            help='Load model sequentially in multiple batches, '
350
            'to avoid RAM OOM when using tensor '
351
            'parallel and large models.')
352
353
354
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
355
            help='If specified, use nsight to profile Ray workers.')
356
        # KV cache arguments
357
358
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
359
                            default=EngineArgs.block_size,
360
                            choices=[8, 16, 32],
361
                            help='Token block size for contiguous chunks of '
362
363
                            'tokens. This is ignored on neuron devices and '
                            'set to max-model-len')
364
365
366

        parser.add_argument('--enable-prefix-caching',
                            action='store_true',
367
                            help='Enables automatic prefix caching.')
368
369
370
371
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
                            'capping to sliding window size')
372
373
374
375
376
377
        parser.add_argument(
            '--use-v2-block-manager',
            default=EngineArgs.use_v2_block_manager,
            action='store_true',
            help='Use BlockSpaceMangerV2. By default this is set to True. '
            'Set to False to use BlockSpaceManagerV1')
378
379
380
381
382
383
384
385
        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.')
386

387
388
389
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
390
                            help='Random seed for operations.')
391
        parser.add_argument('--swap-space',
392
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
393
                            default=EngineArgs.swap_space,
394
                            help='CPU swap space size (GiB) per GPU.')
395
396
397
398
399
400
401
402
403
404
405
406
407
408
        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 '
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight,'
            'which requires at least 26GB GPU memory. Note that this '
            'requires fast CPU-GPU interconnect, as part of the model is'
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
409
410
411
412
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
413
414
415
416
            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, '
            'will use the default value of 0.9.')
417
        parser.add_argument(
418
            '--num-gpu-blocks-override',
419
420
421
422
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
            'of GPU blocks. Used for testing preemption.')
423
424
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
425
                            default=EngineArgs.max_num_batched_tokens,
426
427
                            help='Maximum number of batched tokens per '
                            'iteration.')
428
429
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
430
                            default=EngineArgs.max_num_seqs,
431
                            help='Maximum number of sequences per iteration.')
432
433
434
435
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
436
437
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
438
439
        parser.add_argument('--disable-log-stats',
                            action='store_true',
440
                            help='Disable logging statistics.')
441
442
443
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
444
                            type=nullable_str,
445
                            choices=[*QUANTIZATION_METHODS, None],
446
                            default=EngineArgs.quantization,
447
448
449
450
451
452
                            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.')
453
454
455
456
457
        parser.add_argument('--rope-scaling',
                            default=None,
                            type=json.loads,
                            help='RoPE scaling configuration in JSON format. '
                            'For example, {"type":"dynamic","factor":2.0}')
458
459
460
461
462
463
        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.')
464
465
466
467
468
469
470
471
        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.')
        parser.add_argument('--max-context-len-to-capture',
                            type=int,
                            default=EngineArgs.max_context_len_to_capture,
472
                            help='Maximum context length covered by CUDA '
473
                            'graphs. When a sequence has context length '
474
                            'larger than this, we fall back to eager mode. '
475
                            '(DEPRECATED. Use --max-seq-len-to-capture instead'
476
                            ')')
477
        parser.add_argument('--max-seq-len-to-capture',
478
479
480
481
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
482
483
484
485
                            '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.')
486
487
488
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
489
                            help='See ParallelConfig.')
490
491
492
493
494
495
496
497
498
499
500
501
502
        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',
503
                            type=nullable_str,
504
505
506
507
508
                            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.')
509
510
511
512
513
514
515
516
517
518
519
520
521
522

        # 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.'))
523
524
525
526
527
528
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
            help=('Overrides for the multimodal input mapping/processing,'
                  'e.g., image processor. For example: {"num_crops": 4}.'))
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
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
        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,
            choices=['auto', 'float16', 'bfloat16', 'float32'],
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
556
557
558
559
560
561
562
563
564
565
566
        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.'))
567
568
569
570
571
572
573
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
                  'Must be >= than max_num_seqs. '
                  'Defaults to max_num_seqs.'))
574
575
576
577
578
579
580
581
        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.'))
582
583
584
585
586
587
588
589
590
591
592
        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')
593
594
595
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
596
                            choices=DEVICE_OPTIONS,
597
                            help='Device type for vLLM execution.')
598
599
600
601
602
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
603

604
605
606
607
        parser.add_argument(
            '--multi-step-stream-outputs',
            action='store_true',
            help='If True, then multi-step will stream outputs for every step')
608
609
610
611
612
613
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
            help='Apply a delay (of delay factor multiplied by previous'
            'prompt latency) before scheduling next prompt.')
614
615
        parser.add_argument(
            '--enable-chunked-prefill',
616
617
618
619
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
620
            help='If set, the prefill requests can be chunked based on the '
621
            'max_num_batched_tokens.')
622
623
624

        parser.add_argument(
            '--speculative-model',
625
            type=nullable_str,
626
            default=EngineArgs.speculative_model,
627
628
            help=
            'The name of the draft model to be used in speculative decoding.')
629
630
631
632
633
634
635
636
637
638
639
640
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
            help='Method used to quantize the weights of speculative model.'
            '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.')
641
642
643
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
644
            default=EngineArgs.num_speculative_tokens,
645
            help='The number of speculative tokens to sample from '
646
            'the draft model in speculative decoding.')
647
648
649
650
651
652
        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')
653
654
655
656
657
658
659
        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.')
660

661
662
        parser.add_argument(
            '--speculative-max-model-len',
663
            type=int,
664
665
666
667
668
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

669
670
671
672
673
674
675
        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.')

676
677
678
679
680
681
682
683
684
685
686
687
688
689
        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.')

690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
        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')

722
723
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
724
            action=StoreBoolean,
725
            default=EngineArgs.disable_logprobs_during_spec_decoding,
726
727
            nargs="?",
            const="True",
728
729
730
731
732
733
734
735
            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.')

736
        parser.add_argument('--model-loader-extra-config',
737
                            type=nullable_str,
738
739
740
741
742
743
                            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.')
744
745
746
747
748
749
750
751
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
            "Default to 'original/**/*' to avoid repeated loading of llama's "
            "checkpoints.")
752
        parser.add_argument(
753
            '--preemption-mode',
754
755
            type=str,
            default=None,
756
757
758
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
759

760
761
762
763
764
765
766
767
768
769
770
771
772
773
        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 "
            "same as the `--model` argument. Noted that this name(s)"
            "will also be used in `model_name` tag content of "
            "prometheus metrics, if multiple names provided, metrics"
            "tag will take the first one.")
774
775
776
777
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
778
779
780
781
782
783

        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
784
785
786
787
788
789
790
791
792
793
        parser.add_argument(
            '--collect-detailed-traces',
            type=str,
            default=None,
            help="Valid choices are " +
            ",".join(ALLOWED_DETAILED_TRACE_MODULES) +
            ". It makes sense to set this only if --otlp-traces-endpoint is"
            " 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.")
794

795
796
797
798
799
800
        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.")
801
802
803
804
805
806
807
808
809
810
        parser.add_argument(
            '--override-neuron-config',
            type=lambda configs: {
                str(key): value
                for key, value in
                (config.split(':') for config in configs.split(','))
            },
            default=None,
            help="override or set neuron device configuration.")

811
812
813
814
815
816
817
818
819
820
        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).')

821
        return parser
822
823

    @classmethod
824
    def from_cli_args(cls, args: argparse.Namespace):
825
826
827
        # 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
828
829
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
830

831
832
    def create_model_config(self) -> ModelConfig:
        return ModelConfig(
833
834
835
836
837
838
839
840
841
            model=self.model,
            tokenizer=self.tokenizer,
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
842
            rope_theta=self.rope_theta,
843
844
845
846
847
848
849
850
851
852
            tokenizer_revision=self.tokenizer_revision,
            max_model_len=self.max_model_len,
            quantization=self.quantization,
            quantization_param_path=self.quantization_param_path,
            enforce_eager=self.enforce_eager,
            max_context_len_to_capture=self.max_context_len_to_capture,
            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,
853
            served_model_name=self.served_model_name,
854
            limit_mm_per_prompt=self.limit_mm_per_prompt,
855
            use_async_output_proc=not self.disable_async_output_proc,
856
857
            override_neuron_config=self.override_neuron_config,
            config_format=self.config_format,
858
            mm_processor_kwargs=self.mm_processor_kwargs,
859
860
        )

861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
    def create_load_config(self) -> LoadConfig:
        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,
        )

    def create_engine_config(self) -> EngineConfig:
        # 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"

        # 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}")

        assert self.cpu_offload_gb >= 0, (
            "CPU offload space must be non-negative"
            f", but got {self.cpu_offload_gb}")

        device_config = DeviceConfig(device=self.device)
        model_config = self.create_model_config()

897
898
899
900
901
902
903
        if model_config.is_multimodal_model:
            if self.enable_prefix_caching:
                logger.warning(
                    "--enable-prefix-caching is currently not "
                    "supported for multimodal models and has been disabled.")
            self.enable_prefix_caching = False

904
        cache_config = CacheConfig(
905
906
            block_size=self.block_size if self.device != "neuron" else
            self.max_model_len,  # neuron needs block_size = max_model_len
907
908
909
910
911
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
912
913
914
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
        )
915
        parallel_config = ParallelConfig(
916
917
918
919
920
921
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
            worker_use_ray=self.worker_use_ray,
            max_parallel_loading_workers=self.max_parallel_loading_workers,
            disable_custom_all_reduce=self.disable_custom_all_reduce,
            tokenizer_pool_config=TokenizerPoolConfig.create_config(
922
923
924
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
925
            ),
926
            ray_workers_use_nsight=self.ray_workers_use_nsight,
927
            distributed_executor_backend=self.distributed_executor_backend)
928

929
930
931
932
933
934
        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.
935
936
937
938

            # Chunked prefill is currently disabled for multimodal models by
            # default.
            if use_long_context and not model_config.is_multimodal_model:
939
940
941
942
                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
943
944
945
                has_seqlen_agnostic_layers = (
                    model_config.contains_seqlen_agnostic_layers(
                        parallel_config))
946
947
948
                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
                        and not self.enable_prompt_adapter
949
                        and not has_seqlen_agnostic_layers):
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
                    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)

967
968
969
970
971
972
        if self.num_scheduler_steps > 1 and not self.use_v2_block_manager:
            self.use_v2_block_manager = True
            logger.warning(
                "Enabled BlockSpaceManagerV2 because it is "
                "required for multi-step (--num-scheduler-steps > 1)")

973
974
975
976
977
        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,
978
979
            speculative_model_quantization = \
                self.speculative_model_quantization,
980
981
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
982
            num_speculative_tokens=self.num_speculative_tokens,
983
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
984
985
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
986
987
988
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
            use_v2_block_manager=self.use_v2_block_manager,
989
            disable_log_stats=self.disable_log_stats,
990
991
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
992
993
994
995
996
997
            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,
998
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
999
1000
        )

1001
1002
1003
1004
        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)")
1005
1006
1007
            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")
1008
1009
1010
1011
1012
1013
1014
1015
1016

        # 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

1017
        scheduler_config = SchedulerConfig(
1018
1019
1020
1021
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
            use_v2_block_manager=self.use_v2_block_manager,
1022
            num_lookahead_slots=num_lookahead_slots,
1023
1024
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1025
            embedding_mode=model_config.embedding_mode,
1026
            is_multimodal_model=model_config.is_multimodal_model,
1027
            preemption_mode=self.preemption_mode,
1028
            num_scheduler_steps=self.num_scheduler_steps,
1029
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1030
1031
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1032
            policy=self.scheduling_policy,
1033
        )
1034
1035
1036
        lora_config = LoRAConfig(
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1037
            fully_sharded_loras=self.fully_sharded_loras,
1038
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1039
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1040
1041
1042
            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
1043

1044
1045
1046
1047
1048
1049
1050
        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

1051
        load_config = self.create_load_config()
1052

1053
1054
1055
1056
1057
        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

1058
1059
1060
        decoding_config = DecodingConfig(
            guided_decoding_backend=self.guided_decoding_backend)

1061
1062
1063
1064
1065
1066
1067
1068
        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}")
1069
        observability_config = ObservabilityConfig(
1070
1071
1072
1073
1074
1075
            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,
        )
1076

1077
        if (model_config.get_sliding_window() is not None
1078
1079
                and scheduler_config.chunked_prefill_enabled
                and not scheduler_config.use_v2_block_manager):
1080
            raise ValueError(
1081
1082
                "Chunked prefill is not supported with sliding window. "
                "Set --disable-sliding-window to disable sliding window.")
1083

1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
        return EngineConfig(
            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,
1095
            prompt_adapter_config=prompt_adapter_config,
1096
        )
1097
1098


1099
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1100
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1101
    """Arguments for asynchronous vLLM engine."""
1102
    disable_log_requests: bool = False
1103
1104

    @staticmethod
1105
1106
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1107
1108
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1109
1110
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1111
                            help='Disable logging requests.')
1112
        return parser
1113
1114


1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
class StoreBoolean(argparse.Action):

    def __call__(self, parser, namespace, values, option_string=None):
        if values.lower() == "true":
            setattr(namespace, self.dest, True)
        elif values.lower() == "false":
            setattr(namespace, self.dest, False)
        else:
            raise ValueError(f"Invalid boolean value: {values}. "
                             "Expected 'true' or 'false'.")


1127
1128
# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1129
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1130
1131
1132


def _async_engine_args_parser():
1133
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1134
                                        async_args_only=True)