arg_utils.py 60.6 KB
Newer Older
1
import argparse
2
import dataclasses
3
import json
4
from dataclasses import dataclass
5
from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
6
                    Tuple, Type, Union, cast, get_args)
7

8
9
import torch

10
import vllm.envs as envs
11
from vllm.config import (CacheConfig, CompilationConfig, ConfigFormat,
12
13
14
15
                         DecodingConfig, DeviceConfig, HfOverrides,
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
                         ModelConfig, ObservabilityConfig, ParallelConfig,
                         PoolerConfig, PromptAdapterConfig, SchedulerConfig,
16
17
                         SpeculativeConfig, TaskOption, TokenizerPoolConfig,
                         VllmConfig)
18
from vllm.executor.executor_base import ExecutorBase
19
from vllm.logger import init_logger
20
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
21
from vllm.platforms import current_platform
22
from vllm.transformers_utils.utils import check_gguf_file
23
from vllm.usage.usage_lib import UsageContext
24
from vllm.utils import FlexibleArgumentParser, StoreBoolean
25

26
if TYPE_CHECKING:
27
    from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
28

29
30
logger = init_logger(__name__)

31
32
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]

33
34
35
36
37
38
39
40
DEVICE_OPTIONS = [
    "auto",
    "cuda",
    "neuron",
    "cpu",
    "openvino",
    "tpu",
    "xpu",
41
    "hpu",
42
43
]

44

45
46
47
48
49
50
def nullable_str(val: str):
    if not val or val == "None":
        return None
    return val


51
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
52
53
54
55
56
57
58
59
60
    """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.
    """
61
62
63
64
65
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
66
67
68
69
70
        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
71
72

        try:
73
            parsed_value = int(value)
74
75
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
76
77
78
79
80
81
            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
82
83
84
85

    return out_dict


86
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
87
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
88
    """Arguments for vLLM engine."""
89
    model: str = 'facebook/opt-125m'
90
    served_model_name: Optional[Union[str, List[str]]] = None
91
    tokenizer: Optional[str] = None
92
    task: TaskOption = "auto"
93
    skip_tokenizer_init: bool = False
94
    tokenizer_mode: str = 'auto'
95
    trust_remote_code: bool = False
96
    allowed_local_media_path: str = ""
97
    download_dir: Optional[str] = None
98
    load_format: str = 'auto'
99
    config_format: ConfigFormat = ConfigFormat.AUTO
100
    dtype: str = 'auto'
101
    kv_cache_dtype: str = 'auto'
102
    quantization_param_path: Optional[str] = None
103
    seed: int = 0
104
    max_model_len: Optional[int] = None
105
    worker_use_ray: bool = False
106
107
108
109
110
    # 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
111
    # number of P/D disaggregation (or other disaggregation) workers
112
113
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
114
    max_parallel_loading_workers: Optional[int] = None
115
    block_size: Optional[int] = None
116
    enable_prefix_caching: Optional[bool] = None
117
    disable_sliding_window: bool = False
118
    use_v2_block_manager: bool = True
119
120
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
121
    gpu_memory_utilization: float = 0.90
122
    max_num_batched_tokens: Optional[int] = None
123
    max_num_seqs: Optional[int] = None
124
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
125
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
126
    revision: Optional[str] = None
127
    code_revision: Optional[str] = None
128
    rope_scaling: Optional[Dict[str, Any]] = None
129
    rope_theta: Optional[float] = None
130
    hf_overrides: Optional[HfOverrides] = None
131
    tokenizer_revision: Optional[str] = None
132
    quantization: Optional[str] = None
133
    enforce_eager: Optional[bool] = None
134
    max_seq_len_to_capture: int = 8192
135
    disable_custom_all_reduce: bool = False
136
    tokenizer_pool_size: int = 0
137
138
139
140
    # 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"
141
    tokenizer_pool_extra_config: Optional[Dict[str, Any]] = None
142
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
143
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
144
    mm_cache_preprocessor: bool = False
145
    enable_lora: bool = False
146
    enable_lora_bias: bool = False
147
148
    max_loras: int = 1
    max_lora_rank: int = 16
149
150
151
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
152
    fully_sharded_loras: bool = False
153
    lora_extra_vocab_size: int = 256
154
    long_lora_scaling_factors: Optional[Tuple[float]] = None
155
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
156
    max_cpu_loras: Optional[int] = None
157
158
    merge_lora: bool = False
    lora_target_modules: Optional[List[str]] = None
159
    device: str = 'auto'
160
    num_scheduler_steps: int = 1
161
    multi_step_stream_outputs: bool = True
162
    ray_workers_use_nsight: bool = False
163
    num_gpu_blocks_override: Optional[int] = None
164
    num_lookahead_slots: int = 0
165
    model_loader_extra_config: Optional[dict] = None
166
    ignore_patterns: Optional[Union[str, List[str]]] = None
167
    preemption_mode: Optional[str] = None
168

169
    scheduler_delay_factor: float = 0.0
170
    enable_chunked_prefill: Optional[bool] = None
171

172
    guided_decoding_backend: str = 'xgrammar'
173
    logits_processor_pattern: Optional[str] = None
174
175
    # Speculative decoding configuration.
    speculative_model: Optional[str] = None
176
    speculative_model_quantization: Optional[str] = None
177
    speculative_draft_tensor_parallel_size: Optional[int] = None
178
    num_speculative_tokens: Optional[int] = None
179
    num_speculative_heads: Optional[int] = None
180
    speculative_disable_mqa_scorer: Optional[bool] = False
181
    speculative_max_model_len: Optional[int] = None
182
    speculative_disable_by_batch_size: Optional[int] = None
183
184
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
185
186
187
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
188
    qlora_adapter_name_or_path: Optional[str] = None
189
    disable_logprobs_during_spec_decoding: Optional[bool] = None
190

191
    otlp_traces_endpoint: Optional[str] = None
192
    collect_detailed_traces: Optional[str] = None
193
    disable_async_output_proc: bool = False
194
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
195

196
    override_neuron_config: Optional[Dict[str, Any]] = None
197
    override_pooler_config: Optional[PoolerConfig] = None
198
    compilation_config: Optional[CompilationConfig] = None
199
    worker_cls: str = "auto"
200

201
    kv_transfer_config: Optional[KVTransferConfig] = None
202

203
    def __post_init__(self):
204
        if not self.tokenizer:
205
            self.tokenizer = self.model
206

207
208
209
210
        # 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)
211
212
213
        # 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
214

215
216
217
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
218
        if isinstance(self.compilation_config, (int, dict)):
219
220
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
221

222
        # Setup plugins
223
224
        from vllm.plugins import load_general_plugins
        load_general_plugins()
225
226

    @staticmethod
227
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
228
        """Shared CLI arguments for vLLM engine."""
229

230
        # Model arguments
231
232
233
        parser.add_argument(
            '--model',
            type=str,
234
            default=EngineArgs.model,
235
            help='Name or path of the huggingface model to use.')
236
237
238
239
240
241
242
243
244
        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 '
            'multiple tasks. When the model only supports one task, "auto" '
            'can be used to select it; otherwise, you must specify explicitly '
            'which task to use.')
245
246
        parser.add_argument(
            '--tokenizer',
247
            type=nullable_str,
248
            default=EngineArgs.tokenizer,
249
250
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
251
252
253
254
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
            help='Skip initialization of tokenizer and detokenizer')
Jasmond L's avatar
Jasmond L committed
255
256
        parser.add_argument(
            '--revision',
257
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
258
            default=None,
259
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
260
261
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
262
263
        parser.add_argument(
            '--code-revision',
264
            type=nullable_str,
265
            default=None,
266
            help='The specific revision to use for the model code on '
267
268
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
269
270
        parser.add_argument(
            '--tokenizer-revision',
271
            type=nullable_str,
272
            default=None,
273
274
275
            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.')
276
277
278
279
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
280
            choices=['auto', 'slow', 'mistral'],
281
282
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
283
284
            'always use the slow tokenizer. \n* '
            '"mistral" will always use the `mistral_common` tokenizer.')
285
286
        parser.add_argument('--trust-remote-code',
                            action='store_true',
287
                            help='Trust remote code from huggingface.')
288
289
290
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
291
292
293
294
            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.")
295
        parser.add_argument('--download-dir',
296
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
297
                            default=EngineArgs.download_dir,
298
                            help='Directory to download and load the weights, '
299
                            'default to the default cache dir of '
300
                            'huggingface.')
301
302
303
304
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
305
            choices=[f.value for f in LoadFormat],
306
307
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
308
            'and fall back to the pytorch bin format if safetensors format '
309
310
311
312
313
314
315
316
            '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 '
317
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
318
319
320
            'section for more information.\n'
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
321
322
323
324
325
326
327
        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 ')
328
329
330
331
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
332
333
334
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
335
336
337
338
339
340
341
342
            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.')
343
344
345
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
346
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
347
            default=EngineArgs.kv_cache_dtype,
348
            help='Data type for kv cache storage. If "auto", will use model '
349
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
zhuwenwen's avatar
zhuwenwen committed
350
            'ROCm (hcu) supports fp8 (=fp8_e4m3)')
351
352
        parser.add_argument(
            '--quantization-param-path',
353
            type=nullable_str,
354
355
356
357
358
            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. '
359
            'FP8_E5M2 (without scaling) is only supported on cuda version '
360
            'greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead '
361
            'supported for common inference criteria.')
362
363
        parser.add_argument('--max-model-len',
                            type=int,
364
                            default=EngineArgs.max_model_len,
365
366
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
367
368
369
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
370
371
            default='xgrammar',
            choices=['outlines', 'lm-format-enforcer', 'xgrammar'],
372
            help='Which engine will be used for guided decoding'
373
            ' (JSON schema / regex etc) by default. Currently support '
374
375
            'https://github.com/outlines-dev/outlines,'
            'https://github.com/mlc-ai/xgrammar, and '
376
377
378
            'https://github.com/noamgat/lm-format-enforcer.'
            ' Can be overridden per request via guided_decoding_backend'
            ' parameter.')
379
380
381
382
383
384
385
386
        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.')
387
        # Parallel arguments
388
389
390
391
        parser.add_argument(
            '--distributed-executor-backend',
            choices=['ray', 'mp'],
            default=EngineArgs.distributed_executor_backend,
392
393
394
395
396
397
398
399
            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 '
            'and hpu only support Ray for distributed inference.')

400
401
402
403
        parser.add_argument(
            '--worker-use-ray',
            action='store_true',
            help='Deprecated, use --distributed-executor-backend=ray.')
404
405
406
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
407
                            default=EngineArgs.pipeline_parallel_size,
408
                            help='Number of pipeline stages.')
409
410
411
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
412
                            default=EngineArgs.tensor_parallel_size,
413
                            help='Number of tensor parallel replicas.')
414
415
416
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
417
            default=EngineArgs.max_parallel_loading_workers,
418
            help='Load model sequentially in multiple batches, '
419
            'to avoid RAM OOM when using tensor '
420
            'parallel and large models.')
421
422
423
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
424
            help='If specified, use nsight to profile Ray workers.')
425
        # KV cache arguments
426
427
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
428
                            default=EngineArgs.block_size,
429
                            choices=[8, 16, 32],
430
                            help='Token block size for contiguous chunks of '
431
432
                            'tokens. This is ignored on neuron devices and '
                            'set to max-model-len')
433

434
435
436
437
438
439
440
        parser.add_argument(
            "--enable-prefix-caching",
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.enable_prefix_caching,
            help="Enables automatic prefix caching. "
            "Use --no-enable-prefix-caching to disable explicitly.",
        )
441
442
443
444
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
                            'capping to sliding window size')
445
446
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
447
                            default=True,
448
449
450
451
452
                            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.')
453
454
455
456
457
458
459
460
        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.')
461

462
463
464
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
465
                            help='Random seed for operations.')
466
        parser.add_argument('--swap-space',
467
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
468
                            default=EngineArgs.swap_space,
469
                            help='CPU swap space size (GiB) per GPU.')
470
471
472
473
474
475
476
477
478
        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 '
479
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
480
            'which requires at least 26GB GPU memory. Note that this '
481
            'requires fast CPU-GPU interconnect, as part of the model is '
482
483
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
484
485
486
487
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
488
489
490
            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, '
491
492
493
494
495
496
            '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.')
497
        parser.add_argument(
498
            '--num-gpu-blocks-override',
499
500
501
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
502
            ' of GPU blocks. Used for testing preemption.')
503
504
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
505
                            default=EngineArgs.max_num_batched_tokens,
506
507
                            help='Maximum number of batched tokens per '
                            'iteration.')
508
509
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
510
                            default=EngineArgs.max_num_seqs,
511
                            help='Maximum number of sequences per iteration.')
512
513
514
515
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
516
517
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
518
519
        parser.add_argument('--disable-log-stats',
                            action='store_true',
520
                            help='Disable logging statistics.')
521
522
523
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
524
                            type=nullable_str,
525
                            choices=[*QUANTIZATION_METHODS, None],
526
                            default=EngineArgs.quantization,
527
528
529
530
531
532
                            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.')
533
534
535
536
537
538
        parser.add_argument(
            '--rope-scaling',
            default=None,
            type=json.loads,
            help='RoPE scaling configuration in JSON format. '
            'For example, {"rope_type":"dynamic","factor":2.0}')
539
540
541
542
543
544
        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.')
545
546
547
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
548
                            help='Extra arguments for the HuggingFace config. '
549
550
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
551
552
553
554
555
        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.')
556
        parser.add_argument('--max-seq-len-to-capture',
557
558
559
560
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
561
562
563
564
                            '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.')
565
566
567
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
568
                            help='See ParallelConfig.')
569
570
571
572
573
574
575
576
577
578
579
580
581
        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',
582
                            type=nullable_str,
583
584
585
586
587
                            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.')
588
589
590
591
592
593
594
595
596
597
598
599
600
601

        # 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.'))
602
603
604
605
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
606
            help=('Overrides for the multimodal input mapping/processing, '
607
                  'e.g., image processor. For example: {"num_crops": 4}.'))
608
609
610
611
612
613
        parser.add_argument(
            '--mm-cache-preprocessor',
            action='store_true',
            help='If true, then enables caching of the multi-modal '
            'preprocessor/mapper. Otherwise, the mapper executes each time'
            ', and for better performance consider enabling frontend process.')
614

615
616
617
618
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
619
620
621
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
622
623
624
625
626
627
628
629
        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.')
630
631
632
633
634
635
636
637
        parser.add_argument('--merge-lora',
                            type=bool,
                            default=False,
                            help='If set to True, the weights of the base layer will be merged with the weights of Lora.')
        parser.add_argument('--lora-target-modules',
                            nargs='*',
                            default=None,
                            help='List of lora module name, If not specified, modules will be chosen according to the model architecture.')
638
639
640
641
642
643
644
645
646
647
648
        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,
649
            choices=['auto', 'float16', 'bfloat16'],
650
651
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
652
653
654
655
656
657
658
659
660
661
662
        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.'))
663
664
665
666
667
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
668
669
                  'Must be >= than max_loras. '
                  'Defaults to max_loras.'))
670
671
672
673
674
675
676
677
        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.'))
678
679
680
681
682
683
684
685
686
687
688
        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')
689
690
691
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
692
                            choices=DEVICE_OPTIONS,
693
                            help='Device type for vLLM execution.')
694
695
696
697
698
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
699

700
701
        parser.add_argument(
            '--multi-step-stream-outputs',
702
703
704
705
706
707
            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')
708
709
710
711
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
712
            help='Apply a delay (of delay factor multiplied by previous '
713
            'prompt latency) before scheduling next prompt.')
714
715
        parser.add_argument(
            '--enable-chunked-prefill',
716
717
718
719
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
720
            help='If set, the prefill requests can be chunked based on the '
721
            'max_num_batched_tokens.')
722
723
724

        parser.add_argument(
            '--speculative-model',
725
            type=nullable_str,
726
            default=EngineArgs.speculative_model,
727
728
            help=
            'The name of the draft model to be used in speculative decoding.')
729
730
731
732
733
734
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
735
            help='Method used to quantize the weights of speculative model. '
736
737
738
739
740
            '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.')
741
742
743
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
744
            default=EngineArgs.num_speculative_tokens,
745
            help='The number of speculative tokens to sample from '
746
            'the draft model in speculative decoding.')
747
748
749
750
751
752
        parser.add_argument(
            '--num-speculative-heads',
            type=int,
            default=EngineArgs.num_speculative_heads,
            help='The number of speculative heads to sample from '
                 'the draft model in speculative decoding.')
753
754
755
756
757
758
        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')
759
760
761
762
763
764
765
        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.')
766

767
768
        parser.add_argument(
            '--speculative-max-model-len',
769
            type=int,
770
771
772
773
774
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

775
776
777
778
779
780
781
        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.')

782
783
784
785
786
787
788
789
790
791
792
793
794
795
        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.')

796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
        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')

828
829
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
830
            action=StoreBoolean,
831
            default=EngineArgs.disable_logprobs_during_spec_decoding,
832
833
            nargs="?",
            const="True",
834
835
836
837
838
839
840
841
            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.')

842
        parser.add_argument('--model-loader-extra-config',
843
                            type=nullable_str,
844
845
846
847
848
849
                            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.')
850
851
852
853
854
855
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
856
            "Default to `original/**/*` to avoid repeated loading of llama's "
857
            "checkpoints.")
858
        parser.add_argument(
859
            '--preemption-mode',
860
861
            type=str,
            default=None,
862
863
864
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
865

866
867
868
869
870
871
872
873
874
875
        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 "
876
            "same as the `--model` argument. Noted that this name(s) "
877
            "will also be used in `model_name` tag content of "
878
            "prometheus metrics, if multiple names provided, metrics "
879
            "tag will take the first one.")
880
881
882
883
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
884
885
886
887
888
889

        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
890
891
892
893
894
895
896
897
898
899
        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.")
900

901
902
903
904
905
906
        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.")
907

908
909
910
911
912
913
914
915
916
917
        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).')

918
919
        parser.add_argument(
            '--override-neuron-config',
920
            type=json.loads,
921
            default=None,
922
923
            help="Override or set neuron device configuration. "
            "e.g. {\"cast_logits_dtype\": \"bloat16\"}.'")
924
        parser.add_argument(
925
926
            '--override-pooler-config',
            type=PoolerConfig.from_json,
927
            default=None,
928
            help="Override or set the pooling method for pooling models. "
929
            "e.g. {\"pooling_type\": \"mean\", \"normalize\": false}.'")
930

931
932
933
934
935
936
937
938
939
940
941
942
        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, '
943
944
945
946
                            'use a JSON string.\n'
                            'Following the convention of traditional '
                            'compilers, using -O without space is also '
                            'supported. -O3 is equivalent to -O 3.')
947

948
949
950
951
952
953
        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.')

954
955
956
957
958
        parser.add_argument(
            '--worker-cls',
            type=str,
            default="auto",
            help='The worker class to use for distributed execution.')
959

960
        return parser
961
962

    @classmethod
963
    def from_cli_args(cls, args: argparse.Namespace):
964
965
966
        # 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
967
968
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
969

970
971
    def create_model_config(self) -> ModelConfig:
        return ModelConfig(
972
            model=self.model,
973
            task=self.task,
974
975
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
976
977
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
978
            allowed_local_media_path=self.allowed_local_media_path,
979
980
981
982
983
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
984
            rope_theta=self.rope_theta,
985
            hf_overrides=self.hf_overrides,
986
987
988
989
990
991
992
993
994
            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_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,
995
            served_model_name=self.served_model_name,
996
            limit_mm_per_prompt=self.limit_mm_per_prompt,
997
            use_async_output_proc=not self.disable_async_output_proc,
998
            config_format=self.config_format,
999
            mm_processor_kwargs=self.mm_processor_kwargs,
1000
            mm_cache_preprocessor=self.mm_cache_preprocessor,
1001
1002
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
1003
            logits_processor_pattern=self.logits_processor_pattern)
1004

1005
1006
1007
1008
1009
1010
1011
1012
    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,
        )

1013
1014
1015
1016
1017
1018
    def create_engine_config(self,
                             usage_context: Optional[UsageContext] = None
                             ) -> VllmConfig:
        if envs.VLLM_USE_V1:
            self._override_v1_engine_args(usage_context)

1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
        # 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()

1046
1047
1048
1049
1050
1051
1052
        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

1053
        cache_config = CacheConfig(
1054
            block_size=self.block_size,
1055
1056
1057
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1058
            is_attention_free=model_config.is_attention_free,
1059
1060
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1061
1062
1063
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
        )
1064
        parallel_config = ParallelConfig(
1065
1066
1067
1068
1069
1070
            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(
1071
1072
1073
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1074
            ),
1075
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1076
1077
1078
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
        )
1079

1080
1081
1082
1083
1084
1085
        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.
1086

1087
1088
1089
1090
1091
1092
            # For multimodal models, chunked prefill is disabled by default in
            # V0, but enabled by design in V1
            if model_config.is_multimodal_model:
                self.enable_chunked_prefill = bool(envs.VLLM_USE_V1)

            elif use_long_context:
1093
1094
1095
1096
1097
1098
1099
                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
                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
                        and not self.enable_prompt_adapter
1100
1101
                        and model_config.runner_type != "pooling"
                        and not current_platform.is_rocm()):
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
                    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)
1118
1119
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
1120
            msg = "Chunked prefill is not supported for pooling models"
1121
            raise ValueError(msg)
1122

1123

1124
1125
1126
1127
1128
        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,
1129
1130
            speculative_model_quantization = \
                self.speculative_model_quantization,
1131
1132
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
1133
            num_speculative_tokens=self.num_speculative_tokens,
1134
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
1135
1136
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
1137
1138
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
1139
            disable_log_stats=self.disable_log_stats,
1140
1141
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
1142
1143
1144
1145
1146
1147
            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,
1148
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
1149
            num_speculative_heads=self.num_speculative_heads
1150
1151
        )

1152
        # Reminder: Please update docs/source/usage/compatibility_matrix.rst
1153
        # If the feature combo become valid
1154
1155
1156
1157
        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)")
1158
1159
1160
            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")
1161
1162
1163
1164
1165
1166
1167
1168
1169

        # 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

1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
        if not self.use_v2_block_manager:
            logger.warning(
                "[DEPRECATED] Block manager v1 has been removed, "
                "and setting --use-v2-block-manager to True or False has "
                "no effect on vLLM behavior. Please remove "
                "--use-v2-block-manager in your engine argument. "
                "If your use case is not supported by "
                "SelfAttnBlockSpaceManager (i.e. block manager v2),"
                " please file an issue with detailed information.")

1180
        scheduler_config = SchedulerConfig(
1181
            runner_type=model_config.runner_type,
1182
1183
1184
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1185
            num_lookahead_slots=num_lookahead_slots,
1186
1187
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1188
            is_multimodal_model=model_config.is_multimodal_model,
1189
            preemption_mode=self.preemption_mode,
1190
            num_scheduler_steps=self.num_scheduler_steps,
1191
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1192
1193
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1194
            policy=self.scheduling_policy)
1195
        lora_config = LoRAConfig(
1196
            bias_enabled=self.enable_lora_bias,
1197
1198
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1199
            fully_sharded_loras=self.fully_sharded_loras,
1200
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1201
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1202
1203
            lora_dtype=self.lora_dtype,
            max_cpu_loras=self.max_cpu_loras if self.max_cpu_loras
1204
1205
1206
            and self.max_cpu_loras > 0 else None,
            merge_lora=self.merge_lora,
            lora_target_modules=self.lora_target_modules) if self.enable_lora else None
1207

1208
1209
1210
1211
1212
1213
1214
        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

1215
        load_config = self.create_load_config()
1216

1217
1218
1219
1220
1221
        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

1222
1223
1224
        decoding_config = DecodingConfig(
            guided_decoding_backend=self.guided_decoding_backend)

1225
1226
1227
1228
1229
1230
1231
1232
        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}")
1233
        observability_config = ObservabilityConfig(
1234
1235
1236
1237
1238
1239
            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,
        )
1240

1241
        config = VllmConfig(
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
            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,
1252
            prompt_adapter_config=prompt_adapter_config,
1253
            compilation_config=self.compilation_config,
1254
            kv_transfer_config=self.kv_transfer_config,
1255
        )
1256

1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
        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"

1267
1268
1269
1270
1271
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
        # When no user override, set the default values based on the usage
        # context.
        # TODO(woosuk): Tune the default values for different hardware.
1272
1273
1274
1275
1276
        if self.max_num_batched_tokens is None:
            if usage_context == UsageContext.LLM_CLASS:
                self.max_num_batched_tokens = 8192
            elif usage_context == UsageContext.OPENAI_API_SERVER:
                self.max_num_batched_tokens = 2048
1277
1278
1279
            logger.warning(
                "Setting max_num_batched_tokens to %d for %s usage context.",
                self.max_num_batched_tokens, usage_context.value)
1280
1281
1282
1283
1284
1285
1286

    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"
        if engine_config.model_config.is_multimodal_model:
1287
1288
            # TODO (ywang96): Enable APC by default when VLM supports it.
            assert not engine_config.cache_config.enable_prefix_caching
1289

1290

1291
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1292
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1293
    """Arguments for asynchronous vLLM engine."""
1294
    disable_log_requests: bool = False
1295
1296

    @staticmethod
1297
1298
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1299
1300
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1301
1302
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1303
                            help='Disable logging requests.')
1304
        return parser
1305
1306
1307
1308


# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1309
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1310
1311
1312


def _async_engine_args_parser():
1313
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1314
                                        async_args_only=True)