arg_utils.py 55.4 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
12
13
14
15
16
17
from vllm.config import (CacheConfig, CompilationConfig, ConfigFormat,
                         DecodingConfig, DeviceConfig, HfOverrides, LoadConfig,
                         LoadFormat, LoRAConfig, ModelConfig,
                         ObservabilityConfig, ParallelConfig, PoolerConfig,
                         PromptAdapterConfig, SchedulerConfig,
                         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.utils import FlexibleArgumentParser, StoreBoolean
24

25
if TYPE_CHECKING:
26
    from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
27

28
29
logger = init_logger(__name__)

30
31
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]

32
33
34
35
36
37
38
39
DEVICE_OPTIONS = [
    "auto",
    "cuda",
    "neuron",
    "cpu",
    "openvino",
    "tpu",
    "xpu",
40
    "hpu",
41
42
]

43

44
45
46
47
48
49
def nullable_str(val: str):
    if not val or val == "None":
        return None
    return val


50
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
51
52
53
54
55
56
57
58
59
    """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.
    """
60
61
62
63
64
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
65
66
67
68
69
        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
70
71

        try:
72
            parsed_value = int(value)
73
74
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
75
76
77
78
79
80
            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
81
82
83
84

    return out_dict


85
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
86
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
87
    """Arguments for vLLM engine."""
88
    model: str = 'facebook/opt-125m'
89
    served_model_name: Optional[Union[str, List[str]]] = None
90
    tokenizer: Optional[str] = None
91
    task: TaskOption = "auto"
92
    skip_tokenizer_init: bool = False
93
    tokenizer_mode: str = 'auto'
94
    trust_remote_code: bool = False
95
    allowed_local_media_path: str = ""
96
    download_dir: Optional[str] = None
97
    load_format: str = 'auto'
98
    config_format: ConfigFormat = ConfigFormat.AUTO
99
    dtype: str = 'auto'
100
    kv_cache_dtype: str = 'auto'
101
    quantization_param_path: Optional[str] = None
102
    seed: int = 0
103
    max_model_len: Optional[int] = None
104
    worker_use_ray: bool = False
105
106
107
108
109
    # 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
110
111
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
112
    max_parallel_loading_workers: Optional[int] = None
113
114
115
    # NOTE(kzawora): default block size for Gaudi should be 128
    # smaller sizes still work, but very inefficiently
    block_size: int = 16 if not current_platform.is_hpu() else 128
116
    enable_prefix_caching: bool = False
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: int = 256
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
    enable_lora: bool = False
145
    enable_lora_bias: bool = False
146
147
    max_loras: int = 1
    max_lora_rank: int = 16
148
149
150
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
151
    fully_sharded_loras: bool = False
152
    lora_extra_vocab_size: int = 256
153
    long_lora_scaling_factors: Optional[Tuple[float]] = None
154
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
155
    max_cpu_loras: Optional[int] = None
156
    device: str = 'auto'
157
    num_scheduler_steps: int = 1
158
    multi_step_stream_outputs: bool = True
159
    ray_workers_use_nsight: bool = False
160
    num_gpu_blocks_override: Optional[int] = None
161
    num_lookahead_slots: int = 0
162
    model_loader_extra_config: Optional[dict] = None
163
    ignore_patterns: Optional[Union[str, List[str]]] = None
164
    preemption_mode: Optional[str] = None
165

166
    scheduler_delay_factor: float = 0.0
167
    enable_chunked_prefill: Optional[bool] = None
168

169
    guided_decoding_backend: str = 'outlines'
170
171
    # Speculative decoding configuration.
    speculative_model: Optional[str] = None
172
    speculative_model_quantization: Optional[str] = None
173
    speculative_draft_tensor_parallel_size: Optional[int] = None
174
    num_speculative_tokens: Optional[int] = None
175
    speculative_disable_mqa_scorer: Optional[bool] = False
176
    speculative_max_model_len: Optional[int] = None
177
    speculative_disable_by_batch_size: Optional[int] = None
178
179
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
180
181
182
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
183
    qlora_adapter_name_or_path: Optional[str] = None
184
    disable_logprobs_during_spec_decoding: Optional[bool] = None
185

186
    otlp_traces_endpoint: Optional[str] = None
187
    collect_detailed_traces: Optional[str] = None
188
    disable_async_output_proc: bool = False
189
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
190

191
192
    override_neuron_config: Optional[Dict[str, Any]] = None
    override_pooler_config: Optional[PoolerConfig] = None
193
    compilation_config: Optional[CompilationConfig] = None
194

195
    def __post_init__(self):
196
        if not self.tokenizer:
197
            self.tokenizer = self.model
198
199

        # Setup plugins
200
201
        from vllm.plugins import load_general_plugins
        load_general_plugins()
202
203

    @staticmethod
204
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
205
        """Shared CLI arguments for vLLM engine."""
206

207
        # Model arguments
208
209
210
        parser.add_argument(
            '--model',
            type=str,
211
            default=EngineArgs.model,
212
            help='Name or path of the huggingface model to use.')
213
214
215
216
217
218
219
220
221
        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.')
222
223
        parser.add_argument(
            '--tokenizer',
224
            type=nullable_str,
225
            default=EngineArgs.tokenizer,
226
227
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
228
229
230
231
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
            help='Skip initialization of tokenizer and detokenizer')
Jasmond L's avatar
Jasmond L committed
232
233
        parser.add_argument(
            '--revision',
234
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
235
            default=None,
236
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
237
238
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
239
240
        parser.add_argument(
            '--code-revision',
241
            type=nullable_str,
242
            default=None,
243
            help='The specific revision to use for the model code on '
244
245
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
246
247
        parser.add_argument(
            '--tokenizer-revision',
248
            type=nullable_str,
249
            default=None,
250
251
252
            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.')
253
254
255
256
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
257
            choices=['auto', 'slow', 'mistral'],
258
259
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
260
261
            'always use the slow tokenizer. \n* '
            '"mistral" will always use the `mistral_common` tokenizer.')
262
263
        parser.add_argument('--trust-remote-code',
                            action='store_true',
264
                            help='Trust remote code from huggingface.')
265
266
267
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
268
269
270
271
            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.")
272
        parser.add_argument('--download-dir',
273
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
274
                            default=EngineArgs.download_dir,
275
                            help='Directory to download and load the weights, '
276
                            'default to the default cache dir of '
277
                            'huggingface.')
278
279
280
281
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
282
            choices=[f.value for f in LoadFormat],
283
284
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
285
            'and fall back to the pytorch bin format if safetensors format '
286
287
288
289
290
291
292
293
            '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 '
294
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
295
296
297
            'section for more information.\n'
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
298
299
300
301
302
303
304
        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 ')
305
306
307
308
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
309
310
311
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
312
313
314
315
316
317
318
319
            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.')
320
321
322
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
323
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
324
            default=EngineArgs.kv_cache_dtype,
325
            help='Data type for kv cache storage. If "auto", will use model '
326
327
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
328
329
        parser.add_argument(
            '--quantization-param-path',
330
            type=nullable_str,
331
332
333
334
335
            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. '
336
            'FP8_E5M2 (without scaling) is only supported on cuda version '
337
            'greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead '
338
            'supported for common inference criteria.')
339
340
        parser.add_argument('--max-model-len',
                            type=int,
341
                            default=EngineArgs.max_model_len,
342
343
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
344
345
346
347
348
349
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
            default='outlines',
            choices=['outlines', 'lm-format-enforcer'],
            help='Which engine will be used for guided decoding'
350
351
352
353
354
            ' (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.')
355
        # Parallel arguments
356
357
358
359
        parser.add_argument(
            '--distributed-executor-backend',
            choices=['ray', 'mp'],
            default=EngineArgs.distributed_executor_backend,
360
361
362
363
364
365
366
367
            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.')

368
369
370
371
        parser.add_argument(
            '--worker-use-ray',
            action='store_true',
            help='Deprecated, use --distributed-executor-backend=ray.')
372
373
374
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
375
                            default=EngineArgs.pipeline_parallel_size,
376
                            help='Number of pipeline stages.')
377
378
379
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
380
                            default=EngineArgs.tensor_parallel_size,
381
                            help='Number of tensor parallel replicas.')
382
383
384
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
385
            default=EngineArgs.max_parallel_loading_workers,
386
            help='Load model sequentially in multiple batches, '
387
            'to avoid RAM OOM when using tensor '
388
            'parallel and large models.')
389
390
391
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
392
            help='If specified, use nsight to profile Ray workers.')
393
        # KV cache arguments
394
395
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
396
                            default=EngineArgs.block_size,
397
                            choices=[8, 16, 32, 64, 128],
398
                            help='Token block size for contiguous chunks of '
399
400
                            'tokens. This is ignored on neuron devices and '
                            'set to max-model-len')
401
402
403

        parser.add_argument('--enable-prefix-caching',
                            action='store_true',
404
                            help='Enables automatic prefix caching.')
405
406
407
408
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
                            'capping to sliding window size')
409
410
411
412
413
414
415
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
                            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.')
416
417
418
419
420
421
422
423
        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.')
424

425
426
427
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
428
                            help='Random seed for operations.')
429
        parser.add_argument('--swap-space',
430
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
431
                            default=EngineArgs.swap_space,
432
                            help='CPU swap space size (GiB) per GPU.')
433
434
435
436
437
438
439
440
441
        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 '
442
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
443
            'which requires at least 26GB GPU memory. Note that this '
444
            'requires fast CPU-GPU interconnect, as part of the model is '
445
446
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
447
448
449
450
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
451
452
453
            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, '
454
455
456
457
458
            'will use the default value of 0.9. This is a global gpu memory '
            'utilization limit, for example if 50%% of the gpu memory is '
            'already used before vLLM starts and --gpu-memory-utilization is '
            'set to 0.9, then only 40%% of the gpu memory will be allocated '
            'to the model executor.')
459
        parser.add_argument(
460
            '--num-gpu-blocks-override',
461
462
463
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
464
            ' of GPU blocks. Used for testing preemption.')
465
466
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
467
                            default=EngineArgs.max_num_batched_tokens,
468
469
                            help='Maximum number of batched tokens per '
                            'iteration.')
470
471
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
472
                            default=EngineArgs.max_num_seqs,
473
                            help='Maximum number of sequences per iteration.')
474
475
476
477
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
478
479
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
480
481
        parser.add_argument('--disable-log-stats',
                            action='store_true',
482
                            help='Disable logging statistics.')
483
484
485
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
486
                            type=nullable_str,
487
                            choices=[*QUANTIZATION_METHODS, None],
488
                            default=EngineArgs.quantization,
489
490
491
492
493
494
                            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.')
495
496
497
498
499
500
        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}')
501
502
503
504
505
506
        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.')
507
508
509
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
510
                            help='Extra arguments for the HuggingFace config. '
511
512
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
513
514
515
516
517
        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.')
518
        parser.add_argument('--max-seq-len-to-capture',
519
520
521
522
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
523
524
525
526
                            '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.')
527
528
529
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
530
                            help='See ParallelConfig.')
531
532
533
534
535
536
537
538
539
540
541
542
543
        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',
544
                            type=nullable_str,
545
546
547
548
549
                            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.')
550
551
552
553
554
555
556
557
558
559
560
561
562
563

        # 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.'))
564
565
566
567
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
568
            help=('Overrides for the multimodal input mapping/processing, '
569
                  'e.g., image processor. For example: {"num_crops": 4}.'))
570

571
572
573
574
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
575
576
577
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
        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,
597
            choices=['auto', 'float16', 'bfloat16'],
598
599
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
600
601
602
603
604
605
606
607
608
609
610
        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.'))
611
612
613
614
615
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
616
617
                  'Must be >= than max_loras. '
                  'Defaults to max_loras.'))
618
619
620
621
622
623
624
625
        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.'))
626
627
628
629
630
631
632
633
634
635
636
        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')
637
638
639
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
640
                            choices=DEVICE_OPTIONS,
641
                            help='Device type for vLLM execution.')
642
643
644
645
646
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
647

648
649
        parser.add_argument(
            '--multi-step-stream-outputs',
650
651
652
653
654
655
            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')
656
657
658
659
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
660
            help='Apply a delay (of delay factor multiplied by previous '
661
            'prompt latency) before scheduling next prompt.')
662
663
        parser.add_argument(
            '--enable-chunked-prefill',
664
665
666
667
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
668
            help='If set, the prefill requests can be chunked based on the '
669
            'max_num_batched_tokens.')
670
671
672

        parser.add_argument(
            '--speculative-model',
673
            type=nullable_str,
674
            default=EngineArgs.speculative_model,
675
676
            help=
            'The name of the draft model to be used in speculative decoding.')
677
678
679
680
681
682
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
683
            help='Method used to quantize the weights of speculative model. '
684
685
686
687
688
            '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.')
689
690
691
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
692
            default=EngineArgs.num_speculative_tokens,
693
            help='The number of speculative tokens to sample from '
694
            'the draft model in speculative decoding.')
695
696
697
698
699
700
        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')
701
702
703
704
705
706
707
        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.')
708

709
710
        parser.add_argument(
            '--speculative-max-model-len',
711
            type=int,
712
713
714
715
716
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

717
718
719
720
721
722
723
        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.')

724
725
726
727
728
729
730
731
732
733
734
735
736
737
        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.')

738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
        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')

770
771
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
772
            action=StoreBoolean,
773
            default=EngineArgs.disable_logprobs_during_spec_decoding,
774
775
            nargs="?",
            const="True",
776
777
778
779
780
781
782
783
            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.')

784
        parser.add_argument('--model-loader-extra-config',
785
                            type=nullable_str,
786
787
788
789
790
791
                            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.')
792
793
794
795
796
797
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
798
            "Default to `original/**/*` to avoid repeated loading of llama's "
799
            "checkpoints.")
800
        parser.add_argument(
801
            '--preemption-mode',
802
803
            type=str,
            default=None,
804
805
806
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
807

808
809
810
811
812
813
814
815
816
817
        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 "
818
            "same as the `--model` argument. Noted that this name(s) "
819
            "will also be used in `model_name` tag content of "
820
            "prometheus metrics, if multiple names provided, metrics "
821
            "tag will take the first one.")
822
823
824
825
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
826
827
828
829
830
831

        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
832
833
834
835
836
837
838
839
840
841
        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.")
842

843
844
845
846
847
848
        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.")
849

850
851
852
853
854
855
856
857
858
859
        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).')

860
        parser.add_argument(
861
862
            '--override-neuron-config',
            type=json.loads,
863
            default=None,
864
865
            help="Override or set neuron device configuration. "
            "e.g. {\"cast_logits_dtype\": \"bloat16\"}.'")
866
        parser.add_argument(
867
868
            '--override-pooler-config',
            type=PoolerConfig.from_json,
869
            default=None,
870
871
            help="Override or set the pooling method in the embedding model. "
            "e.g. {\"pooling_type\": \"mean\", \"normalize\": false}.'")
872

873
874
875
876
877
878
879
880
881
882
883
884
        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, '
885
886
887
888
                            'use a JSON string.\n'
                            'Following the convention of traditional '
                            'compilers, using -O without space is also '
                            'supported. -O3 is equivalent to -O 3.')
889

890
        return parser
891
892

    @classmethod
893
    def from_cli_args(cls, args: argparse.Namespace):
894
895
896
        # 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
897
898
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
899

900
901
    def create_model_config(self) -> ModelConfig:
        return ModelConfig(
902
            model=self.model,
903
            task=self.task,
904
905
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
906
907
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
908
            allowed_local_media_path=self.allowed_local_media_path,
909
910
911
912
913
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
914
            rope_theta=self.rope_theta,
915
            hf_overrides=self.hf_overrides,
916
917
918
919
920
921
922
923
924
            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,
925
            served_model_name=self.served_model_name,
926
            limit_mm_per_prompt=self.limit_mm_per_prompt,
927
            use_async_output_proc=not self.disable_async_output_proc,
928
            config_format=self.config_format,
929
            mm_processor_kwargs=self.mm_processor_kwargs,
930
931
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
932
933
        )

934
935
936
937
938
939
940
941
    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,
        )

942
    def create_engine_config(self) -> VllmConfig:
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
        # 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()

970
971
972
973
974
975
976
        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

977
        cache_config = CacheConfig(
978
            # neuron needs block_size = max_model_len
979
            block_size=self.block_size if self.device != "neuron" else
980
            (self.max_model_len if self.max_model_len is not None else 0),
981
982
983
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
984
            is_attention_free=model_config.is_attention_free,
985
986
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
987
988
989
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
        )
990
        parallel_config = ParallelConfig(
991
992
993
994
995
996
            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(
997
998
999
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1000
            ),
1001
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1002
            distributed_executor_backend=self.distributed_executor_backend)
1003

1004
1005
1006
1007
1008
1009
        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.
1010
1011
1012
1013

            # Chunked prefill is currently disabled for multimodal models by
            # default.
            if use_long_context and not model_config.is_multimodal_model:
1014
1015
1016
1017
1018
1019
                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
1020
1021
                        and not self.enable_prompt_adapter
                        and model_config.task != "embedding"):
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
                    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)
1038
1039
1040
        elif self.enable_chunked_prefill and model_config.task == "embedding":
            msg = "Chunked prefill is not supported for embedding models"
            raise ValueError(msg)
1041

1042
1043
1044
1045
1046
        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,
1047
1048
            speculative_model_quantization = \
                self.speculative_model_quantization,
1049
1050
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
1051
            num_speculative_tokens=self.num_speculative_tokens,
1052
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
1053
1054
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
1055
1056
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
1057
            disable_log_stats=self.disable_log_stats,
1058
1059
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
1060
1061
1062
1063
1064
1065
            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,
1066
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
1067
1068
        )

1069
1070
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
1071
1072
1073
1074
        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)")
1075
1076
1077
            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")
1078
1079
1080
1081
1082
1083
1084
1085
1086

        # 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

1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
        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.")

1097
        scheduler_config = SchedulerConfig(
1098
            task=model_config.task,
1099
1100
1101
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1102
            num_lookahead_slots=num_lookahead_slots,
1103
1104
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1105
            is_multimodal_model=model_config.is_multimodal_model,
1106
            preemption_mode=self.preemption_mode,
1107
            num_scheduler_steps=self.num_scheduler_steps,
1108
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1109
1110
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1111
            policy=self.scheduling_policy)
1112
        lora_config = LoRAConfig(
1113
            bias_enabled=self.enable_lora_bias,
1114
1115
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1116
            fully_sharded_loras=self.fully_sharded_loras,
1117
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1118
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1119
1120
1121
            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
1122

1123
1124
1125
1126
1127
1128
1129
        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

1130
        load_config = self.create_load_config()
1131

1132
1133
1134
1135
1136
        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

1137
1138
1139
        decoding_config = DecodingConfig(
            guided_decoding_backend=self.guided_decoding_backend)

1140
1141
1142
1143
1144
1145
1146
1147
        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}")
1148
        observability_config = ObservabilityConfig(
1149
1150
1151
1152
1153
1154
            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,
        )
1155

1156
        return VllmConfig(
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
            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,
1167
            prompt_adapter_config=prompt_adapter_config,
1168
            compilation_config=self.compilation_config,
1169
        )
1170
1171


1172
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1173
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1174
    """Arguments for asynchronous vLLM engine."""
1175
    disable_log_requests: bool = False
1176
1177

    @staticmethod
1178
1179
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1180
1181
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1182
1183
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1184
                            help='Disable logging requests.')
1185
        return parser
1186
1187
1188
1189


# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1190
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1191
1192
1193


def _async_engine_args_parser():
1194
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1195
                                        async_args_only=True)