arg_utils.py 56.6 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, ConfigFormat, DecodingConfig,
12
13
14
15
16
                         DeviceConfig, LoadConfig, LoadFormat, LoRAConfig,
                         ModelConfig, ObservabilityConfig, ParallelConfig,
                         PromptAdapterConfig, SchedulerConfig,
                         SpeculativeConfig, TaskOption, TokenizerPoolConfig,
                         VllmConfig)
17
from vllm.executor.executor_base import ExecutorBase
18
from vllm.logger import init_logger
19
from vllm.model_executor.layers.pooler import PoolingType
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
    chat_template_text_format: str = 'string'
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
112
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
113
    max_parallel_loading_workers: Optional[int] = None
114
115
116
    # 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
117
    enable_prefix_caching: bool = False
118
    disable_sliding_window: bool = False
119
    use_v2_block_manager: bool = True
120
121
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
122
    gpu_memory_utilization: float = 0.90
123
    max_num_batched_tokens: Optional[int] = None
124
    max_num_seqs: int = 256
125
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
126
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
127
    revision: Optional[str] = None
128
    code_revision: Optional[str] = None
129
    rope_scaling: Optional[Dict[str, Any]] = None
130
    rope_theta: Optional[float] = None
131
    hf_overrides: Optional[Dict[str, Any]] = None
132
    tokenizer_revision: Optional[str] = None
133
    quantization: Optional[str] = None
134
    enforce_eager: Optional[bool] = None
135
    max_seq_len_to_capture: int = 8192
136
    disable_custom_all_reduce: bool = False
137
    tokenizer_pool_size: int = 0
138
139
140
141
    # 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"
142
    tokenizer_pool_extra_config: Optional[Dict[str, Any]] = None
143
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
144
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
145
146
147
    enable_lora: bool = False
    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
    override_neuron_config: Optional[Dict[str, Any]] = None
190
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
191

192
193
194
195
196
197
198
    # Pooling configuration.
    pooling_type: Optional[str] = None
    pooling_norm: Optional[bool] = None
    pooling_softmax: Optional[bool] = None
    pooling_step_tag_id: Optional[int] = None
    pooling_returned_token_ids: Optional[List[int]] = None

199
    def __post_init__(self):
200
        if not self.tokenizer:
201
            self.tokenizer = self.model
202
203

        # Setup plugins
204
205
        from vllm.plugins import load_general_plugins
        load_general_plugins()
206
207

    @staticmethod
208
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
209
        """Shared CLI arguments for vLLM engine."""
210

211
        # Model arguments
212
213
214
        parser.add_argument(
            '--model',
            type=str,
215
            default=EngineArgs.model,
216
            help='Name or path of the huggingface model to use.')
217
218
219
220
221
222
223
224
225
        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.')
226
227
        parser.add_argument(
            '--tokenizer',
228
            type=nullable_str,
229
            default=EngineArgs.tokenizer,
230
231
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
232
233
234
235
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
            help='Skip initialization of tokenizer and detokenizer')
Jasmond L's avatar
Jasmond L committed
236
237
        parser.add_argument(
            '--revision',
238
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
239
            default=None,
240
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
241
242
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
243
244
        parser.add_argument(
            '--code-revision',
245
            type=nullable_str,
246
            default=None,
247
            help='The specific revision to use for the model code on '
248
249
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
250
251
        parser.add_argument(
            '--tokenizer-revision',
252
            type=nullable_str,
253
            default=None,
254
255
256
            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.')
257
258
259
260
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
261
            choices=['auto', 'slow', 'mistral'],
262
263
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
264
265
            'always use the slow tokenizer. \n* '
            '"mistral" will always use the `mistral_common` tokenizer.')
266
267
268
269
270
271
272
273
        parser.add_argument(
            '--chat-template-text-format',
            type=str,
            default=EngineArgs.chat_template_text_format,
            choices=['string', 'openai'],
            help='The format to render text content within a chat template. '
            '"string" will keep the content field as a string whereas '
            '"openai" will parse content in the current OpenAI format.')
274
275
        parser.add_argument('--trust-remote-code',
                            action='store_true',
276
                            help='Trust remote code from huggingface.')
277
278
279
280
281
282
283
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
            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")
284
        parser.add_argument('--download-dir',
285
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
286
                            default=EngineArgs.download_dir,
287
                            help='Directory to download and load the weights, '
288
                            'default to the default cache dir of '
289
                            'huggingface.')
290
291
292
293
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
294
            choices=[f.value for f in LoadFormat],
295
296
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
297
            'and fall back to the pytorch bin format if safetensors format '
298
299
300
301
302
303
304
305
            '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 '
306
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
307
308
309
            'section for more information.\n'
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
310
311
312
313
314
315
316
        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 ')
317
318
319
320
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
321
322
323
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
324
325
326
327
328
329
330
331
            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.')
332
333
334
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
335
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
336
            default=EngineArgs.kv_cache_dtype,
337
            help='Data type for kv cache storage. If "auto", will use model '
338
339
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
340
341
        parser.add_argument(
            '--quantization-param-path',
342
            type=nullable_str,
343
344
345
346
347
348
349
            default=None,
            help='Path to the JSON file containing the KV cache '
            'scaling factors. This should generally be supplied, when '
            'KV cache dtype is FP8. Otherwise, KV cache scaling factors '
            'default to 1.0, which may cause accuracy issues. '
            'FP8_E5M2 (without scaling) is only supported on cuda version'
            'greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead '
350
            'supported for common inference criteria.')
351
352
        parser.add_argument('--max-model-len',
                            type=int,
353
                            default=EngineArgs.max_model_len,
354
355
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
356
357
358
359
360
361
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
            default='outlines',
            choices=['outlines', 'lm-format-enforcer'],
            help='Which engine will be used for guided decoding'
362
363
364
365
366
            ' (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.')
367
        # Parallel arguments
368
369
370
371
        parser.add_argument(
            '--distributed-executor-backend',
            choices=['ray', 'mp'],
            default=EngineArgs.distributed_executor_backend,
372
373
374
375
376
377
378
379
            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.')

380
381
382
383
        parser.add_argument(
            '--worker-use-ray',
            action='store_true',
            help='Deprecated, use --distributed-executor-backend=ray.')
384
385
386
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
387
                            default=EngineArgs.pipeline_parallel_size,
388
                            help='Number of pipeline stages.')
389
390
391
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
392
                            default=EngineArgs.tensor_parallel_size,
393
                            help='Number of tensor parallel replicas.')
394
395
396
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
397
            default=EngineArgs.max_parallel_loading_workers,
398
            help='Load model sequentially in multiple batches, '
399
            'to avoid RAM OOM when using tensor '
400
            'parallel and large models.')
401
402
403
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
404
            help='If specified, use nsight to profile Ray workers.')
405
        # KV cache arguments
406
407
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
408
                            default=EngineArgs.block_size,
409
                            choices=[8, 16, 32, 64, 128],
410
                            help='Token block size for contiguous chunks of '
411
412
                            'tokens. This is ignored on neuron devices and '
                            'set to max-model-len')
413
414
415

        parser.add_argument('--enable-prefix-caching',
                            action='store_true',
416
                            help='Enables automatic prefix caching.')
417
418
419
420
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
                            'capping to sliding window size')
421
422
423
424
425
426
427
        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.')
428
429
430
431
432
433
434
435
        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.')
436

437
438
439
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
440
                            help='Random seed for operations.')
441
        parser.add_argument('--swap-space',
442
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
443
                            default=EngineArgs.swap_space,
444
                            help='CPU swap space size (GiB) per GPU.')
445
446
447
448
449
450
451
452
453
454
455
456
457
458
        parser.add_argument(
            '--cpu-offload-gb',
            type=float,
            default=0,
            help='The space in GiB to offload to CPU, per GPU. '
            'Default is 0, which means no offloading. Intuitively, '
            'this argument can be seen as a virtual way to increase '
            'the GPU memory size. For example, if you have one 24 GB '
            'GPU and set this to 10, virtually you can think of it as '
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight,'
            'which requires at least 26GB GPU memory. Note that this '
            'requires fast CPU-GPU interconnect, as part of the model is'
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
459
460
461
462
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
463
464
465
            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, '
466
467
468
469
470
            '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.')
471
        parser.add_argument(
472
            '--num-gpu-blocks-override',
473
474
475
476
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
            'of GPU blocks. Used for testing preemption.')
477
478
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
479
                            default=EngineArgs.max_num_batched_tokens,
480
481
                            help='Maximum number of batched tokens per '
                            'iteration.')
482
483
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
484
                            default=EngineArgs.max_num_seqs,
485
                            help='Maximum number of sequences per iteration.')
486
487
488
489
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
490
491
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
492
493
        parser.add_argument('--disable-log-stats',
                            action='store_true',
494
                            help='Disable logging statistics.')
495
496
497
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
498
                            type=nullable_str,
499
                            choices=[*QUANTIZATION_METHODS, None],
500
                            default=EngineArgs.quantization,
501
502
503
504
505
506
                            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.')
507
508
509
510
511
512
        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}')
513
514
515
516
517
518
        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.')
519
520
521
522
523
524
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
                            help='Extra arguments for the HuggingFace config.'
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
525
526
527
528
529
        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.')
530
        parser.add_argument('--max-seq-len-to-capture',
531
532
533
534
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
535
536
537
538
                            '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.')
539
540
541
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
542
                            help='See ParallelConfig.')
543
544
545
546
547
548
549
550
551
552
553
554
555
        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',
556
                            type=nullable_str,
557
558
559
560
561
                            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.')
562
563
564
565
566
567
568
569
570
571
572
573
574
575

        # 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.'))
576
577
578
579
580
581
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
            help=('Overrides for the multimodal input mapping/processing,'
                  'e.g., image processor. For example: {"num_crops": 4}.'))
582

583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
        parser.add_argument('--max-loras',
                            type=int,
                            default=EngineArgs.max_loras,
                            help='Max number of LoRAs in a single batch.')
        parser.add_argument('--max-lora-rank',
                            type=int,
                            default=EngineArgs.max_lora_rank,
                            help='Max LoRA rank.')
        parser.add_argument(
            '--lora-extra-vocab-size',
            type=int,
            default=EngineArgs.lora_extra_vocab_size,
            help=('Maximum size of extra vocabulary that can be '
                  'present in a LoRA adapter (added to the base '
                  'model vocabulary).'))
        parser.add_argument(
            '--lora-dtype',
            type=str,
            default=EngineArgs.lora_dtype,
            choices=['auto', 'float16', 'bfloat16', 'float32'],
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
609
610
611
612
613
614
615
616
617
618
619
        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.'))
620
621
622
623
624
625
626
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
                  'Must be >= than max_num_seqs. '
                  'Defaults to max_num_seqs.'))
627
628
629
630
631
632
633
634
        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.'))
635
636
637
638
639
640
641
642
643
644
645
        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')
646
647
648
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
649
                            choices=DEVICE_OPTIONS,
650
                            help='Device type for vLLM execution.')
651
652
653
654
655
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
656

657
658
        parser.add_argument(
            '--multi-step-stream-outputs',
659
660
661
662
663
664
            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')
665
666
667
668
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
669
            help='Apply a delay (of delay factor multiplied by previous '
670
            'prompt latency) before scheduling next prompt.')
671
672
        parser.add_argument(
            '--enable-chunked-prefill',
673
674
675
676
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
677
            help='If set, the prefill requests can be chunked based on the '
678
            'max_num_batched_tokens.')
679
680
681

        parser.add_argument(
            '--speculative-model',
682
            type=nullable_str,
683
            default=EngineArgs.speculative_model,
684
685
            help=
            'The name of the draft model to be used in speculative decoding.')
686
687
688
689
690
691
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
692
            help='Method used to quantize the weights of speculative model. '
693
694
695
696
697
            '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.')
698
699
700
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
701
            default=EngineArgs.num_speculative_tokens,
702
            help='The number of speculative tokens to sample from '
703
            'the draft model in speculative decoding.')
704
705
706
707
708
709
        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')
710
711
712
713
714
715
716
        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.')
717

718
719
        parser.add_argument(
            '--speculative-max-model-len',
720
            type=int,
721
722
723
724
725
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

726
727
728
729
730
731
732
        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.')

733
734
735
736
737
738
739
740
741
742
743
744
745
746
        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.')

747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
        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')

779
780
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
781
            action=StoreBoolean,
782
            default=EngineArgs.disable_logprobs_during_spec_decoding,
783
784
            nargs="?",
            const="True",
785
786
787
788
789
790
791
792
            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.')

793
        parser.add_argument('--model-loader-extra-config',
794
                            type=nullable_str,
795
796
797
798
799
800
                            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.')
801
802
803
804
805
806
807
808
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
            "Default to 'original/**/*' to avoid repeated loading of llama's "
            "checkpoints.")
809
        parser.add_argument(
810
            '--preemption-mode',
811
812
            type=str,
            default=None,
813
814
815
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
816

817
818
819
820
821
822
823
824
825
826
827
828
829
830
        parser.add_argument(
            "--served-model-name",
            nargs="+",
            type=str,
            default=None,
            help="The model name(s) used in the API. If multiple "
            "names are provided, the server will respond to any "
            "of the provided names. The model name in the model "
            "field of a response will be the first name in this "
            "list. If not specified, the model name will be the "
            "same as the `--model` argument. Noted that this name(s)"
            "will also be used in `model_name` tag content of "
            "prometheus metrics, if multiple names provided, metrics"
            "tag will take the first one.")
831
832
833
834
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
835
836
837
838
839
840

        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
841
842
843
844
845
846
847
848
849
850
        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.")
851

852
853
854
855
856
857
        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.")
858
859
        parser.add_argument(
            '--override-neuron-config',
860
            type=json.loads,
861
            default=None,
862
863
            help="Override or set neuron device configuration. "
            "e.g. {\"cast_logits_dtype\": \"bloat16\"}.'")
864

865
866
867
868
869
870
871
872
873
874
        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).')

875
876
        parser.add_argument(
            '--pooling-type',
877
            choices=[pt.name for pt in PoolingType],
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
            default=None,
            help='Used to configure the pooling method in the embedding model.'
        )

        parser.add_argument('--pooling-norm',
                            default=None,
                            action='store_true',
                            help="Used to determine whether to normalize "
                            "the pooled data in the embedding model.")

        parser.add_argument('--no-pooling-norm',
                            default=None,
                            action='store_false',
                            dest='pooling_norm',
                            help="Used to determine whether to normalize "
                            "the pooled data in the embedding model.")

        parser.add_argument('--pooling-softmax',
                            default=None,
                            action='store_true',
                            help="Used to determine whether to softmax "
                            "the pooled data in the embedding model.")

        parser.add_argument('--no-pooling-softmax',
                            default=None,
                            action='store_false',
                            dest='pooling_softmax',
                            help="Used to determine whether to softmax "
                            "the pooled data in the embedding model.")

        parser.add_argument(
            '--pooling-step-tag-id',
            type=int,
            default=None,
            help="When pooling-step-tag-id is not -1, it indicates "
            "that the score corresponding to the step-tag-ids in the "
            "generated sentence should be returned. Otherwise, it "
            "returns the scores for all tokens.")

        parser.add_argument(
            '--pooling-returned-token-ids',
            nargs='+',
            type=int,
            default=None,
            help="pooling-returned-token-ids represents a list of "
            "indices for the vocabulary dimensions to be extracted, "
            "such as the token IDs of good_token and bad_token in "
            "the math-shepherd-mistral-7b-prm model.")

927
        return parser
928
929

    @classmethod
930
    def from_cli_args(cls, args: argparse.Namespace):
931
932
933
        # 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
934
935
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
936

937
938
    def create_model_config(self) -> ModelConfig:
        return ModelConfig(
939
            model=self.model,
940
            task=self.task,
941
942
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
943
            tokenizer_mode=self.tokenizer_mode,
944
            chat_template_text_format=self.chat_template_text_format,
945
            trust_remote_code=self.trust_remote_code,
946
            allowed_local_media_path=self.allowed_local_media_path,
947
948
949
950
951
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
952
            rope_theta=self.rope_theta,
953
            hf_overrides=self.hf_overrides,
954
955
956
957
958
959
960
961
962
            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,
963
            served_model_name=self.served_model_name,
964
            limit_mm_per_prompt=self.limit_mm_per_prompt,
965
            use_async_output_proc=not self.disable_async_output_proc,
966
967
            override_neuron_config=self.override_neuron_config,
            config_format=self.config_format,
968
            mm_processor_kwargs=self.mm_processor_kwargs,
969
970
971
972
973
            pooling_type=self.pooling_type,
            pooling_norm=self.pooling_norm,
            pooling_softmax=self.pooling_softmax,
            pooling_step_tag_id=self.pooling_step_tag_id,
            pooling_returned_token_ids=self.pooling_returned_token_ids,
974
975
        )

976
977
978
979
980
981
982
983
    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,
        )

984
    def create_engine_config(self) -> VllmConfig:
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
        # 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()

1012
1013
1014
1015
1016
1017
1018
        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

1019
        cache_config = CacheConfig(
1020
            # neuron needs block_size = max_model_len
1021
            block_size=self.block_size if self.device != "neuron" else
1022
            (self.max_model_len if self.max_model_len is not None else 0),
1023
1024
1025
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1026
            is_attention_free=model_config.is_attention_free,
1027
1028
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1029
1030
1031
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
        )
1032
        parallel_config = ParallelConfig(
1033
1034
1035
1036
1037
1038
            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(
1039
1040
1041
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1042
            ),
1043
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1044
            distributed_executor_backend=self.distributed_executor_backend)
1045

1046
1047
1048
1049
1050
1051
        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.
1052
1053
1054
1055

            # Chunked prefill is currently disabled for multimodal models by
            # default.
            if use_long_context and not model_config.is_multimodal_model:
1056
1057
1058
1059
1060
1061
                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
1062
                        and not self.enable_prompt_adapter):
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
                    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)

1080
1081
1082
1083
1084
        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,
1085
1086
            speculative_model_quantization = \
                self.speculative_model_quantization,
1087
1088
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
1089
            num_speculative_tokens=self.num_speculative_tokens,
1090
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
1091
1092
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
1093
1094
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
1095
            disable_log_stats=self.disable_log_stats,
1096
1097
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
1098
1099
1100
1101
1102
1103
            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,
1104
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
1105
1106
        )

1107
1108
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
1109
1110
1111
1112
        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)")
1113
1114
1115
            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")
1116
1117
1118
1119
1120
1121
1122
1123
1124

        # 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

1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
        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.")

1135
        scheduler_config = SchedulerConfig(
1136
            task=model_config.task,
1137
1138
1139
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1140
            num_lookahead_slots=num_lookahead_slots,
1141
1142
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1143
            is_multimodal_model=model_config.is_multimodal_model,
1144
            preemption_mode=self.preemption_mode,
1145
            num_scheduler_steps=self.num_scheduler_steps,
1146
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1147
1148
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1149
            policy=self.scheduling_policy)
1150
1151
1152
        lora_config = LoRAConfig(
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1153
            fully_sharded_loras=self.fully_sharded_loras,
1154
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1155
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1156
1157
1158
            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
1159

1160
1161
1162
1163
1164
1165
1166
        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

1167
        load_config = self.create_load_config()
1168

1169
1170
1171
1172
1173
        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

1174
1175
1176
        decoding_config = DecodingConfig(
            guided_decoding_backend=self.guided_decoding_backend)

1177
1178
1179
1180
1181
1182
1183
1184
        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}")
1185
        observability_config = ObservabilityConfig(
1186
1187
1188
1189
1190
1191
            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,
        )
1192

1193
        return VllmConfig(
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
            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,
1204
            prompt_adapter_config=prompt_adapter_config,
1205
        )
1206
1207


1208
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1209
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1210
    """Arguments for asynchronous vLLM engine."""
1211
    disable_log_requests: bool = False
1212
1213

    @staticmethod
1214
1215
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1216
1217
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1218
1219
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1220
                            help='Disable logging requests.')
1221
        return parser
1222
1223
1224
1225


# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1226
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1227
1228
1229


def _async_engine_args_parser():
1230
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1231
                                        async_args_only=True)