arg_utils.py 63.6 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

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

10
11
import torch

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

27
if TYPE_CHECKING:
28
    from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
29

30
31
logger = init_logger(__name__)

32
33
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]

34
35
36
37
38
39
40
41
DEVICE_OPTIONS = [
    "auto",
    "cuda",
    "neuron",
    "cpu",
    "openvino",
    "tpu",
    "xpu",
42
    "hpu",
43
44
]

45

46
47
48
49
50
51
def nullable_str(val: str):
    if not val or val == "None":
        return None
    return val


52
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
53
54
55
56
57
58
59
60
61
    """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.
    """
62
63
64
65
66
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
67
68
69
70
71
        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
72
73

        try:
74
            parsed_value = int(value)
75
76
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
77
78
79
80
81
82
            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
83
84
85
86

    return out_dict


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

168
    scheduler_delay_factor: float = 0.0
169
    enable_chunked_prefill: Optional[bool] = None
170

171
    guided_decoding_backend: str = 'xgrammar'
172
    logits_processor_pattern: Optional[str] = None
173
174
    # Speculative decoding configuration.
    speculative_model: Optional[str] = None
175
    speculative_model_quantization: Optional[str] = None
176
    speculative_draft_tensor_parallel_size: Optional[int] = None
177
    num_speculative_tokens: Optional[int] = None
178
    num_speculative_heads: Optional[int] = None
179
    speculative_disable_mqa_scorer: Optional[bool] = False
180
    speculative_max_model_len: Optional[int] = None
181
    speculative_disable_by_batch_size: Optional[int] = None
182
183
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
184
185
186
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
187
    qlora_adapter_name_or_path: Optional[str] = None
188
    disable_logprobs_during_spec_decoding: Optional[bool] = None
189

190
    otlp_traces_endpoint: Optional[str] = None
191
    collect_detailed_traces: Optional[str] = None
192
    disable_async_output_proc: bool = False
193
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
194

195
    override_neuron_config: Optional[Dict[str, Any]] = None
196
    override_pooler_config: Optional[PoolerConfig] = None
197
    compilation_config: Optional[CompilationConfig] = None
198
    worker_cls: str = "auto"
199

200
    kv_transfer_config: Optional[KVTransferConfig] = None
201

202
    generation_config: Optional[str] = None
203
    override_generation_config: Optional[Dict[str, Any]] = None
204
    enable_sleep_mode: bool = False
205
    model_impl: str = "auto"
206

207
    calculate_kv_scales: Optional[bool] = None
208

209
    def __post_init__(self):
210
        if not self.tokenizer:
211
            self.tokenizer = self.model
212

213
214
215
216
        # Override the default value of enable_prefix_caching if it's not set
        # by user.
        if self.enable_prefix_caching is None:
            self.enable_prefix_caching = bool(envs.VLLM_USE_V1)
217

218
219
220
        # Override max_num_seqs if it's not set by user.
        if self.max_num_seqs is None:
            self.max_num_seqs = 256 if not envs.VLLM_USE_V1 else 1024
221

222
223
224
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
225
        if isinstance(self.compilation_config, (int, dict)):
226
227
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
228

229
        # Setup plugins
230
231
        from vllm.plugins import load_general_plugins
        load_general_plugins()
232
233

    @staticmethod
234
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
235
        """Shared CLI arguments for vLLM engine."""
236

237
        # Model arguments
238
239
240
        parser.add_argument(
            '--model',
            type=str,
241
            default=EngineArgs.model,
242
            help='Name or path of the huggingface model to use.')
243
244
245
246
247
248
        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 '
249
            'multiple tasks. When the model only supports one task, ``"auto"`` '
250
251
            'can be used to select it; otherwise, you must specify explicitly '
            'which task to use.')
252
253
        parser.add_argument(
            '--tokenizer',
254
            type=nullable_str,
255
            default=EngineArgs.tokenizer,
256
257
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
258
259
260
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
261
            help='Skip initialization of tokenizer and detokenizer.')
Jasmond L's avatar
Jasmond L committed
262
263
        parser.add_argument(
            '--revision',
264
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
265
            default=None,
266
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
267
268
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
269
270
        parser.add_argument(
            '--code-revision',
271
            type=nullable_str,
272
            default=None,
273
            help='The specific revision to use for the model code on '
274
275
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
276
277
        parser.add_argument(
            '--tokenizer-revision',
278
            type=nullable_str,
279
            default=None,
280
281
282
            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.')
283
284
285
286
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
287
            choices=['auto', 'slow', 'mistral'],
288
289
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
290
291
            'always use the slow tokenizer. \n* '
            '"mistral" will always use the `mistral_common` tokenizer.')
292
293
        parser.add_argument('--trust-remote-code',
                            action='store_true',
294
                            help='Trust remote code from huggingface.')
295
296
297
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
298
299
300
301
            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.")
302
        parser.add_argument('--download-dir',
303
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
304
                            default=EngineArgs.download_dir,
305
                            help='Directory to download and load the weights, '
306
                            'default to the default cache dir of '
307
                            'huggingface.')
308
309
310
311
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
312
            choices=[f.value for f in LoadFormat],
313
314
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
315
            'and fall back to the pytorch bin format if safetensors format '
316
317
318
319
320
321
322
323
            '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 '
324
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
325
            'section for more information.\n'
326
327
            '* "runai_streamer" will load the Safetensors weights using Run:ai'
            'Model Streamer \n'
328
329
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
330
331
332
333
334
335
336
        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 ')
337
338
339
340
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
341
342
343
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
344
345
346
347
348
349
350
351
            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.')
352
353
354
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
355
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
356
            default=EngineArgs.kv_cache_dtype,
357
            help='Data type for kv cache storage. If "auto", will use model '
358
359
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
360
361
        parser.add_argument('--max-model-len',
                            type=int,
362
                            default=EngineArgs.max_model_len,
363
364
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
365
366
367
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
368
369
            default='xgrammar',
            choices=['outlines', 'lm-format-enforcer', 'xgrammar'],
370
            help='Which engine will be used for guided decoding'
371
            ' (JSON schema / regex etc) by default. Currently support '
372
            'https://github.com/outlines-dev/outlines, '
373
            'https://github.com/mlc-ai/xgrammar, and '
374
375
376
            'https://github.com/noamgat/lm-format-enforcer.'
            ' Can be overridden per request via guided_decoding_backend'
            ' parameter.')
377
378
379
380
381
382
383
384
        parser.add_argument(
            '--logits-processor-pattern',
            type=nullable_str,
            default=None,
            help='Optional regex pattern specifying valid logits processor '
            'qualified names that can be passed with the `logits_processors` '
            'extra completion argument. Defaults to None, which allows no '
            'processors.')
385
386
387
388
389
390
391
392
393
394
395
396
        parser.add_argument(
            '--model-impl',
            type=str,
            default=EngineArgs.model_impl,
            choices=[f.value for f in ModelImpl],
            help='Which implementation of the model to use.\n\n'
            '* "auto" will try to use the vLLM implementation if it exists '
            'and fall back to the Transformers implementation if no vLLM '
            'implementation is available.\n'
            '* "vllm" will use the vLLM model implementation.\n'
            '* "transformers" will use the Transformers model '
            'implementation.\n')
397
        # Parallel arguments
398
399
        parser.add_argument(
            '--distributed-executor-backend',
400
            choices=['ray', 'mp', 'uni', 'external_launcher'],
401
            default=EngineArgs.distributed_executor_backend,
402
403
404
405
406
407
            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 '
408
            'only supports Ray for distributed inference.')
409

410
411
412
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
413
                            default=EngineArgs.pipeline_parallel_size,
414
                            help='Number of pipeline stages.')
415
416
417
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
418
                            default=EngineArgs.tensor_parallel_size,
419
                            help='Number of tensor parallel replicas.')
420
421
422
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
423
            default=EngineArgs.max_parallel_loading_workers,
424
            help='Load model sequentially in multiple batches, '
425
            'to avoid RAM OOM when using tensor '
426
            'parallel and large models.')
427
428
429
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
430
            help='If specified, use nsight to profile Ray workers.')
431
        # KV cache arguments
432
433
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
434
                            default=EngineArgs.block_size,
435
                            choices=[8, 16, 32, 64, 128],
436
                            help='Token block size for contiguous chunks of '
437
                            'tokens. This is ignored on neuron devices and '
438
                            'set to ``--max-model-len``. On CUDA devices, '
439
440
                            'only block sizes up to 32 are supported. '
                            'On HPU devices, block size defaults to 128.')
441

442
443
444
445
446
        parser.add_argument(
            "--enable-prefix-caching",
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.enable_prefix_caching,
            help="Enables automatic prefix caching. "
447
            "Use ``--no-enable-prefix-caching`` to disable explicitly.",
448
        )
449
450
451
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
452
                            'capping to sliding window size.')
453
454
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
455
                            default=True,
456
457
458
459
460
                            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.')
461
462
463
464
465
466
467
468
        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.')
469

470
471
472
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
473
                            help='Random seed for operations.')
474
        parser.add_argument('--swap-space',
475
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
476
                            default=EngineArgs.swap_space,
477
                            help='CPU swap space size (GiB) per GPU.')
478
479
480
481
482
483
484
485
486
        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 '
487
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
488
            'which requires at least 26GB GPU memory. Note that this '
489
            'requires fast CPU-GPU interconnect, as part of the model is '
490
491
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
492
493
494
495
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
496
497
498
            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, '
499
500
501
502
503
504
            'will use the default value of 0.9. This is a per-instance '
            'limit, and only applies to the current vLLM instance.'
            'It does not matter if you have another vLLM instance running '
            'on the same GPU. For example, if you have two vLLM instances '
            'running on the same GPU, you can set the GPU memory utilization '
            'to 0.5 for each instance.')
505
        parser.add_argument(
506
            '--num-gpu-blocks-override',
507
508
509
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
510
            ' of GPU blocks. Used for testing preemption.')
511
512
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
513
                            default=EngineArgs.max_num_batched_tokens,
514
515
                            help='Maximum number of batched tokens per '
                            'iteration.')
516
517
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
518
                            default=EngineArgs.max_num_seqs,
519
                            help='Maximum number of sequences per iteration.')
520
521
522
523
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
524
525
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
526
527
        parser.add_argument('--disable-log-stats',
                            action='store_true',
528
                            help='Disable logging statistics.')
529
530
531
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
532
                            type=nullable_str,
533
                            choices=[*QUANTIZATION_METHODS, None],
534
                            default=EngineArgs.quantization,
535
536
537
538
539
540
                            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.')
541
542
543
544
545
        parser.add_argument(
            '--rope-scaling',
            default=None,
            type=json.loads,
            help='RoPE scaling configuration in JSON format. '
546
            'For example, ``{"rope_type":"dynamic","factor":2.0}``')
547
548
549
550
551
552
        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.')
553
554
555
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
556
                            help='Extra arguments for the HuggingFace config. '
557
558
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
559
560
561
562
563
        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.')
564
        parser.add_argument('--max-seq-len-to-capture',
565
566
567
568
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
569
570
571
572
                            '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.')
573
574
575
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
576
                            help='See ParallelConfig.')
577
578
579
580
581
582
583
584
585
586
587
588
589
        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',
590
                            type=nullable_str,
591
592
593
594
595
                            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.')
596
597
598
599
600
601
602
603
604
605
606
607
608
609

        # 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.'))
610
611
612
613
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
614
            help=('Overrides for the multimodal input mapping/processing, '
615
                  'e.g., image processor. For example: ``{"num_crops": 4}``.'))
616
        parser.add_argument(
617
            '--disable-mm-preprocessor-cache',
618
            action='store_true',
619
620
            help='If true, then disables caching of the multi-modal '
            'preprocessor/mapper. (not recommended)')
621

622
623
624
625
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
626
627
628
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
629
630
631
632
633
634
635
636
        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.')
637
638
639
640
641
642
643
644
        parser.add_argument('--merge-lora',
                            type=bool,
                            default=False,
                            help='If set to True, the weights of the base layer will be merged with the weights of Lora.')
        parser.add_argument('--lora-target-modules',
                            nargs='*',
                            default=None,
                            help='List of lora module name, If not specified, modules will be chosen according to the model architecture.')
645
646
647
648
649
650
651
652
653
654
655
        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,
656
            choices=['auto', 'float16', 'bfloat16'],
657
658
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
659
660
661
662
663
664
665
666
667
668
669
        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.'))
670
671
672
673
674
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
675
676
                  'Must be >= than max_loras. '
                  'Defaults to max_loras.'))
677
678
679
680
681
682
683
684
        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.'))
685
686
687
688
689
690
691
692
693
694
695
        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')
696
697
698
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
699
                            choices=DEVICE_OPTIONS,
700
                            help='Device type for vLLM execution.')
701
702
703
704
705
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
706

707
708
        parser.add_argument(
            '--multi-step-stream-outputs',
709
710
711
712
713
714
            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')
715
716
717
718
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
719
            help='Apply a delay (of delay factor multiplied by previous '
720
            'prompt latency) before scheduling next prompt.')
721
722
        parser.add_argument(
            '--enable-chunked-prefill',
723
724
725
726
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
727
            help='If set, the prefill requests can be chunked based on the '
728
            'max_num_batched_tokens.')
729
730
731

        parser.add_argument(
            '--speculative-model',
732
            type=nullable_str,
733
            default=EngineArgs.speculative_model,
734
735
            help=
            'The name of the draft model to be used in speculative decoding.')
736
737
738
739
740
741
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
742
            help='Method used to quantize the weights of speculative model. '
743
744
745
746
747
            '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.')
748
749
750
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
751
            default=EngineArgs.num_speculative_tokens,
752
            help='The number of speculative tokens to sample from '
753
            'the draft model in speculative decoding.')
754
755
756
757
758
759
        parser.add_argument(
            '--num-speculative-heads',
            type=int,
            default=EngineArgs.num_speculative_heads,
            help='The number of speculative heads to sample from '
                 'the draft model in speculative decoding.')
760
761
762
763
764
765
        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')
766
767
768
769
770
771
772
        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.')
773

774
775
        parser.add_argument(
            '--speculative-max-model-len',
776
            type=int,
777
778
779
780
781
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

782
783
784
785
786
787
788
        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.')

789
790
791
792
793
794
795
796
797
798
799
800
801
802
        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.')

803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
        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')

835
836
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
837
            action=StoreBoolean,
838
            default=EngineArgs.disable_logprobs_during_spec_decoding,
839
840
            nargs="?",
            const="True",
841
842
843
844
845
846
847
848
            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.')

849
        parser.add_argument('--model-loader-extra-config',
850
                            type=nullable_str,
851
852
853
854
855
856
                            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.')
857
858
859
860
861
862
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
863
            "Default to `original/**/*` to avoid repeated loading of llama's "
864
            "checkpoints.")
865
        parser.add_argument(
866
            '--preemption-mode',
867
868
            type=str,
            default=None,
869
870
871
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
872

873
874
875
876
877
878
879
880
881
882
        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 "
883
            "same as the ``--model`` argument. Noted that this name(s) "
884
            "will also be used in `model_name` tag content of "
885
            "prometheus metrics, if multiple names provided, metrics "
886
            "tag will take the first one.")
887
888
889
890
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
891
892
893
894
895
896

        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
897
898
899
900
901
902
        parser.add_argument(
            '--collect-detailed-traces',
            type=str,
            default=None,
            help="Valid choices are " +
            ",".join(ALLOWED_DETAILED_TRACE_MODULES) +
903
            ". It makes sense to set this only if ``--otlp-traces-endpoint`` is"
904
905
906
            " 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.")
907

908
909
910
911
912
913
        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.")
914

915
916
917
918
919
920
921
922
923
924
        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).')

925
926
        parser.add_argument(
            '--override-neuron-config',
927
            type=json.loads,
928
            default=None,
929
            help="Override or set neuron device configuration. "
930
            "e.g. ``{\"cast_logits_dtype\": \"bloat16\"}``.")
931
        parser.add_argument(
932
933
            '--override-pooler-config',
            type=PoolerConfig.from_json,
934
            default=None,
935
            help="Override or set the pooling method for pooling models. "
936
            "e.g. ``{\"pooling_type\": \"mean\", \"normalize\": false}``.")
937

938
939
940
941
942
943
944
945
946
947
948
949
        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, '
950
951
952
953
                            'use a JSON string.\n'
                            'Following the convention of traditional '
                            'compilers, using -O without space is also '
                            'supported. -O3 is equivalent to -O 3.')
954

955
956
957
958
959
960
        parser.add_argument('--kv-transfer-config',
                            type=KVTransferConfig.from_cli,
                            default=None,
                            help='The configurations for distributed KV cache '
                            'transfer. Should be a JSON string.')

961
962
963
964
965
        parser.add_argument(
            '--worker-cls',
            type=str,
            default="auto",
            help='The worker class to use for distributed execution.')
966
967
968
969
970
        parser.add_argument(
            "--generation-config",
            type=nullable_str,
            default=None,
            help="The folder path to the generation config. "
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
            "Defaults to None, no generation config is loaded, vLLM defaults "
            "will be used. If set to 'auto', the generation config will be "
            "loaded from model path. If set to a folder path, the generation "
            "config will be loaded from the specified folder path. If "
            "`max_new_tokens` is specified in generation config, then "
            "it sets a server-wide limit on the number of output tokens "
            "for all requests.")

        parser.add_argument(
            "--override-generation-config",
            type=json.loads,
            default=None,
            help="Overrides or sets generation config in JSON format. "
            "e.g. ``{\"temperature\": 0.5}``. If used with "
            "--generation-config=auto, the override parameters will be merged "
            "with the default config from the model. If generation-config is "
            "None, only the override parameters are used.")
988

989
990
991
992
993
994
        parser.add_argument("--enable-sleep-mode",
                            action="store_true",
                            default=False,
                            help="Enable sleep mode for the engine. "
                            "(only cuda platform is supported)")

995
996
997
998
999
1000
1001
1002
        parser.add_argument(
            '--calculate-kv-scales',
            action='store_true',
            help='This enables dynamic calculation of '
            'k_scale and v_scale when kv-cache-dtype is fp8. '
            'If calculate-kv-scales is false, the scales will '
            'be loaded from the model checkpoint if available. '
            'Otherwise, the scales will default to 1.0.')
1003

1004
        return parser
1005
1006

    @classmethod
1007
    def from_cli_args(cls, args: argparse.Namespace):
1008
1009
1010
        # 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
1011
1012
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
1013

1014
1015
    def create_model_config(self) -> ModelConfig:
        return ModelConfig(
1016
            model=self.model,
1017
            task=self.task,
1018
1019
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
1020
1021
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1022
            allowed_local_media_path=self.allowed_local_media_path,
1023
1024
1025
1026
1027
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
1028
            rope_theta=self.rope_theta,
1029
            hf_overrides=self.hf_overrides,
1030
1031
1032
1033
1034
1035
1036
1037
            tokenizer_revision=self.tokenizer_revision,
            max_model_len=self.max_model_len,
            quantization=self.quantization,
            enforce_eager=self.enforce_eager,
            max_seq_len_to_capture=self.max_seq_len_to_capture,
            max_logprobs=self.max_logprobs,
            disable_sliding_window=self.disable_sliding_window,
            skip_tokenizer_init=self.skip_tokenizer_init,
1038
            served_model_name=self.served_model_name,
1039
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1040
            use_async_output_proc=not self.disable_async_output_proc,
1041
            config_format=self.config_format,
1042
            mm_processor_kwargs=self.mm_processor_kwargs,
1043
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
1044
1045
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
1046
            logits_processor_pattern=self.logits_processor_pattern,
1047
            generation_config=self.generation_config,
1048
            override_generation_config=self.override_generation_config,
1049
            enable_sleep_mode=self.enable_sleep_mode,
1050
            model_impl=self.model_impl,
1051
        )
1052

1053
1054
1055
1056
1057
1058
1059
1060
    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,
        )

1061
1062
1063
1064
1065
1066
    def create_engine_config(self,
                             usage_context: Optional[UsageContext] = None
                             ) -> VllmConfig:
        if envs.VLLM_USE_V1:
            self._override_v1_engine_args(usage_context)

1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
        # 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()

1094
1095
1096
1097
1098
        if (model_config.is_multimodal_model and not envs.VLLM_USE_V1
                and self.enable_prefix_caching):
            logger.warning("--enable-prefix-caching is currently not "
                           "supported for multimodal models in v0 and "
                           "has been disabled.")
1099
1100
            self.enable_prefix_caching = False

1101
        cache_config = CacheConfig(
1102
            block_size=self.block_size,
1103
1104
1105
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1106
            is_attention_free=model_config.is_attention_free,
1107
1108
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1109
1110
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
1111
            calculate_kv_scales=self.calculate_kv_scales,
1112
        )
1113
        parallel_config = ParallelConfig(
1114
1115
1116
1117
1118
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
            max_parallel_loading_workers=self.max_parallel_loading_workers,
            disable_custom_all_reduce=self.disable_custom_all_reduce,
            tokenizer_pool_config=TokenizerPoolConfig.create_config(
1119
1120
1121
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1122
            ),
1123
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1124
1125
1126
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
        )
1127

1128
1129
1130
1131
1132
1133
        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.
1134

1135
1136
1137
1138
1139
1140
            # For multimodal models, chunked prefill is disabled by default in
            # V0, but enabled by design in V1
            if model_config.is_multimodal_model:
                self.enable_chunked_prefill = bool(envs.VLLM_USE_V1)

            elif use_long_context:
1141
1142
1143
1144
                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
1145
                from vllm.platforms import current_platform
1146
1147
1148
                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
                        and not self.enable_prompt_adapter
1149
1150
                        and model_config.runner_type != "pooling"
                        and not current_platform.is_rocm()):
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
                    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)
1167
1168
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
1169
            msg = "Chunked prefill is not supported for pooling models"
1170
            raise ValueError(msg)
1171

1172

1173
1174
1175
1176
1177
        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,
1178
1179
            speculative_model_quantization = \
                self.speculative_model_quantization,
1180
1181
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
1182
            num_speculative_tokens=self.num_speculative_tokens,
1183
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
1184
1185
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
1186
1187
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
1188
            disable_log_stats=self.disable_log_stats,
1189
1190
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
1191
1192
1193
1194
1195
1196
            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,
1197
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
1198
            num_speculative_heads=self.num_speculative_heads
1199
1200
        )

1201
        # Reminder: Please update docs/source/features/compatibility_matrix.md
1202
        # If the feature combo become valid
1203
1204
1205
1206
        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)")
1207
1208
1209
            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")
1210
1211
1212
1213
1214
1215
            from vllm.platforms import current_platform
            if current_platform.is_cpu():
                logger.warning("Multi-Step (--num-scheduler-steps > 1) is "
                               "currently not supported for CPUs and has been "
                               "disabled.")
                self.num_scheduler_steps = 1
1216
1217
1218
1219
1220
1221
1222
1223
1224

        # 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

1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
        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.")

1235
        scheduler_config = SchedulerConfig(
1236
            runner_type=model_config.runner_type,
1237
1238
1239
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1240
            num_lookahead_slots=num_lookahead_slots,
1241
1242
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1243
            is_multimodal_model=model_config.is_multimodal_model,
1244
            preemption_mode=self.preemption_mode,
1245
            num_scheduler_steps=self.num_scheduler_steps,
1246
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1247
1248
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1249
            policy=self.scheduling_policy)
1250
        lora_config = LoRAConfig(
1251
            bias_enabled=self.enable_lora_bias,
1252
1253
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1254
            fully_sharded_loras=self.fully_sharded_loras,
1255
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1256
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1257
1258
            lora_dtype=self.lora_dtype,
            max_cpu_loras=self.max_cpu_loras if self.max_cpu_loras
1259
1260
1261
            and self.max_cpu_loras > 0 else None,
            merge_lora=self.merge_lora,
            lora_target_modules=self.lora_target_modules) if self.enable_lora else None
1262

1263
1264
1265
1266
1267
1268
1269
        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

1270
        load_config = self.create_load_config()
1271

1272
1273
1274
1275
1276
        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

1277
1278
1279
        decoding_config = DecodingConfig(
            guided_decoding_backend=self.guided_decoding_backend)

1280
1281
1282
1283
1284
1285
1286
1287
        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}")
1288
        observability_config = ObservabilityConfig(
1289
1290
1291
1292
1293
1294
            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,
        )
1295

1296
        config = VllmConfig(
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
            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,
1307
            prompt_adapter_config=prompt_adapter_config,
1308
            compilation_config=self.compilation_config,
1309
            kv_transfer_config=self.kv_transfer_config,
1310
        )
1311

1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
        if envs.VLLM_USE_V1:
            self._override_v1_engine_config(config)
        return config

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

1322
1323
1324
1325
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
        # When no user override, set the default values based on the usage
        # context.
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
        # Use different default values for different hardware.
        from vllm.platforms import current_platform
        device_name = current_platform.get_device_name().lower()
        if "h100" in device_name or "h200" in device_name:
            # For H100 and H200, we use larger default values.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }

1342
1343
1344
1345
        if (self.max_num_batched_tokens is None
                and usage_context in default_max_num_batched_tokens):
            self.max_num_batched_tokens = default_max_num_batched_tokens[
                usage_context]
1346
1347
1348
            logger.warning(
                "Setting max_num_batched_tokens to %d for %s usage context.",
                self.max_num_batched_tokens, usage_context.value)
1349
1350
1351
1352
1353
1354
1355

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

1356

1357
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1358
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1359
    """Arguments for asynchronous vLLM engine."""
1360
    disable_log_requests: bool = False
1361
1362

    @staticmethod
1363
1364
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1365
1366
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1367
1368
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1369
                            help='Disable logging requests.')
1370
        return parser
1371
1372
1373
1374


# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1375
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1376
1377
1378


def _async_engine_args_parser():
1379
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1380
                                        async_args_only=True)