arg_utils.py 86.7 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import argparse
4
import dataclasses
5
import json
6
import threading
7
from dataclasses import dataclass
8
from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
9
                    Tuple, Type, Union, cast, get_args)
10

11
12
import torch

13
import vllm.envs as envs
14
from vllm import version
15
from vllm.config import (CacheConfig, CompilationConfig, ConfigFormat,
16
17
                         DecodingConfig, DeviceConfig, HfOverrides,
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
18
19
20
21
                         ModelConfig, ModelImpl, ObservabilityConfig,
                         ParallelConfig, PoolerConfig, PromptAdapterConfig,
                         SchedulerConfig, SpeculativeConfig, TaskOption,
                         TokenizerPoolConfig, VllmConfig)
22
from vllm.executor.executor_base import ExecutorBase
23
from vllm.logger import init_logger
24
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
25
from vllm.plugins import load_general_plugins
26
from vllm.reasoning import ReasoningParserManager
27
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
28
from vllm.transformers_utils.utils import check_gguf_file
29
from vllm.usage.usage_lib import UsageContext
30
from vllm.utils import FlexibleArgumentParser, StoreBoolean, is_in_ray_actor
31

32
if TYPE_CHECKING:
33
    from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
34

35
36
logger = init_logger(__name__)

37
38
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]

39
40
41
42
43
44
45
DEVICE_OPTIONS = [
    "auto",
    "cuda",
    "neuron",
    "cpu",
    "tpu",
    "xpu",
46
    "hpu",
47
48
]

49

50
51
52
53
54
55
def nullable_str(val: str):
    if not val or val == "None":
        return None
    return val


56
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
57
58
59
60
61
62
63
64
65
    """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.
    """
66
67
68
69
70
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
71
72
73
74
75
        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
76
77

        try:
78
            parsed_value = int(value)
79
80
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
81
82
83
84
85
86
            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
87
88
89
90

    return out_dict


91
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
92
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
93
    """Arguments for vLLM engine."""
94
    model: str = 'facebook/opt-125m'
95
    served_model_name: Optional[Union[str, List[str]]] = None
96
    tokenizer: Optional[str] = None
97
    hf_config_path: Optional[str] = None
98
    task: TaskOption = "auto"
99
    skip_tokenizer_init: bool = False
100
    tokenizer_mode: str = 'auto'
101
    trust_remote_code: bool = False
102
    allowed_local_media_path: str = ""
103
    download_dir: Optional[str] = None
104
    load_format: str = 'auto'
105
    config_format: ConfigFormat = ConfigFormat.AUTO
106
    dtype: str = 'auto'
107
    kv_cache_dtype: str = 'auto'
108
    seed: Optional[int] = None
109
    max_model_len: Optional[int] = None
110
111
112
113
114
    # 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
115
    # number of P/D disaggregation (or other disaggregation) workers
116
117
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
118
    data_parallel_size: int = 1
119
    enable_expert_parallel: bool = False
120
    max_parallel_loading_workers: Optional[int] = None
121
    block_size: Optional[int] = None
122
    enable_prefix_caching: Optional[bool] = None
123
    prefix_caching_hash_algo: str = "builtin"
124
    disable_sliding_window: bool = False
125
    disable_cascade_attn: bool = False
126
    use_v2_block_manager: bool = True
127
128
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
129
    gpu_memory_utilization: float = 0.90
130
    max_num_batched_tokens: Optional[int] = None
131
132
133
    max_num_partial_prefills: Optional[int] = 1
    max_long_partial_prefills: Optional[int] = 1
    long_prefill_token_threshold: Optional[int] = 0
134
    max_num_seqs: Optional[int] = None
135
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
136
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
137
    revision: Optional[str] = None
138
    code_revision: Optional[str] = None
139
    rope_scaling: Optional[Dict[str, Any]] = None
140
    rope_theta: Optional[float] = None
141
    hf_overrides: Optional[HfOverrides] = None
142
    tokenizer_revision: Optional[str] = None
143
    quantization: Optional[str] = None
144
    enforce_eager: Optional[bool] = None
145
    max_seq_len_to_capture: int = 8192
146
    disable_custom_all_reduce: bool = False
147
    tokenizer_pool_size: int = 0
148
149
150
151
    # 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"
152
    tokenizer_pool_extra_config: Optional[Dict[str, Any]] = None
153
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
154
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
155
    disable_mm_preprocessor_cache: bool = False
156
    enable_lora: bool = False
157
    enable_lora_bias: bool = False
158
159
    max_loras: int = 1
    max_lora_rank: int = 16
160
161
162
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
163
    fully_sharded_loras: bool = False
164
    lora_extra_vocab_size: int = 256
165
    long_lora_scaling_factors: Optional[Tuple[float]] = None
166
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
167
    max_cpu_loras: Optional[int] = None
168
    device: str = 'auto'
169
    num_scheduler_steps: int = 1
170
    multi_step_stream_outputs: bool = True
171
    ray_workers_use_nsight: bool = False
172
    num_gpu_blocks_override: Optional[int] = None
173
    num_lookahead_slots: int = 0
174
    model_loader_extra_config: Optional[dict] = None
175
    ignore_patterns: Optional[Union[str, List[str]]] = None
176
    preemption_mode: Optional[str] = None
177

178
    scheduler_delay_factor: float = 0.0
179
    enable_chunked_prefill: Optional[bool] = None
180

181
    guided_decoding_backend: str = 'xgrammar'
182
    logits_processor_pattern: Optional[str] = None
183
184
185
186

    speculative_config: Optional[Union[str, Dict[str, Any]]] = None

    # TODO(Shangming): Deprecate these out-of-date params after next release
187
    speculative_model: Optional[str] = None
188
    speculative_model_quantization: Optional[str] = None
189
    speculative_draft_tensor_parallel_size: Optional[int] = None
190
    num_speculative_tokens: Optional[int] = None
191
    speculative_disable_mqa_scorer: Optional[bool] = False
192
    speculative_max_model_len: Optional[int] = None
193
    speculative_disable_by_batch_size: Optional[int] = None
194
195
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
196
197
198
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
199
    disable_logprobs_during_spec_decoding: Optional[bool] = None
200

201
    qlora_adapter_name_or_path: Optional[str] = None
202
    show_hidden_metrics_for_version: Optional[str] = None
203
    otlp_traces_endpoint: Optional[str] = None
204
    collect_detailed_traces: Optional[str] = None
205
    disable_async_output_proc: bool = False
206
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
207
    scheduler_cls: Union[str, Type[object]] = "vllm.core.scheduler.Scheduler"
208

209
210
    override_neuron_config: Optional[Dict[str, Any]] = None
    override_pooler_config: Optional[PoolerConfig] = None
211
    compilation_config: Optional[CompilationConfig] = None
212
    worker_cls: str = "auto"
213
    worker_extension_cls: str = ""
214

215
216
    kv_transfer_config: Optional[KVTransferConfig] = None

217
    generation_config: Optional[str] = "auto"
218
    override_generation_config: Optional[Dict[str, Any]] = None
219
    enable_sleep_mode: bool = False
220
    model_impl: str = "auto"
221

222
223
    calculate_kv_scales: Optional[bool] = None

224
    additional_config: Optional[Dict[str, Any]] = None
225
226
    enable_reasoning: Optional[bool] = None
    reasoning_parser: Optional[str] = None
227
    use_tqdm_on_load: bool = True
228

229
    def __post_init__(self):
230
        if not self.tokenizer:
231
            self.tokenizer = self.model
232

233
234
235
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
236
        if isinstance(self.compilation_config, (int, dict)):
237
238
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
239

240
        # Setup plugins
241
242
        from vllm.plugins import load_general_plugins
        load_general_plugins()
243
244

    @staticmethod
245
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
246
        """Shared CLI arguments for vLLM engine."""
247
        # Model arguments
248
249
250
        parser.add_argument(
            '--model',
            type=str,
251
            default=EngineArgs.model,
252
            help='Name or path of the huggingface model to use.')
253
254
255
256
257
258
        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 '
259
            'multiple tasks. When the model only supports one task, ``"auto"`` '
260
261
            'can be used to select it; otherwise, you must specify explicitly '
            'which task to use.')
262
263
        parser.add_argument(
            '--tokenizer',
264
            type=nullable_str,
265
            default=EngineArgs.tokenizer,
266
267
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
268
269
270
271
272
273
        parser.add_argument(
            "--hf-config-path",
            type=nullable_str,
            default=EngineArgs.hf_config_path,
            help='Name or path of the huggingface config to use. '
            'If unspecified, model name or path will be used.')
274
275
276
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
277
278
279
            help='Skip initialization of tokenizer and detokenizer. '
            'Expects valid prompt_token_ids and None for prompt from '
            'the input. The generated output will contain token ids.')
Jasmond L's avatar
Jasmond L committed
280
281
        parser.add_argument(
            '--revision',
282
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
283
            default=None,
284
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
285
286
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
287
288
        parser.add_argument(
            '--code-revision',
289
            type=nullable_str,
290
            default=None,
291
            help='The specific revision to use for the model code on '
292
293
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
294
295
        parser.add_argument(
            '--tokenizer-revision',
296
            type=nullable_str,
297
            default=None,
298
299
300
            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.')
301
302
303
304
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
305
            choices=['auto', 'slow', 'mistral', 'custom'],
306
307
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
308
            'always use the slow tokenizer. \n* '
309
310
311
            '"mistral" will always use the `mistral_common` tokenizer. \n* '
            '"custom" will use --tokenizer to select the '
            'preregistered tokenizer.')
312
313
        parser.add_argument('--trust-remote-code',
                            action='store_true',
314
                            help='Trust remote code from huggingface.')
315
316
317
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
318
319
320
321
            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.")
322
        parser.add_argument('--download-dir',
323
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
324
                            default=EngineArgs.download_dir,
325
                            help='Directory to download and load the weights, '
326
                            'default to the default cache dir of '
327
                            'huggingface.')
328
329
330
331
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
332
            choices=[f.value for f in LoadFormat],
333
334
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
335
            'and fall back to the pytorch bin format if safetensors format '
336
337
338
339
340
341
342
343
            '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 '
344
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
345
            'section for more information.\n'
346
            '* "runai_streamer" will load the Safetensors weights using Run:ai'
347
            'Model Streamer.\n'
348
            '* "bitsandbytes" will load the weights using bitsandbytes '
349
350
351
352
353
354
355
            'quantization.\n'
            '* "sharded_state" will load weights from pre-sharded checkpoint '
            'files, supporting efficient loading of tensor-parallel models\n'
            '* "gguf" will load weights from GGUF format files (details '
            'specified in https://github.com/ggml-org/ggml/blob/master/docs/gguf.md).\n'
            '* "mistral" will load weights from consolidated safetensors files '
            'used by Mistral models.\n')
356
357
358
359
360
361
362
        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 ')
363
364
365
366
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
367
368
369
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
370
371
372
373
374
375
376
377
            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.')
378
379
380
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
381
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
382
            default=EngineArgs.kv_cache_dtype,
383
            help='Data type for kv cache storage. If "auto", will use model '
384
385
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
386
387
        parser.add_argument('--max-model-len',
                            type=int,
388
                            default=EngineArgs.max_model_len,
389
390
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
391
392
393
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
394
            default='xgrammar',
395
            help='Which engine will be used for guided decoding'
396
            ' (JSON schema / regex etc) by default. Currently support '
397
398
399
400
401
402
403
            'https://github.com/mlc-ai/xgrammar and '
            'https://github.com/guidance-ai/llguidance.'
            'Valid backend values are "xgrammar", "guidance", and "auto". '
            'With "auto", we will make opinionated choices based on request'
            'contents and what the backend libraries currently support, so '
            'the behavior is subject to change in each release. '
            'The default is xgrammar.')
404
405
406
407
408
409
410
411
        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.')
412
413
414
415
416
417
418
419
420
421
422
423
        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')
424
        # Parallel arguments
425
426
        parser.add_argument(
            '--distributed-executor-backend',
427
            choices=['ray', 'mp', 'uni', 'external_launcher'],
428
            default=EngineArgs.distributed_executor_backend,
429
430
431
432
433
434
            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 '
435
            'only supports Ray for distributed inference.')
436

437
438
439
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
440
                            default=EngineArgs.pipeline_parallel_size,
441
                            help='Number of pipeline stages.')
442
443
444
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
445
                            default=EngineArgs.tensor_parallel_size,
446
                            help='Number of tensor parallel replicas.')
447
448
449
450
451
452
453
454
        parser.add_argument('--data-parallel-size',
                            '-dp',
                            type=int,
                            default=EngineArgs.data_parallel_size,
                            help='Number of data parallel replicas. '
                            'MoE layers will be sharded according to the '
                            'product of the tensor-parallel-size and '
                            'data-parallel-size.')
455
456
457
458
459
        parser.add_argument(
            '--enable-expert-parallel',
            action='store_true',
            help='Use expert parallelism instead of tensor parallelism '
            'for MoE layers.')
460
461
462
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
463
            default=EngineArgs.max_parallel_loading_workers,
464
            help='Load model sequentially in multiple batches, '
465
            'to avoid RAM OOM when using tensor '
466
            'parallel and large models.')
467
468
469
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
470
            help='If specified, use nsight to profile Ray workers.')
471
        # KV cache arguments
472
473
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
474
                            default=EngineArgs.block_size,
475
                            choices=[8, 16, 32, 64, 128],
476
                            help='Token block size for contiguous chunks of '
477
                            'tokens. This is ignored on neuron devices and '
478
                            'set to ``--max-model-len``. On CUDA devices, '
479
480
                            'only block sizes up to 32 are supported. '
                            'On HPU devices, block size defaults to 128.')
481

482
483
484
485
486
        parser.add_argument(
            "--enable-prefix-caching",
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.enable_prefix_caching,
            help="Enables automatic prefix caching. "
487
            "Use ``--no-enable-prefix-caching`` to disable explicitly.",
488
        )
489
490
491
492
493
494
495
496
497
498
        parser.add_argument(
            "--prefix-caching-hash-algo",
            type=str,
            choices=["builtin", "sha256"],
            default=EngineArgs.prefix_caching_hash_algo,
            help="Set the hash algorithm for prefix caching. "
            "Options are 'builtin' (Python's built-in hash) or 'sha256' "
            "(collision resistant but with certain overheads). Defaults "
            "to 'builtin'.",
        )
499
500
501
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
502
                            'capping to sliding window size.')
503
504
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
505
                            default=True,
506
507
508
509
510
                            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.')
511
512
513
514
515
516
517
518
        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.')
519

520
521
522
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
523
                            help='Random seed for operations.')
524
        parser.add_argument('--swap-space',
525
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
526
                            default=EngineArgs.swap_space,
527
                            help='CPU swap space size (GiB) per GPU.')
528
529
530
531
532
533
534
535
536
        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 '
537
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
538
            'which requires at least 26GB GPU memory. Note that this '
539
            'requires fast CPU-GPU interconnect, as part of the model is '
540
541
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
542
543
544
545
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
546
547
548
            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, '
549
550
551
552
553
554
            '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.')
555
        parser.add_argument(
556
            '--num-gpu-blocks-override',
557
558
559
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
560
            ' of GPU blocks. Used for testing preemption.')
561
562
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
563
                            default=EngineArgs.max_num_batched_tokens,
564
565
                            help='Maximum number of batched tokens per '
                            'iteration.')
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
        parser.add_argument(
            "--max-num-partial-prefills",
            type=int,
            default=EngineArgs.max_num_partial_prefills,
            help="For chunked prefill, the max number of concurrent \
            partial prefills."
            "Defaults to 1",
        )
        parser.add_argument(
            "--max-long-partial-prefills",
            type=int,
            default=EngineArgs.max_long_partial_prefills,
            help="For chunked prefill, the maximum number of prompts longer "
            "than --long-prefill-token-threshold that will be prefilled "
            "concurrently. Setting this less than --max-num-partial-prefills "
            "will allow shorter prompts to jump the queue in front of longer "
            "prompts in some cases, improving latency. Defaults to 1.")
        parser.add_argument(
            "--long-prefill-token-threshold",
            type=float,
            default=EngineArgs.long_prefill_token_threshold,
            help="For chunked prefill, a request is considered long if the "
            "prompt is longer than this number of tokens. Defaults to 4%% of "
            "the model's context length.",
        )
591
592
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
593
                            default=EngineArgs.max_num_seqs,
594
                            help='Maximum number of sequences per iteration.')
595
596
597
598
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
599
600
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
601
602
        parser.add_argument('--disable-log-stats',
                            action='store_true',
603
                            help='Disable logging statistics.')
604
605
606
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
607
                            type=nullable_str,
608
                            choices=[*QUANTIZATION_METHODS, None],
609
                            default=EngineArgs.quantization,
610
611
612
613
614
615
                            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.')
616
617
618
619
620
        parser.add_argument(
            '--rope-scaling',
            default=None,
            type=json.loads,
            help='RoPE scaling configuration in JSON format. '
621
            'For example, ``{"rope_type":"dynamic","factor":2.0}``')
622
623
624
625
626
627
        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.')
628
629
630
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
631
                            help='Extra arguments for the HuggingFace config. '
632
633
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
634
635
636
637
638
        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.')
639
        parser.add_argument('--max-seq-len-to-capture',
640
641
642
643
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
644
645
646
647
                            '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.')
648
649
650
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
651
                            help='See ParallelConfig.')
652
653
654
655
656
657
658
659
660
661
662
663
664
        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',
665
                            type=nullable_str,
666
667
668
669
670
                            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.')
671
672
673
674
675
676
677

        # 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
678
            # MultiModalConfig.get_limit_per_prompt
679
680
681
682
683
684
            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.'))
685
686
687
688
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
689
            help=('Overrides for the multimodal input mapping/processing, '
690
                  'e.g., image processor. For example: ``{"num_crops": 4}``.'))
691
        parser.add_argument(
692
            '--disable-mm-preprocessor-cache',
693
            action='store_true',
694
695
            help='If true, then disables caching of the multi-modal '
            'preprocessor/mapper. (not recommended)')
696

697
698
699
700
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
701
702
703
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
        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,
723
            choices=['auto', 'float16', 'bfloat16'],
724
725
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
726
727
728
729
730
731
732
733
734
735
736
        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.'))
737
738
739
740
741
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
742
743
                  'Must be >= than max_loras. '
                  'Defaults to max_loras.'))
744
745
746
747
748
749
750
751
        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.'))
752
753
754
755
756
757
758
759
760
761
762
        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')
763
764
765
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
766
                            choices=DEVICE_OPTIONS,
767
                            help='Device type for vLLM execution.')
768
769
770
771
772
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
773
774
775
776
777
778
779
780
        parser.add_argument(
            '--use-tqdm-on-load',
            dest='use_tqdm_on_load',
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.use_tqdm_on_load,
            help='Whether to enable/disable progress bar '
            'when loading model weights.',
        )
781

782
783
        parser.add_argument(
            '--multi-step-stream-outputs',
784
785
786
787
788
789
            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')
790
791
792
793
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
794
            help='Apply a delay (of delay factor multiplied by previous '
795
            'prompt latency) before scheduling next prompt.')
796
797
        parser.add_argument(
            '--enable-chunked-prefill',
798
799
800
801
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
802
            help='If set, the prefill requests can be chunked based on the '
803
            'max_num_batched_tokens.')
804
805
806
807
808
        parser.add_argument('--speculative-config',
                            type=nullable_str,
                            default=None,
                            help='The configurations for speculative decoding.'
                            ' Should be a JSON string.')
809
810
        parser.add_argument(
            '--speculative-model',
811
            type=nullable_str,
812
            default=EngineArgs.speculative_model,
813
814
            help=
            'The name of the draft model to be used in speculative decoding.')
815
816
817
818
819
820
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
821
            help='Method used to quantize the weights of speculative model. '
822
823
824
825
826
            '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.')
827
828
829
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
830
            default=EngineArgs.num_speculative_tokens,
831
            help='The number of speculative tokens to sample from '
832
            'the draft model in speculative decoding.')
833
834
835
836
837
838
        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')
839
840
841
842
843
844
845
        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.')
846

847
848
        parser.add_argument(
            '--speculative-max-model-len',
849
            type=int,
850
851
852
853
854
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

855
856
857
858
859
860
861
        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.')

862
863
864
865
866
867
868
869
870
871
872
873
874
875
        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.')

876
877
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
        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')

908
909
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
910
            action=StoreBoolean,
911
            default=EngineArgs.disable_logprobs_during_spec_decoding,
912
913
            nargs="?",
            const="True",
914
915
916
917
918
919
920
921
            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.')

922
        parser.add_argument('--model-loader-extra-config',
923
                            type=nullable_str,
924
925
926
927
928
929
                            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.')
930
931
932
933
934
935
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
936
            "Default to `original/**/*` to avoid repeated loading of llama's "
937
            "checkpoints.")
938
        parser.add_argument(
939
            '--preemption-mode',
940
941
            type=str,
            default=None,
942
943
944
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
945

946
947
948
949
950
951
952
953
954
955
        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 "
956
            "same as the ``--model`` argument. Noted that this name(s) "
957
            "will also be used in `model_name` tag content of "
958
            "prometheus metrics, if multiple names provided, metrics "
959
            "tag will take the first one.")
960
961
962
963
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
964

965
966
967
968
969
970
971
972
973
974
975
976
        parser.add_argument('--show-hidden-metrics-for-version',
                            type=str,
                            default=None,
                            help='Enable deprecated Prometheus metrics that '
                            'have been hidden since the specified version. '
                            'For example, if a previously deprecated metric '
                            'has been hidden since the v0.7.0 release, you '
                            'use --show-hidden-metrics-for-version=0.7 as a '
                            'temporary escape hatch while you migrate to new '
                            'metrics. The metric is likely to be removed '
                            'completely in an upcoming release.')

977
978
979
980
981
        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
982
983
984
985
986
987
        parser.add_argument(
            '--collect-detailed-traces',
            type=str,
            default=None,
            help="Valid choices are " +
            ",".join(ALLOWED_DETAILED_TRACE_MODULES) +
988
            ". It makes sense to set this only if ``--otlp-traces-endpoint`` is"
989
990
991
            " 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.")
992

993
994
995
996
997
998
        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.")
999

1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
        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).')

1010
1011
1012
1013
1014
1015
1016
        parser.add_argument(
            '--scheduler-cls',
            default=EngineArgs.scheduler_cls,
            help='The scheduler class to use. "vllm.core.scheduler.Scheduler" '
            'is the default scheduler. Can be a class directly or the path to '
            'a class of form "mod.custom_class".')

1017
        parser.add_argument(
1018
1019
            '--override-neuron-config',
            type=json.loads,
1020
            default=None,
1021
            help="Override or set neuron device configuration. "
1022
            "e.g. ``{\"cast_logits_dtype\": \"bloat16\"}``.")
1023
        parser.add_argument(
1024
1025
            '--override-pooler-config',
            type=PoolerConfig.from_json,
1026
            default=None,
1027
            help="Override or set the pooling method for pooling models. "
1028
            "e.g. ``{\"pooling_type\": \"mean\", \"normalize\": false}``.")
1029

1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
        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, '
1042
1043
1044
1045
                            'use a JSON string.\n'
                            'Following the convention of traditional '
                            'compilers, using -O without space is also '
                            'supported. -O3 is equivalent to -O 3.')
1046

1047
1048
1049
1050
1051
1052
        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.')

1053
1054
1055
1056
1057
        parser.add_argument(
            '--worker-cls',
            type=str,
            default="auto",
            help='The worker class to use for distributed execution.')
1058
1059
1060
1061
1062
1063
1064
        parser.add_argument(
            '--worker-extension-cls',
            type=str,
            default="",
            help='The worker extension class on top of the worker cls, '
            'it is useful if you just want to add new functions to the worker '
            'class without changing the existing functions.')
1065
1066
1067
        parser.add_argument(
            "--generation-config",
            type=nullable_str,
1068
            default="auto",
1069
            help="The folder path to the generation config. "
1070
1071
1072
1073
1074
            "Defaults to 'auto', the generation config will be loaded from "
            "model path. If set to 'vllm', no generation config is loaded, "
            "vLLM defaults will be used. 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 "
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
            "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.")
1087

1088
1089
1090
1091
1092
1093
        parser.add_argument("--enable-sleep-mode",
                            action="store_true",
                            default=False,
                            help="Enable sleep mode for the engine. "
                            "(only cuda platform is supported)")

1094
1095
1096
1097
1098
1099
1100
1101
1102
        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.')

1103
1104
1105
1106
1107
1108
1109
1110
        parser.add_argument(
            "--additional-config",
            type=json.loads,
            default=None,
            help="Additional config for specified platform in JSON format. "
            "Different platforms may support different configs. Make sure the "
            "configs are valid for the platform you are using. The input format"
            " is like '{\"config_key\":\"config_value\"}'")
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122

        parser.add_argument(
            "--enable-reasoning",
            action="store_true",
            default=False,
            help="Whether to enable reasoning_content for the model. "
            "If enabled, the model will be able to generate reasoning content."
        )

        parser.add_argument(
            "--reasoning-parser",
            type=str,
1123
            choices=list(ReasoningParserManager.reasoning_parsers),
1124
1125
1126
1127
1128
1129
            default=None,
            help=
            "Select the reasoning parser depending on the model that you're "
            "using. This is used to parse the reasoning content into OpenAI "
            "API format. Required for ``--enable-reasoning``.")

1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
        parser.add_argument(
            "--disable-cascade-attn",
            action="store_true",
            default=False,
            help="Disable cascade attention for V1. While cascade attention "
            "does not change the mathematical correctness, disabling it "
            "could be useful for preventing potential numerical issues. "
            "Note that even if this is set to False, cascade attention will be "
            "only used when the heuristic tells that it's beneficial.")

1140
        return parser
1141
1142

    @classmethod
1143
    def from_cli_args(cls, args: argparse.Namespace):
1144
1145
1146
        # 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
1147
1148
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
1149

1150
    def create_model_config(self) -> ModelConfig:
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
        # 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"

        # NOTE: This is to allow model loading from S3 in CI
        if (not isinstance(self, AsyncEngineArgs) and envs.VLLM_CI_USE_S3
                and self.model in MODELS_ON_S3
                and self.load_format == LoadFormat.AUTO):  # noqa: E501
            self.model = f"{MODEL_WEIGHTS_S3_BUCKET}/{self.model}"
            self.load_format = LoadFormat.RUNAI_STREAMER

1162
        return ModelConfig(
1163
            model=self.model,
1164
            hf_config_path=self.hf_config_path,
1165
            task=self.task,
1166
1167
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
1168
1169
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1170
            allowed_local_media_path=self.allowed_local_media_path,
1171
1172
1173
1174
1175
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
1176
            rope_theta=self.rope_theta,
1177
            hf_overrides=self.hf_overrides,
1178
1179
1180
1181
1182
1183
1184
            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,
1185
            disable_cascade_attn=self.disable_cascade_attn,
1186
            skip_tokenizer_init=self.skip_tokenizer_init,
1187
            served_model_name=self.served_model_name,
1188
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1189
            use_async_output_proc=not self.disable_async_output_proc,
1190
            config_format=self.config_format,
1191
            mm_processor_kwargs=self.mm_processor_kwargs,
1192
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
1193
1194
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
1195
            logits_processor_pattern=self.logits_processor_pattern,
1196
            generation_config=self.generation_config,
1197
            override_generation_config=self.override_generation_config,
1198
            enable_sleep_mode=self.enable_sleep_mode,
1199
            model_impl=self.model_impl,
1200
        )
1201

1202
1203
    def create_load_config(self) -> LoadConfig:

1204
        if(self.qlora_adapter_name_or_path is not None) and \
1205
1206
            self.quantization != "bitsandbytes":
            raise ValueError(
1207
                "QLoRA adapter only support "
1208
1209
                f"'bitsandbytes' quantization, but got {self.quantization}")

1210
1211
        if self.quantization == "bitsandbytes":
            self.load_format = "bitsandbytes"
1212
1213
1214
1215
1216
        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,
1217
            use_tqdm_on_load=self.use_tqdm_on_load,
1218
        )
1219

1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
    def create_speculative_config(
        self,
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        enable_chunked_prefill: bool,
        disable_log_stats: bool,
    ) -> Optional["SpeculativeConfig"]:
        """Initializes and returns a SpeculativeConfig object based on
        `speculative_config`.

        This function utilizes `speculative_config` to create a
        SpeculativeConfig object. The `speculative_config` can either be
        provided as a JSON string input via CLI arguments or directly as a
        dictionary from the engine. If `speculative_config` is not set, this
        function will attempt to construct a configuration dictionary using
        certain parameters, which are scheduled for deprecation in the next
        release. Note that in next releases, `speculative_config` must be
        provided, and the deprecated standalone speculative-related parameters
        will be removed.
        """
        if self.speculative_config is None:
            if (self.speculative_model is None
                    and self.num_speculative_tokens is None):
                return None

            # TODO(Shangming): Deprecate this way of setting SpeculativeConfig,
            # only allow '--speculative-config' after next release
            logger.warning_once(
                "Please use '--speculative-config' to set all configurations "
                "related to speculative decoding. The current method of "
                "specifying the model through '--speculative-model' and "
                "adding related parameters (e.g., '--num-speculative-tokens') "
                "separately will be deprecated in the next release.")

            spec_config_dict = {
                "model": self.speculative_model,
                "quantization": self.speculative_model_quantization,
                "max_model_len": self.speculative_max_model_len,
                "draft_tensor_parallel_size":
                self.speculative_draft_tensor_parallel_size,
                "num_speculative_tokens": self.num_speculative_tokens,
                "disable_mqa_scorer": self.speculative_disable_mqa_scorer,
                "disable_by_batch_size":
                self.speculative_disable_by_batch_size,
                "prompt_lookup_max": self.ngram_prompt_lookup_max,
                "prompt_lookup_min": self.ngram_prompt_lookup_min,
                "acceptance_method": self.spec_decoding_acceptance_method,
                "posterior_threshold":
                self.typical_acceptance_sampler_posterior_threshold,
                "posterior_alpha":
                self.typical_acceptance_sampler_posterior_alpha,
                "disable_logprobs": self.disable_logprobs_during_spec_decoding,
            }

            self.speculative_config = spec_config_dict
        else:
            if isinstance(self.speculative_config, str):
                import ast
                self.speculative_config = ast.literal_eval(
                    self.speculative_config)
        # Note(Shangming): These parameters are not obtained from the cli arg
        # '--speculative-config' and must be passed in when creating the engine
        # config.

        assert isinstance(self.speculative_config, dict)
        self.speculative_config.update({
            "target_model_config": target_model_config,
            "target_parallel_config": target_parallel_config,
            "enable_chunked_prefill": enable_chunked_prefill,
            "disable_log_stats": disable_log_stats,
        })
        speculative_config = SpeculativeConfig.from_dict(
            self.speculative_config)

        return speculative_config

1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
    def create_engine_config(
        self,
        usage_context: Optional[UsageContext] = None,
    ) -> VllmConfig:
        """
        Create the VllmConfig.

        NOTE: for autoselection of V0 vs V1 engine, we need to
        create the ModelConfig first, since ModelConfig's attrs
        (e.g. the model arch) are needed to make the decision.
Simon Mo's avatar
Simon Mo committed
1306

1307
1308
1309
1310
1311
1312
        This function set VLLM_USE_V1=X if VLLM_USE_V1 is
        unspecified by the user.

        If VLLM_USE_V1 is specified by the user but the VllmConfig
        is incompatible, we raise an error.
        """
1313
1314
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update()
1315

1316
        device_config = DeviceConfig(device=self.device)
1317
1318
        model_config = self.create_model_config()

1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
        # * If VLLM_USE_V1 is unset, we enable V1 for "supported features"
        #   and fall back to V0 for experimental or unsupported features.
        # * If VLLM_USE_V1=1, we enable V1 for supported + experimental
        #   features and raise error for unsupported features.
        # * If VLLM_USE_V1=0, we disable V1.
        use_v1 = False
        try_v1 = envs.VLLM_USE_V1 or not envs.is_set("VLLM_USE_V1")
        if try_v1 and self._is_v1_supported_oracle(model_config):
            use_v1 = True

        # If user explicitly set VLLM_USE_V1, sanity check we respect it.
        if envs.is_set("VLLM_USE_V1"):
            assert use_v1 == envs.VLLM_USE_V1
        # Otherwise, set the VLLM_USE_V1 variable globally.
        else:
            envs.set_vllm_use_v1(use_v1)

        # Set default arguments for V0 or V1 Engine.
        if use_v1:
            self._set_default_args_v1(usage_context)
        else:
            self._set_default_args_v0(model_config)
1341

1342
1343
        assert self.enable_chunked_prefill is not None

1344
        cache_config = CacheConfig(
1345
            block_size=self.block_size,
1346
1347
1348
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1349
            is_attention_free=model_config.is_attention_free,
1350
1351
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1352
            enable_prefix_caching=self.enable_prefix_caching,
1353
            prefix_caching_hash_algo=self.prefix_caching_hash_algo,
1354
            cpu_offload_gb=self.cpu_offload_gb,
1355
            calculate_kv_scales=self.calculate_kv_scales,
1356
        )
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368

        # Get the current placement group if Ray is initialized and
        # we are in a Ray actor. If so, then the placement group will be
        # passed to spawned processes.
        placement_group = None
        if is_in_ray_actor():
            import ray

            # This call initializes Ray automatically if it is not initialized,
            # but we should not do this here.
            placement_group = ray.util.get_current_placement_group()

1369
        parallel_config = ParallelConfig(
1370
1371
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
1372
            data_parallel_size=self.data_parallel_size,
1373
            enable_expert_parallel=self.enable_expert_parallel,
1374
1375
1376
            max_parallel_loading_workers=self.max_parallel_loading_workers,
            disable_custom_all_reduce=self.disable_custom_all_reduce,
            tokenizer_pool_config=TokenizerPoolConfig.create_config(
1377
1378
1379
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1380
            ),
1381
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1382
            placement_group=placement_group,
1383
1384
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
1385
            worker_extension_cls=self.worker_extension_cls,
1386
        )
1387

1388
        speculative_config = self.create_speculative_config(
1389
1390
            target_model_config=model_config,
            target_parallel_config=parallel_config,
1391
            enable_chunked_prefill=self.enable_chunked_prefill,
1392
            disable_log_stats=self.disable_log_stats,
1393
1394
        )

1395
        # Reminder: Please update docs/source/features/compatibility_matrix.md
1396
        # If the feature combo become valid
1397
1398
1399
1400
        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)")
1401
1402
1403
            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")
1404
1405
1406
1407
1408
1409
            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
1410
1411
1412
1413
1414
1415
1416
1417
1418

        # 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

1419
        scheduler_config = SchedulerConfig(
1420
            runner_type=model_config.runner_type,
1421
1422
1423
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1424
            num_lookahead_slots=num_lookahead_slots,
1425
1426
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1427
            is_multimodal_model=model_config.is_multimodal_model,
1428
            preemption_mode=self.preemption_mode,
1429
            num_scheduler_steps=self.num_scheduler_steps,
1430
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1431
1432
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1433
            policy=self.scheduling_policy,
1434
            scheduler_cls=self.scheduler_cls,
1435
1436
1437
1438
            max_num_partial_prefills=self.max_num_partial_prefills,
            max_long_partial_prefills=self.max_long_partial_prefills,
            long_prefill_token_threshold=self.long_prefill_token_threshold,
        )
1439

1440
        lora_config = LoRAConfig(
1441
            bias_enabled=self.enable_lora_bias,
1442
1443
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1444
            fully_sharded_loras=self.fully_sharded_loras,
1445
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1446
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1447
1448
1449
            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
1450

1451
1452
1453
1454
1455
1456
1457
        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

1458
        load_config = self.create_load_config()
1459

1460
1461
1462
1463
1464
        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

1465
        decoding_config = DecodingConfig(
1466
1467
1468
1469
            guided_decoding_backend=self.guided_decoding_backend,
            reasoning_backend=self.reasoning_parser
            if self.enable_reasoning else None,
        )
1470

1471
1472
1473
1474
1475
        show_hidden_metrics = False
        if self.show_hidden_metrics_for_version is not None:
            show_hidden_metrics = version._prev_minor_version_was(
                self.show_hidden_metrics_for_version)

1476
1477
1478
1479
1480
1481
1482
1483
        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}")
1484
        observability_config = ObservabilityConfig(
1485
            show_hidden_metrics=show_hidden_metrics,
1486
1487
1488
1489
1490
1491
            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,
        )
1492

1493
        config = VllmConfig(
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
            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,
1504
            prompt_adapter_config=prompt_adapter_config,
1505
            compilation_config=self.compilation_config,
1506
            kv_transfer_config=self.kv_transfer_config,
1507
            additional_config=self.additional_config,
1508
        )
1509

1510
1511
        return config

1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
    def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
        """Oracle for whether to use V0 or V1 Engine by default."""

        #############################################################
        # Unsupported Feature Flags on V1.

        if (self.load_format == LoadFormat.TENSORIZER.value
                or self.load_format == LoadFormat.SHARDED_STATE.value):
            _raise_or_fallback(
                feature_name=f"--load_format {self.load_format}",
                recommend_to_remove=False)
            return False

        if (self.logits_processor_pattern
                != EngineArgs.logits_processor_pattern):
            _raise_or_fallback(feature_name="--logits-processor-pattern",
                               recommend_to_remove=False)
            return False

        if self.preemption_mode != EngineArgs.preemption_mode:
            _raise_or_fallback(feature_name="--preemption-mode",
                               recommend_to_remove=True)
            return False

        if (self.disable_async_output_proc
                != EngineArgs.disable_async_output_proc):
            _raise_or_fallback(feature_name="--disable-async-output-proc",
                               recommend_to_remove=True)
            return False

        if self.scheduling_policy != EngineArgs.scheduling_policy:
            _raise_or_fallback(feature_name="--scheduling-policy",
                               recommend_to_remove=False)
            return False

        if self.num_scheduler_steps != EngineArgs.num_scheduler_steps:
            _raise_or_fallback(feature_name="--num-scheduler-steps",
                               recommend_to_remove=True)
            return False

        if self.scheduler_delay_factor != EngineArgs.scheduler_delay_factor:
            _raise_or_fallback(feature_name="--scheduler-delay-factor",
                               recommend_to_remove=True)
            return False

        if self.additional_config != EngineArgs.additional_config:
            _raise_or_fallback(feature_name="--additional-config",
                               recommend_to_remove=False)
            return False

1562
        # Xgrammar and Guidance are supported.
1563
        SUPPORTED_GUIDED_DECODING = [
1564
1565
            "xgrammar", "xgrammar:disable-any-whitespace", "guidance",
            "guidance:disable-any-whitespace", "auto"
1566
        ]
1567
1568
1569
1570
1571
1572
        if self.guided_decoding_backend not in SUPPORTED_GUIDED_DECODING:
            _raise_or_fallback(feature_name="--guided-decoding-backend",
                               recommend_to_remove=False)
            return False

        # Need at least Ampere for now (FA support required).
1573
1574
1575
        # Skip this check if we are running on a non-GPU platform,
        # or if the device capability is not available
        # (e.g. in a Ray actor without GPUs).
1576
1577
        from vllm.platforms import current_platform
        if (current_platform.is_cuda()
1578
                and current_platform.get_device_capability()
1579
1580
1581
1582
1583
1584
1585
                and current_platform.get_device_capability().major < 8):
            _raise_or_fallback(feature_name="Compute Capability < 8.0",
                               recommend_to_remove=False)
            return False

        # No Fp8 KV cache so far.
        if self.kv_cache_dtype != "auto":
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
            fp8_attention = self.kv_cache_dtype.startswith("fp8")
            will_use_fa = (
                current_platform.is_cuda()
                and not envs.is_set("VLLM_ATTENTION_BACKEND")
            ) or envs.VLLM_ATTENTION_BACKEND == "FLASH_ATTN_VLLM_V1"
            supported = False
            if fp8_attention and will_use_fa:
                from vllm.vllm_flash_attn.fa_utils import (
                    flash_attn_supports_fp8)
                supported = flash_attn_supports_fp8()
            if not supported:
                _raise_or_fallback(feature_name="--kv-cache-dtype",
                                   recommend_to_remove=False)
                return False
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620

        # No Prompt Adapter so far.
        if self.enable_prompt_adapter:
            _raise_or_fallback(feature_name="--enable-prompt-adapter",
                               recommend_to_remove=False)
            return False

        # No CPU offloading yet.
        if self.cpu_offload_gb != EngineArgs.cpu_offload_gb:
            _raise_or_fallback(feature_name="--cpu-offload-gb",
                               recommend_to_remove=False)
            return False

        # Only Fp16 and Bf16 dtypes since we only support FA.
        V1_SUPPORTED_DTYPES = [torch.bfloat16, torch.float16]
        if model_config.dtype not in V1_SUPPORTED_DTYPES:
            _raise_or_fallback(feature_name=f"--dtype {model_config.dtype}",
                               recommend_to_remove=False)
            return False

        # Some quantization is not compatible with torch.compile.
1621
        V1_UNSUPPORTED_QUANT = ["gguf"]
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
        if model_config.quantization in V1_UNSUPPORTED_QUANT:
            _raise_or_fallback(
                feature_name=f"--quantization {model_config.quantization}",
                recommend_to_remove=False)
            return False

        # No Embedding Models so far.
        if model_config.task not in ["generate"]:
            _raise_or_fallback(feature_name=f"--task {model_config.task}",
                               recommend_to_remove=False)
            return False

        # No Mamba or Encoder-Decoder so far.
        if not model_config.is_v1_compatible:
            _raise_or_fallback(feature_name=model_config.architectures,
                               recommend_to_remove=False)
            return False

        # No Concurrent Partial Prefills so far.
        if (self.max_num_partial_prefills
                != EngineArgs.max_num_partial_prefills
                or self.max_long_partial_prefills
1644
                != EngineArgs.max_long_partial_prefills):
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
            _raise_or_fallback(feature_name="Concurrent Partial Prefill",
                               recommend_to_remove=False)
            return False

        # No OTLP observability so far.
        if (self.otlp_traces_endpoint or self.collect_detailed_traces):
            _raise_or_fallback(feature_name="--otlp-traces-endpoint",
                               recommend_to_remove=False)
            return False

        # Only Ngram speculative decoding so far.
        if (self.speculative_model is not None
                or self.num_speculative_tokens is not None):
            # This is supported but experimental (handled below).
1659
            if self.speculative_model in ("ngram", "[ngram]"):
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
                pass
            else:
                _raise_or_fallback(feature_name="Speculative Decoding",
                                   recommend_to_remove=False)
                return False

        # No Disaggregated Prefill so far.
        if self.kv_transfer_config != EngineArgs.kv_transfer_config:
            _raise_or_fallback(feature_name="--kv-transfer-config",
                               recommend_to_remove=False)
            return False

        # No FlashInfer or XFormers so far.
        V1_BACKENDS = [
            "FLASH_ATTN_VLLM_V1", "FLASH_ATTN", "PALLAS", "PALLAS_VLLM_V1",
1675
            "TRITON_ATTN_VLLM_V1", "TRITON_MLA", "FLASHMLA"
1676
1677
1678
1679
1680
1681
1682
        ]
        if (envs.is_set("VLLM_ATTENTION_BACKEND")
                and envs.VLLM_ATTENTION_BACKEND not in V1_BACKENDS):
            name = f"VLLM_ATTENTION_BACKEND={envs.VLLM_ATTENTION_BACKEND}"
            _raise_or_fallback(feature_name=name, recommend_to_remove=True)
            return False

1683
1684
        # Platforms must decide if they can support v1 for this model
        if not current_platform.supports_v1(model_config=model_config):
1685
1686
1687
1688
            _raise_or_fallback(
                feature_name=f"device type={current_platform.device_type}",
                recommend_to_remove=False)
            return False
1689
1690
1691
        #############################################################
        # Experimental Features - allow users to opt in.

1692
1693
1694
1695
1696
        # Signal Handlers requires running in main thread.
        if (threading.current_thread() != threading.main_thread()
                and _warn_or_fallback("Engine in background thread")):
            return False

1697
1698
1699
1700
        # LoRA is supported on V1, but off by default for now.
        if self.enable_lora and _warn_or_fallback("LORA"):
            return False

1701
1702
1703
1704
1705
        # PP is supported on V1 with Ray distributed executor,
        # but off for MP distributed executor for now.
        if (self.pipeline_parallel_size > 1
                and self.distributed_executor_backend == "mp"
                and _warn_or_fallback("PP (MP distributed executor)")):
1706
1707
1708
            return False

        # ngram is supported on V1, but off by default for now.
1709
1710
        if self.speculative_model in (
                "ngram", "[ngram]") and _warn_or_fallback("ngram"):
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
            return False

        # Non-CUDA is supported on V1, but off by default for now.
        not_cuda = not current_platform.is_cuda()
        if not_cuda and _warn_or_fallback(  # noqa: SIM103
                current_platform.device_type):
            return False
        #############################################################

        return True

    def _set_default_args_v0(self, model_config: ModelConfig) -> None:
        """Set Default Arguments for V0 Engine."""

        max_model_len = model_config.max_model_len
        use_long_context = max_model_len > 32768
        if self.enable_chunked_prefill is None:
            # Chunked prefill not supported for Multimodal or MLA in V0.
            if model_config.is_multimodal_model or model_config.use_mla:
                self.enable_chunked_prefill = False

            # Enable chunked prefill by default for long context (> 32K)
            # models to avoid OOM errors in initial memory profiling phase.
            elif use_long_context:
                from vllm.platforms import current_platform
                is_gpu = current_platform.is_cuda()
                use_sliding_window = (model_config.get_sliding_window()
                                      is not None)
                use_spec_decode = self.speculative_model is not None

                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
                        and not self.enable_prompt_adapter
                        and model_config.runner_type != "pooling"):
                    self.enable_chunked_prefill = True
                    logger.warning(
                        "Chunked prefill is enabled by default for models "
                        "with max_model_len > 32K. Chunked prefill might "
                        "not work with some features or models. If you "
                        "encounter any issues, please disable by launching "
                        "with --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 during the initial memory profiling phase, or result "
                "in low performance due to small KV cache size. Consider "
                "setting --max-model-len to a smaller value.", max_model_len)
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
            msg = "Chunked prefill is not supported for pooling models"
            raise ValueError(msg)

1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
        # if using prefix caching, we must set a hash algo
        if self.enable_prefix_caching:
            # Disable prefix caching for multimodal models for VLLM_V0.
            if model_config.is_multimodal_model:
                logger.warning(
                    "--enable-prefix-caching is not supported for multimodal "
                    "models in V0 and has been disabled.")
                self.enable_prefix_caching = False

            # VLLM_V0 only supports builtin hash algo for prefix caching.
            if self.prefix_caching_hash_algo is None:
                self.prefix_caching_hash_algo = "builtin"
            elif self.prefix_caching_hash_algo == "sha256":
                raise ValueError(
                    "sha256 is not supported for prefix caching in V0 engine. "
                    "Please use 'builtin'.")
1783
1784
1785
1786
1787
1788
1789

        # Set max_num_seqs to 256 for VLLM_V0.
        if self.max_num_seqs is None:
            self.max_num_seqs = 256

    def _set_default_args_v1(self, usage_context: UsageContext) -> None:
        """Set Default Arguments for V1 Engine."""
1790

1791
1792
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
1793
1794
1795
1796
1797

        # V1 enables prefix caching by default.
        if self.enable_prefix_caching is None:
            self.enable_prefix_caching = True

1798
1799
1800
1801
        # if using prefix caching, we must set a hash algo
        if self.enable_prefix_caching and self.prefix_caching_hash_algo is None:
            self.prefix_caching_hash_algo = "builtin"

1802
1803
1804
        # V1 should use the new scheduler by default.
        # Swap it only if this arg is set to the original V0 default
        if self.scheduler_cls == EngineArgs.scheduler_cls:
1805
            self.scheduler_cls = "vllm.v1.core.sched.scheduler.Scheduler"
1806

1807
1808
        # When no user override, set the default values based on the usage
        # context.
1809
        # Use different default values for different hardware.
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822

        # Try to query the device name on the current platform. If it fails,
        # it may be because the platform that imports vLLM is not the same
        # as the platform that vLLM is running on (e.g. the case of scaling
        # vLLM with Ray) and has no GPUs. In this case we use the default
        # values for non-H100/H200 GPUs.
        try:
            from vllm.platforms import current_platform
            device_name = current_platform.get_device_name().lower()
        except Exception:
            # This is only used to set default_max_num_batched_tokens
            device_name = "no-device"

1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
        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,
            }

1836
        use_context_value = usage_context.value if usage_context else None
1837
1838
1839
1840
        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]
1841
            logger.debug(
1842
                "Setting max_num_batched_tokens to %d for %s usage context.",
1843
                self.max_num_batched_tokens, use_context_value)
1844

1845
1846
1847
1848
1849
1850
        default_max_num_seqs = 1024
        if self.max_num_seqs is None:
            self.max_num_seqs = default_max_num_seqs

            logger.debug("Setting max_num_seqs to %d for %s usage context.",
                         self.max_num_seqs, use_context_value)
1851

1852

1853
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1854
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1855
    """Arguments for asynchronous vLLM engine."""
1856
    disable_log_requests: bool = False
1857
1858

    @staticmethod
1859
1860
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1861
1862
1863
1864
        # Initialize plugin to update the parser, for example, The plugin may
        # adding a new kind of quantization method to --quantization argument or
        # a new device to --device argument.
        load_general_plugins()
1865
1866
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1867
1868
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1869
                            help='Disable logging requests.')
1870
1871
        from vllm.platforms import current_platform
        current_platform.pre_register_and_update(parser)
1872
        return parser
1873
1874


1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
def _raise_or_fallback(feature_name: str, recommend_to_remove: bool):
    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
        raise NotImplementedError(
            f"VLLM_USE_V1=1 is not supported with {feature_name}.")
    msg = f"{feature_name} is not supported by the V1 Engine. "
    msg += "Falling back to V0. "
    if recommend_to_remove:
        msg += f"We recommend to remove {feature_name} from your config "
        msg += "in favor of the V1 Engine."
    logger.warning(msg)


def _warn_or_fallback(feature_name: str) -> bool:
    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
        logger.warning(
            "Detected VLLM_USE_V1=1 with %s. Usage should "
            "be considered experimental. Please report any "
            "issues on Github.", feature_name)
        should_exit = False
    else:
        logger.info(
            "%s is experimental on VLLM_USE_V1=1. "
            "Falling back to V0 Engine.", feature_name)
        should_exit = True
    return should_exit


1902
1903
# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1904
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1905
1906
1907


def _async_engine_args_parser():
1908
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1909
                                        async_args_only=True)