arg_utils.py 61.9 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

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

10
11
import torch

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

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

30
31
logger = init_logger(__name__)

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

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

45

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


52
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
53
54
55
56
57
58
59
60
61
    """Parses a string containing comma separate key [str] to value [int]
    pairs into a dictionary.

    Args:
        val: String value to be parsed.

    Returns:
        Dictionary with parsed values.
    """
62
63
64
65
66
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
67
68
69
70
71
        kv_parts = [part.lower().strip() for part in item.split("=")]
        if len(kv_parts) != 2:
            raise argparse.ArgumentTypeError(
                "Each item should be in the form KEY=VALUE")
        key, value = kv_parts
72
73

        try:
74
            parsed_value = int(value)
75
76
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
77
78
79
80
81
82
            raise argparse.ArgumentTypeError(msg) from exc

        if key in out_dict and out_dict[key] != parsed_value:
            raise argparse.ArgumentTypeError(
                f"Conflicting values specified for key: {key}")
        out_dict[key] = parsed_value
83
84
85
86

    return out_dict


87
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
88
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
89
    """Arguments for vLLM engine."""
90
    model: str = 'facebook/opt-125m'
91
    served_model_name: Optional[Union[str, List[str]]] = None
92
    tokenizer: Optional[str] = None
93
    task: TaskOption = "auto"
94
    skip_tokenizer_init: bool = False
95
    tokenizer_mode: str = 'auto'
96
    trust_remote_code: bool = False
97
    allowed_local_media_path: str = ""
98
    download_dir: Optional[str] = None
99
    load_format: str = 'auto'
100
    config_format: ConfigFormat = ConfigFormat.AUTO
101
    dtype: str = 'auto'
102
    kv_cache_dtype: str = 'auto'
103
    seed: int = 0
104
    max_model_len: Optional[int] = None
105
106
107
108
109
    # Note: Specifying a custom executor backend by passing a class
    # is intended for expert use only. The API may change without
    # notice.
    distributed_executor_backend: Optional[Union[str,
                                                 Type[ExecutorBase]]] = None
110
    # number of P/D disaggregation (or other disaggregation) workers
111
112
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
113
    max_parallel_loading_workers: Optional[int] = None
114
    block_size: Optional[int] = None
115
    enable_prefix_caching: Optional[bool] = None
116
    disable_sliding_window: bool = False
117
    use_v2_block_manager: bool = True
118
119
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
120
    gpu_memory_utilization: float = 0.90
121
    max_num_batched_tokens: Optional[int] = None
122
    max_num_seqs: Optional[int] = None
123
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
124
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
125
    revision: Optional[str] = None
126
    code_revision: Optional[str] = None
127
    rope_scaling: Optional[Dict[str, Any]] = None
128
    rope_theta: Optional[float] = None
129
    hf_overrides: Optional[HfOverrides] = None
130
    tokenizer_revision: Optional[str] = None
131
    quantization: Optional[str] = None
132
    enforce_eager: Optional[bool] = None
133
    max_seq_len_to_capture: int = 8192
134
    disable_custom_all_reduce: bool = False
135
    tokenizer_pool_size: int = 0
136
137
138
139
    # Note: Specifying a tokenizer pool by passing a class
    # is intended for expert use only. The API may change without
    # notice.
    tokenizer_pool_type: Union[str, Type["BaseTokenizerGroup"]] = "ray"
140
    tokenizer_pool_extra_config: Optional[Dict[str, Any]] = None
141
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
142
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
143
    disable_mm_preprocessor_cache: bool = False
144
    enable_lora: bool = False
145
    enable_lora_bias: bool = False
146
147
    max_loras: int = 1
    max_lora_rank: int = 16
148
149
150
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
151
    fully_sharded_loras: bool = False
152
    lora_extra_vocab_size: int = 256
153
    long_lora_scaling_factors: Optional[Tuple[float]] = None
154
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
155
    max_cpu_loras: Optional[int] = None
156
    device: str = 'auto'
157
    num_scheduler_steps: int = 1
158
    multi_step_stream_outputs: bool = True
159
    ray_workers_use_nsight: bool = False
160
    num_gpu_blocks_override: Optional[int] = None
161
    num_lookahead_slots: int = 0
162
    model_loader_extra_config: Optional[dict] = None
163
    ignore_patterns: Optional[Union[str, List[str]]] = None
164
    preemption_mode: Optional[str] = None
165

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

169
    guided_decoding_backend: str = 'xgrammar'
170
    logits_processor_pattern: Optional[str] = None
171
172
    # Speculative decoding configuration.
    speculative_model: Optional[str] = None
173
    speculative_model_quantization: Optional[str] = None
174
    speculative_draft_tensor_parallel_size: Optional[int] = None
175
    num_speculative_tokens: Optional[int] = None
176
    speculative_disable_mqa_scorer: Optional[bool] = False
177
    speculative_max_model_len: Optional[int] = None
178
    speculative_disable_by_batch_size: Optional[int] = None
179
180
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
181
182
183
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
184
    qlora_adapter_name_or_path: Optional[str] = None
185
    disable_logprobs_during_spec_decoding: Optional[bool] = None
186

187
    otlp_traces_endpoint: Optional[str] = None
188
    collect_detailed_traces: Optional[str] = None
189
    disable_async_output_proc: bool = False
190
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
191

192
193
    override_neuron_config: Optional[Dict[str, Any]] = None
    override_pooler_config: Optional[PoolerConfig] = None
194
    compilation_config: Optional[CompilationConfig] = None
195
    worker_cls: str = "auto"
196

197
198
    kv_transfer_config: Optional[KVTransferConfig] = None

199
    generation_config: Optional[str] = None
200
    override_generation_config: Optional[Dict[str, Any]] = None
201
    enable_sleep_mode: bool = False
202

203
204
    calculate_kv_scales: Optional[bool] = None

205
    def __post_init__(self):
206
        if not self.tokenizer:
207
            self.tokenizer = self.model
208

209
210
211
212
        # Override the default value of enable_prefix_caching if it's not set
        # by user.
        if self.enable_prefix_caching is None:
            self.enable_prefix_caching = bool(envs.VLLM_USE_V1)
213

214
215
216
        # Override max_num_seqs if it's not set by user.
        if self.max_num_seqs is None:
            self.max_num_seqs = 256 if not envs.VLLM_USE_V1 else 1024
217

218
219
220
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
221
        if isinstance(self.compilation_config, (int, dict)):
222
223
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
224

225
        # Setup plugins
226
227
        from vllm.plugins import load_general_plugins
        load_general_plugins()
228
229

    @staticmethod
230
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
231
        """Shared CLI arguments for vLLM engine."""
232

233
        # Model arguments
234
235
236
        parser.add_argument(
            '--model',
            type=str,
237
            default=EngineArgs.model,
238
            help='Name or path of the huggingface model to use.')
239
240
241
242
243
244
        parser.add_argument(
            '--task',
            default=EngineArgs.task,
            choices=get_args(TaskOption),
            help='The task to use the model for. Each vLLM instance only '
            'supports one task, even if the same model can be used for '
245
            'multiple tasks. When the model only supports one task, ``"auto"`` '
246
247
            'can be used to select it; otherwise, you must specify explicitly '
            'which task to use.')
248
249
        parser.add_argument(
            '--tokenizer',
250
            type=nullable_str,
251
            default=EngineArgs.tokenizer,
252
253
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
254
255
256
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
257
            help='Skip initialization of tokenizer and detokenizer.')
Jasmond L's avatar
Jasmond L committed
258
259
        parser.add_argument(
            '--revision',
260
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
261
            default=None,
262
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
263
264
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
265
266
        parser.add_argument(
            '--code-revision',
267
            type=nullable_str,
268
            default=None,
269
            help='The specific revision to use for the model code on '
270
271
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
272
273
        parser.add_argument(
            '--tokenizer-revision',
274
            type=nullable_str,
275
            default=None,
276
277
278
            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.')
279
280
281
282
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
283
            choices=['auto', 'slow', 'mistral'],
284
285
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
286
287
            'always use the slow tokenizer. \n* '
            '"mistral" will always use the `mistral_common` tokenizer.')
288
289
        parser.add_argument('--trust-remote-code',
                            action='store_true',
290
                            help='Trust remote code from huggingface.')
291
292
293
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
294
295
296
297
            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.")
298
        parser.add_argument('--download-dir',
299
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
300
                            default=EngineArgs.download_dir,
301
                            help='Directory to download and load the weights, '
302
                            'default to the default cache dir of '
303
                            'huggingface.')
304
305
306
307
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
308
            choices=[f.value for f in LoadFormat],
309
310
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
311
            'and fall back to the pytorch bin format if safetensors format '
312
313
314
315
316
317
318
319
            '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 '
320
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
321
            'section for more information.\n'
322
323
            '* "runai_streamer" will load the Safetensors weights using Run:ai'
            'Model Streamer \n'
324
325
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
326
327
328
329
330
331
332
        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 ')
333
334
335
336
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
337
338
339
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
340
341
342
343
344
345
346
347
            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.')
348
349
350
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
351
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
352
            default=EngineArgs.kv_cache_dtype,
353
            help='Data type for kv cache storage. If "auto", will use model '
354
355
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
356
357
        parser.add_argument('--max-model-len',
                            type=int,
358
                            default=EngineArgs.max_model_len,
359
360
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
361
362
363
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
364
365
            default='xgrammar',
            choices=['outlines', 'lm-format-enforcer', 'xgrammar'],
366
            help='Which engine will be used for guided decoding'
367
            ' (JSON schema / regex etc) by default. Currently support '
368
            'https://github.com/outlines-dev/outlines, '
369
            'https://github.com/mlc-ai/xgrammar, and '
370
371
372
            'https://github.com/noamgat/lm-format-enforcer.'
            ' Can be overridden per request via guided_decoding_backend'
            ' parameter.')
373
374
375
376
377
378
379
380
        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.')
381
        # Parallel arguments
382
383
        parser.add_argument(
            '--distributed-executor-backend',
384
            choices=['ray', 'mp', 'uni', 'external_launcher'],
385
            default=EngineArgs.distributed_executor_backend,
386
387
388
389
390
391
            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 '
392
            'only supports Ray for distributed inference.')
393

394
395
396
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
397
                            default=EngineArgs.pipeline_parallel_size,
398
                            help='Number of pipeline stages.')
399
400
401
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
402
                            default=EngineArgs.tensor_parallel_size,
403
                            help='Number of tensor parallel replicas.')
404
405
406
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
407
            default=EngineArgs.max_parallel_loading_workers,
408
            help='Load model sequentially in multiple batches, '
409
            'to avoid RAM OOM when using tensor '
410
            'parallel and large models.')
411
412
413
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
414
            help='If specified, use nsight to profile Ray workers.')
415
        # KV cache arguments
416
417
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
418
                            default=EngineArgs.block_size,
419
                            choices=[8, 16, 32, 64, 128],
420
                            help='Token block size for contiguous chunks of '
421
                            'tokens. This is ignored on neuron devices and '
422
                            'set to ``--max-model-len``. On CUDA devices, '
423
424
                            'only block sizes up to 32 are supported. '
                            'On HPU devices, block size defaults to 128.')
425

426
427
428
429
430
        parser.add_argument(
            "--enable-prefix-caching",
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.enable_prefix_caching,
            help="Enables automatic prefix caching. "
431
            "Use ``--no-enable-prefix-caching`` to disable explicitly.",
432
        )
433
434
435
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
436
                            'capping to sliding window size.')
437
438
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
439
                            default=True,
440
441
442
443
444
                            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.')
445
446
447
448
449
450
451
452
        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.')
453

454
455
456
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
457
                            help='Random seed for operations.')
458
        parser.add_argument('--swap-space',
459
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
460
                            default=EngineArgs.swap_space,
461
                            help='CPU swap space size (GiB) per GPU.')
462
463
464
465
466
467
468
469
470
        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 '
471
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
472
            'which requires at least 26GB GPU memory. Note that this '
473
            'requires fast CPU-GPU interconnect, as part of the model is '
474
475
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
476
477
478
479
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
480
481
482
            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, '
483
484
485
486
487
488
            '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.')
489
        parser.add_argument(
490
            '--num-gpu-blocks-override',
491
492
493
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
494
            ' of GPU blocks. Used for testing preemption.')
495
496
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
497
                            default=EngineArgs.max_num_batched_tokens,
498
499
                            help='Maximum number of batched tokens per '
                            'iteration.')
500
501
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
502
                            default=EngineArgs.max_num_seqs,
503
                            help='Maximum number of sequences per iteration.')
504
505
506
507
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
508
509
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
510
511
        parser.add_argument('--disable-log-stats',
                            action='store_true',
512
                            help='Disable logging statistics.')
513
514
515
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
516
                            type=nullable_str,
517
                            choices=[*QUANTIZATION_METHODS, None],
518
                            default=EngineArgs.quantization,
519
520
521
522
523
524
                            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.')
525
526
527
528
529
        parser.add_argument(
            '--rope-scaling',
            default=None,
            type=json.loads,
            help='RoPE scaling configuration in JSON format. '
530
            'For example, ``{"rope_type":"dynamic","factor":2.0}``')
531
532
533
534
535
536
        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.')
537
538
539
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
540
                            help='Extra arguments for the HuggingFace config. '
541
542
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
543
544
545
546
547
        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.')
548
        parser.add_argument('--max-seq-len-to-capture',
549
550
551
552
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
553
554
555
556
                            '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.')
557
558
559
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
560
                            help='See ParallelConfig.')
561
562
563
564
565
566
567
568
569
570
571
572
573
        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',
574
                            type=nullable_str,
575
576
577
578
579
                            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.')
580
581
582
583
584
585
586
587
588
589
590
591
592
593

        # Multimodal related configs
        parser.add_argument(
            '--limit-mm-per-prompt',
            type=nullable_kvs,
            default=EngineArgs.limit_mm_per_prompt,
            # The default value is given in
            # MultiModalRegistry.init_mm_limits_per_prompt
            help=('For each multimodal plugin, limit how many '
                  'input instances to allow for each prompt. '
                  'Expects a comma-separated list of items, '
                  'e.g.: `image=16,video=2` allows a maximum of 16 '
                  'images and 2 videos per prompt. Defaults to 1 for '
                  'each modality.'))
594
595
596
597
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
598
            help=('Overrides for the multimodal input mapping/processing, '
599
                  'e.g., image processor. For example: ``{"num_crops": 4}``.'))
600
        parser.add_argument(
601
            '--disable-mm-preprocessor-cache',
602
            action='store_true',
603
604
            help='If true, then disables caching of the multi-modal '
            'preprocessor/mapper. (not recommended)')
605

606
607
608
609
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
610
611
612
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
        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,
632
            choices=['auto', 'float16', 'bfloat16'],
633
634
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
635
636
637
638
639
640
641
642
643
644
645
        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.'))
646
647
648
649
650
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
651
652
                  'Must be >= than max_loras. '
                  'Defaults to max_loras.'))
653
654
655
656
657
658
659
660
        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.'))
661
662
663
664
665
666
667
668
669
670
671
        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')
672
673
674
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
675
                            choices=DEVICE_OPTIONS,
676
                            help='Device type for vLLM execution.')
677
678
679
680
681
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
682

683
684
        parser.add_argument(
            '--multi-step-stream-outputs',
685
686
687
688
689
690
            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')
691
692
693
694
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
695
            help='Apply a delay (of delay factor multiplied by previous '
696
            'prompt latency) before scheduling next prompt.')
697
698
        parser.add_argument(
            '--enable-chunked-prefill',
699
700
701
702
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
703
            help='If set, the prefill requests can be chunked based on the '
704
            'max_num_batched_tokens.')
705
706
707

        parser.add_argument(
            '--speculative-model',
708
            type=nullable_str,
709
            default=EngineArgs.speculative_model,
710
711
            help=
            'The name of the draft model to be used in speculative decoding.')
712
713
714
715
716
717
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
718
            help='Method used to quantize the weights of speculative model. '
719
720
721
722
723
            '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.')
724
725
726
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
727
            default=EngineArgs.num_speculative_tokens,
728
            help='The number of speculative tokens to sample from '
729
            'the draft model in speculative decoding.')
730
731
732
733
734
735
        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')
736
737
738
739
740
741
742
        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.')
743

744
745
        parser.add_argument(
            '--speculative-max-model-len',
746
            type=int,
747
748
749
750
751
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

752
753
754
755
756
757
758
        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.')

759
760
761
762
763
764
765
766
767
768
769
770
771
772
        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.')

773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
        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')

805
806
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
807
            action=StoreBoolean,
808
            default=EngineArgs.disable_logprobs_during_spec_decoding,
809
810
            nargs="?",
            const="True",
811
812
813
814
815
816
817
818
            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.')

819
        parser.add_argument('--model-loader-extra-config',
820
                            type=nullable_str,
821
822
823
824
825
826
                            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.')
827
828
829
830
831
832
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
833
            "Default to `original/**/*` to avoid repeated loading of llama's "
834
            "checkpoints.")
835
        parser.add_argument(
836
            '--preemption-mode',
837
838
            type=str,
            default=None,
839
840
841
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
842

843
844
845
846
847
848
849
850
851
852
        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 "
853
            "same as the ``--model`` argument. Noted that this name(s) "
854
            "will also be used in `model_name` tag content of "
855
            "prometheus metrics, if multiple names provided, metrics "
856
            "tag will take the first one.")
857
858
859
860
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
861
862
863
864
865
866

        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
867
868
869
870
871
872
        parser.add_argument(
            '--collect-detailed-traces',
            type=str,
            default=None,
            help="Valid choices are " +
            ",".join(ALLOWED_DETAILED_TRACE_MODULES) +
873
            ". It makes sense to set this only if ``--otlp-traces-endpoint`` is"
874
875
876
            " 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.")
877

878
879
880
881
882
883
        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.")
884

885
886
887
888
889
890
891
892
893
894
        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).')

895
        parser.add_argument(
896
897
            '--override-neuron-config',
            type=json.loads,
898
            default=None,
899
            help="Override or set neuron device configuration. "
900
            "e.g. ``{\"cast_logits_dtype\": \"bloat16\"}``.")
901
        parser.add_argument(
902
903
            '--override-pooler-config',
            type=PoolerConfig.from_json,
904
            default=None,
905
            help="Override or set the pooling method for pooling models. "
906
            "e.g. ``{\"pooling_type\": \"mean\", \"normalize\": false}``.")
907

908
909
910
911
912
913
914
915
916
917
918
919
        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, '
920
921
922
923
                            'use a JSON string.\n'
                            'Following the convention of traditional '
                            'compilers, using -O without space is also '
                            'supported. -O3 is equivalent to -O 3.')
924

925
926
927
928
929
930
        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.')

931
932
933
934
935
        parser.add_argument(
            '--worker-cls',
            type=str,
            default="auto",
            help='The worker class to use for distributed execution.')
936
937
938
939
940
        parser.add_argument(
            "--generation-config",
            type=nullable_str,
            default=None,
            help="The folder path to the generation config. "
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
            "Defaults to None, no generation config is loaded, vLLM defaults "
            "will be used. If set to 'auto', the generation config will be "
            "loaded from model path. If set to a folder path, the generation "
            "config will be loaded from the specified folder path. If "
            "`max_new_tokens` is specified in generation config, then "
            "it sets a server-wide limit on the number of output tokens "
            "for all requests.")

        parser.add_argument(
            "--override-generation-config",
            type=json.loads,
            default=None,
            help="Overrides or sets generation config in JSON format. "
            "e.g. ``{\"temperature\": 0.5}``. If used with "
            "--generation-config=auto, the override parameters will be merged "
            "with the default config from the model. If generation-config is "
            "None, only the override parameters are used.")
958

959
960
961
962
963
964
        parser.add_argument("--enable-sleep-mode",
                            action="store_true",
                            default=False,
                            help="Enable sleep mode for the engine. "
                            "(only cuda platform is supported)")

965
966
967
968
969
970
971
972
973
        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.')

974
        return parser
975
976

    @classmethod
977
    def from_cli_args(cls, args: argparse.Namespace):
978
979
980
        # 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
981
982
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
983

984
985
    def create_model_config(self) -> ModelConfig:
        return ModelConfig(
986
            model=self.model,
987
            task=self.task,
988
989
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
990
991
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
992
            allowed_local_media_path=self.allowed_local_media_path,
993
994
995
996
997
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
998
            rope_theta=self.rope_theta,
999
            hf_overrides=self.hf_overrides,
1000
1001
1002
1003
1004
1005
1006
1007
            tokenizer_revision=self.tokenizer_revision,
            max_model_len=self.max_model_len,
            quantization=self.quantization,
            enforce_eager=self.enforce_eager,
            max_seq_len_to_capture=self.max_seq_len_to_capture,
            max_logprobs=self.max_logprobs,
            disable_sliding_window=self.disable_sliding_window,
            skip_tokenizer_init=self.skip_tokenizer_init,
1008
            served_model_name=self.served_model_name,
1009
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1010
            use_async_output_proc=not self.disable_async_output_proc,
1011
            config_format=self.config_format,
1012
            mm_processor_kwargs=self.mm_processor_kwargs,
1013
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
1014
1015
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
1016
            logits_processor_pattern=self.logits_processor_pattern,
1017
            generation_config=self.generation_config,
1018
            override_generation_config=self.override_generation_config,
1019
1020
            enable_sleep_mode=self.enable_sleep_mode,
        )
1021

1022
1023
1024
1025
1026
1027
1028
1029
    def create_load_config(self) -> LoadConfig:
        return LoadConfig(
            load_format=self.load_format,
            download_dir=self.download_dir,
            model_loader_extra_config=self.model_loader_extra_config,
            ignore_patterns=self.ignore_patterns,
        )

1030
1031
1032
1033
1034
1035
    def create_engine_config(self,
                             usage_context: Optional[UsageContext] = None
                             ) -> VllmConfig:
        if envs.VLLM_USE_V1:
            self._override_v1_engine_args(usage_context)

1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
        # gguf file needs a specific model loader and doesn't use hf_repo
        if check_gguf_file(self.model):
            self.quantization = self.load_format = "gguf"

        # bitsandbytes quantization needs a specific model loader
        # so we make sure the quant method and the load format are consistent
        if (self.quantization == "bitsandbytes" or
           self.qlora_adapter_name_or_path is not None) and \
           self.load_format != "bitsandbytes":
            raise ValueError(
                "BitsAndBytes quantization and QLoRA adapter only support "
                f"'bitsandbytes' load format, but got {self.load_format}")

        if (self.load_format == "bitsandbytes" or
            self.qlora_adapter_name_or_path is not None) and \
            self.quantization != "bitsandbytes":
            raise ValueError(
                "BitsAndBytes load format and QLoRA adapter only support "
                f"'bitsandbytes' quantization, but got {self.quantization}")

        assert self.cpu_offload_gb >= 0, (
            "CPU offload space must be non-negative"
            f", but got {self.cpu_offload_gb}")

        device_config = DeviceConfig(device=self.device)
        model_config = self.create_model_config()

1063
1064
1065
1066
1067
        if (model_config.is_multimodal_model and not envs.VLLM_USE_V1
                and self.enable_prefix_caching):
            logger.warning("--enable-prefix-caching is currently not "
                           "supported for multimodal models in v0 and "
                           "has been disabled.")
1068
1069
            self.enable_prefix_caching = False

1070
        cache_config = CacheConfig(
1071
            block_size=self.block_size,
1072
1073
1074
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1075
            is_attention_free=model_config.is_attention_free,
1076
1077
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1078
1079
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
1080
            calculate_kv_scales=self.calculate_kv_scales,
1081
        )
1082
        parallel_config = ParallelConfig(
1083
1084
1085
1086
1087
            pipeline_parallel_size=self.pipeline_parallel_size,
            tensor_parallel_size=self.tensor_parallel_size,
            max_parallel_loading_workers=self.max_parallel_loading_workers,
            disable_custom_all_reduce=self.disable_custom_all_reduce,
            tokenizer_pool_config=TokenizerPoolConfig.create_config(
1088
1089
1090
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1091
            ),
1092
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1093
1094
1095
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
        )
1096

1097
1098
1099
1100
1101
1102
        max_model_len = model_config.max_model_len
        use_long_context = max_model_len > 32768
        if self.enable_chunked_prefill is None:
            # If not explicitly set, enable chunked prefill by default for
            # long context (> 32K) models. This is to avoid OOM errors in the
            # initial memory profiling phase.
1103

1104
1105
1106
1107
1108
1109
            # For multimodal models, chunked prefill is disabled by default in
            # V0, but enabled by design in V1
            if model_config.is_multimodal_model:
                self.enable_chunked_prefill = bool(envs.VLLM_USE_V1)

            elif use_long_context:
1110
1111
1112
1113
                is_gpu = device_config.device_type == "cuda"
                use_sliding_window = (model_config.get_sliding_window()
                                      is not None)
                use_spec_decode = self.speculative_model is not None
1114
                from vllm.platforms import current_platform
1115
1116
                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
1117
                        and not self.enable_prompt_adapter
1118
1119
                        and model_config.runner_type != "pooling"
                        and not current_platform.is_rocm()):
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
                    self.enable_chunked_prefill = True
                    logger.warning(
                        "Chunked prefill is enabled by default for models with "
                        "max_model_len > 32K. Currently, chunked prefill might "
                        "not work with some features or models. If you "
                        "encounter any issues, please disable chunked prefill "
                        "by setting --enable-chunked-prefill=False.")
            if self.enable_chunked_prefill is None:
                self.enable_chunked_prefill = False

        if not self.enable_chunked_prefill and use_long_context:
            logger.warning(
                "The model has a long context length (%s). This may cause OOM "
                "errors during the initial memory profiling phase, or result "
                "in low performance due to small KV cache space. Consider "
                "setting --max-model-len to a smaller value.", max_model_len)
1136
1137
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
1138
            msg = "Chunked prefill is not supported for pooling models"
1139
            raise ValueError(msg)
1140

1141

1142
1143
1144
1145
1146
        speculative_config = SpeculativeConfig.maybe_create_spec_config(
            target_model_config=model_config,
            target_parallel_config=parallel_config,
            target_dtype=self.dtype,
            speculative_model=self.speculative_model,
1147
1148
            speculative_model_quantization = \
                self.speculative_model_quantization,
1149
1150
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
1151
            num_speculative_tokens=self.num_speculative_tokens,
1152
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
1153
1154
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
1155
1156
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
1157
            disable_log_stats=self.disable_log_stats,
1158
1159
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
1160
1161
1162
1163
1164
1165
            draft_token_acceptance_method=\
                self.spec_decoding_acceptance_method,
            typical_acceptance_sampler_posterior_threshold=self.
            typical_acceptance_sampler_posterior_threshold,
            typical_acceptance_sampler_posterior_alpha=self.
            typical_acceptance_sampler_posterior_alpha,
1166
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
1167
1168
        )

1169
        # Reminder: Please update docs/source/features/compatibility_matrix.md
1170
        # If the feature combo become valid
1171
1172
1173
1174
        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)")
1175
1176
1177
            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")
1178
1179
1180
1181
1182
1183
            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
1184
1185
1186
1187
1188
1189
1190
1191
1192

        # 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

1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
        if not self.use_v2_block_manager:
            logger.warning(
                "[DEPRECATED] Block manager v1 has been removed, "
                "and setting --use-v2-block-manager to True or False has "
                "no effect on vLLM behavior. Please remove "
                "--use-v2-block-manager in your engine argument. "
                "If your use case is not supported by "
                "SelfAttnBlockSpaceManager (i.e. block manager v2),"
                " please file an issue with detailed information.")

1203
        scheduler_config = SchedulerConfig(
1204
            runner_type=model_config.runner_type,
1205
1206
1207
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1208
            num_lookahead_slots=num_lookahead_slots,
1209
1210
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1211
            is_multimodal_model=model_config.is_multimodal_model,
1212
            preemption_mode=self.preemption_mode,
1213
            num_scheduler_steps=self.num_scheduler_steps,
1214
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1215
1216
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1217
            policy=self.scheduling_policy)
1218
        lora_config = LoRAConfig(
1219
            bias_enabled=self.enable_lora_bias,
1220
1221
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1222
            fully_sharded_loras=self.fully_sharded_loras,
1223
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1224
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1225
1226
1227
            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
1228

1229
1230
1231
1232
1233
1234
1235
        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

1236
        load_config = self.create_load_config()
1237

1238
1239
1240
1241
1242
        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

1243
1244
1245
        decoding_config = DecodingConfig(
            guided_decoding_backend=self.guided_decoding_backend)

1246
1247
1248
1249
1250
1251
1252
1253
        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}")
1254
        observability_config = ObservabilityConfig(
1255
1256
1257
1258
1259
1260
            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,
        )
1261

1262
        config = VllmConfig(
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
            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,
1273
            prompt_adapter_config=prompt_adapter_config,
1274
            compilation_config=self.compilation_config,
1275
            kv_transfer_config=self.kv_transfer_config,
1276
        )
1277

1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
        if envs.VLLM_USE_V1:
            self._override_v1_engine_config(config)
        return config

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

1288
1289
1290
1291
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
        # When no user override, set the default values based on the usage
        # context.
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
        # Use different default values for different hardware.
        from vllm.platforms import current_platform
        device_name = current_platform.get_device_name().lower()
        if "h100" in device_name or "h200" in device_name:
            # For H100 and H200, we use larger default values.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 16384,
                UsageContext.OPENAI_API_SERVER: 8192,
            }
        else:
            # TODO(woosuk): Tune the default values for other hardware.
            default_max_num_batched_tokens = {
                UsageContext.LLM_CLASS: 8192,
                UsageContext.OPENAI_API_SERVER: 2048,
            }

1308
1309
1310
1311
        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]
1312
1313
1314
            logger.warning(
                "Setting max_num_batched_tokens to %d for %s usage context.",
                self.max_num_batched_tokens, usage_context.value)
1315
1316
1317
1318
1319
1320
1321

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

1322

1323
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1324
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1325
    """Arguments for asynchronous vLLM engine."""
1326
    disable_log_requests: bool = False
1327
1328

    @staticmethod
1329
1330
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1331
1332
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1333
1334
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1335
                            help='Disable logging requests.')
1336
        return parser
1337
1338
1339
1340


# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1341
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1342
1343
1344


def _async_engine_args_parser():
1345
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1346
                                        async_args_only=True)