arg_utils.py 62.9 KB
Newer Older
1
import argparse
2
import dataclasses
3
import json
4
from dataclasses import dataclass
5
from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
6
                    Tuple, Type, Union, cast, get_args)
7

8
9
import torch

10
import vllm.envs as envs
11
from vllm.config import (CacheConfig, CompilationConfig, ConfigFormat,
12
13
14
15
                         DecodingConfig, DeviceConfig, HfOverrides,
                         KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
                         ModelConfig, ObservabilityConfig, ParallelConfig,
                         PoolerConfig, PromptAdapterConfig, SchedulerConfig,
16
17
                         SpeculativeConfig, TaskOption, TokenizerPoolConfig,
                         VllmConfig)
18
from vllm.executor.executor_base import ExecutorBase
19
from vllm.logger import init_logger
20
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
21
from vllm.transformers_utils.utils import check_gguf_file
22
from vllm.usage.usage_lib import UsageContext
23
from vllm.utils import FlexibleArgumentParser, StoreBoolean
24

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

28
29
logger = init_logger(__name__)

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

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

43

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


50
def nullable_kvs(val: str) -> Optional[Mapping[str, int]]:
51
52
53
54
55
56
57
58
59
    """Parses a string containing comma separate key [str] to value [int]
    pairs into a dictionary.

    Args:
        val: String value to be parsed.

    Returns:
        Dictionary with parsed values.
    """
60
61
62
63
64
    if len(val) == 0:
        return None

    out_dict: Dict[str, int] = {}
    for item in val.split(","):
65
66
67
68
69
        kv_parts = [part.lower().strip() for part in item.split("=")]
        if len(kv_parts) != 2:
            raise argparse.ArgumentTypeError(
                "Each item should be in the form KEY=VALUE")
        key, value = kv_parts
70
71

        try:
72
            parsed_value = int(value)
73
74
        except ValueError as exc:
            msg = f"Failed to parse value of item {key}={value}"
75
76
77
78
79
80
            raise argparse.ArgumentTypeError(msg) from exc

        if key in out_dict and out_dict[key] != parsed_value:
            raise argparse.ArgumentTypeError(
                f"Conflicting values specified for key: {key}")
        out_dict[key] = parsed_value
81
82
83
84

    return out_dict


85
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
86
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
87
    """Arguments for vLLM engine."""
88
    model: str = 'facebook/opt-125m'
89
    served_model_name: Optional[Union[str, List[str]]] = None
90
    tokenizer: Optional[str] = None
91
    task: TaskOption = "auto"
92
    skip_tokenizer_init: bool = False
93
    tokenizer_mode: str = 'auto'
94
    trust_remote_code: bool = False
95
    allowed_local_media_path: str = ""
96
    download_dir: Optional[str] = None
97
    load_format: str = 'auto'
98
    config_format: ConfigFormat = ConfigFormat.AUTO
99
    dtype: str = 'auto'
100
    kv_cache_dtype: str = 'auto'
101
    seed: int = 0
102
    max_model_len: Optional[int] = None
103
104
105
106
107
    # 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
108
    # number of P/D disaggregation (or other disaggregation) workers
109
110
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
111
    max_parallel_loading_workers: Optional[int] = None
112
    block_size: Optional[int] = None
113
    enable_prefix_caching: Optional[bool] = None
114
    disable_sliding_window: bool = False
115
    use_v2_block_manager: bool = True
116
117
    swap_space: float = 4  # GiB
    cpu_offload_gb: float = 0  # GiB
118
    gpu_memory_utilization: float = 0.90
119
    max_num_batched_tokens: Optional[int] = None
120
    max_num_seqs: Optional[int] = None
121
    max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
122
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
123
    revision: Optional[str] = None
124
    code_revision: Optional[str] = None
125
    rope_scaling: Optional[Dict[str, Any]] = None
126
    rope_theta: Optional[float] = None
127
    hf_overrides: Optional[HfOverrides] = None
128
    tokenizer_revision: Optional[str] = None
129
    quantization: Optional[str] = None
130
    enforce_eager: Optional[bool] = None
131
    max_seq_len_to_capture: int = 8192
132
    disable_custom_all_reduce: bool = False
133
    tokenizer_pool_size: int = 0
134
135
136
137
    # 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"
138
    tokenizer_pool_extra_config: Optional[Dict[str, Any]] = None
139
    limit_mm_per_prompt: Optional[Mapping[str, int]] = None
140
    mm_processor_kwargs: Optional[Dict[str, Any]] = None
141
    disable_mm_preprocessor_cache: bool = False
142
    enable_lora: bool = False
143
    enable_lora_bias: bool = False
144
145
    max_loras: int = 1
    max_lora_rank: int = 16
146
147
148
    enable_prompt_adapter: bool = False
    max_prompt_adapters: int = 1
    max_prompt_adapter_token: int = 0
149
    fully_sharded_loras: bool = False
150
    lora_extra_vocab_size: int = 256
151
    long_lora_scaling_factors: Optional[Tuple[float]] = None
152
    lora_dtype: Optional[Union[str, torch.dtype]] = 'auto'
153
    max_cpu_loras: Optional[int] = None
154
155
    merge_lora: bool = False
    lora_target_modules: Optional[List[str]] = 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
    num_speculative_heads: Optional[int] = None
177
    speculative_disable_mqa_scorer: Optional[bool] = False
178
    speculative_max_model_len: Optional[int] = None
179
    speculative_disable_by_batch_size: Optional[int] = None
180
181
    ngram_prompt_lookup_max: Optional[int] = None
    ngram_prompt_lookup_min: Optional[int] = None
182
183
184
    spec_decoding_acceptance_method: str = 'rejection_sampler'
    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
185
    qlora_adapter_name_or_path: Optional[str] = None
186
    disable_logprobs_during_spec_decoding: Optional[bool] = None
187

188
    otlp_traces_endpoint: Optional[str] = None
189
    collect_detailed_traces: Optional[str] = None
190
    disable_async_output_proc: bool = False
191
    scheduling_policy: Literal["fcfs", "priority"] = "fcfs"
192

193
    override_neuron_config: Optional[Dict[str, Any]] = None
194
    override_pooler_config: Optional[PoolerConfig] = None
195
    compilation_config: Optional[CompilationConfig] = None
196
    worker_cls: str = "auto"
197

198
    kv_transfer_config: Optional[KVTransferConfig] = None
199

200
    generation_config: Optional[str] = None
201
    override_generation_config: Optional[Dict[str, Any]] = None
202
    enable_sleep_mode: bool = False
203

204
    calculate_kv_scales: Optional[bool] = None
205

206
    def __post_init__(self):
207
        if not self.tokenizer:
208
            self.tokenizer = self.model
209

210
211
212
213
        # 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)
214

215
216
217
        # 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
218

219
220
221
        # support `EngineArgs(compilation_config={...})`
        # without having to manually construct a
        # CompilationConfig object
222
        if isinstance(self.compilation_config, (int, dict)):
223
224
            self.compilation_config = CompilationConfig.from_cli(
                str(self.compilation_config))
225

226
        # Setup plugins
227
228
        from vllm.plugins import load_general_plugins
        load_general_plugins()
229
230

    @staticmethod
231
    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
232
        """Shared CLI arguments for vLLM engine."""
233

234
        # Model arguments
235
236
237
        parser.add_argument(
            '--model',
            type=str,
238
            default=EngineArgs.model,
239
            help='Name or path of the huggingface model to use.')
240
241
242
243
244
245
        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 '
246
            'multiple tasks. When the model only supports one task, ``"auto"`` '
247
248
            'can be used to select it; otherwise, you must specify explicitly '
            'which task to use.')
249
250
        parser.add_argument(
            '--tokenizer',
251
            type=nullable_str,
252
            default=EngineArgs.tokenizer,
253
254
            help='Name or path of the huggingface tokenizer to use. '
            'If unspecified, model name or path will be used.')
255
256
257
        parser.add_argument(
            '--skip-tokenizer-init',
            action='store_true',
258
            help='Skip initialization of tokenizer and detokenizer.')
Jasmond L's avatar
Jasmond L committed
259
260
        parser.add_argument(
            '--revision',
261
            type=nullable_str,
Jasmond L's avatar
Jasmond L committed
262
            default=None,
263
            help='The specific model version to use. It can be a branch '
Jasmond L's avatar
Jasmond L committed
264
265
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
266
267
        parser.add_argument(
            '--code-revision',
268
            type=nullable_str,
269
            default=None,
270
            help='The specific revision to use for the model code on '
271
272
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
273
274
        parser.add_argument(
            '--tokenizer-revision',
275
            type=nullable_str,
276
            default=None,
277
278
279
            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.')
280
281
282
283
        parser.add_argument(
            '--tokenizer-mode',
            type=str,
            default=EngineArgs.tokenizer_mode,
284
            choices=['auto', 'slow', 'mistral'],
285
286
            help='The tokenizer mode.\n\n* "auto" will use the '
            'fast tokenizer if available.\n* "slow" will '
287
288
            'always use the slow tokenizer. \n* '
            '"mistral" will always use the `mistral_common` tokenizer.')
289
290
        parser.add_argument('--trust-remote-code',
                            action='store_true',
291
                            help='Trust remote code from huggingface.')
292
293
294
        parser.add_argument(
            '--allowed-local-media-path',
            type=str,
295
296
297
298
            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.")
299
        parser.add_argument('--download-dir',
300
                            type=nullable_str,
Zhuohan Li's avatar
Zhuohan Li committed
301
                            default=EngineArgs.download_dir,
302
                            help='Directory to download and load the weights, '
303
                            'default to the default cache dir of '
304
                            'huggingface.')
305
306
307
308
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
309
            choices=[f.value for f in LoadFormat],
310
311
            help='The format of the model weights to load.\n\n'
            '* "auto" will try to load the weights in the safetensors format '
312
            'and fall back to the pytorch bin format if safetensors format '
313
314
315
316
317
318
319
320
            '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 '
321
            'CoreWeave. See the Tensorize vLLM Model script in the Examples '
322
            'section for more information.\n'
323
324
            '* "runai_streamer" will load the Safetensors weights using Run:ai'
            'Model Streamer \n'
325
326
            '* "bitsandbytes" will load the weights using bitsandbytes '
            'quantization.\n')
327
328
329
330
331
332
333
        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 ')
334
335
336
337
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
338
339
340
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
341
342
343
344
345
346
347
348
            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.')
349
350
351
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
352
            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
353
            default=EngineArgs.kv_cache_dtype,
354
            help='Data type for kv cache storage. If "auto", will use model '
355
356
            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
357
358
        parser.add_argument('--max-model-len',
                            type=int,
359
                            default=EngineArgs.max_model_len,
360
361
                            help='Model context length. If unspecified, will '
                            'be automatically derived from the model config.')
362
363
364
        parser.add_argument(
            '--guided-decoding-backend',
            type=str,
365
366
            default='xgrammar',
            choices=['outlines', 'lm-format-enforcer', 'xgrammar'],
367
            help='Which engine will be used for guided decoding'
368
            ' (JSON schema / regex etc) by default. Currently support '
369
            'https://github.com/outlines-dev/outlines, '
370
            'https://github.com/mlc-ai/xgrammar, and '
371
372
373
            'https://github.com/noamgat/lm-format-enforcer.'
            ' Can be overridden per request via guided_decoding_backend'
            ' parameter.')
374
375
376
377
378
379
380
381
        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.')
382
        # Parallel arguments
383
384
        parser.add_argument(
            '--distributed-executor-backend',
385
            choices=['ray', 'mp', 'uni', 'external_launcher'],
386
            default=EngineArgs.distributed_executor_backend,
387
388
389
390
391
392
            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 '
393
            'only supports Ray for distributed inference.')
394

395
396
397
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
398
                            default=EngineArgs.pipeline_parallel_size,
399
                            help='Number of pipeline stages.')
400
401
402
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
403
                            default=EngineArgs.tensor_parallel_size,
404
                            help='Number of tensor parallel replicas.')
405
406
407
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
408
            default=EngineArgs.max_parallel_loading_workers,
409
            help='Load model sequentially in multiple batches, '
410
            'to avoid RAM OOM when using tensor '
411
            'parallel and large models.')
412
413
414
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
415
            help='If specified, use nsight to profile Ray workers.')
416
        # KV cache arguments
417
418
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
419
                            default=EngineArgs.block_size,
420
                            choices=[8, 16, 32, 64, 128],
421
                            help='Token block size for contiguous chunks of '
422
                            'tokens. This is ignored on neuron devices and '
423
                            'set to ``--max-model-len``. On CUDA devices, '
424
425
                            'only block sizes up to 32 are supported. '
                            'On HPU devices, block size defaults to 128.')
426

427
428
429
430
431
        parser.add_argument(
            "--enable-prefix-caching",
            action=argparse.BooleanOptionalAction,
            default=EngineArgs.enable_prefix_caching,
            help="Enables automatic prefix caching. "
432
            "Use ``--no-enable-prefix-caching`` to disable explicitly.",
433
        )
434
435
436
        parser.add_argument('--disable-sliding-window',
                            action='store_true',
                            help='Disables sliding window, '
437
                            'capping to sliding window size.')
438
439
        parser.add_argument('--use-v2-block-manager',
                            action='store_true',
440
                            default=True,
441
442
443
444
445
                            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.')
446
447
448
449
450
451
452
453
        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.')
454

455
456
457
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
458
                            help='Random seed for operations.')
459
        parser.add_argument('--swap-space',
460
                            type=float,
Zhuohan Li's avatar
Zhuohan Li committed
461
                            default=EngineArgs.swap_space,
462
                            help='CPU swap space size (GiB) per GPU.')
463
464
465
466
467
468
469
470
471
        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 '
472
            'a 34 GB GPU. Then you can load a 13B model with BF16 weight, '
473
            'which requires at least 26GB GPU memory. Note that this '
474
            'requires fast CPU-GPU interconnect, as part of the model is '
475
476
            'loaded from CPU memory to GPU memory on the fly in each '
            'model forward pass.')
477
478
479
480
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
481
482
483
            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, '
484
485
486
487
488
489
            '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.')
490
        parser.add_argument(
491
            '--num-gpu-blocks-override',
492
493
494
            type=int,
            default=None,
            help='If specified, ignore GPU profiling result and use this number'
495
            ' of GPU blocks. Used for testing preemption.')
496
497
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
498
                            default=EngineArgs.max_num_batched_tokens,
499
500
                            help='Maximum number of batched tokens per '
                            'iteration.')
501
502
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
503
                            default=EngineArgs.max_num_seqs,
504
                            help='Maximum number of sequences per iteration.')
505
506
507
508
        parser.add_argument(
            '--max-logprobs',
            type=int,
            default=EngineArgs.max_logprobs,
509
510
            help=('Max number of log probs to return logprobs is specified in'
                  ' SamplingParams.'))
511
512
        parser.add_argument('--disable-log-stats',
                            action='store_true',
513
                            help='Disable logging statistics.')
514
515
516
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
517
                            type=nullable_str,
518
                            choices=[*QUANTIZATION_METHODS, None],
519
                            default=EngineArgs.quantization,
520
521
522
523
524
525
                            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.')
526
527
528
529
530
        parser.add_argument(
            '--rope-scaling',
            default=None,
            type=json.loads,
            help='RoPE scaling configuration in JSON format. '
531
            'For example, ``{"rope_type":"dynamic","factor":2.0}``')
532
533
534
535
536
537
        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.')
538
539
540
        parser.add_argument('--hf-overrides',
                            type=json.loads,
                            default=EngineArgs.hf_overrides,
541
                            help='Extra arguments for the HuggingFace config. '
542
543
                            'This should be a JSON string that will be '
                            'parsed into a dictionary.')
544
545
546
547
548
        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.')
549
        parser.add_argument('--max-seq-len-to-capture',
550
551
552
553
                            type=int,
                            default=EngineArgs.max_seq_len_to_capture,
                            help='Maximum sequence length covered by CUDA '
                            'graphs. When a sequence has context length '
554
555
556
557
                            '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.')
558
559
560
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
561
                            help='See ParallelConfig.')
562
563
564
565
566
567
568
569
570
571
572
573
574
        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',
575
                            type=nullable_str,
576
577
578
579
580
                            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.')
581
582
583
584
585
586
587
588
589
590
591
592
593
594

        # 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.'))
595
596
597
598
        parser.add_argument(
            '--mm-processor-kwargs',
            default=None,
            type=json.loads,
599
            help=('Overrides for the multimodal input mapping/processing, '
600
                  'e.g., image processor. For example: ``{"num_crops": 4}``.'))
601
        parser.add_argument(
602
            '--disable-mm-preprocessor-cache',
603
            action='store_true',
604
605
            help='If true, then disables caching of the multi-modal '
            'preprocessor/mapper. (not recommended)')
606

607
608
609
610
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
611
612
613
        parser.add_argument('--enable-lora-bias',
                            action='store_true',
                            help='If True, enable bias for LoRA adapters.')
614
615
616
617
618
619
620
621
        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.')
622
623
624
625
626
627
628
629
        parser.add_argument('--merge-lora',
                            type=bool,
                            default=False,
                            help='If set to True, the weights of the base layer will be merged with the weights of Lora.')
        parser.add_argument('--lora-target-modules',
                            nargs='*',
                            default=None,
                            help='List of lora module name, If not specified, modules will be chosen according to the model architecture.')
630
631
632
633
634
635
636
637
638
639
640
        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,
641
            choices=['auto', 'float16', 'bfloat16'],
642
643
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
644
645
646
647
648
649
650
651
652
653
654
        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.'))
655
656
657
658
659
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
660
661
                  'Must be >= than max_loras. '
                  'Defaults to max_loras.'))
662
663
664
665
666
667
668
669
        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.'))
670
671
672
673
674
675
676
677
678
679
680
        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')
681
682
683
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
684
                            choices=DEVICE_OPTIONS,
685
                            help='Device type for vLLM execution.')
686
687
688
689
690
        parser.add_argument('--num-scheduler-steps',
                            type=int,
                            default=1,
                            help=('Maximum number of forward steps per '
                                  'scheduler call.'))
691

692
693
        parser.add_argument(
            '--multi-step-stream-outputs',
694
695
696
697
698
699
            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')
700
701
702
703
        parser.add_argument(
            '--scheduler-delay-factor',
            type=float,
            default=EngineArgs.scheduler_delay_factor,
704
            help='Apply a delay (of delay factor multiplied by previous '
705
            'prompt latency) before scheduling next prompt.')
706
707
        parser.add_argument(
            '--enable-chunked-prefill',
708
709
710
711
            action=StoreBoolean,
            default=EngineArgs.enable_chunked_prefill,
            nargs="?",
            const="True",
712
            help='If set, the prefill requests can be chunked based on the '
713
            'max_num_batched_tokens.')
714
715
716

        parser.add_argument(
            '--speculative-model',
717
            type=nullable_str,
718
            default=EngineArgs.speculative_model,
719
720
            help=
            'The name of the draft model to be used in speculative decoding.')
721
722
723
724
725
726
        # Quantization settings for speculative model.
        parser.add_argument(
            '--speculative-model-quantization',
            type=nullable_str,
            choices=[*QUANTIZATION_METHODS, None],
            default=EngineArgs.speculative_model_quantization,
727
            help='Method used to quantize the weights of speculative model. '
728
729
730
731
732
            '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.')
733
734
735
        parser.add_argument(
            '--num-speculative-tokens',
            type=int,
736
            default=EngineArgs.num_speculative_tokens,
737
            help='The number of speculative tokens to sample from '
738
            'the draft model in speculative decoding.')
739
740
741
742
743
744
        parser.add_argument(
            '--num-speculative-heads',
            type=int,
            default=EngineArgs.num_speculative_heads,
            help='The number of speculative heads to sample from '
                 'the draft model in speculative decoding.')
745
746
747
748
749
750
        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')
751
752
753
754
755
756
757
        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.')
758

759
760
        parser.add_argument(
            '--speculative-max-model-len',
761
            type=int,
762
763
764
765
766
            default=EngineArgs.speculative_max_model_len,
            help='The maximum sequence length supported by the '
            'draft model. Sequences over this length will skip '
            'speculation.')

767
768
769
770
771
772
773
        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.')

774
775
776
777
778
779
780
781
782
783
784
785
786
787
        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.')

788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
        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')

820
821
        parser.add_argument(
            '--disable-logprobs-during-spec-decoding',
822
            action=StoreBoolean,
823
            default=EngineArgs.disable_logprobs_during_spec_decoding,
824
825
            nargs="?",
            const="True",
826
827
828
829
830
831
832
833
            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.')

834
        parser.add_argument('--model-loader-extra-config',
835
                            type=nullable_str,
836
837
838
839
840
841
                            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.')
842
843
844
845
846
847
        parser.add_argument(
            '--ignore-patterns',
            action="append",
            type=str,
            default=[],
            help="The pattern(s) to ignore when loading the model."
848
            "Default to `original/**/*` to avoid repeated loading of llama's "
849
            "checkpoints.")
850
        parser.add_argument(
851
            '--preemption-mode',
852
853
            type=str,
            default=None,
854
855
856
            help='If \'recompute\', the engine performs preemption by '
            'recomputing; If \'swap\', the engine performs preemption by '
            'block swapping.')
857

858
859
860
861
862
863
864
865
866
867
        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 "
868
            "same as the ``--model`` argument. Noted that this name(s) "
869
            "will also be used in `model_name` tag content of "
870
            "prometheus metrics, if multiple names provided, metrics "
871
            "tag will take the first one.")
872
873
874
875
        parser.add_argument('--qlora-adapter-name-or-path',
                            type=str,
                            default=None,
                            help='Name or path of the QLoRA adapter.')
876
877
878
879
880
881

        parser.add_argument(
            '--otlp-traces-endpoint',
            type=str,
            default=None,
            help='Target URL to which OpenTelemetry traces will be sent.')
882
883
884
885
886
887
        parser.add_argument(
            '--collect-detailed-traces',
            type=str,
            default=None,
            help="Valid choices are " +
            ",".join(ALLOWED_DETAILED_TRACE_MODULES) +
888
            ". It makes sense to set this only if ``--otlp-traces-endpoint`` is"
889
890
891
            " 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.")
892

893
894
895
896
897
898
        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.")
899

900
901
902
903
904
905
906
907
908
909
        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).')

910
911
        parser.add_argument(
            '--override-neuron-config',
912
            type=json.loads,
913
            default=None,
914
            help="Override or set neuron device configuration. "
915
            "e.g. ``{\"cast_logits_dtype\": \"bloat16\"}``.")
916
        parser.add_argument(
917
918
            '--override-pooler-config',
            type=PoolerConfig.from_json,
919
            default=None,
920
            help="Override or set the pooling method for pooling models. "
921
            "e.g. ``{\"pooling_type\": \"mean\", \"normalize\": false}``.")
922

923
924
925
926
927
928
929
930
931
932
933
934
        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, '
935
936
937
938
                            'use a JSON string.\n'
                            'Following the convention of traditional '
                            'compilers, using -O without space is also '
                            'supported. -O3 is equivalent to -O 3.')
939

940
941
942
943
944
945
        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.')

946
947
948
949
950
        parser.add_argument(
            '--worker-cls',
            type=str,
            default="auto",
            help='The worker class to use for distributed execution.')
951
952
953
954
955
        parser.add_argument(
            "--generation-config",
            type=nullable_str,
            default=None,
            help="The folder path to the generation config. "
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
            "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.")
973

974
975
976
977
978
979
        parser.add_argument("--enable-sleep-mode",
                            action="store_true",
                            default=False,
                            help="Enable sleep mode for the engine. "
                            "(only cuda platform is supported)")

980
981
982
983
984
985
986
987
        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.')
988

989
        return parser
990
991

    @classmethod
992
    def from_cli_args(cls, args: argparse.Namespace):
993
994
995
        # 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
996
997
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
998

999
1000
    def create_model_config(self) -> ModelConfig:
        return ModelConfig(
1001
            model=self.model,
1002
            task=self.task,
1003
1004
            # We know this is not None because we set it in __post_init__
            tokenizer=cast(str, self.tokenizer),
1005
1006
            tokenizer_mode=self.tokenizer_mode,
            trust_remote_code=self.trust_remote_code,
1007
            allowed_local_media_path=self.allowed_local_media_path,
1008
1009
1010
1011
1012
            dtype=self.dtype,
            seed=self.seed,
            revision=self.revision,
            code_revision=self.code_revision,
            rope_scaling=self.rope_scaling,
1013
            rope_theta=self.rope_theta,
1014
            hf_overrides=self.hf_overrides,
1015
1016
1017
1018
1019
1020
1021
1022
            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,
1023
            served_model_name=self.served_model_name,
1024
            limit_mm_per_prompt=self.limit_mm_per_prompt,
1025
            use_async_output_proc=not self.disable_async_output_proc,
1026
            config_format=self.config_format,
1027
            mm_processor_kwargs=self.mm_processor_kwargs,
1028
            disable_mm_preprocessor_cache=self.disable_mm_preprocessor_cache,
1029
1030
            override_neuron_config=self.override_neuron_config,
            override_pooler_config=self.override_pooler_config,
1031
            logits_processor_pattern=self.logits_processor_pattern,
1032
            generation_config=self.generation_config,
1033
            override_generation_config=self.override_generation_config,
1034
1035
            enable_sleep_mode=self.enable_sleep_mode,
        )
1036

1037
1038
1039
1040
1041
1042
1043
1044
    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,
        )

1045
1046
1047
1048
1049
1050
    def create_engine_config(self,
                             usage_context: Optional[UsageContext] = None
                             ) -> VllmConfig:
        if envs.VLLM_USE_V1:
            self._override_v1_engine_args(usage_context)

1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
        # 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()

1078
1079
1080
1081
1082
        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.")
1083
1084
            self.enable_prefix_caching = False

1085
        cache_config = CacheConfig(
1086
            block_size=self.block_size,
1087
1088
1089
            gpu_memory_utilization=self.gpu_memory_utilization,
            swap_space=self.swap_space,
            cache_dtype=self.kv_cache_dtype,
1090
            is_attention_free=model_config.is_attention_free,
1091
1092
            num_gpu_blocks_override=self.num_gpu_blocks_override,
            sliding_window=model_config.get_sliding_window(),
1093
1094
            enable_prefix_caching=self.enable_prefix_caching,
            cpu_offload_gb=self.cpu_offload_gb,
1095
            calculate_kv_scales=self.calculate_kv_scales,
1096
        )
1097
        parallel_config = ParallelConfig(
1098
1099
1100
1101
1102
            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(
1103
1104
1105
                self.tokenizer_pool_size,
                self.tokenizer_pool_type,
                self.tokenizer_pool_extra_config,
1106
            ),
1107
            ray_workers_use_nsight=self.ray_workers_use_nsight,
1108
1109
1110
            distributed_executor_backend=self.distributed_executor_backend,
            worker_cls=self.worker_cls,
        )
1111

1112
1113
1114
1115
1116
1117
        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.
1118

1119
1120
1121
1122
1123
1124
            # 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:
1125
1126
1127
1128
                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
1129
                from vllm.platforms import current_platform
1130
1131
1132
                if (is_gpu and not use_sliding_window and not use_spec_decode
                        and not self.enable_lora
                        and not self.enable_prompt_adapter
1133
1134
                        and model_config.runner_type != "pooling"
                        and not current_platform.is_rocm()):
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
                    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)
1151
1152
        elif (self.enable_chunked_prefill
              and model_config.runner_type == "pooling"):
1153
            msg = "Chunked prefill is not supported for pooling models"
1154
            raise ValueError(msg)
1155

1156

1157
1158
1159
1160
1161
        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,
1162
1163
            speculative_model_quantization = \
                self.speculative_model_quantization,
1164
1165
            speculative_draft_tensor_parallel_size = \
                self.speculative_draft_tensor_parallel_size,
1166
            num_speculative_tokens=self.num_speculative_tokens,
1167
            speculative_disable_mqa_scorer=self.speculative_disable_mqa_scorer,
1168
1169
            speculative_disable_by_batch_size=self.
            speculative_disable_by_batch_size,
1170
1171
            speculative_max_model_len=self.speculative_max_model_len,
            enable_chunked_prefill=self.enable_chunked_prefill,
1172
            disable_log_stats=self.disable_log_stats,
1173
1174
            ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
            ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
1175
1176
1177
1178
1179
1180
            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,
1181
            disable_logprobs=self.disable_logprobs_during_spec_decoding,
1182
            num_speculative_heads=self.num_speculative_heads
1183
1184
        )

1185
        # Reminder: Please update docs/source/features/compatibility_matrix.md
1186
        # If the feature combo become valid
1187
1188
1189
1190
        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)")
1191
1192
1193
            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")
1194
1195
1196
1197
1198
1199
            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
1200
1201
1202
1203
1204
1205
1206
1207
1208

        # 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

1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
        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.")

1219
        scheduler_config = SchedulerConfig(
1220
            runner_type=model_config.runner_type,
1221
1222
1223
            max_num_batched_tokens=self.max_num_batched_tokens,
            max_num_seqs=self.max_num_seqs,
            max_model_len=model_config.max_model_len,
1224
            num_lookahead_slots=num_lookahead_slots,
1225
1226
            delay_factor=self.scheduler_delay_factor,
            enable_chunked_prefill=self.enable_chunked_prefill,
1227
            is_multimodal_model=model_config.is_multimodal_model,
1228
            preemption_mode=self.preemption_mode,
1229
            num_scheduler_steps=self.num_scheduler_steps,
1230
            multi_step_stream_outputs=self.multi_step_stream_outputs,
1231
1232
            send_delta_data=(envs.VLLM_USE_RAY_SPMD_WORKER
                             and parallel_config.use_ray),
1233
            policy=self.scheduling_policy)
1234
        lora_config = LoRAConfig(
1235
            bias_enabled=self.enable_lora_bias,
1236
1237
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
1238
            fully_sharded_loras=self.fully_sharded_loras,
1239
            lora_extra_vocab_size=self.lora_extra_vocab_size,
1240
            long_lora_scaling_factors=self.long_lora_scaling_factors,
1241
1242
            lora_dtype=self.lora_dtype,
            max_cpu_loras=self.max_cpu_loras if self.max_cpu_loras
1243
1244
1245
            and self.max_cpu_loras > 0 else None,
            merge_lora=self.merge_lora,
            lora_target_modules=self.lora_target_modules) if self.enable_lora else None
1246

1247
1248
1249
1250
1251
1252
1253
        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

1254
        load_config = self.create_load_config()
1255

1256
1257
1258
1259
1260
        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

1261
1262
1263
        decoding_config = DecodingConfig(
            guided_decoding_backend=self.guided_decoding_backend)

1264
1265
1266
1267
1268
1269
1270
1271
        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}")
1272
        observability_config = ObservabilityConfig(
1273
1274
1275
1276
1277
1278
            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,
        )
1279

1280
        config = VllmConfig(
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
            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,
1291
            prompt_adapter_config=prompt_adapter_config,
1292
            compilation_config=self.compilation_config,
1293
            kv_transfer_config=self.kv_transfer_config,
1294
        )
1295

1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
        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"

1306
1307
1308
1309
        # V1 always uses chunked prefills.
        self.enable_chunked_prefill = True
        # When no user override, set the default values based on the usage
        # context.
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
        # 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,
            }

1326
1327
1328
1329
        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]
1330
1331
1332
            logger.warning(
                "Setting max_num_batched_tokens to %d for %s usage context.",
                self.max_num_batched_tokens, usage_context.value)
1333
1334
1335
1336
1337
1338
1339

    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"

1340

1341
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
1342
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
1343
    """Arguments for asynchronous vLLM engine."""
1344
    disable_log_requests: bool = False
1345
1346

    @staticmethod
1347
1348
    def add_cli_args(parser: FlexibleArgumentParser,
                     async_args_only: bool = False) -> FlexibleArgumentParser:
1349
1350
        if not async_args_only:
            parser = EngineArgs.add_cli_args(parser)
1351
1352
        parser.add_argument('--disable-log-requests',
                            action='store_true',
1353
                            help='Disable logging requests.')
1354
        return parser
1355
1356
1357
1358


# These functions are used by sphinx to build the documentation
def _engine_args_parser():
1359
    return EngineArgs.add_cli_args(FlexibleArgumentParser())
1360
1361
1362


def _async_engine_args_parser():
1363
    return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
1364
                                        async_args_only=True)