arg_utils.py 16.2 KB
Newer Older
1
import argparse
2
3
4
import dataclasses
from dataclasses import dataclass
from typing import Optional, Tuple
5

6
7
from vllm.config import (CacheConfig, DeviceConfig, ModelConfig,
                         ParallelConfig, SchedulerConfig, LoRAConfig)
8
9


10
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
11
class EngineArgs:
Woosuk Kwon's avatar
Woosuk Kwon committed
12
    """Arguments for vLLM engine."""
13
    model: str
14
    tokenizer: Optional[str] = None
15
    tokenizer_mode: str = 'auto'
16
    trust_remote_code: bool = False
17
    download_dir: Optional[str] = None
18
    load_format: str = 'auto'
19
    dtype: str = 'auto'
20
    kv_cache_dtype: str = 'auto'
21
    seed: int = 0
22
    max_model_len: Optional[int] = None
23
    worker_use_ray: bool = False
24
25
    pipeline_parallel_size: int = 1
    tensor_parallel_size: int = 1
26
    max_parallel_loading_workers: Optional[int] = None
27
    block_size: int = 16
28
    enable_prefix_caching: bool = False
29
    swap_space: int = 4  # GiB
30
    gpu_memory_utilization: float = 0.90
31
    max_num_batched_tokens: Optional[int] = None
32
    max_num_seqs: int = 256
33
    max_paddings: int = 256
34
    disable_log_stats: bool = False
Jasmond L's avatar
Jasmond L committed
35
    revision: Optional[str] = None
36
    code_revision: Optional[str] = None
37
    tokenizer_revision: Optional[str] = None
38
    quantization: Optional[str] = None
39
40
    enforce_eager: bool = False
    max_context_len_to_capture: int = 8192
41
    disable_custom_all_reduce: bool = False
42
43
44
45
46
47
    enable_lora: bool = False
    max_loras: int = 1
    max_lora_rank: int = 16
    lora_extra_vocab_size: int = 256
    lora_dtype = 'auto'
    max_cpu_loras: Optional[int] = None
48
    device: str = 'auto'
49
    ray_workers_use_nsight: bool = False
50

51
    def __post_init__(self):
52
53
        if self.tokenizer is None:
            self.tokenizer = self.model
54
55
56

    @staticmethod
    def add_cli_args(
57
            parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
Woosuk Kwon's avatar
Woosuk Kwon committed
58
        """Shared CLI arguments for vLLM engine."""
59
60
61
62

        # NOTE: If you update any of the arguments below, please also
        # make sure to update docs/source/models/engine_args.rst

63
        # Model arguments
64
65
66
67
68
69
70
71
72
73
        parser.add_argument(
            '--model',
            type=str,
            default='facebook/opt-125m',
            help='name or path of the huggingface model to use')
        parser.add_argument(
            '--tokenizer',
            type=str,
            default=EngineArgs.tokenizer,
            help='name or path of the huggingface tokenizer to use')
Jasmond L's avatar
Jasmond L committed
74
75
76
77
78
79
80
        parser.add_argument(
            '--revision',
            type=str,
            default=None,
            help='the specific model version to use. It can be a branch '
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
81
82
83
84
85
86
87
        parser.add_argument(
            '--code-revision',
            type=str,
            default=None,
            help='the specific revision to use for the model code on '
            'Hugging Face Hub. It can be a branch name, a tag name, or a '
            'commit id. If unspecified, will use the default version.')
88
89
90
91
92
93
94
        parser.add_argument(
            '--tokenizer-revision',
            type=str,
            default=None,
            help='the specific tokenizer version to use. It can be a branch '
            'name, a tag name, or a commit id. If unspecified, will use '
            'the default version.')
95
96
        parser.add_argument('--tokenizer-mode',
                            type=str,
97
98
99
                            default=EngineArgs.tokenizer_mode,
                            choices=['auto', 'slow'],
                            help='tokenizer mode. "auto" will use the fast '
100
101
                            'tokenizer if available, and "slow" will '
                            'always use the slow tokenizer.')
102
103
104
        parser.add_argument('--trust-remote-code',
                            action='store_true',
                            help='trust remote code from huggingface')
105
106
        parser.add_argument('--download-dir',
                            type=str,
Zhuohan Li's avatar
Zhuohan Li committed
107
                            default=EngineArgs.download_dir,
108
                            help='directory to download and load the weights, '
109
110
                            'default to the default cache dir of '
                            'huggingface')
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
        parser.add_argument(
            '--load-format',
            type=str,
            default=EngineArgs.load_format,
            choices=['auto', 'pt', 'safetensors', 'npcache', 'dummy'],
            help='The format of the model weights to load. '
            '"auto" will try to load the weights in the safetensors format '
            'and fall back to the pytorch bin format if safetensors format '
            'is not available. '
            '"pt" will load the weights in the pytorch bin format. '
            '"safetensors" will load the weights in the safetensors format. '
            '"npcache" will load the weights in pytorch format and store '
            'a numpy cache to speed up the loading. '
            '"dummy" will initialize the weights with random values, '
            'which is mainly for profiling.')
126
127
128
129
        parser.add_argument(
            '--dtype',
            type=str,
            default=EngineArgs.dtype,
Woosuk Kwon's avatar
Woosuk Kwon committed
130
131
132
            choices=[
                'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
            ],
133
134
135
136
            help='data type for model weights and activations. '
            'The "auto" option will use FP16 precision '
            'for FP32 and FP16 models, and BF16 precision '
            'for BF16 models.')
137
138
139
140
        parser.add_argument(
            '--kv-cache-dtype',
            type=str,
            choices=['auto', 'fp8_e5m2'],
141
            default=EngineArgs.kv_cache_dtype,
142
143
144
            help='Data type for kv cache storage. If "auto", will use model '
            'data type. Note FP8 is not supported when cuda version is '
            'lower than 11.8.')
145
146
        parser.add_argument('--max-model-len',
                            type=int,
147
                            default=EngineArgs.max_model_len,
148
149
                            help='model context length. If unspecified, '
                            'will be automatically derived from the model.')
150
        # Parallel arguments
151
152
        parser.add_argument('--worker-use-ray',
                            action='store_true',
153
                            help='use Ray for distributed serving, will be '
154
155
156
157
                            'automatically set when using more than 1 GPU')
        parser.add_argument('--pipeline-parallel-size',
                            '-pp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
158
                            default=EngineArgs.pipeline_parallel_size,
159
                            help='number of pipeline stages')
160
161
162
        parser.add_argument('--tensor-parallel-size',
                            '-tp',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
163
                            default=EngineArgs.tensor_parallel_size,
164
                            help='number of tensor parallel replicas')
165
166
167
        parser.add_argument(
            '--max-parallel-loading-workers',
            type=int,
168
            default=EngineArgs.max_parallel_loading_workers,
169
170
171
            help='load model sequentially in multiple batches, '
            'to avoid RAM OOM when using tensor '
            'parallel and large models')
172
173
174
175
        parser.add_argument(
            '--ray-workers-use-nsight',
            action='store_true',
            help='If specified, use nsight to profile ray workers')
176
        # KV cache arguments
177
178
        parser.add_argument('--block-size',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
179
                            default=EngineArgs.block_size,
180
                            choices=[8, 16, 32, 128],
181
                            help='token block size')
182
183
184
185
186

        parser.add_argument('--enable-prefix-caching',
                            action='store_true',
                            help='Enables automatic prefix caching')

187
188
189
        parser.add_argument('--seed',
                            type=int,
                            default=EngineArgs.seed,
190
                            help='random seed')
191
192
        parser.add_argument('--swap-space',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
193
                            default=EngineArgs.swap_space,
194
                            help='CPU swap space size (GiB) per GPU')
195
196
197
198
199
200
201
        parser.add_argument(
            '--gpu-memory-utilization',
            type=float,
            default=EngineArgs.gpu_memory_utilization,
            help='the fraction of GPU memory to be used for '
            'the model executor, which can range from 0 to 1.'
            'If unspecified, will use the default value of 0.9.')
202
203
        parser.add_argument('--max-num-batched-tokens',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
204
                            default=EngineArgs.max_num_batched_tokens,
205
                            help='maximum number of batched tokens per '
206
207
208
                            'iteration')
        parser.add_argument('--max-num-seqs',
                            type=int,
Zhuohan Li's avatar
Zhuohan Li committed
209
                            default=EngineArgs.max_num_seqs,
210
                            help='maximum number of sequences per iteration')
211
212
213
214
        parser.add_argument('--max-paddings',
                            type=int,
                            default=EngineArgs.max_paddings,
                            help='maximum number of paddings in a batch')
215
216
        parser.add_argument('--disable-log-stats',
                            action='store_true',
217
                            help='disable logging statistics')
218
219
220
221
        # Quantization settings.
        parser.add_argument('--quantization',
                            '-q',
                            type=str,
CHU Tianxiang's avatar
CHU Tianxiang committed
222
                            choices=['awq', 'gptq', 'squeezellm', None],
223
                            default=EngineArgs.quantization,
224
225
226
227
228
229
                            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.')
230
231
232
233
234
235
236
237
238
239
240
        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.')
        parser.add_argument('--max-context-len-to-capture',
                            type=int,
                            default=EngineArgs.max_context_len_to_capture,
                            help='maximum context length covered by CUDA '
                            'graphs. When a sequence has context length '
                            'larger than this, we fall back to eager mode.')
241
242
243
244
        parser.add_argument('--disable-custom-all-reduce',
                            action='store_true',
                            default=EngineArgs.disable_custom_all_reduce,
                            help='See ParallelConfig')
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
        # LoRA related configs
        parser.add_argument('--enable-lora',
                            action='store_true',
                            help='If True, enable handling of LoRA adapters.')
        parser.add_argument('--max-loras',
                            type=int,
                            default=EngineArgs.max_loras,
                            help='Max number of LoRAs in a single batch.')
        parser.add_argument('--max-lora-rank',
                            type=int,
                            default=EngineArgs.max_lora_rank,
                            help='Max LoRA rank.')
        parser.add_argument(
            '--lora-extra-vocab-size',
            type=int,
            default=EngineArgs.lora_extra_vocab_size,
            help=('Maximum size of extra vocabulary that can be '
                  'present in a LoRA adapter (added to the base '
                  'model vocabulary).'))
        parser.add_argument(
            '--lora-dtype',
            type=str,
            default=EngineArgs.lora_dtype,
            choices=['auto', 'float16', 'bfloat16', 'float32'],
            help=('Data type for LoRA. If auto, will default to '
                  'base model dtype.'))
        parser.add_argument(
            '--max-cpu-loras',
            type=int,
            default=EngineArgs.max_cpu_loras,
            help=('Maximum number of LoRAs to store in CPU memory. '
                  'Must be >= than max_num_seqs. '
                  'Defaults to max_num_seqs.'))
278
279
280
281
282
        parser.add_argument("--device",
                            type=str,
                            default=EngineArgs.device,
                            choices=["auto", "cuda", "neuron"],
                            help='Device type for vLLM execution.')
283
        return parser
284
285

    @classmethod
286
    def from_cli_args(cls, args: argparse.Namespace) -> 'EngineArgs':
287
288
289
        # 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
290
291
        engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
        return engine_args
292

Zhuohan Li's avatar
Zhuohan Li committed
293
    def create_engine_configs(
294
        self,
295
    ) -> Tuple[ModelConfig, CacheConfig, ParallelConfig, SchedulerConfig,
296
297
               DeviceConfig, Optional[LoRAConfig]]:
        device_config = DeviceConfig(self.device)
298
299
300
301
302
303
        model_config = ModelConfig(
            self.model, self.tokenizer, self.tokenizer_mode,
            self.trust_remote_code, self.download_dir, self.load_format,
            self.dtype, self.seed, self.revision, self.code_revision,
            self.tokenizer_revision, self.max_model_len, self.quantization,
            self.enforce_eager, self.max_context_len_to_capture)
304
305
        cache_config = CacheConfig(self.block_size,
                                   self.gpu_memory_utilization,
306
                                   self.swap_space, self.kv_cache_dtype,
307
308
                                   model_config.get_sliding_window(),
                                   self.enable_prefix_caching)
309
310
        parallel_config = ParallelConfig(self.pipeline_parallel_size,
                                         self.tensor_parallel_size,
311
                                         self.worker_use_ray,
312
                                         self.max_parallel_loading_workers,
313
314
                                         self.disable_custom_all_reduce,
                                         self.ray_workers_use_nsight)
315
        scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
316
                                           self.max_num_seqs,
317
318
                                           model_config.max_model_len,
                                           self.max_paddings)
319
320
321
322
323
324
325
        lora_config = LoRAConfig(
            max_lora_rank=self.max_lora_rank,
            max_loras=self.max_loras,
            lora_extra_vocab_size=self.lora_extra_vocab_size,
            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
326
327
        return (model_config, cache_config, parallel_config, scheduler_config,
                device_config, lora_config)
328
329


330
@dataclass
Zhuohan Li's avatar
Zhuohan Li committed
331
class AsyncEngineArgs(EngineArgs):
Woosuk Kwon's avatar
Woosuk Kwon committed
332
    """Arguments for asynchronous vLLM engine."""
Zhuohan Li's avatar
Zhuohan Li committed
333
    engine_use_ray: bool = False
334
    disable_log_requests: bool = False
335
    max_log_len: Optional[int] = None
336
337
338

    @staticmethod
    def add_cli_args(
339
            parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
Zhuohan Li's avatar
Zhuohan Li committed
340
        parser = EngineArgs.add_cli_args(parser)
341
342
        parser.add_argument('--engine-use-ray',
                            action='store_true',
Zhuohan Li's avatar
Zhuohan Li committed
343
                            help='use Ray to start the LLM engine in a '
344
345
346
                            'separate process as the server process.')
        parser.add_argument('--disable-log-requests',
                            action='store_true',
347
                            help='disable logging requests')
348
349
350
351
352
353
        parser.add_argument('--max-log-len',
                            type=int,
                            default=None,
                            help='max number of prompt characters or prompt '
                            'ID numbers being printed in log. '
                            'Default: unlimited.')
354
        return parser