config.py 95.3 KB
Newer Older
1
import copy
2
import enum
3
import json
4
import warnings
5
from dataclasses import dataclass, field, replace
6
7
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Dict, Final, List,
                    Literal, Mapping, Optional, Set, Tuple, Type, Union)
8
9

import torch
10
from transformers import PretrainedConfig
11

12
import vllm.envs as envs
Woosuk Kwon's avatar
Woosuk Kwon committed
13
from vllm.logger import init_logger
14
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
15
from vllm.model_executor.models import ModelRegistry
16
from vllm.platforms import current_platform
17
from vllm.tracing import is_otel_available, otel_import_error_traceback
18
19
20
21
from vllm.transformers_utils.config import (
    ConfigFormat, get_config, get_hf_image_processor_config,
    get_hf_text_config, get_pooling_config,
    get_sentence_transformer_tokenizer_config, is_encoder_decoder, uses_mrope)
22
from vllm.utils import (GiB_bytes, cuda_device_count_stateless, get_cpu_memory,
23
                        identity, print_warning_once)
24

25
26
27
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

28
    from vllm.executor.executor_base import ExecutorBase
29
30
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
31
    from vllm.model_executor.model_loader.loader import BaseModelLoader
32
33
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
34
35
else:
    QuantizationConfig = None
36

37
38
logger = init_logger(__name__)

39
_EMBEDDING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
40
_MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
41

42
43
44
45
TaskOption = Literal["auto", "generate", "embedding"]

# "draft" is only used internally for speculative decoding
_Task = Literal["generate", "embedding", "draft"]
46

47
48
49
HfOverrides = Union[Dict[str, Any], Callable[[PretrainedConfig],
                                             PretrainedConfig]]

50
51

class ModelConfig:
52
53
54
55
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
56
            It is also used as the content for `model_name` tag in metrics
57
58
59
60
61
            output when `served_model_name` is not specified.
        task: The task to use the model for. Each vLLM instance only supports
            one task, even if the same model can be used for multiple tasks.
            When the model only supports one task, "auto" can be used to select
            it; otherwise, you must specify explicitly which task to use.
62
        tokenizer: Name or path of the huggingface tokenizer to use.
63
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
64
65
            available, "slow" will always use the slow tokenizer, and
            "mistral" will always use the tokenizer from `mistral_common`.
66
67
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
68
69
70
71
        allowed_local_media_path: 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.
72
73
74
75
        dtype: 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.
        seed: Random seed for reproducibility.
Jasmond L's avatar
Jasmond L committed
76
77
78
        revision: 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.
79
        code_revision: The specific revision to use for the model code on
80
            Hugging Face Hub. It can be a branch name, a tag name, or a
81
            commit id. If unspecified, will use the default version.
82
83
84
        tokenizer_revision: 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.
85
86
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
87
88
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
89
90
        quantization_param_path: Path to JSON file containing scaling factors.
            Used to load KV cache scaling factors into the model when KV cache
91
92
            type is FP8_E4M3 on ROCm (AMD GPU). In the future these will also
            be used to load activation and weight scaling factors when the
93
            model dtype is FP8_E4M3 on ROCm.
94
95
96
        enforce_eager: Whether to enforce eager execution. If True, we will
            disable CUDA graph and always execute the model in eager mode.
            If False, we will use CUDA graph and eager execution in hybrid.
97
            If None, the user did not specify, so default to False.
98
99
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
100
101
102
            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.
103
104
105
106
        disable_sliding_window: Whether to disable sliding window. If True,
            we will disable the sliding window functionality of the model.
            If the model does not support sliding window, this argument is
            ignored.
107
108
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
109
        served_model_name: The model name used in metrics tag `model_name`,
110
111
            matches the model name exposed via the APIs. If multiple model
            names provided, the first name will be used. If not specified,
112
            the model name will be the same as `model`.
113
        limit_mm_per_prompt: Maximum number of data items per modality
114
            per prompt. Only applicable for multimodal models.
115
116
        config_format: The config format which shall be loaded.
            Defaults to 'auto' which defaults to 'hf'.
117
118
119
        hf_overrides: If a dictionary, contains arguments to be forwarded to the
            HuggingFace config. If a callable, it is called to update the
            HuggingFace config.
120
121
        mm_processor_kwargs: Arguments to be forwarded to the model's processor
            for multi-modal data, e.g., image processor.
122
123
124
125
126
127
        override_neuron_config: Initialize non default neuron config or
            override default neuron config that are specific to Neuron devices,
            this argument will be used to configure the neuron config that
            can not be gathered from the vllm arguments.
        override_pooling_config: Initialize non default pooling config or
            override default pooling config for the embedding model.
128
    """
129

130
131
132
133
134
135
136
137
138
    def __init__(
            self,
            model: str,
            task: Union[TaskOption, _Task],
            tokenizer: str,
            tokenizer_mode: str,
            trust_remote_code: bool,
            dtype: Union[str, torch.dtype],
            seed: int,
139
            allowed_local_media_path: str = "",
140
141
            revision: Optional[str] = None,
            code_revision: Optional[str] = None,
142
            rope_scaling: Optional[Dict[str, Any]] = None,
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
            rope_theta: Optional[float] = None,
            tokenizer_revision: Optional[str] = None,
            max_model_len: Optional[int] = None,
            spec_target_max_model_len: Optional[int] = None,
            quantization: Optional[str] = None,
            quantization_param_path: Optional[str] = None,
            enforce_eager: Optional[bool] = None,
            max_seq_len_to_capture: Optional[int] = None,
            max_logprobs: int = 20,
            disable_sliding_window: bool = False,
            skip_tokenizer_init: bool = False,
            served_model_name: Optional[Union[str, List[str]]] = None,
            limit_mm_per_prompt: Optional[Mapping[str, int]] = None,
            use_async_output_proc: bool = True,
            config_format: ConfigFormat = ConfigFormat.AUTO,
158
            hf_overrides: Optional[HfOverrides] = None,
159
            mm_processor_kwargs: Optional[Dict[str, Any]] = None,
160
161
            override_neuron_config: Optional[Dict[str, Any]] = None,
            override_pooler_config: Optional["PoolerConfig"] = None) -> None:
162
        self.model = model
163
        self.tokenizer = tokenizer
164
        self.tokenizer_mode = tokenizer_mode
165
        self.trust_remote_code = trust_remote_code
166
        self.allowed_local_media_path = allowed_local_media_path
167
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
168
        self.revision = revision
169
        self.code_revision = code_revision
170
171
172

        if hf_overrides is None:
            hf_overrides = {}
173
174
175
176
177
178
179
180

        if callable(hf_overrides):
            hf_overrides_kw = {}
            hf_overrides_fn = hf_overrides
        else:
            hf_overrides_kw = hf_overrides
            hf_overrides_fn = identity

181
182
        if rope_scaling is not None:
            hf_override: Dict[str, Any] = {"rope_scaling": rope_scaling}
183
            hf_overrides_kw.update(hf_override)
184
185
186
187
188
            msg = ("`--rope-scaling` will be removed in a future release. "
                   f"'Please instead use `--hf-overrides '{hf_override!r}'`")
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
        if rope_theta is not None:
            hf_override = {"rope_theta": rope_theta}
189
            hf_overrides_kw.update(hf_override)
190
191
192
193
            msg = ("`--rope-theta` will be removed in a future release. "
                   f"'Please instead use `--hf-overrides '{hf_override!r}'`")
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

194
195
196
197
198
        # The tokenizer version is consistent with the model version by default.
        if tokenizer_revision is None:
            self.tokenizer_revision = revision
        else:
            self.tokenizer_revision = tokenizer_revision
199
        self.quantization = quantization
200
        self.quantization_param_path = quantization_param_path
201
        self.enforce_eager = enforce_eager
202
        self.max_seq_len_to_capture = max_seq_len_to_capture
203
        self.max_logprobs = max_logprobs
204
        self.disable_sliding_window = disable_sliding_window
205
        self.skip_tokenizer_init = skip_tokenizer_init
206
207
208
209
210
211

        hf_config = get_config(self.model, trust_remote_code, revision,
                               code_revision, config_format, **hf_overrides_kw)
        hf_config = hf_overrides_fn(hf_config)
        self.hf_config = hf_config

212
        self.hf_text_config = get_hf_text_config(self.hf_config)
213
        self.encoder_config = self._get_encoder_config()
214
215
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, revision)
216
        self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
217
        self.use_async_output_proc = use_async_output_proc
218
        self.mm_processor_kwargs = mm_processor_kwargs
Woosuk Kwon's avatar
Woosuk Kwon committed
219

220
221
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
222
223
            self.enforce_eager = False

224
225
226
227
228
229
230
231
232
        sliding_window = getattr(self.hf_text_config, "sliding_window", None)
        has_interleaved_attention = (sliding_window is not None) and (
            isinstance(sliding_window, list) or
            (self.hf_text_config.model_type in ["gemma2"]))

        if (not self.disable_sliding_window and has_interleaved_attention):
            sliding_window_len_min = get_min_sliding_window(
                self.hf_text_config.sliding_window)

Woosuk Kwon's avatar
Woosuk Kwon committed
233
            print_warning_once(
234
                f"{self.hf_text_config.model_type} has interleaved attention, "
Woosuk Kwon's avatar
Woosuk Kwon committed
235
236
                "which is currently not supported by vLLM. Disabling sliding "
                "window and capping the max length to the sliding window size "
237
                f"({sliding_window_len_min}).")
Woosuk Kwon's avatar
Woosuk Kwon committed
238
239
            self.disable_sliding_window = True

240
241
242
243
        self.max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
244
            sliding_window_len=self.get_hf_config_sliding_window(),
245
246
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
247
248
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
249
250
        self.multimodal_config = self._init_multimodal_config(
            limit_mm_per_prompt)
251
252
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
253

254
255
256
        self.is_attention_free = self._init_attention_free()
        self.has_inner_state = self._init_has_inner_state()

257
258
259
260
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
261
262
263
264

        supported_tasks, task = self._resolve_task(task, self.hf_config)
        self.supported_tasks = supported_tasks
        self.task: Final = task
265
        self.pooler_config = self._init_pooler_config(override_pooler_config)
266

267
        self._verify_quantization()
268
        self._verify_cuda_graph()
269
        self._verify_bnb_config()
270

271
272
273
274
    def _init_multimodal_config(
        self, limit_mm_per_prompt: Optional[Mapping[str, int]]
    ) -> Optional["MultiModalConfig"]:
        architectures = getattr(self.hf_config, "architectures", [])
275
        if ModelRegistry.is_multimodal_model(architectures):
276
            return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
277
278
279
280
281
282

        if limit_mm_per_prompt:
            raise ValueError("`limit_mm_per_prompt` is only supported for "
                             "multimodal models.")

        return None
283

284
285
286
287
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

288
289
    def _init_pooler_config(
        self,
290
        override_pooler_config: Optional["PoolerConfig"],
291
    ) -> Optional["PoolerConfig"]:
292

293
        if self.task == "embedding":
294
295
296
297
298
299
300
301
302
303
304
            user_config = override_pooler_config or PoolerConfig()

            base_config = get_pooling_config(self.model, self.revision)
            if base_config is not None:
                # Only set values that are not overridden by the user
                for k, v in base_config.items():
                    if getattr(user_config, k) is None:
                        setattr(user_config, k, v)

            return user_config

305
306
        return None

307
308
309
310
311
312
313
314
    def _init_attention_free(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_attention_free_model(architectures)

    def _init_has_inner_state(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.model_has_inner_state(architectures)

315
316
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
317
        if tokenizer_mode not in ["auto", "slow", "mistral"]:
318
319
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
320
                "either 'auto', 'slow' or 'mistral'.")
321
        self.tokenizer_mode = tokenizer_mode
322

323
324
    def _resolve_task(
        self,
325
        task_option: Union[TaskOption, _Task],
326
        hf_config: PretrainedConfig,
327
328
329
330
    ) -> Tuple[Set[_Task], _Task]:
        if task_option == "draft":
            return {"draft"}, "draft"

331
332
        architectures = getattr(hf_config, "architectures", [])

333
        task_support: Dict[_Task, bool] = {
334
335
336
337
338
            # NOTE: Listed from highest to lowest priority,
            # in case the model supports multiple of them
            "generate": ModelRegistry.is_text_generation_model(architectures),
            "embedding": ModelRegistry.is_embedding_model(architectures),
        }
339
        supported_tasks_lst: List[_Task] = [
340
341
342
343
344
345
            task for task, is_supported in task_support.items() if is_supported
        ]
        supported_tasks = set(supported_tasks_lst)

        if task_option == "auto":
            selected_task = next(iter(supported_tasks_lst))
346

347
348
349
350
            if len(supported_tasks) > 1:
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
351
        else:
352
353
354
355
356
357
358
            if task_option not in supported_tasks:
                msg = (
                    f"This model does not support the '{task_option}' task. "
                    f"Supported tasks: {supported_tasks}")
                raise ValueError(msg)

            selected_task = task_option
359

360
        return supported_tasks, selected_task
361

362
363
364
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
365
            # compressed-tensors uses a "compression_config" key
366
            quant_cfg = getattr(self.hf_config, "compression_config", None)
367
368
        return quant_cfg

369
    def _verify_quantization(self) -> None:
370
        supported_quantization = [*QUANTIZATION_METHODS]
371
372
373
374
        rocm_supported_quantization = [
            "awq", "gptq", "fp8", "compressed_tensors", "compressed-tensors",
            "fbgemm_fp8"
        ]
375
        optimized_quantization_methods = [
376
377
378
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
            "awq_marlin", "fbgemm_fp8", "compressed_tensors",
            "compressed-tensors", "experts_int8"
379
        ]
380
        tpu_supported_quantization = ["tpu_int8"]
381
        neuron_supported_quantization = ["neuron_quant"]
382
383
384
385
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
386
387
        quant_cfg = self._parse_quant_hf_config()

388
389
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
390
391

            # Detect which checkpoint is it
392
            for _, method in QUANTIZATION_METHODS.items():
393
394
395
396
397
398
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
399

400
            # Verify quantization configurations.
401
            if self.quantization is None:
402
403
                self.quantization = quant_method
            elif self.quantization != quant_method:
404
405
                raise ValueError(
                    "Quantization method specified in the model config "
406
                    f"({quant_method}) does not match the quantization "
407
408
409
410
411
412
413
414
                    f"method specified in the `quantization` argument "
                    f"({self.quantization}).")

        if self.quantization is not None:
            if self.quantization not in supported_quantization:
                raise ValueError(
                    f"Unknown quantization method: {self.quantization}. Must "
                    f"be one of {supported_quantization}.")
415
            if current_platform.is_rocm(
416
            ) and self.quantization not in rocm_supported_quantization:
417
                raise ValueError(
418
419
                    f"{self.quantization} quantization is currently not "
                    f"supported in ROCm.")
420
            if current_platform.is_tpu(
421
422
423
424
            ) and self.quantization not in tpu_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in TPU Backend.")
425
            if self.quantization not in optimized_quantization_methods:
426
                logger.warning(
427
                    "%s quantization is not fully "
428
                    "optimized yet. The speed can be slower than "
429
                    "non-quantized models.", self.quantization)
430
            if (self.quantization == "awq" and current_platform.is_rocm()
431
432
433
434
435
                    and not envs.VLLM_USE_TRITON_AWQ):
                logger.warning(
                    "Using AWQ quantization with ROCm, but VLLM_USE_TRITON_AWQ"
                    " is not set, enabling VLLM_USE_TRITON_AWQ.")
                envs.VLLM_USE_TRITON_AWQ = True
436
            if current_platform.is_neuron(
437
438
439
440
            ) and self.quantization not in neuron_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in Neuron Backend.")
441

442
    def _verify_cuda_graph(self) -> None:
443
444
445
446
        if self.max_seq_len_to_capture is None:
            self.max_seq_len_to_capture = self.max_model_len
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
447

448
449
    def _verify_bnb_config(self) -> None:
        """
450
        The current version of bitsandbytes (0.44.0) with 8-bit models does not
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
        yet support CUDA graph.
        """
        is_bitsandbytes = self.quantization == "bitsandbytes"
        has_quantization_config = (getattr(self.hf_config,
                                           "quantization_config", None)
                                   is not None)
        is_8bit = (self.hf_config.quantization_config.get(
            "load_in_8bit", False) if has_quantization_config else False)
        if all([
                is_bitsandbytes,
                has_quantization_config,
                is_8bit,
                not self.enforce_eager,
        ]):
            logger.warning(
                "CUDA graph is not supported on BitAndBytes 8bit yet, "
                "fallback to the eager mode.")
            self.enforce_eager = True

470
471
472
473
474
475
476
477
478
479
480
481
    def verify_async_output_proc(self, parallel_config, speculative_config,
                                 device_config) -> None:
        if not self.use_async_output_proc:
            # Nothing to check
            return

        if parallel_config.pipeline_parallel_size > 1:
            logger.warning("Async output processing can not be enabled "
                           "with pipeline parallel")
            self.use_async_output_proc = False
            return

482
483
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
484
        if device_config.device_type not in ("cuda", "tpu", "xpu", "hpu"):
485
            logger.warning(
486
487
                "Async output processing is only supported for CUDA, TPU, XPU "
                "and HPU."
488
                "Disabling it for other platforms.")
489
490
491
492
493
494
495
496
497
            self.use_async_output_proc = False
            return

        if envs.VLLM_USE_RAY_SPMD_WORKER:
            logger.warning(
                "Async output processing can not be enabled with ray spmd")
            self.use_async_output_proc = False
            return

498
499
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
500
        if device_config.device_type == "cuda" and self.enforce_eager:
501
502
503
504
505
506
507
508
509
            logger.warning(
                "To see benefits of async output processing, enable CUDA "
                "graph. Since, enforce-eager is enabled, async output "
                "processor cannot be used")
            self.use_async_output_proc = not self.enforce_eager
            return

        # Async postprocessor is not necessary with embedding mode
        # since there is no token generation
510
        if self.task == "embedding":
511
512
            self.use_async_output_proc = False

513
514
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
515
516
517
518
519
        if speculative_config:
            logger.warning("Async output processing is not supported with"
                           " speculative decoding currently.")
            self.use_async_output_proc = False

520
521
522
523
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
524
525
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
526
527
528
529
530
531
532
533
        tensor_parallel_size = parallel_config.tensor_parallel_size
        if total_num_attention_heads % tensor_parallel_size != 0:
            raise ValueError(
                f"Total number of attention heads ({total_num_attention_heads})"
                " must be divisible by tensor parallel size "
                f"({tensor_parallel_size}).")

        pipeline_parallel_size = parallel_config.pipeline_parallel_size
534
535
536
537
538
539
540
541
542
543
544
        if pipeline_parallel_size > 1:
            architectures = getattr(self.hf_config, "architectures", [])
            if not ModelRegistry.is_pp_supported_model(architectures):
                raise NotImplementedError(
                    "Pipeline parallelism is not supported for this model. "
                    "Supported models implement the `SupportsPP` interface.")

            if self.use_async_output_proc:
                logger.warning("Async output processor is not supported with "
                               "pipeline parallelism currently. Disabling it.")
                self.use_async_output_proc = False
545

546
547
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], List[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
548
        """Get the sliding window size, or None if disabled."""
549
550
551
552

        # Some models, like Qwen2 and Qwen1.5, use `use_sliding_window` in
        # addition to sliding window size. We check if that field is present
        # and if it's False, return None.
553
554
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
555
            return None
556
        return getattr(self.hf_text_config, "sliding_window", None)
557

558
    def get_sliding_window(self) -> Optional[Union[int, List[Optional[int]]]]:
559
560
561
562
563
564
565
566
        """Get the sliding window size, or None if disabled.
        """
        # If user disables sliding window, return None.
        if self.disable_sliding_window:
            return None
        # Otherwise get the value from the hf config.
        return self.get_hf_config_sliding_window()

567
    def get_vocab_size(self) -> int:
568
        return self.hf_text_config.vocab_size
569

570
    def get_hidden_size(self) -> int:
571
        return self.hf_text_config.hidden_size
572
573

    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
574
575
576
577
578
579
        # TODO remove hard code
        if hasattr(self.hf_text_config, "model_type"
                   ) and self.hf_text_config.model_type == 'deepseek_v2':
            # FlashAttention supports only head_size 32, 64, 128, 256,
            # we need to pad head_size 192 to 256
            return 256
580
581
582
583

        if self.is_attention_free:
            return 0

584
585
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
586
        # FIXME(woosuk): This may not be true for all models.
587
588
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
589

590
591
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
592
        # For GPTBigCode & Falcon:
593
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
594
595
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
596
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
597
        new_decoder_arch_falcon = (
598
            self.hf_config.model_type in falcon_model_types
599
            and getattr(self.hf_config, "new_decoder_architecture", False))
600
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
601
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
602
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
603
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
604
            return 1
605

606
        # For DBRX and MPT
607
608
609
610
611
        if self.hf_config.model_type == "mpt":
            if "kv_n_heads" in self.hf_config.attn_config:
                return self.hf_config.attn_config["kv_n_heads"]
            return self.hf_config.num_attention_heads
        if self.hf_config.model_type == "dbrx":
612
613
614
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

615
616
617
        if self.is_attention_free:
            return 0

618
619
620
621
622
623
624
625
626
627
        attributes = [
            # For Falcon:
            "n_head_kv",
            "num_kv_heads",
            # For LLaMA-2:
            "num_key_value_heads",
            # For ChatGLM:
            "multi_query_group_num",
        ]
        for attr in attributes:
628
            num_kv_heads = getattr(self.hf_text_config, attr, None)
629
630
631
632
633
            if num_kv_heads is not None:
                return num_kv_heads

        # For non-grouped-query attention models, the number of KV heads is
        # equal to the number of attention heads.
634
        return self.hf_text_config.num_attention_heads
635
636
637
638
639
640
641
642
643
644

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
        total_num_kv_heads = self.get_total_num_kv_heads()
        # If tensor parallelism is used, we divide the number of KV heads by
        # the tensor parallel size. We will replicate the KV heads in the
        # case where the number of KV heads is smaller than the tensor
        # parallel size so each GPU has at least one KV head.
        return max(1,
                   total_num_kv_heads // parallel_config.tensor_parallel_size)
645

646
647
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
648
649
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
650

651
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
652
        from vllm.distributed.utils import get_pp_indices
Mor Zusman's avatar
Mor Zusman committed
653
654
        total_num_hidden_layers = getattr(self.hf_text_config,
                                          "num_hidden_layers", 0)
655
656
657
658
        pp_rank = parallel_config.rank // parallel_config.tensor_parallel_size
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
        return end - start
659

660
661
662
663
    def get_num_attention_layers(self,
                                 parallel_config: "ParallelConfig") -> int:
        if self.is_attention_free:
            return 0
Mor Zusman's avatar
Mor Zusman committed
664
665
666

        num_layers = self.get_num_layers(parallel_config)

667
668
669
670
        # Transformers supports layers_block_type @property
        layers = getattr(self.hf_config, "layers_block_type",
                         ["attention"] * num_layers)
        return len([t for t in layers if t == "attention"])
Mor Zusman's avatar
Mor Zusman committed
671

672
673
674
675
676
677
678
679
680
681
682
683
    def get_multimodal_config(self) -> "MultiModalConfig":
        """
        Get the multimodal configuration of the model.

        Raises:
            ValueError: If the model is not multimodal.
        """
        if self.multimodal_config is None:
            raise ValueError("The model is not multimodal.")

        return self.multimodal_config

684
    @property
685
    def is_encoder_decoder(self) -> bool:
686
        """Extract the HF encoder/decoder model flag."""
687
688
689
690
691
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
692

693
694
695
696
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

697
698

class CacheConfig:
699
700
701
702
703
    """Configuration for the KV cache.

    Args:
        block_size: Size of a cache block in number of tokens.
        gpu_memory_utilization: Fraction of GPU memory to use for the
Woosuk Kwon's avatar
Woosuk Kwon committed
704
            vLLM execution.
705
        swap_space: Size of the CPU swap space per GPU (in GiB).
706
        cache_dtype: Data type for kv cache storage.
707
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
708
            profiled num_gpu_blocks if specified. Does nothing if None.
709
    """
710

711
712
713
714
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
715
        swap_space: float,
716
        cache_dtype: str,
717
        is_attention_free: bool = False,
718
        num_gpu_blocks_override: Optional[int] = None,
719
        sliding_window: Optional[int] = None,
720
        enable_prefix_caching: bool = False,
721
        cpu_offload_gb: float = 0,
722
723
724
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
725
        self.swap_space_bytes = swap_space * GiB_bytes
726
        self.num_gpu_blocks_override = num_gpu_blocks_override
727
        self.cache_dtype = cache_dtype
728
        self.is_attention_free = is_attention_free
729
        self.sliding_window = sliding_window
730
        self.enable_prefix_caching = enable_prefix_caching
731
        self.cpu_offload_gb = cpu_offload_gb
732

733
        self._verify_args()
734
        self._verify_cache_dtype()
735
        self._verify_prefix_caching()
736
737

        # Will be set after profiling.
738
739
        self.num_gpu_blocks: Optional[int] = None
        self.num_cpu_blocks: Optional[int] = None
740

741
    def metrics_info(self):
742
743
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
744
745
        return {key: str(value) for key, value in self.__dict__.items()}

746
747
748
749
750
751
    def _verify_args(self) -> None:
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

752
753
754
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
755
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
756
            logger.info(
757
758
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
759
760
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
761
762
763
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

764
765
766
767
768
769
770
771
772
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

        if self.sliding_window is not None:
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

773
774
775
776
777
778
779
780
781
782
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
        total_cpu_memory = get_cpu_memory()
        # FIXME(woosuk): Here, it is assumed that the GPUs in a tensor parallel
        # group are in the same node. However, the GPUs may span multiple nodes.
        num_gpus_per_node = parallel_config.tensor_parallel_size
        cpu_memory_usage = self.swap_space_bytes * num_gpus_per_node

783
784
785
        msg = (f"{cpu_memory_usage / GiB_bytes:.2f} GiB out of the "
               f"{total_cpu_memory / GiB_bytes:.2f} GiB total CPU memory "
               "is allocated for the swap space.")
786
787
788
        if cpu_memory_usage > 0.7 * total_cpu_memory:
            raise ValueError("Too large swap space. " + msg)
        elif cpu_memory_usage > 0.4 * total_cpu_memory:
789
            logger.warning("Possibly too large swap space. %s", msg)
790

791

792
793
794
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
795

796
797
798
799
800
801
802
803
    Args:
        pool_size: Number of tokenizer workers in the pool.
        pool_type: Type of the pool.
        extra_config: Additional config for the pool.
            The way the config will be used depends on the
            pool type.
    """
    pool_size: int
804
    pool_type: Union[str, Type["BaseTokenizerGroup"]]
805
806
807
    extra_config: dict

    def __post_init__(self):
808
809
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
810
811
812
813
814
815
            raise ValueError(f"Unknown pool type: {self.pool_type}")
        if not isinstance(self.extra_config, dict):
            raise ValueError("extra_config must be a dictionary.")

    @classmethod
    def create_config(
816
817
        cls, tokenizer_pool_size: int,
        tokenizer_pool_type: Union[str, Type["BaseTokenizerGroup"]],
818
819
820
        tokenizer_pool_extra_config: Optional[Union[str, dict]]
    ) -> Optional["TokenizerPoolConfig"]:
        """Create a TokenizerPoolConfig from the given parameters.
821

822
        If tokenizer_pool_size is 0, return None.
823

824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
        Args:
            tokenizer_pool_size: Number of tokenizer workers in the pool.
            tokenizer_pool_type: Type of the pool.
            tokenizer_pool_extra_config: Additional config for the pool.
                The way the config will be used depends on the
                pool type. This can be a JSON string (will be parsed).
        """
        if tokenizer_pool_size:
            if isinstance(tokenizer_pool_extra_config, str):
                tokenizer_pool_extra_config_parsed = json.loads(
                    tokenizer_pool_extra_config)
            else:
                tokenizer_pool_extra_config_parsed = (
                    tokenizer_pool_extra_config or {})
            tokenizer_pool_config = cls(tokenizer_pool_size,
                                        tokenizer_pool_type,
                                        tokenizer_pool_extra_config_parsed)
        else:
            tokenizer_pool_config = None
        return tokenizer_pool_config


846
847
848
849
850
851
852
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
853
    SHARDED_STATE = "sharded_state"
854
    GGUF = "gguf"
855
    BITSANDBYTES = "bitsandbytes"
856
    MISTRAL = "mistral"
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875


@dataclass
class LoadConfig:
    """
        download_dir: Directory to download and load the weights, default to the
            default cache directory of huggingface.
        load_format: 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.
            "tensorizer" will use CoreWeave's tensorizer library for
                fast weight loading.
876
            "bitsandbytes" will load nf4 type weights.
877
        ignore_patterns: The list of patterns to ignore when loading the model.
878
            Default to "original/**/*" to avoid repeated loading of llama's
879
            checkpoints.
880
881
882
883
884
885
    """

    load_format: Union[str, LoadFormat, "BaseModelLoader"] = LoadFormat.AUTO
    download_dir: Optional[str] = None
    model_loader_extra_config: Optional[Union[str, dict]] = field(
        default_factory=dict)
886
    ignore_patterns: Optional[Union[List[str], str]] = None
887
888
889
890
891
892
893
894

    def __post_init__(self):
        model_loader_extra_config = self.model_loader_extra_config or {}
        if isinstance(model_loader_extra_config, str):
            self.model_loader_extra_config = json.loads(
                model_loader_extra_config)
        self._verify_load_format()

895
896
897
898
899
900
901
        if self.ignore_patterns is not None and len(self.ignore_patterns) > 0:
            logger.info(
                "Ignoring the following patterns when downloading weights: %s",
                self.ignore_patterns)
        else:
            self.ignore_patterns = ["original/**/*"]

902
903
904
905
906
907
908
909
    def _verify_load_format(self) -> None:
        if not isinstance(self.load_format, str):
            return

        load_format = self.load_format.lower()
        self.load_format = LoadFormat(load_format)

        rocm_not_supported_load_format: List[str] = []
910
911
        if current_platform.is_rocm(
        ) and load_format in rocm_not_supported_load_format:
912
913
914
915
916
917
918
919
920
921
            rocm_supported_load_format = [
                f for f in LoadFormat.__members__
                if (f not in rocm_not_supported_load_format)
            ]
            raise ValueError(
                f"load format '{load_format}' is not supported in ROCm. "
                f"Supported load formats are "
                f"{rocm_supported_load_format}")


922
class ParallelConfig:
923
924
925
926
927
    """Configuration for the distributed execution.

    Args:
        pipeline_parallel_size: Number of pipeline parallel groups.
        tensor_parallel_size: Number of tensor parallel groups.
928
        worker_use_ray: Deprecated, use distributed_executor_backend instead.
zspo's avatar
zspo committed
929
930
931
        max_parallel_loading_workers: Maximum number of multiple batches
            when load model sequentially. To avoid RAM OOM when using tensor
            parallel and large models.
932
933
        disable_custom_all_reduce: Disable the custom all-reduce kernel and
            fall back to NCCL.
934
935
        tokenizer_pool_config: Config for the tokenizer pool.
            If None, will use synchronous tokenization.
936
937
        ray_workers_use_nsight: Whether to profile Ray workers with nsight, see
            https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler.
938
        placement_group: ray distributed model workers placement group.
939
        distributed_executor_backend: Backend to use for distributed model
940
941
942
943
944
945
            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
            and hpu only support Ray for distributed inference.
946
    """
947

948
949
950
951
    def __init__(
        self,
        pipeline_parallel_size: int,
        tensor_parallel_size: int,
952
        worker_use_ray: Optional[bool] = None,
953
        max_parallel_loading_workers: Optional[int] = None,
954
        disable_custom_all_reduce: bool = False,
955
        tokenizer_pool_config: Optional[TokenizerPoolConfig] = None,
956
        ray_workers_use_nsight: bool = False,
957
        placement_group: Optional["PlacementGroup"] = None,
958
959
        distributed_executor_backend: Optional[Union[
            str, Type["ExecutorBase"]]] = None,
960
961
    ) -> None:
        self.pipeline_parallel_size = pipeline_parallel_size
962
        self.tensor_parallel_size = tensor_parallel_size
963
        self.distributed_executor_backend = distributed_executor_backend
964
        self.max_parallel_loading_workers = max_parallel_loading_workers
965
        self.disable_custom_all_reduce = disable_custom_all_reduce
966
        self.tokenizer_pool_config = tokenizer_pool_config
967
        self.ray_workers_use_nsight = ray_workers_use_nsight
968
        self.placement_group = placement_group
969
        self.world_size = pipeline_parallel_size * self.tensor_parallel_size
970

971
972
973
        if worker_use_ray:
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
974
            elif not self.use_ray:
975
976
977
978
                raise ValueError(f"worker-use-ray can't be used with "
                                 f"distributed executor backend "
                                 f"'{self.distributed_executor_backend}'.")

979
980
981
982
983
984
985
        if current_platform.is_tpu() and self.world_size > 1:
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
                    "TPU backend only supports Ray for distributed inference.")

986
987
988
989
990
991
992
        if current_platform.is_hpu() and self.world_size > 1:
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
                    "HPU backend only supports Ray for distributed inference.")

993
        if self.distributed_executor_backend is None and self.world_size > 1:
994
995
996
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

997
            from vllm.executor import ray_utils
998
            backend = "mp"
999
            ray_found = ray_utils.ray_is_available()
1000
            if (current_platform.is_cuda()
1001
                    and cuda_device_count_stateless() < self.world_size):
1002
1003
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1004
1005
1006
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1007
1008
                backend = "ray"
            elif ray_found:
1009
                if self.placement_group:
1010
                    backend = "ray"
1011
1012
1013
1014
1015
1016
                else:
                    from ray import is_initialized as ray_is_initialized
                    if ray_is_initialized():
                        from ray.util import get_current_placement_group
                        if get_current_placement_group():
                            backend = "ray"
1017
1018
1019
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1020

1021
        self._verify_args()
1022
        self.rank: int = 0
1023

1024
1025
1026
1027
1028
1029
    @property
    def use_ray(self) -> bool:
        return self.distributed_executor_backend == "ray" or (
            isinstance(self.distributed_executor_backend, type)
            and self.distributed_executor_backend.uses_ray)

1030
    def _verify_args(self) -> None:
1031
1032
1033
1034
1035
1036
1037
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase

        if self.distributed_executor_backend not in (
                "ray", "mp", None) and not (isinstance(
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
1038
            raise ValueError(
1039
1040
1041
1042
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
                "values are 'ray', 'mp' or custom ExecutorBase subclass.")
        if self.use_ray:
1043
1044
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1045
        if current_platform.is_rocm():
1046
1047
1048
1049
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
1050
        if self.ray_workers_use_nsight and not self.use_ray:
1051
1052
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1053

1054
1055

class SchedulerConfig:
1056
1057
1058
    """Scheduler configuration.

    Args:
1059
        task: The task to use the model for.
1060
1061
1062
1063
        max_num_batched_tokens: Maximum number of tokens to be processed in
            a single iteration.
        max_num_seqs: Maximum number of sequences to be processed in a single
            iteration.
Chaofan Lin's avatar
Chaofan Lin committed
1064
        max_model_len: Maximum length of a sequence (including prompt
Lily Liu's avatar
Lily Liu committed
1065
            and generated text).
1066
1067
1068
1069
        num_lookahead_slots: The number of slots to allocate per sequence per
            step, beyond the known token ids. This is used in speculative
            decoding to store KV activations of tokens which may or may not be
            accepted.
1070
1071
        delay_factor: Apply a delay (of delay factor multiplied by previous
            prompt latency) before scheduling next prompt.
1072
1073
        enable_chunked_prefill: If True, prefill requests can be chunked based
            on the remaining max_num_batched_tokens.
1074
        preemption_mode: Whether to perform preemption by swapping or
1075
1076
1077
1078
1079
            recomputation. If not specified, we determine the mode as follows:
            We use recomputation by default since it incurs lower overhead than
            swapping. However, when the sequence group has multiple sequences
            (e.g., beam search), recomputation is not currently supported. In
            such a case, we use swapping instead.
1080
1081
1082
1083
        send_delta_data: Private API. If used, scheduler sends delta data to
            workers instead of an entire data. It should be enabled only
            when SPMD worker architecture is enabled. I.e.,
            VLLM_USE_RAY_SPMD_WORKER=1
1084
        policy: The scheduling policy to use. "fcfs" (default) or "priority".
1085
    """
1086

1087
    def __init__(self,
1088
                 task: _Task,
1089
1090
1091
1092
1093
1094
                 max_num_batched_tokens: Optional[int],
                 max_num_seqs: int,
                 max_model_len: int,
                 num_lookahead_slots: int = 0,
                 delay_factor: float = 0.0,
                 enable_chunked_prefill: bool = False,
1095
                 is_multimodal_model: bool = False,
1096
                 preemption_mode: Optional[str] = None,
1097
                 num_scheduler_steps: int = 1,
1098
                 multi_step_stream_outputs: bool = False,
1099
1100
                 send_delta_data: bool = False,
                 policy: str = "fcfs") -> None:
1101
        if max_num_batched_tokens is None:
1102
            if enable_chunked_prefill:
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
                if num_scheduler_steps > 1:
                    # Multi-step Chunked-Prefill doesn't allow prompt-chunking
                    # for now. Have max_num_batched_tokens set to max_model_len
                    # so we don't reject sequences on account of a short
                    # max_num_batched_tokens.
                    max_num_batched_tokens = max(max_model_len, 2048)
                else:
                    # It is the values that have the best balance between ITL
                    # and TTFT on A100. Note it is not optimized for throughput.
                    max_num_batched_tokens = 512
1113
1114
1115
            else:
                # If max_model_len is too short, use 2048 as the default value
                # for higher throughput.
1116
1117
                max_num_batched_tokens = max(max_model_len, 2048)

1118
            if task == "embedding":
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
                # For embedding, choose specific value for higher throughput
                max_num_batched_tokens = max(
                    max_num_batched_tokens,
                    _EMBEDDING_MODEL_MAX_NUM_BATCHED_TOKENS,
                )
            if is_multimodal_model:
                # The value needs to be at least the number of multimodal tokens
                max_num_batched_tokens = max(
                    max_num_batched_tokens,
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

        self.max_num_batched_tokens = max_num_batched_tokens

1133
        if enable_chunked_prefill:
1134
1135
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
1136
                self.max_num_batched_tokens)
1137

1138
        self.task: Final = task
1139
        self.max_num_seqs = max_num_seqs
Lily Liu's avatar
Lily Liu committed
1140
        self.max_model_len = max_model_len
1141
1142
        self.num_lookahead_slots = num_lookahead_slots
        self.delay_factor = delay_factor
1143
        self.chunked_prefill_enabled = enable_chunked_prefill
1144
        self.preemption_mode = preemption_mode
1145
        self.num_scheduler_steps = num_scheduler_steps
1146
        self.multi_step_stream_outputs = multi_step_stream_outputs
1147
        self.send_delta_data = send_delta_data
1148
        self.policy = policy
1149
1150
1151
        self._verify_args()

    def _verify_args(self) -> None:
1152
1153
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
1154
1155
1156
1157
1158
1159
1160
            raise ValueError(
                f"max_num_batched_tokens ({self.max_num_batched_tokens}) is "
                f"smaller than max_model_len ({self.max_model_len}). "
                "This effectively limits the maximum sequence length to "
                "max_num_batched_tokens and makes vLLM reject longer "
                "sequences. Please increase max_num_batched_tokens or "
                "decrease max_model_len.")
1161

1162
1163
1164
1165
1166
        if self.max_num_batched_tokens < self.max_num_seqs:
            raise ValueError(
                f"max_num_batched_tokens ({self.max_num_batched_tokens}) must "
                "be greater than or equal to max_num_seqs "
                f"({self.max_num_seqs}).")
1167

1168
1169
1170
1171
1172
1173
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
        if self.num_scheduler_steps < 1:
            raise ValueError(
                "num_scheduler_steps "
                f"({self.num_scheduler_steps}) must be greater than or "
                "equal to 1.")

    @property
    def is_multi_step(self) -> bool:
        return self.num_scheduler_steps > 1

1184

1185
class DeviceConfig:
1186
    device: Optional[torch.device]
1187

1188
1189
1190
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
1191
1192
            if current_platform.is_cuda_alike():
                self.device_type = "cuda"
1193
            elif current_platform.is_neuron():
1194
                self.device_type = "neuron"
1195
1196
            elif current_platform.is_hpu():
                self.device_type = "hpu"
1197
            elif current_platform.is_openvino():
1198
                self.device_type = "openvino"
1199
            elif current_platform.is_tpu():
1200
                self.device_type = "tpu"
1201
            elif current_platform.is_cpu():
1202
                self.device_type = "cpu"
1203
            elif current_platform.is_xpu():
1204
                self.device_type = "xpu"
1205
            else:
1206
                raise RuntimeError("Failed to infer device type")
1207
1208
1209
1210
1211
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
1212
        if self.device_type in ["neuron", "openvino"]:
1213
            self.device = torch.device("cpu")
1214
1215
        elif self.device_type in ["tpu"]:
            self.device = None
1216
1217
1218
1219
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1220

1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
class SpeculativeConfig:
    """Configuration for speculative decoding.

    The configuration is currently specialized to draft-model speculative
    decoding with top-1 proposals.
    """

    @staticmethod
    def maybe_create_spec_config(
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        target_dtype: str,
        speculative_model: Optional[str],
1234
        speculative_model_quantization: Optional[str],
1235
        speculative_draft_tensor_parallel_size: Optional[int],
1236
        num_speculative_tokens: Optional[int],
1237
        speculative_disable_mqa_scorer: Optional[bool],
1238
1239
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
1240
        disable_log_stats: bool,
1241
        speculative_disable_by_batch_size: Optional[int],
1242
1243
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1244
1245
1246
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
1247
        disable_logprobs: Optional[bool],
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
    ) -> Optional["SpeculativeConfig"]:
        """Create a SpeculativeConfig if possible, else return None.

        This function attempts to create a SpeculativeConfig object based on the
        provided parameters. If the necessary conditions are met, it returns an
        instance of SpeculativeConfig. Otherwise, it returns None.

        Args:
            target_model_config (ModelConfig): The configuration of the target
                model.
            target_parallel_config (ParallelConfig): The parallel configuration
                for the target model.
            target_dtype (str): The data type used for the target model.
            speculative_model (Optional[str]): The name of the speculative
                model, if provided.
1263
1264
1265
            speculative_model_quantization (Optional[str]): Quantization method
                that was used to quantize the speculative model weights. If
                None, we assume the model weights are not quantized.
1266
1267
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
1268
            num_speculative_tokens (Optional[int]): The number of speculative
1269
1270
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
1271
1272
1273
            speculative_disable_mqa_scorer (Optional[bool]): Disable the MQA
                scorer for the speculative model and fall back to batch
                expansion for scoring.
1274
1275
1276
1277
1278
1279
            speculative_max_model_len (Optional[int]): The maximum model len of
                the speculative model. Used when testing the ability to skip
                speculation for some sequences.
            enable_chunked_prefill (bool): Whether vLLM is configured to use
                chunked prefill or not. Used for raising an error since its not
                yet compatible with spec decode.
1280
1281
1282
            speculative_disable_by_batch_size (Optional[int]): Disable
                speculative decoding for new incoming requests when the number
                of enqueue requests  is larger than this value, if provided.
1283
1284
1285
1286
            ngram_prompt_lookup_max (Optional[int]): Max size of ngram token
                window, if provided.
            ngram_prompt_lookup_min (Optional[int]): Min size of ngram token
                window, if provided.
1287
1288
1289
1290
1291
1292
1293
1294
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
1295
                accepted. This threshold is used only when we use the
1296
1297
1298
1299
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1300
1301
1302
1303
1304
            disable_logprobs (Optional[bool]): If set to True, token log
                probabilities are not returned during speculative decoding.
                If set to False, token log probabilities are returned
                according to the log probability settings in SamplingParams.
                If not specified, it defaults to True.
1305

1306
1307
1308
1309
1310
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """

1311
1312
1313
1314
        if speculative_model is None:
            if num_speculative_tokens is not None:
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative_model.")
1315
1316
            return None

1317
1318
1319
1320
1321
1322
        if (speculative_disable_by_batch_size is not None
                and speculative_disable_by_batch_size < 2):
            raise ValueError("Expect the batch size threshold of disabling "
                             "speculative decoding is > 1, but got "
                             f"{speculative_disable_by_batch_size=}")

1323
1324
        # TODO: The user should be able to specify revision/max model len
        # for the draft model. It is not currently supported.
1325
1326
        draft_revision = None
        draft_code_revision = None
1327
        draft_quantization = speculative_model_quantization
1328

1329
1330
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1331
1332
1333
1334
1335
1336
1337
1338
                ngram_prompt_lookup_min = 1
            if ngram_prompt_lookup_max is None or ngram_prompt_lookup_max < 1:
                raise ValueError(f"{ngram_prompt_lookup_max=} must be > 0")
            if ngram_prompt_lookup_min < 1:
                raise ValueError(f"{ngram_prompt_lookup_min=} must be > 0")
            if ngram_prompt_lookup_min > ngram_prompt_lookup_max:
                raise ValueError(f"{ngram_prompt_lookup_min=} cannot be "
                                 f"larger than {ngram_prompt_lookup_max=}")
1339

1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
            # TODO: current we still need extract vocab_size from target model
            # config, in future, we may try refactor it out, and set
            # draft related config as None here.
            draft_model_config = target_model_config
            draft_parallel_config = target_parallel_config
        else:
            ngram_prompt_lookup_max = 0
            ngram_prompt_lookup_min = 0
            draft_model_config = ModelConfig(
                model=speculative_model,
1350
                task="draft",
1351
1352
1353
                tokenizer=target_model_config.tokenizer,
                tokenizer_mode=target_model_config.tokenizer_mode,
                trust_remote_code=target_model_config.trust_remote_code,
1354
1355
                allowed_local_media_path=target_model_config.
                allowed_local_media_path,
1356
1357
1358
1359
1360
1361
                dtype=target_model_config.dtype,
                seed=target_model_config.seed,
                revision=draft_revision,
                code_revision=draft_code_revision,
                tokenizer_revision=target_model_config.tokenizer_revision,
                max_model_len=None,
1362
                spec_target_max_model_len=target_model_config.max_model_len,
1363
1364
                quantization=draft_quantization,
                enforce_eager=target_model_config.enforce_eager,
1365
1366
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1367
1368
1369
                max_logprobs=target_model_config.max_logprobs,
            )

1370
            draft_hf_config = draft_model_config.hf_config
1371

1372
1373
1374
1375
1376
            if (num_speculative_tokens is not None
                    and hasattr(draft_hf_config, "num_lookahead_tokens")):
                draft_hf_config.num_lookahead_tokens = num_speculative_tokens

            n_predict = getattr(draft_hf_config, "n_predict", None)
1377
1378
1379
1380
1381
1382
1383
1384
            if n_predict is not None:
                if num_speculative_tokens is None:
                    # Default to max value defined in draft model config.
                    num_speculative_tokens = n_predict
                elif num_speculative_tokens > n_predict:
                    # Verify provided value doesn't exceed the maximum
                    # supported by the draft model.
                    raise ValueError(
1385
1386
1387
                        "This speculative model supports a maximum of "
                        f"num_speculative_tokens={n_predict}, but "
                        f"{num_speculative_tokens=} was provided.")
1388

1389
1390
1391
1392
1393
1394
            if enable_chunked_prefill and draft_hf_config.model_type in (
                    "medusa", "mlp_speculator", "eagle"):
                raise ValueError(
                    "Chunked prefill and hidden-state based draft models are "
                    "not compatible.")

1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
            speculative_draft_tensor_parallel_size = \
                SpeculativeConfig._verify_and_get_draft_model_tensor_parallel_size(
                    target_parallel_config,
                    speculative_draft_tensor_parallel_size,
                    draft_hf_config
            )

            if (enable_chunked_prefill and \
                 speculative_draft_tensor_parallel_size != 1):
                # TODO - Investigate why the error reported in
                # https://github.com/vllm-project/vllm/pull/9291#issuecomment-2463266258
                # is happening and re-enable it.
                raise ValueError(
                    "Chunked prefill and speculative decoding can be enabled "
                    "simultaneously only for draft models with tensor "
                    "parallel size 1.")

1412
1413
1414
1415
1416
1417
1418
1419
1420
            draft_model_config.max_model_len = (
                SpeculativeConfig._maybe_override_draft_max_model_len(
                    speculative_max_model_len,
                    draft_model_config.max_model_len,
                    target_model_config.max_model_len,
                ))

            draft_parallel_config = (
                SpeculativeConfig.create_draft_parallel_config(
1421
                    target_parallel_config,
1422
                    speculative_draft_tensor_parallel_size, draft_hf_config))
1423

1424
1425
1426
1427
1428
1429
        if num_speculative_tokens is None:
            raise ValueError(
                "num_speculative_tokens must be provided with "
                "speculative_model unless the draft model config contains an "
                "n_predict parameter.")

1430
1431
1432
1433
        if typical_acceptance_sampler_posterior_threshold is None:
            typical_acceptance_sampler_posterior_threshold = 0.09
        if typical_acceptance_sampler_posterior_alpha is None:
            typical_acceptance_sampler_posterior_alpha = 0.3
1434
1435
        if disable_logprobs is None:
            disable_logprobs = True
1436

1437
1438
1439
1440
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
1441
            speculative_disable_mqa_scorer,
1442
            speculative_disable_by_batch_size,
1443
1444
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
1445
1446
1447
1448
1449
            draft_token_acceptance_method=draft_token_acceptance_method,
            typical_acceptance_sampler_posterior_threshold=\
                typical_acceptance_sampler_posterior_threshold,
            typical_acceptance_sampler_posterior_alpha=\
                typical_acceptance_sampler_posterior_alpha,
1450
1451
            disable_logprobs=disable_logprobs,
            disable_log_stats=disable_log_stats,
1452
1453
        )

1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
    @staticmethod
    def _maybe_override_draft_max_model_len(
        speculative_max_model_len: Optional[int],
        draft_max_model_len: int,
        target_max_model_len: int,
    ) -> int:
        """Determine the max sequence len for the draft model. This is usually
        the draft_max_model_len, but may be the target_max_model_len if it is
        less than the draft_max_model_len, or may be speculative_max_model_len
        if it is specified.

        This is necessary so that sequences do not exceed the capacity of the
        draft model or the target model.

        speculative_max_model_len is mainly used for testing that sequences can
        skip speculation.
        """

        if speculative_max_model_len is not None:

            if speculative_max_model_len > draft_max_model_len:
                raise ValueError(f"{speculative_max_model_len=} cannot be "
                                 f"larger than {draft_max_model_len=}")

            if speculative_max_model_len > target_max_model_len:
                raise ValueError(f"{speculative_max_model_len=} cannot be "
                                 f"larger than {target_max_model_len=}")

            return speculative_max_model_len

        return min(
            draft_max_model_len,
            target_max_model_len,
        )

1489
    @staticmethod
1490
1491
1492
1493
1494
1495
1496
    def _verify_and_get_draft_model_tensor_parallel_size(
            target_parallel_config: ParallelConfig,
            speculative_draft_tensor_parallel_size: Optional[int],
            draft_hf_config: PretrainedConfig) -> int:
        """
        Verifies and adjusts the tensor parallel size for a draft model
        specified using speculative_draft_tensor_parallel_size.
1497
        """
1498
1499
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
1500
        if speculative_draft_tensor_parallel_size is None:
1501
1502
1503
1504
1505
1506
1507
1508
1509
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
                        "MLPSpeculator cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1")
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
1510
1511
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
1512
            raise ValueError(
1513
                f"{speculative_draft_tensor_parallel_size=} cannot be "
1514
                f"other value than 1 or target model tensor_parallel_size")
1515
        return speculative_draft_tensor_parallel_size
1516

1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
    @staticmethod
    def create_draft_parallel_config(
        target_parallel_config: ParallelConfig,
        speculative_draft_tensor_parallel_size: int,
        draft_hf_config: PretrainedConfig,
    ) -> ParallelConfig:
        """Create a parallel config for use by the draft worker.

        This is mostly a copy of the target parallel config, except the tp_size.
        """
1527
1528
1529
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
1530
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
1531
1532
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
            max_parallel_loading_workers=target_parallel_config.
            max_parallel_loading_workers,
            disable_custom_all_reduce=target_parallel_config.
            disable_custom_all_reduce,
            tokenizer_pool_config=target_parallel_config.tokenizer_pool_config,
            ray_workers_use_nsight=target_parallel_config.
            ray_workers_use_nsight,
            placement_group=target_parallel_config.placement_group,
        )

        return draft_parallel_config

    def __init__(
        self,
        draft_model_config: ModelConfig,
        draft_parallel_config: ParallelConfig,
        num_speculative_tokens: int,
1550
        speculative_disable_mqa_scorer: Optional[bool],
1551
1552
1553
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1554
1555
1556
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
1557
        disable_logprobs: bool,
1558
        disable_log_stats: bool,
1559
1560
1561
1562
1563
1564
1565
1566
    ):
        """Create a SpeculativeConfig object.

        Args:
            draft_model_config: ModelConfig for the draft model.
            draft_parallel_config: ParallelConfig for the draft model.
            num_speculative_tokens: The number of tokens to sample from the
                draft model before scoring with the target model.
1567
1568
1569
1570
1571
            speculative_disable_by_batch_size: Disable speculative
                decoding for new incoming requests when the number of
                enqueue requests is larger than this value.
            ngram_prompt_lookup_max: Max size of ngram token window.
            ngram_prompt_lookup_min: Min size of ngram token window.
1572
1573
1574
1575
1576
1577
1578
1579
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
1580
                accepted. This threshold is used only when we use the
1581
1582
1583
1584
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1585
            disable_logprobs: If set to True, token log probabilities will not
1586
                be returned even if requested by sampling parameters. This
1587
1588
1589
1590
                reduces latency by skipping logprob calculation in proposal
                sampling, target sampling, and after accepted tokens are
                determined. If set to False, log probabilities will be
                returned.
1591
1592
            disable_log_stats: Whether to disable periodic printing of stage
                times in speculative decoding.
1593
1594
1595
1596
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
1597
        self.speculative_disable_mqa_scorer = speculative_disable_mqa_scorer
1598
1599
1600
1601
        self.speculative_disable_by_batch_size = \
            speculative_disable_by_batch_size
        self.ngram_prompt_lookup_max = ngram_prompt_lookup_max or 0
        self.ngram_prompt_lookup_min = ngram_prompt_lookup_min or 0
1602
1603
1604
1605
1606
        self.draft_token_acceptance_method = draft_token_acceptance_method
        self.typical_acceptance_sampler_posterior_threshold = \
            typical_acceptance_sampler_posterior_threshold
        self.typical_acceptance_sampler_posterior_alpha = \
            typical_acceptance_sampler_posterior_alpha
1607
        self.disable_logprobs = disable_logprobs
1608
        self.disable_log_stats = disable_log_stats
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619

        self._verify_args()

    def _verify_args(self) -> None:
        if self.num_speculative_tokens <= 0:
            raise ValueError("Expected num_speculative_tokens to be greater "
                             f"than zero ({self.num_speculative_tokens}).")

        if self.draft_model_config:
            self.draft_model_config.verify_with_parallel_config(
                self.draft_parallel_config)
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
            # Validate and set draft token acceptance related settings.

        if (self.draft_token_acceptance_method is None):
            raise ValueError("draft_token_acceptance_method is not set. "
                             "Expected values are rejection_sampler or "
                             "typical_acceptance_sampler.")

        if (self.draft_token_acceptance_method != 'rejection_sampler'
                and self.draft_token_acceptance_method !=
                'typical_acceptance_sampler'):
            raise ValueError(
                "Expected draft_token_acceptance_method to be either "
                "rejection_sampler or typical_acceptance_sampler. Instead it "
                f"is {self.draft_token_acceptance_method}")

        if (self.typical_acceptance_sampler_posterior_threshold < 0
                or self.typical_acceptance_sampler_posterior_alpha < 0):
            raise ValueError(
                "Expected typical_acceptance_sampler_posterior_threshold "
                "and typical_acceptance_sampler_posterior_alpha to be > 0. "
                "Instead found "
                f"typical_acceptance_sampler_posterior_threshold = "
                f"{self.typical_acceptance_sampler_posterior_threshold} and "
                f"typical_acceptance_sampler_posterior_alpha = "
                f"{self.typical_acceptance_sampler_posterior_alpha}")
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656

    @property
    def num_lookahead_slots(self) -> int:
        """The number of additional slots the scheduler should allocate per
        step, in addition to the slots allocated for each known token.

        This is equal to the number of speculative tokens, as each speculative
        token must be scored.
        """
        return self.num_speculative_tokens

    def __repr__(self) -> str:
1657
1658
1659
1660
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
1661
1662
1663
1664
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


1665
1666
1667
1668
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
1669
    fully_sharded_loras: bool = False
1670
    max_cpu_loras: Optional[int] = None
1671
    lora_dtype: Optional[Union[torch.dtype, str]] = None
1672
1673
1674
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
1675
    long_lora_scaling_factors: Optional[Tuple[float]] = None
1676
    bias_enabled: bool = False
1677
1678

    def __post_init__(self):
1679
1680
1681
        # Setting the maximum rank to 256 should be able to satisfy the vast
        # majority of applications.
        possible_max_ranks = (8, 16, 32, 64, 128, 256)
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
        possible_lora_extra_vocab_size = (0, 256, 512)
        if self.max_lora_rank not in possible_max_ranks:
            raise ValueError(
                f"max_lora_rank ({self.max_lora_rank}) must be one of "
                f"{possible_max_ranks}.")
        if self.lora_extra_vocab_size not in possible_lora_extra_vocab_size:
            raise ValueError(
                f"lora_extra_vocab_size ({self.lora_extra_vocab_size}) "
                f"must be one of {possible_lora_extra_vocab_size}.")
        if self.max_loras < 1:
            raise ValueError(f"max_loras ({self.max_loras}) must be >= 1.")
        if self.max_cpu_loras is None:
            self.max_cpu_loras = self.max_loras
        elif self.max_cpu_loras < self.max_loras:
            raise ValueError(
                f"max_cpu_loras ({self.max_cpu_loras}) must be >= "
zspo's avatar
zspo committed
1698
                f"max_loras ({self.max_loras})")
1699
1700
1701
1702
1703
1704

    def verify_with_model_config(self, model_config: ModelConfig):
        if self.lora_dtype in (None, "auto"):
            self.lora_dtype = model_config.dtype
        elif isinstance(self.lora_dtype, str):
            self.lora_dtype = getattr(torch, self.lora_dtype)
1705
1706
1707
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
1708
            # TODO support marlin
1709
1710
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
1711
1712

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
1713
1714
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
1715
1716
        if scheduler_config.chunked_prefill_enabled:
            raise ValueError("LoRA is not supported with chunked prefill yet.")
1717
1718


1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
@dataclass
class PromptAdapterConfig:
    max_prompt_adapters: int
    max_prompt_adapter_token: int
    max_cpu_prompt_adapters: Optional[int] = None
    prompt_adapter_dtype: Optional[torch.dtype] = None

    def __post_init__(self):

        if self.max_prompt_adapters < 1:
            raise ValueError(f"max_prompt_adapters "
                             f"({self.max_prompt_adapters}) must be >= 1.")
        if self.max_prompt_adapter_token == 0:
            raise ValueError("max_prompt_adapter_token must be set.")
        if self.max_cpu_prompt_adapters is None:
            self.max_cpu_prompt_adapters = self.max_prompt_adapters

    def verify_with_model_config(self, model_config: ModelConfig):
        if self.prompt_adapter_dtype in (None, "auto"):
            self.prompt_adapter_dtype = model_config.dtype
        elif isinstance(self.prompt_adapter_dtype, str):
            self.prompt_adapter_dtype = getattr(torch,
                                                self.prompt_adapter_dtype)


1744
@dataclass
1745
class MultiModalConfig:
1746
1747
    """Controls the behavior of multimodal models."""

1748
    limit_per_prompt: Mapping[str, int] = field(default_factory=dict)
1749
1750
1751
1752
1753
    """
    The maximum number of multi-modal input instances allowed per prompt
    for each :class:`~vllm.multimodal.MultiModalPlugin`.
    """

1754
    # TODO: Add configs to init vision tower or not.
1755

1756

1757
1758
@dataclass
class PoolerConfig:
1759
    """Controls the behavior of output pooling in embedding models."""
1760
1761

    pooling_type: Optional[str] = None
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
    """
    The pooling method of the embedding model. This should be a key in
    :class:`vllm.model_executor.layers.pooler.PoolingType`.
    """

    normalize: Optional[bool] = None
    """
    Whether to normalize the pooled outputs. Usually, this should be set to
    ``True`` for embedding outputs.
    """

    softmax: Optional[bool] = None
    """
    Whether to apply softmax to the pooled outputs. Usually, this should be set
    to ``True`` for classification outputs.
    """

    step_tag_id: Optional[int] = None
    """
    If set, only the score corresponding to the ``step_tag_id`` in the 
    generated sentence should be returned. Otherwise, the scores for all tokens
    are returned.
    """

    returned_token_ids: Optional[List[int]] = None
    """
    A list of indices for the vocabulary dimensions to be extracted, 
    such as the token IDs of ``good_token`` and ``bad_token`` in the 
    ``math-shepherd-mistral-7b-prm`` model.
    """

    @staticmethod
    def from_json(json_str: str) -> "PoolerConfig":
        return PoolerConfig(**json.loads(json_str))
1796
1797


1798
1799
1800
1801
1802
1803
1804
1805
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

1806
_ROCM_NOT_SUPPORTED_DTYPE: List[str] = []  #
1807

1808
1809
1810

def _get_and_verify_dtype(
    config: PretrainedConfig,
1811
    dtype: Union[str, torch.dtype],
1812
1813
1814
1815
1816
1817
1818
) -> torch.dtype:
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
    config_dtype = getattr(config, "torch_dtype", None)
    if config_dtype is None:
        config_dtype = torch.float32

1819
1820
1821
1822
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
Woosuk Kwon's avatar
Woosuk Kwon committed
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
                if config.model_type == "gemma2":
                    logger.info(
                        "For Gemma 2, we downcast float32 to bfloat16 instead "
                        "of float16 by default. Please specify `dtype` if you "
                        "want to use float16.")
                    torch_dtype = torch.bfloat16
                else:
                    # Following the common practice, we use float16 for float32
                    # models.
                    torch_dtype = torch.float16
1833
1834
            else:
                torch_dtype = config_dtype
1835
1836
1837
1838
1839
1840
1841

            if current_platform.is_hpu() and config_dtype == torch.float16:
                logger.info(
                    "For HPU, we cast models to bfloat16 instead of"
                    "using float16 by default. Please specify `dtype` if you "
                    "want to use float16.")
                torch_dtype = torch.bfloat16
1842
        else:
1843
1844
1845
1846
1847
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
                raise ValueError(f"Unknown dtype: {dtype}")
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
1848
    else:
1849
        raise ValueError(f"Unknown dtype: {dtype}")
1850
1851
1852
1853
1854

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
1855
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
1856
1857
1858
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
1859
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
1860
1861
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
1862
            # Casting between float16 and bfloat16 is allowed with a warning.
1863
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
1864
1865

    return torch_dtype
1866
1867
1868
1869
1870


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
1871
    disable_sliding_window: bool,
1872
    sliding_window_len: Optional[Union[int, List[Optional[int]]]],
1873
    spec_target_max_model_len: Optional[int] = None,
1874
    encoder_config: Optional[Any] = None,
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
) -> int:
    """Get and verify the model's maximum length."""
    derived_max_model_len = float("inf")
    possible_keys = [
        # OPT
        "max_position_embeddings",
        # GPT-2
        "n_positions",
        # MPT
        "max_seq_len",
1885
1886
        # ChatGLM2
        "seq_length",
1887
1888
        # Command-R
        "model_max_length",
1889
1890
1891
1892
1893
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
1894
    # Choose the smallest "max_length" from the possible keys.
1895
    max_len_key = None
1896
    for key in possible_keys:
1897
1898
1899
1900
1901
        max_len = getattr(hf_config, key, None)
        if max_len is not None:
            max_len_key = key if max_len < derived_max_model_len \
                else max_len_key
            derived_max_model_len = min(derived_max_model_len, max_len)
1902
1903
1904
1905

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
    if disable_sliding_window and sliding_window_len is not None:
1906
1907

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
1908
        max_len_key = "sliding_window" \
1909
1910
1911
            if sliding_window_len_min < derived_max_model_len else max_len_key
        derived_max_model_len = min(derived_max_model_len,
                                    sliding_window_len_min)
1912
1913
1914

    # If none of the keys were found in the config, use a default and
    # log a warning.
1915
    if derived_max_model_len == float("inf"):
1916
1917
1918
1919
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

1920
1921
1922
1923
1924
        if spec_target_max_model_len is not None:
            # If this is a speculative draft model, we use the max model len
            # from the target model.
            return spec_target_max_model_len

1925
1926
1927
1928
        default_max_len = 2048
        logger.warning(
            "The model's config.json does not contain any of the following "
            "keys to determine the original maximum length of the model: "
1929
            "%s. Assuming the model's maximum length is %d.", possible_keys,
1930
            default_max_len)
1931
        derived_max_model_len = default_max_len
1932

1933
    rope_scaling = getattr(hf_config, "rope_scaling", None)
1934
    if rope_scaling is not None:
1935
1936
1937
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947

        if rope_type not in ("su", "longrope", "llama3"):
            if disable_sliding_window:
                # TODO(robertgshaw): Find a model that supports rope_scaling
                # with sliding window to see if this case should be allowed.
                raise NotImplementedError(
                    "Disabling sliding window is not supported for models "
                    "with rope_scaling. Please raise an issue so we can "
                    "investigate.")

1948
1949
1950
1951
            # NOTE: rope_type == "default" does not define factor
            # https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/modeling_rope_utils.py
            scaling_factor = rope_scaling.get("factor", 1.0)

1952
1953
1954
1955
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
1956

1957
1958
1959
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

1960
1961
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
1962
    if max_model_len is None:
1963
        max_model_len = int(derived_max_model_len)
1964
    elif max_model_len > derived_max_model_len:
1965
1966
1967
1968
1969
        # Some models might have a separate key for specifying model_max_length
        # that will be bigger than derived_max_model_len. We compare user input
        # with model_max_length and allow this override when it's smaller.
        model_max_length = getattr(hf_config, "model_max_length", None)
        if model_max_length is not None and max_model_len <= model_max_length:
1970
1971
1972
1973
1974
1975
1976
            if disable_sliding_window:
                # TODO(robertgshaw): Find a model that has model_max_length
                # with sliding window to see if this case should be allowed.
                raise NotImplementedError(
                    "Disabling sliding window is not supported for models "
                    "model_max_length in the config. Please raise an issue "
                    "so we can investigate.")
1977
        else:
1978
            msg = (
1979
                f"User-specified max_model_len ({max_model_len}) is greater "
1980
1981
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
1982
                f"{model_max_length} in model's config.json). This may lead "
1983
1984
1985
1986
1987
1988
1989
1990
1991
                "to incorrect model outputs or CUDA errors.")
            if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
                logger.warning(
                    "%s Make sure the value is correct and within the "
                    "model context size.", msg)
            else:
                raise ValueError(
                    f"{msg} To allow overriding this maximum, set "
                    "the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1")
1992
    return int(max_model_len)
1993
1994


1995
1996
1997
1998
1999
2000
2001
2002
def get_min_sliding_window(
        sliding_window: Union[int, List[Optional[int]]]) -> int:
    if isinstance(sliding_window, list):
        return min(s for s in sliding_window if s is not None)

    return sliding_window


2003
2004
2005
def get_served_model_name(model: str,
                          served_model_name: Optional[Union[str, List[str]]]):
    """
2006
2007
2008
2009
    If the input is a non-empty list, the first model_name in
    `served_model_name` is taken.
    If the input is a non-empty string, it is used directly.
    For cases where the input is either an empty string or an
2010
2011
2012
2013
2014
2015
2016
2017
2018
    empty list, the fallback is to use `self.model`.
    """
    if not served_model_name:
        return model
    if isinstance(served_model_name, list):
        return served_model_name[0]
    return served_model_name


2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
@dataclass
class DecodingConfig:
    """Dataclass which contains the decoding strategy of the engine"""

    # Which guided decoding algo to use. 'outlines' / 'lm-format-enforcer'
    guided_decoding_backend: str = 'outlines'

    def __post_init__(self):
        valid_guided_backends = ['outlines', 'lm-format-enforcer']
        backend = self.guided_decoding_backend
        if backend not in valid_guided_backends:
            raise ValueError(f"Invalid guided_decoding_backend '{backend},"
                             f"must be one of {valid_guided_backends}")


2034
2035
2036
2037
2038
@dataclass
class ObservabilityConfig:
    """Configuration for observability."""
    otlp_traces_endpoint: Optional[str] = None

2039
2040
2041
2042
2043
2044
2045
2046
    # Collecting detailed timing information for each request can be expensive.

    # If set, collects the model forward time for the request.
    collect_model_forward_time: bool = False

    # If set, collects the model execute time for the request.
    collect_model_execute_time: bool = False

2047
    def __post_init__(self):
2048
2049
2050
2051
2052
        if not is_otel_available() and self.otlp_traces_endpoint is not None:
            raise ValueError(
                "OpenTelemetry is not available. Unable to configure "
                "'otlp_traces_endpoint'. Ensure OpenTelemetry packages are "
                f"installed. Original error:\n{otel_import_error_traceback}")
2053
2054


2055
2056
2057
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
2058
2059
2060
    simplifies passing around the distinct configurations in the codebase.
    """

2061
2062
2063
2064
2065
2066
2067
2068
2069
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
    parallel_config: ParallelConfig = field(default=None,
                                            init=True)  # type: ignore
    scheduler_config: SchedulerConfig = field(default=None,
                                              init=True)  # type: ignore
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
2070
2071
2072
2073
2074
    lora_config: Optional[LoRAConfig] = None
    speculative_config: Optional[SpeculativeConfig] = None
    decoding_config: Optional[DecodingConfig] = None
    observability_config: Optional[ObservabilityConfig] = None
    prompt_adapter_config: Optional[PromptAdapterConfig] = None
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
    quant_config: Optional[QuantizationConfig] = None

    @staticmethod
    def _get_quantization_config(
            model_config: ModelConfig,
            load_config: LoadConfig) -> Optional[QuantizationConfig]:
        """Get the quantization config."""
        if model_config.quantization is not None:
            from vllm.model_executor.model_loader.weight_utils import (
                get_quant_config)
            quant_config = get_quant_config(model_config, load_config)
            capability_tuple = current_platform.get_device_capability()

            if capability_tuple is not None:
                capability = capability_tuple.to_int()
                if capability < quant_config.get_min_capability():
                    raise ValueError(
                        f"The quantization method {model_config.quantization} "
                        "is not supported for the current GPU. Minimum "
                        f"capability: {quant_config.get_min_capability()}. "
                        f"Current capability: {capability}.")
            supported_dtypes = quant_config.get_supported_act_dtypes()
            if model_config.dtype not in supported_dtypes:
                raise ValueError(
                    f"{model_config.dtype} is not supported for quantization "
                    f"method {model_config.quantization}. Supported dtypes: "
                    f"{supported_dtypes}")
            return quant_config
        return None
2104

2105
2106
2107
2108
2109
2110
    def with_hf_config(self, hf_config: PretrainedConfig) -> "VllmConfig":
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config

        return replace(self, model_config=model_config)

2111
2112
2113
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
2114
2115
2116
2117
2118
2119
2120
2121
        if self.model_config is not None:
            self.model_config.verify_async_output_proc(self.parallel_config,
                                                       self.speculative_config,
                                                       self.device_config)
            self.model_config.verify_with_parallel_config(self.parallel_config)

        if self.cache_config is not None:
            self.cache_config.verify_with_parallel_config(self.parallel_config)
2122
2123
2124
2125
2126

        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
2127
2128
2129
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
2130
2131
2132
2133
2134

        if self.quant_config is None and \
            self.model_config is not None and self.load_config is not None:
            self.quant_config = VllmConfig._get_quantization_config(
                self.model_config, self.load_config)
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174

    def __str__(self):
        return ("model=%r, speculative_config=%r, tokenizer=%r, "
        "skip_tokenizer_init=%s, tokenizer_mode=%s, revision=%s, "
        "override_neuron_config=%s, tokenizer_revision=%s, "
        "trust_remote_code=%s, dtype=%s, max_seq_len=%d, "
        "download_dir=%r, load_format=%s, tensor_parallel_size=%d, "
        "pipeline_parallel_size=%d, "
        "disable_custom_all_reduce=%s, quantization=%s, "
        "enforce_eager=%s, kv_cache_dtype=%s, "
        "quantization_param_path=%s, device_config=%s, "
        "decoding_config=%r, observability_config=%r, "
        "seed=%d, served_model_name=%s, "
        "num_scheduler_steps=%d, enable_prefix_caching=%s, "
        "use_async_output_proc=%s, mm_processor_kwargs=%s") % \
        (self.model_config.model, self.speculative_config,
        self.model_config.tokenizer,
        self.model_config.skip_tokenizer_init,
        self.model_config.tokenizer_mode,
        self.model_config.revision,
        self.model_config.override_neuron_config,
        self.model_config.tokenizer_revision,
        self.model_config.trust_remote_code,
        self.model_config.dtype,
        self.model_config.max_model_len,
        self.load_config.download_dir,
        self.load_config.load_format,
        self.parallel_config.tensor_parallel_size,
        self.parallel_config.pipeline_parallel_size,
        self.parallel_config.disable_custom_all_reduce,
        self.model_config.quantization,
        self.model_config.enforce_eager,
        self.cache_config.cache_dtype,
        self.model_config.quantization_param_path,
        self.device_config.device, self.decoding_config,
        self.observability_config, self.model_config.seed,
        self.model_config.served_model_name,
        self.scheduler_config.num_scheduler_steps,
        self.cache_config.enable_prefix_caching,
        self.model_config.use_async_output_proc,
2175
        self.model_config.mm_processor_kwargs)