config.py 95.5 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, 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
                        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

class ModelConfig:
49
50
51
52
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
53
            It is also used as the content for `model_name` tag in metrics
54
55
56
57
58
            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.
59
        tokenizer: Name or path of the huggingface tokenizer to use.
60
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
61
62
            available, "slow" will always use the slow tokenizer, and
            "mistral" will always use the tokenizer from `mistral_common`.
63
64
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
65
66
67
68
        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.
69
70
71
72
        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
73
74
75
        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.
76
        code_revision: The specific revision to use for the model code on
77
            Hugging Face Hub. It can be a branch name, a tag name, or a
78
            commit id. If unspecified, will use the default version.
79
80
81
        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.
82
83
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
84
85
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
86
87
        quantization_param_path: Path to JSON file containing scaling factors.
            Used to load KV cache scaling factors into the model when KV cache
88
89
            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
90
            model dtype is FP8_E4M3 on ROCm.
91
92
93
        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.
94
            If None, the user did not specify, so default to False.
95
96
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
97
98
99
            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.
100
101
102
103
        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.
104
105
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
106
        served_model_name: The model name used in metrics tag `model_name`,
107
108
            matches the model name exposed via the APIs. If multiple model
            names provided, the first name will be used. If not specified,
109
            the model name will be the same as `model`.
110
        limit_mm_per_prompt: Maximum number of data instances per modality
111
            per prompt. Only applicable for multimodal models.
112
113
114
115
        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.
116
117
        config_format: The config format which shall be loaded.
            Defaults to 'auto' which defaults to 'hf'.
118
        hf_overrides: Arguments to be forwarded to the HuggingFace config.
119
120
        mm_processor_kwargs: Arguments to be forwarded to the model's processor
            for multi-modal data, e.g., image processor.
121
122
123
124
125
126
127
128
129
130
131
132
133
134
        pooling_type: Used to configure the pooling method in the embedding 
            model.
        pooling_norm: Used to determine whether to normalize the pooled 
            data in the embedding model.
        pooling_softmax: Used to determine whether to softmax the pooled 
            data in the embedding model.
        pooling_step_tag_id: When pooling_step_tag_id is not -1, it indicates 
            that the score corresponding to the pooling_step_tag_id in the 
            generated sentence should be returned. Otherwise, it returns 
            the scores for all tokens.
        pooling_returned_token_ids: pooling_returned_token_ids represents 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.
135
    """
136

137
138
139
140
141
142
143
144
145
    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,
146
            allowed_local_media_path: str = "",
147
148
            revision: Optional[str] = None,
            code_revision: Optional[str] = None,
149
            rope_scaling: Optional[Dict[str, Any]] = None,
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
            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,
            override_neuron_config: Optional[Dict[str, Any]] = None,
            config_format: ConfigFormat = ConfigFormat.AUTO,
            chat_template_text_format: str = "string",
167
            hf_overrides: Optional[Dict[str, Any]] = None,
168
169
170
171
172
173
            mm_processor_kwargs: Optional[Dict[str, Any]] = None,
            pooling_type: Optional[str] = None,
            pooling_norm: Optional[bool] = None,
            pooling_softmax: Optional[bool] = None,
            pooling_step_tag_id: Optional[int] = None,
            pooling_returned_token_ids: Optional[List[int]] = None) -> None:
174
        self.model = model
175
        self.tokenizer = tokenizer
176
        self.tokenizer_mode = tokenizer_mode
177
        self.trust_remote_code = trust_remote_code
178
        self.allowed_local_media_path = allowed_local_media_path
179
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
180
        self.revision = revision
181
        self.code_revision = code_revision
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197

        if hf_overrides is None:
            hf_overrides = {}
        if rope_scaling is not None:
            hf_override: Dict[str, Any] = {"rope_scaling": rope_scaling}
            hf_overrides.update(hf_override)
            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}
            hf_overrides.update(hf_override)
            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)

198
199
200
201
202
        # 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
203
        self.quantization = quantization
204
        self.quantization_param_path = quantization_param_path
205
        self.enforce_eager = enforce_eager
206
        self.max_seq_len_to_capture = max_seq_len_to_capture
207
        self.max_logprobs = max_logprobs
208
        self.disable_sliding_window = disable_sliding_window
209
        self.skip_tokenizer_init = skip_tokenizer_init
210
        self.hf_config = get_config(self.model, trust_remote_code, revision,
211
212
                                    code_revision, config_format,
                                    **hf_overrides)
213
        self.hf_text_config = get_hf_text_config(self.hf_config)
214
        self.encoder_config = self._get_encoder_config()
215
216
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, revision)
217
        self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
218
        self.use_async_output_proc = use_async_output_proc
219
        self.chat_template_text_format = chat_template_text_format
220
        self.mm_processor_kwargs = mm_processor_kwargs
Woosuk Kwon's avatar
Woosuk Kwon committed
221

222
223
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
224
225
            self.enforce_eager = False

226
227
228
229
230
231
232
233
234
        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
235
            print_warning_once(
236
                f"{self.hf_text_config.model_type} has interleaved attention, "
Woosuk Kwon's avatar
Woosuk Kwon committed
237
238
                "which is currently not supported by vLLM. Disabling sliding "
                "window and capping the max length to the sliding window size "
239
                f"({sliding_window_len_min}).")
Woosuk Kwon's avatar
Woosuk Kwon committed
240
241
            self.disable_sliding_window = True

242
243
244
245
        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,
246
            sliding_window_len=self.get_hf_config_sliding_window(),
247
248
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
249
250
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
251
252
        self.multimodal_config = self._init_multimodal_config(
            limit_mm_per_prompt)
253
254
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
255

256
257
258
        self.is_attention_free = self._init_attention_free()
        self.has_inner_state = self._init_has_inner_state()

259
260
261
262
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
263
264
265
266

        supported_tasks, task = self._resolve_task(task, self.hf_config)
        self.supported_tasks = supported_tasks
        self.task: Final = task
267
268
269
270
271
272
273
        self.pooler_config = self._init_pooler_config(
            pooling_type,
            pooling_norm,
            pooling_softmax,
            pooling_step_tag_id,
            pooling_returned_token_ids,
        )
274

275
        self._verify_quantization()
276
        self._verify_cuda_graph()
277
        self._verify_bnb_config()
278

279
280
281
282
    def _init_multimodal_config(
        self, limit_mm_per_prompt: Optional[Mapping[str, int]]
    ) -> Optional["MultiModalConfig"]:
        architectures = getattr(self.hf_config, "architectures", [])
283
        if ModelRegistry.is_multimodal_model(architectures):
284
            return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
285
286
287
288
289
290

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

        return None
291

292
293
294
295
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

296
297
298
299
300
301
302
303
304
    def _init_pooler_config(
        self,
        pooling_type: Optional[str] = None,
        pooling_norm: Optional[bool] = None,
        pooling_softmax: Optional[bool] = None,
        pooling_step_tag_id: Optional[int] = None,
        pooling_returned_token_ids: Optional[List[int]] = None
    ) -> Optional["PoolerConfig"]:
        if self.task == "embedding":
305
306
307
308
309
310
311
312
            pooling_config = get_pooling_config(self.model, self.revision)
            if pooling_config is not None:
                # override if user does not
                # specifies pooling_type and/or pooling_norm
                if pooling_type is None:
                    pooling_type = pooling_config["pooling_type"]
                if pooling_norm is None:
                    pooling_norm = pooling_config["normalize"]
313
314
315
316
317
318
319
320
            return PoolerConfig(
                pooling_type=pooling_type,
                pooling_norm=pooling_norm,
                pooling_softmax=pooling_softmax,
                pooling_step_tag_id=pooling_step_tag_id,
                pooling_returned_token_ids=pooling_returned_token_ids)
        return None

321
322
323
324
325
326
327
328
    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)

329
330
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
331
        if tokenizer_mode not in ["auto", "slow", "mistral"]:
332
333
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
334
                "either 'auto', 'slow' or 'mistral'.")
335
        self.tokenizer_mode = tokenizer_mode
336

337
338
    def _resolve_task(
        self,
339
        task_option: Union[TaskOption, _Task],
340
        hf_config: PretrainedConfig,
341
342
343
344
    ) -> Tuple[Set[_Task], _Task]:
        if task_option == "draft":
            return {"draft"}, "draft"

345
346
        architectures = getattr(hf_config, "architectures", [])

347
        task_support: Dict[_Task, bool] = {
348
349
350
351
352
            # 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),
        }
353
        supported_tasks_lst: List[_Task] = [
354
355
356
357
358
359
            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))
360

361
362
363
364
            if len(supported_tasks) > 1:
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
365
        else:
366
367
368
369
370
371
372
            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
373

374
        return supported_tasks, selected_task
375

376
377
378
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
379
            # compressed-tensors uses a "compression_config" key
380
            quant_cfg = getattr(self.hf_config, "compression_config", None)
381
382
        return quant_cfg

383
    def _verify_quantization(self) -> None:
384
        supported_quantization = [*QUANTIZATION_METHODS]
385
386
387
388
        rocm_supported_quantization = [
            "awq", "gptq", "fp8", "compressed_tensors", "compressed-tensors",
            "fbgemm_fp8"
        ]
389
        optimized_quantization_methods = [
390
391
392
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
            "awq_marlin", "fbgemm_fp8", "compressed_tensors",
            "compressed-tensors", "experts_int8"
393
        ]
394
        tpu_supported_quantization = ["tpu_int8"]
395
        neuron_supported_quantization = ["neuron_quant"]
396
397
398
399
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
400
401
        quant_cfg = self._parse_quant_hf_config()

402
403
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
404
405

            # Detect which checkpoint is it
406
            for _, method in QUANTIZATION_METHODS.items():
407
408
409
410
411
412
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
413

414
            # Verify quantization configurations.
415
            if self.quantization is None:
416
417
                self.quantization = quant_method
            elif self.quantization != quant_method:
418
419
                raise ValueError(
                    "Quantization method specified in the model config "
420
                    f"({quant_method}) does not match the quantization "
421
422
423
424
425
426
427
428
                    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}.")
429
            if current_platform.is_rocm(
430
            ) and self.quantization not in rocm_supported_quantization:
431
                raise ValueError(
432
433
                    f"{self.quantization} quantization is currently not "
                    f"supported in ROCm.")
434
            if current_platform.is_tpu(
435
436
437
438
            ) and self.quantization not in tpu_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in TPU Backend.")
439
            if self.quantization not in optimized_quantization_methods:
440
                logger.warning(
441
                    "%s quantization is not fully "
442
                    "optimized yet. The speed can be slower than "
443
                    "non-quantized models.", self.quantization)
444
            if (self.quantization == "awq" and current_platform.is_rocm()
445
446
447
448
449
                    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
450
            if current_platform.is_neuron(
451
452
453
454
            ) and self.quantization not in neuron_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in Neuron Backend.")
455

456
    def _verify_cuda_graph(self) -> None:
457
458
459
460
        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)
461

462
463
    def _verify_bnb_config(self) -> None:
        """
464
        The current version of bitsandbytes (0.44.0) with 8-bit models does not
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
        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

484
485
486
487
488
489
490
491
492
493
494
495
    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

496
497
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
498
        if device_config.device_type not in ("cuda", "tpu", "xpu", "hpu"):
499
            logger.warning(
500
501
                "Async output processing is only supported for CUDA, TPU, XPU "
                "and HPU."
502
                "Disabling it for other platforms.")
503
504
505
506
507
508
509
510
511
            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

512
513
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
514
        if device_config.device_type == "cuda" and self.enforce_eager:
515
516
517
518
519
520
521
522
523
            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
524
        if self.task == "embedding":
525
526
            self.use_async_output_proc = False

527
528
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
529
530
531
532
533
        if speculative_config:
            logger.warning("Async output processing is not supported with"
                           " speculative decoding currently.")
            self.use_async_output_proc = False

534
535
536
537
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
538
539
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
540
541
542
543
544
545
546
547
        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
548
549
550
551
552
553
554
555
556
557
558
        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
559

560
561
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], List[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
562
        """Get the sliding window size, or None if disabled."""
563
564
565
566

        # 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.
567
568
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
569
            return None
570
        return getattr(self.hf_text_config, "sliding_window", None)
571

572
    def get_sliding_window(self) -> Optional[Union[int, List[Optional[int]]]]:
573
574
575
576
577
578
579
580
        """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()

581
    def get_vocab_size(self) -> int:
582
        return self.hf_text_config.vocab_size
583

584
    def get_hidden_size(self) -> int:
585
        return self.hf_text_config.hidden_size
586
587

    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
588
589
590
591
592
593
        # 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
594
595
596
597

        if self.is_attention_free:
            return 0

598
599
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
600
        # FIXME(woosuk): This may not be true for all models.
601
602
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
603

604
605
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
606
        # For GPTBigCode & Falcon:
607
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
608
609
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
610
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
611
        new_decoder_arch_falcon = (
612
            self.hf_config.model_type in falcon_model_types
613
            and getattr(self.hf_config, "new_decoder_architecture", False))
614
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
615
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
616
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
617
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
618
            return 1
619

620
        # For DBRX and MPT
621
622
623
624
625
        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":
626
627
628
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

629
630
631
        if self.is_attention_free:
            return 0

632
633
634
635
636
637
638
639
640
641
        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:
642
            num_kv_heads = getattr(self.hf_text_config, attr, None)
643
644
645
646
647
            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.
648
        return self.hf_text_config.num_attention_heads
649
650
651
652
653
654
655
656
657
658

    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)
659

660
661
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
662
663
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
664

665
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
666
        from vllm.distributed.utils import get_pp_indices
Mor Zusman's avatar
Mor Zusman committed
667
668
        total_num_hidden_layers = getattr(self.hf_text_config,
                                          "num_hidden_layers", 0)
669
670
671
672
        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
673

674
675
676
677
    def get_num_attention_layers(self,
                                 parallel_config: "ParallelConfig") -> int:
        if self.is_attention_free:
            return 0
Mor Zusman's avatar
Mor Zusman committed
678
679
680

        num_layers = self.get_num_layers(parallel_config)

681
682
683
684
        # 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
685

686
687
688
689
690
691
692
693
694
695
696
697
    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

698
    @property
699
    def is_encoder_decoder(self) -> bool:
700
        """Extract the HF encoder/decoder model flag."""
701
702
703
704
705
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
706

707
708
709
710
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

711
712

class CacheConfig:
713
714
715
716
717
    """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
718
            vLLM execution.
719
        swap_space: Size of the CPU swap space per GPU (in GiB).
720
        cache_dtype: Data type for kv cache storage.
721
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
722
            profiled num_gpu_blocks if specified. Does nothing if None.
723
    """
724

725
726
727
728
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
729
        swap_space: float,
730
        cache_dtype: str,
731
        is_attention_free: bool = False,
732
        num_gpu_blocks_override: Optional[int] = None,
733
        sliding_window: Optional[int] = None,
734
        enable_prefix_caching: bool = False,
735
        cpu_offload_gb: float = 0,
736
737
738
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
739
        self.swap_space_bytes = swap_space * GiB_bytes
740
        self.num_gpu_blocks_override = num_gpu_blocks_override
741
        self.cache_dtype = cache_dtype
742
        self.is_attention_free = is_attention_free
743
        self.sliding_window = sliding_window
744
        self.enable_prefix_caching = enable_prefix_caching
745
        self.cpu_offload_gb = cpu_offload_gb
746

747
        self._verify_args()
748
        self._verify_cache_dtype()
749
        self._verify_prefix_caching()
750
751

        # Will be set after profiling.
752
753
        self.num_gpu_blocks: Optional[int] = None
        self.num_cpu_blocks: Optional[int] = None
754

755
    def metrics_info(self):
756
757
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
758
759
        return {key: str(value) for key, value in self.__dict__.items()}

760
761
762
763
764
765
    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}.")

766
767
768
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
769
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
770
            logger.info(
771
772
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
773
774
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
775
776
777
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

778
779
780
781
782
783
784
785
786
    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.")

787
788
789
790
791
792
793
794
795
796
    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

797
798
799
        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.")
800
801
802
        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:
803
            logger.warning("Possibly too large swap space. %s", msg)
804

805

806
807
808
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
809

810
811
812
813
814
815
816
817
    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
818
    pool_type: Union[str, Type["BaseTokenizerGroup"]]
819
820
821
    extra_config: dict

    def __post_init__(self):
822
823
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
824
825
826
827
828
829
            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(
830
831
        cls, tokenizer_pool_size: int,
        tokenizer_pool_type: Union[str, Type["BaseTokenizerGroup"]],
832
833
834
        tokenizer_pool_extra_config: Optional[Union[str, dict]]
    ) -> Optional["TokenizerPoolConfig"]:
        """Create a TokenizerPoolConfig from the given parameters.
835

836
        If tokenizer_pool_size is 0, return None.
837

838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
        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


860
861
862
863
864
865
866
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
867
    SHARDED_STATE = "sharded_state"
868
    GGUF = "gguf"
869
    BITSANDBYTES = "bitsandbytes"
870
    MISTRAL = "mistral"
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889


@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.
890
            "bitsandbytes" will load nf4 type weights.
891
        ignore_patterns: The list of patterns to ignore when loading the model.
892
            Default to "original/**/*" to avoid repeated loading of llama's
893
            checkpoints.
894
895
896
897
898
899
    """

    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)
900
    ignore_patterns: Optional[Union[List[str], str]] = None
901
902
903
904
905
906
907
908

    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()

909
910
911
912
913
914
915
        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/**/*"]

916
917
918
919
920
921
922
923
    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] = []
924
925
        if current_platform.is_rocm(
        ) and load_format in rocm_not_supported_load_format:
926
927
928
929
930
931
932
933
934
935
            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}")


936
class ParallelConfig:
937
938
939
940
941
    """Configuration for the distributed execution.

    Args:
        pipeline_parallel_size: Number of pipeline parallel groups.
        tensor_parallel_size: Number of tensor parallel groups.
942
        worker_use_ray: Deprecated, use distributed_executor_backend instead.
zspo's avatar
zspo committed
943
944
945
        max_parallel_loading_workers: Maximum number of multiple batches
            when load model sequentially. To avoid RAM OOM when using tensor
            parallel and large models.
946
947
        disable_custom_all_reduce: Disable the custom all-reduce kernel and
            fall back to NCCL.
948
949
        tokenizer_pool_config: Config for the tokenizer pool.
            If None, will use synchronous tokenization.
950
951
        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.
952
        placement_group: ray distributed model workers placement group.
953
        distributed_executor_backend: Backend to use for distributed model
954
955
956
957
958
959
            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.
960
    """
961

962
963
964
965
    def __init__(
        self,
        pipeline_parallel_size: int,
        tensor_parallel_size: int,
966
        worker_use_ray: Optional[bool] = None,
967
        max_parallel_loading_workers: Optional[int] = None,
968
        disable_custom_all_reduce: bool = False,
969
        tokenizer_pool_config: Optional[TokenizerPoolConfig] = None,
970
        ray_workers_use_nsight: bool = False,
971
        placement_group: Optional["PlacementGroup"] = None,
972
973
        distributed_executor_backend: Optional[Union[
            str, Type["ExecutorBase"]]] = None,
974
975
    ) -> None:
        self.pipeline_parallel_size = pipeline_parallel_size
976
        self.tensor_parallel_size = tensor_parallel_size
977
        self.distributed_executor_backend = distributed_executor_backend
978
        self.max_parallel_loading_workers = max_parallel_loading_workers
979
        self.disable_custom_all_reduce = disable_custom_all_reduce
980
        self.tokenizer_pool_config = tokenizer_pool_config
981
        self.ray_workers_use_nsight = ray_workers_use_nsight
982
        self.placement_group = placement_group
983
        self.world_size = pipeline_parallel_size * self.tensor_parallel_size
984

985
986
987
        if worker_use_ray:
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
988
            elif not self.use_ray:
989
990
991
992
                raise ValueError(f"worker-use-ray can't be used with "
                                 f"distributed executor backend "
                                 f"'{self.distributed_executor_backend}'.")

993
994
995
996
997
998
999
        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.")

1000
1001
1002
1003
1004
1005
1006
        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.")

1007
        if self.distributed_executor_backend is None and self.world_size > 1:
1008
1009
1010
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1011
            from vllm.executor import ray_utils
1012
            backend = "mp"
1013
            ray_found = ray_utils.ray_is_available()
1014
            if (current_platform.is_cuda()
1015
                    and cuda_device_count_stateless() < self.world_size):
1016
1017
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1018
1019
1020
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1021
1022
                backend = "ray"
            elif ray_found:
1023
                if self.placement_group:
1024
                    backend = "ray"
1025
1026
1027
1028
1029
1030
                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"
1031
1032
1033
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1034

1035
        self._verify_args()
1036
        self.rank: int = 0
1037

1038
1039
1040
1041
1042
1043
    @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)

1044
    def _verify_args(self) -> None:
1045
1046
1047
1048
1049
1050
1051
        # 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)):
1052
            raise ValueError(
1053
1054
1055
1056
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
                "values are 'ray', 'mp' or custom ExecutorBase subclass.")
        if self.use_ray:
1057
1058
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1059
        if current_platform.is_rocm():
1060
1061
1062
1063
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
1064
        if self.ray_workers_use_nsight and not self.use_ray:
1065
1066
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1067

1068
1069

class SchedulerConfig:
1070
1071
1072
    """Scheduler configuration.

    Args:
1073
        task: The task to use the model for.
1074
1075
1076
1077
        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
1078
        max_model_len: Maximum length of a sequence (including prompt
Lily Liu's avatar
Lily Liu committed
1079
            and generated text).
1080
1081
1082
1083
        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.
1084
1085
        delay_factor: Apply a delay (of delay factor multiplied by previous
            prompt latency) before scheduling next prompt.
1086
1087
        enable_chunked_prefill: If True, prefill requests can be chunked based
            on the remaining max_num_batched_tokens.
1088
        preemption_mode: Whether to perform preemption by swapping or
1089
1090
1091
1092
1093
            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.
1094
1095
1096
1097
        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
1098
        policy: The scheduling policy to use. "fcfs" (default) or "priority".
1099
    """
1100

1101
    def __init__(self,
1102
                 task: _Task,
1103
1104
1105
1106
1107
1108
                 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,
1109
                 is_multimodal_model: bool = False,
1110
                 preemption_mode: Optional[str] = None,
1111
                 num_scheduler_steps: int = 1,
1112
                 multi_step_stream_outputs: bool = False,
1113
1114
                 send_delta_data: bool = False,
                 policy: str = "fcfs") -> None:
1115
        if max_num_batched_tokens is None:
1116
            if enable_chunked_prefill:
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
                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
1127
1128
1129
            else:
                # If max_model_len is too short, use 2048 as the default value
                # for higher throughput.
1130
1131
                max_num_batched_tokens = max(max_model_len, 2048)

1132
            if task == "embedding":
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
                # 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

1147
        if enable_chunked_prefill:
1148
1149
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
1150
                self.max_num_batched_tokens)
1151

1152
        self.task: Final = task
1153
        self.max_num_seqs = max_num_seqs
Lily Liu's avatar
Lily Liu committed
1154
        self.max_model_len = max_model_len
1155
1156
        self.num_lookahead_slots = num_lookahead_slots
        self.delay_factor = delay_factor
1157
        self.chunked_prefill_enabled = enable_chunked_prefill
1158
        self.preemption_mode = preemption_mode
1159
        self.num_scheduler_steps = num_scheduler_steps
1160
        self.multi_step_stream_outputs = multi_step_stream_outputs
1161
        self.send_delta_data = send_delta_data
1162
        self.policy = policy
1163
1164
1165
        self._verify_args()

    def _verify_args(self) -> None:
1166
1167
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
1168
1169
1170
1171
1172
1173
1174
            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.")
1175

1176
1177
1178
1179
1180
        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}).")
1181

1182
1183
1184
1185
1186
1187
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
        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

1198

1199
class DeviceConfig:
1200
    device: Optional[torch.device]
1201

1202
1203
1204
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
1205
1206
            if current_platform.is_cuda_alike():
                self.device_type = "cuda"
1207
            elif current_platform.is_neuron():
1208
                self.device_type = "neuron"
1209
1210
            elif current_platform.is_hpu():
                self.device_type = "hpu"
1211
            elif current_platform.is_openvino():
1212
                self.device_type = "openvino"
1213
            elif current_platform.is_tpu():
1214
                self.device_type = "tpu"
1215
            elif current_platform.is_cpu():
1216
                self.device_type = "cpu"
1217
            elif current_platform.is_xpu():
1218
                self.device_type = "xpu"
1219
            else:
1220
                raise RuntimeError("Failed to infer device type")
1221
1222
1223
1224
1225
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
1226
        if self.device_type in ["neuron", "openvino"]:
1227
            self.device = torch.device("cpu")
1228
1229
        elif self.device_type in ["tpu"]:
            self.device = None
1230
1231
1232
1233
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1234

1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
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],
1248
        speculative_model_quantization: Optional[str],
1249
        speculative_draft_tensor_parallel_size: Optional[int],
1250
        num_speculative_tokens: Optional[int],
1251
        speculative_disable_mqa_scorer: Optional[bool],
1252
1253
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
1254
        disable_log_stats: bool,
1255
        speculative_disable_by_batch_size: Optional[int],
1256
1257
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1258
1259
1260
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
1261
        disable_logprobs: Optional[bool],
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
    ) -> 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.
1277
1278
1279
            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.
1280
1281
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
1282
            num_speculative_tokens (Optional[int]): The number of speculative
1283
1284
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
1285
1286
1287
            speculative_disable_mqa_scorer (Optional[bool]): Disable the MQA
                scorer for the speculative model and fall back to batch
                expansion for scoring.
1288
1289
1290
1291
1292
1293
            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.
1294
1295
1296
            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.
1297
1298
1299
1300
            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.
1301
1302
1303
1304
1305
1306
1307
1308
            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
1309
                accepted. This threshold is used only when we use the
1310
1311
1312
1313
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1314
1315
1316
1317
1318
            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.
1319

1320
1321
1322
1323
1324
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """

1325
1326
1327
1328
        if speculative_model is None:
            if num_speculative_tokens is not None:
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative_model.")
1329
1330
            return None

1331
1332
1333
1334
1335
1336
        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=}")

1337
1338
        # TODO: The user should be able to specify revision/max model len
        # for the draft model. It is not currently supported.
1339
1340
        draft_revision = None
        draft_code_revision = None
1341
        draft_quantization = speculative_model_quantization
1342

1343
1344
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1345
1346
1347
1348
1349
1350
1351
1352
                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=}")
1353

1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
            # 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,
1364
                task="draft",
1365
1366
1367
                tokenizer=target_model_config.tokenizer,
                tokenizer_mode=target_model_config.tokenizer_mode,
                trust_remote_code=target_model_config.trust_remote_code,
1368
1369
                allowed_local_media_path=target_model_config.
                allowed_local_media_path,
1370
1371
1372
1373
1374
1375
                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,
1376
                spec_target_max_model_len=target_model_config.max_model_len,
1377
1378
                quantization=draft_quantization,
                enforce_eager=target_model_config.enforce_eager,
1379
1380
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1381
1382
1383
                max_logprobs=target_model_config.max_logprobs,
            )

1384
            draft_hf_config = draft_model_config.hf_config
1385

1386
1387
1388
1389
1390
            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)
1391
1392
1393
1394
1395
1396
1397
1398
            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(
1399
1400
1401
                        "This speculative model supports a maximum of "
                        f"num_speculative_tokens={n_predict}, but "
                        f"{num_speculative_tokens=} was provided.")
1402

1403
1404
1405
1406
1407
1408
            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.")

1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
            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.")

1426
1427
1428
1429
1430
1431
1432
1433
1434
            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(
1435
                    target_parallel_config,
1436
                    speculative_draft_tensor_parallel_size, draft_hf_config))
1437

1438
1439
1440
1441
1442
1443
        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.")

1444
1445
1446
1447
        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
1448
1449
        if disable_logprobs is None:
            disable_logprobs = True
1450

1451
1452
1453
1454
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
1455
            speculative_disable_mqa_scorer,
1456
            speculative_disable_by_batch_size,
1457
1458
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
1459
1460
1461
1462
1463
            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,
1464
1465
            disable_logprobs=disable_logprobs,
            disable_log_stats=disable_log_stats,
1466
1467
        )

1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
    @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,
        )

1503
    @staticmethod
1504
1505
1506
1507
1508
1509
1510
    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.
1511
        """
1512
1513
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
1514
        if speculative_draft_tensor_parallel_size is None:
1515
1516
1517
1518
1519
1520
1521
1522
1523
            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
1524
1525
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
1526
            raise ValueError(
1527
                f"{speculative_draft_tensor_parallel_size=} cannot be "
1528
                f"other value than 1 or target model tensor_parallel_size")
1529
        return speculative_draft_tensor_parallel_size
1530

1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
    @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.
        """
1541
1542
1543
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
1544
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
1545
1546
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
            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,
1564
        speculative_disable_mqa_scorer: Optional[bool],
1565
1566
1567
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1568
1569
1570
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
1571
        disable_logprobs: bool,
1572
        disable_log_stats: bool,
1573
1574
1575
1576
1577
1578
1579
1580
    ):
        """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.
1581
1582
1583
1584
1585
            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.
1586
1587
1588
1589
1590
1591
1592
1593
            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
1594
                accepted. This threshold is used only when we use the
1595
1596
1597
1598
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1599
            disable_logprobs: If set to True, token log probabilities will not
1600
                be returned even if requested by sampling parameters. This
1601
1602
1603
1604
                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.
1605
1606
            disable_log_stats: Whether to disable periodic printing of stage
                times in speculative decoding.
1607
1608
1609
1610
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
1611
        self.speculative_disable_mqa_scorer = speculative_disable_mqa_scorer
1612
1613
1614
1615
        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
1616
1617
1618
1619
1620
        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
1621
        self.disable_logprobs = disable_logprobs
1622
        self.disable_log_stats = disable_log_stats
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633

        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)
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
            # 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}")
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670

    @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:
1671
1672
1673
1674
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
1675
1676
1677
1678
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


1679
1680
1681
1682
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
1683
    fully_sharded_loras: bool = False
1684
    max_cpu_loras: Optional[int] = None
1685
    lora_dtype: Optional[Union[torch.dtype, str]] = None
1686
1687
1688
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
1689
    long_lora_scaling_factors: Optional[Tuple[float]] = None
1690
1691

    def __post_init__(self):
1692
1693
1694
        # 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)
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
        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
1711
                f"max_loras ({self.max_loras})")
1712
1713
1714
1715
1716
1717

    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)
1718
1719
1720
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
1721
            # TODO support marlin
1722
1723
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
1724
1725

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
1726
1727
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
1728
1729
        if scheduler_config.chunked_prefill_enabled:
            raise ValueError("LoRA is not supported with chunked prefill yet.")
1730
1731


1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
@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)


1757
@dataclass
1758
class MultiModalConfig:
1759
1760
    """Controls the behavior of multimodal models."""

1761
    limit_per_prompt: Mapping[str, int] = field(default_factory=dict)
1762
1763
1764
1765
1766
    """
    The maximum number of multi-modal input instances allowed per prompt
    for each :class:`~vllm.multimodal.MultiModalPlugin`.
    """

1767
    # TODO: Add configs to init vision tower or not.
1768

1769

1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
@dataclass
class PoolerConfig:
    """Controls the behavior of pooler in embedding model"""

    pooling_type: Optional[str] = None
    pooling_norm: Optional[bool] = None
    pooling_softmax: Optional[bool] = None
    pooling_step_tag_id: Optional[int] = None
    pooling_returned_token_ids: Optional[List[int]] = None


1781
1782
1783
1784
1785
1786
1787
1788
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

1789
_ROCM_NOT_SUPPORTED_DTYPE: List[str] = []  #
1790

1791
1792
1793

def _get_and_verify_dtype(
    config: PretrainedConfig,
1794
    dtype: Union[str, torch.dtype],
1795
1796
1797
1798
1799
1800
1801
) -> 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

1802
1803
1804
1805
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
Woosuk Kwon's avatar
Woosuk Kwon committed
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
                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
1816
1817
            else:
                torch_dtype = config_dtype
1818
1819
1820
1821
1822
1823
1824

            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
1825
        else:
1826
1827
1828
1829
1830
            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
1831
    else:
1832
        raise ValueError(f"Unknown dtype: {dtype}")
1833
1834
1835
1836
1837

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
1838
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
1839
1840
1841
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
1842
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
1843
1844
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
1845
            # Casting between float16 and bfloat16 is allowed with a warning.
1846
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
1847
1848

    return torch_dtype
1849
1850
1851
1852
1853


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
1854
    disable_sliding_window: bool,
1855
    sliding_window_len: Optional[Union[int, List[Optional[int]]]],
1856
    spec_target_max_model_len: Optional[int] = None,
1857
    encoder_config: Optional[Any] = None,
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
) -> 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",
1868
1869
        # ChatGLM2
        "seq_length",
1870
1871
        # Command-R
        "model_max_length",
1872
1873
1874
1875
1876
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
1877
    # Choose the smallest "max_length" from the possible keys.
1878
    max_len_key = None
1879
    for key in possible_keys:
1880
1881
1882
1883
1884
        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)
1885
1886
1887
1888

    # 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:
1889
1890

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
1891
        max_len_key = "sliding_window" \
1892
1893
1894
            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)
1895
1896
1897

    # If none of the keys were found in the config, use a default and
    # log a warning.
1898
    if derived_max_model_len == float("inf"):
1899
1900
1901
1902
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

1903
1904
1905
1906
1907
        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

1908
1909
1910
1911
        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: "
1912
            "%s. Assuming the model's maximum length is %d.", possible_keys,
1913
            default_max_len)
1914
        derived_max_model_len = default_max_len
1915

1916
    rope_scaling = getattr(hf_config, "rope_scaling", None)
1917
    if rope_scaling is not None:
1918
1919
1920
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930

        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.")

1931
1932
1933
1934
            # 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)

1935
1936
1937
1938
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
1939

1940
1941
1942
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

1943
1944
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
1945
    if max_model_len is None:
1946
        max_model_len = int(derived_max_model_len)
1947
    elif max_model_len > derived_max_model_len:
1948
1949
1950
1951
1952
        # 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:
1953
1954
1955
1956
1957
1958
1959
            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.")
1960
        else:
1961
            msg = (
1962
                f"User-specified max_model_len ({max_model_len}) is greater "
1963
1964
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
1965
                f"{model_max_length} in model's config.json). This may lead "
1966
1967
1968
1969
1970
1971
1972
1973
1974
                "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")
1975
    return int(max_model_len)
1976
1977


1978
1979
1980
1981
1982
1983
1984
1985
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


1986
1987
1988
def get_served_model_name(model: str,
                          served_model_name: Optional[Union[str, List[str]]]):
    """
1989
1990
1991
1992
    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
1993
1994
1995
1996
1997
1998
1999
2000
2001
    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


2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
@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}")


2017
2018
2019
2020
2021
@dataclass
class ObservabilityConfig:
    """Configuration for observability."""
    otlp_traces_endpoint: Optional[str] = None

2022
2023
2024
2025
2026
2027
2028
2029
    # 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

2030
    def __post_init__(self):
2031
2032
2033
2034
2035
        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}")
2036
2037


2038
2039
2040
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
2041
2042
2043
    simplifies passing around the distinct configurations in the codebase.
    """

2044
2045
2046
2047
2048
2049
2050
2051
2052
    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
2053
2054
2055
2056
2057
    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
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
    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
2087

2088
2089
2090
2091
2092
2093
    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)

2094
2095
2096
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
2097
2098
2099
2100
2101
2102
2103
2104
        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)
2105
2106
2107
2108
2109

        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
2110
2111
2112
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
2113
2114
2115
2116
2117

        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)
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157

    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,
2158
        self.model_config.mm_processor_kwargs)