config.py 110 KB
Newer Older
1
import copy
2
import enum
3
import hashlib
4
import json
5
import warnings
6
from contextlib import contextmanager
7
from dataclasses import dataclass, field, replace
8
from pathlib import Path
9
10
11
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Counter, Dict,
                    Final, List, Literal, Mapping, Optional, Set, Tuple, Type,
                    Union)
12
13

import torch
14
from pydantic import BaseModel, Field, PrivateAttr
15
from transformers import PretrainedConfig
16

17
import vllm.envs as envs
18
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
Woosuk Kwon's avatar
Woosuk Kwon committed
19
from vllm.logger import init_logger
20
21
from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
                                                     get_quantization_config)
22
from vllm.model_executor.models import ModelRegistry
23
from vllm.platforms import current_platform
24
from vllm.tracing import is_otel_available, otel_import_error_traceback
25
26
27
28
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)
29
from vllm.utils import (GiB_bytes, cuda_device_count_stateless, get_cpu_memory,
30
                        identity, print_warning_once, resolve_obj_by_qualname)
31

32
33
34
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

35
    from vllm.executor.executor_base import ExecutorBase
36
37
    from vllm.model_executor.layers.quantization.base_config import (
        QuantizationConfig)
38
    from vllm.model_executor.model_loader.loader import BaseModelLoader
39
40
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
41
42
else:
    QuantizationConfig = None
43

44
45
logger = init_logger(__name__)

46
_EMBEDDING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
47
_MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
48

49
50
51
52
TaskOption = Literal["auto", "generate", "embedding"]

# "draft" is only used internally for speculative decoding
_Task = Literal["generate", "embedding", "draft"]
53

54
55
56
HfOverrides = Union[Dict[str, Any], Callable[[PretrainedConfig],
                                             PretrainedConfig]]

57
58

class ModelConfig:
59
60
61
62
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
63
            It is also used as the content for `model_name` tag in metrics
64
65
66
67
68
            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.
69
        tokenizer: Name or path of the huggingface tokenizer to use.
70
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
71
72
            available, "slow" will always use the slow tokenizer, and
            "mistral" will always use the tokenizer from `mistral_common`.
73
74
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
75
76
77
78
        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.
79
80
81
82
        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
83
84
85
        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.
86
        code_revision: The specific revision to use for the model code on
87
            Hugging Face Hub. It can be a branch name, a tag name, or a
88
            commit id. If unspecified, will use the default version.
89
90
91
        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.
92
93
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
94
95
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
96
97
        quantization_param_path: Path to JSON file containing scaling factors.
            Used to load KV cache scaling factors into the model when KV cache
98
99
            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
100
            model dtype is FP8_E4M3 on ROCm.
101
102
103
        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.
104
            If None, the user did not specify, so default to False.
105
106
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
107
108
109
            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.
110
111
112
113
        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.
114
115
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
116
        served_model_name: The model name used in metrics tag `model_name`,
117
118
            matches the model name exposed via the APIs. If multiple model
            names provided, the first name will be used. If not specified,
119
            the model name will be the same as `model`.
120
        limit_mm_per_prompt: Maximum number of data items per modality
121
            per prompt. Only applicable for multimodal models.
122
123
        config_format: The config format which shall be loaded.
            Defaults to 'auto' which defaults to 'hf'.
124
125
126
        hf_overrides: If a dictionary, contains arguments to be forwarded to the
            HuggingFace config. If a callable, it is called to update the
            HuggingFace config.
127
128
        mm_processor_kwargs: Arguments to be forwarded to the model's processor
            for multi-modal data, e.g., image processor.
129
130
131
132
133
134
        override_neuron_config: Initialize non default neuron config or
            override default neuron config that are specific to Neuron devices,
            this argument will be used to configure the neuron config that
            can not be gathered from the vllm arguments.
        override_pooling_config: Initialize non default pooling config or
            override default pooling config for the embedding model.
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
            rope_theta: Optional[float] = None,
            tokenizer_revision: Optional[str] = None,
            max_model_len: Optional[int] = None,
            spec_target_max_model_len: Optional[int] = None,
            quantization: Optional[str] = None,
            quantization_param_path: Optional[str] = None,
            enforce_eager: Optional[bool] = None,
            max_seq_len_to_capture: Optional[int] = None,
            max_logprobs: int = 20,
            disable_sliding_window: bool = False,
            skip_tokenizer_init: bool = False,
            served_model_name: Optional[Union[str, List[str]]] = None,
            limit_mm_per_prompt: Optional[Mapping[str, int]] = None,
            use_async_output_proc: bool = True,
            config_format: ConfigFormat = ConfigFormat.AUTO,
165
            hf_overrides: Optional[HfOverrides] = None,
166
            mm_processor_kwargs: Optional[Dict[str, Any]] = None,
167
168
            override_neuron_config: Optional[Dict[str, Any]] = None,
            override_pooler_config: Optional["PoolerConfig"] = None) -> None:
169
        self.model = model
170
        self.tokenizer = tokenizer
171
        self.tokenizer_mode = tokenizer_mode
172
        self.trust_remote_code = trust_remote_code
173
        self.allowed_local_media_path = allowed_local_media_path
174
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
175
        self.revision = revision
176
        self.code_revision = code_revision
177
178
179

        if hf_overrides is None:
            hf_overrides = {}
180
181
182
183
184
185
186
187

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

188
189
        if rope_scaling is not None:
            hf_override: Dict[str, Any] = {"rope_scaling": rope_scaling}
190
            hf_overrides_kw.update(hf_override)
191
192
193
194
195
            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}
196
            hf_overrides_kw.update(hf_override)
197
198
199
200
            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)

201
202
203
204
205
        # 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
206
        self.quantization = quantization
207
        self.quantization_param_path = quantization_param_path
208
        self.enforce_eager = enforce_eager
209
        self.max_seq_len_to_capture = max_seq_len_to_capture
210
        self.max_logprobs = max_logprobs
211
        self.disable_sliding_window = disable_sliding_window
212
        self.skip_tokenizer_init = skip_tokenizer_init
213
214
215
216
217
218

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

219
        self.hf_text_config = get_hf_text_config(self.hf_config)
220
        self.encoder_config = self._get_encoder_config()
221
222
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, revision)
223
        self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
224
        self.use_async_output_proc = use_async_output_proc
225
        self.mm_processor_kwargs = mm_processor_kwargs
Woosuk Kwon's avatar
Woosuk Kwon committed
226

227
228
        # Set enforce_eager to False if the value is unset.
        if self.enforce_eager is None:
229
230
            self.enforce_eager = False

231
232
233
234
235
236
        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):
237
238
239
            if envs.VLLM_ATTENTION_BACKEND == "XFORMERS":
                sliding_window_len_min = get_min_sliding_window(
                    self.hf_text_config.sliding_window)
240

241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
                print_warning_once(
                    f"{self.hf_text_config.model_type} has interleaved "
                    "attention, which is currently not supported by the "
                    "XFORMERS backend. Disabling sliding window and capping "
                    "the max length to the sliding window size "
                    f"({sliding_window_len_min}).")
                self.disable_sliding_window = True
            else:
                # for a model with interleaved attention,
                # the scheduler and the model treat it as full attention
                # (i.e., not dropping any tokens outside the window).
                # only the attention layer itself is aware of the sliding
                # window, and use the window size to compute the attention.
                self.hf_text_config.interleaved_sliding_window = sliding_window
                delattr(self.hf_text_config, "sliding_window")
                sliding_window = None
Woosuk Kwon's avatar
Woosuk Kwon committed
257

258
259
260
261
        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,
262
            sliding_window_len=self.get_hf_config_sliding_window(),
263
264
            spec_target_max_model_len=spec_target_max_model_len,
            encoder_config=self.encoder_config)
265
266
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
267
268
        self.multimodal_config = self._init_multimodal_config(
            limit_mm_per_prompt)
269
270
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
271

272
273
274
        self.is_attention_free = self._init_attention_free()
        self.has_inner_state = self._init_has_inner_state()

275
276
277
278
        if current_platform.is_neuron():
            self.override_neuron_config = override_neuron_config
        else:
            self.override_neuron_config = None
279
280
281
282

        supported_tasks, task = self._resolve_task(task, self.hf_config)
        self.supported_tasks = supported_tasks
        self.task: Final = task
283
        self.pooler_config = self._init_pooler_config(override_pooler_config)
284

285
        self._verify_quantization()
286
        self._verify_cuda_graph()
287
        self._verify_bnb_config()
288

289
290
291
292
    def _init_multimodal_config(
        self, limit_mm_per_prompt: Optional[Mapping[str, int]]
    ) -> Optional["MultiModalConfig"]:
        architectures = getattr(self.hf_config, "architectures", [])
293
        if ModelRegistry.is_multimodal_model(architectures):
294
            return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
295
296
297
298
299
300

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

        return None
301

302
303
304
305
    def _get_encoder_config(self):
        return get_sentence_transformer_tokenizer_config(
            self.model, self.revision)

306
307
    def _init_pooler_config(
        self,
308
        override_pooler_config: Optional["PoolerConfig"],
309
    ) -> Optional["PoolerConfig"]:
310

311
        if self.task == "embedding":
312
313
314
315
316
317
318
319
320
321
322
            user_config = override_pooler_config or PoolerConfig()

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

            return user_config

323
324
        return None

325
326
327
328
329
330
331
332
    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)

333
334
    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
335
        if tokenizer_mode not in ["auto", "slow", "mistral"]:
336
337
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
338
                "either 'auto', 'slow' or 'mistral'.")
339
        self.tokenizer_mode = tokenizer_mode
340

341
342
    def _resolve_task(
        self,
343
        task_option: Union[TaskOption, _Task],
344
        hf_config: PretrainedConfig,
345
346
347
348
    ) -> Tuple[Set[_Task], _Task]:
        if task_option == "draft":
            return {"draft"}, "draft"

349
350
        architectures = getattr(hf_config, "architectures", [])

351
        task_support: Dict[_Task, bool] = {
352
353
354
355
356
            # 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),
        }
357
        supported_tasks_lst: List[_Task] = [
358
359
360
361
362
363
            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))
364

365
366
367
368
            if len(supported_tasks) > 1:
                logger.info(
                    "This model supports multiple tasks: %s. "
                    "Defaulting to '%s'.", supported_tasks, selected_task)
369
        else:
370
371
372
373
374
375
376
            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
377

378
        return supported_tasks, selected_task
379

380
381
382
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
383
            # compressed-tensors uses a "compression_config" key
384
            quant_cfg = getattr(self.hf_config, "compression_config", None)
385
386
        return quant_cfg

387
    def _verify_quantization(self) -> None:
388
        supported_quantization = QUANTIZATION_METHODS
389
390
        rocm_supported_quantization = [
            "awq", "gptq", "fp8", "compressed_tensors", "compressed-tensors",
391
            "fbgemm_fp8", "gguf"
392
        ]
393
        optimized_quantization_methods = [
394
395
396
            "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
            "awq_marlin", "fbgemm_fp8", "compressed_tensors",
            "compressed-tensors", "experts_int8"
397
        ]
398
        tpu_supported_quantization = ["tpu_int8"]
399
        neuron_supported_quantization = ["neuron_quant"]
400
401
402
403
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
404
405
        quant_cfg = self._parse_quant_hf_config()

406
407
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
408
409

            # Detect which checkpoint is it
410
411
            for name in QUANTIZATION_METHODS:
                method = get_quantization_config(name)
412
413
414
415
416
417
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
418

419
            # Verify quantization configurations.
420
            if self.quantization is None:
421
422
                self.quantization = quant_method
            elif self.quantization != quant_method:
423
424
                raise ValueError(
                    "Quantization method specified in the model config "
425
                    f"({quant_method}) does not match the quantization "
426
427
428
429
430
431
432
433
                    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}.")
434
            if current_platform.is_rocm(
435
            ) and self.quantization not in rocm_supported_quantization:
436
                raise ValueError(
437
438
                    f"{self.quantization} quantization is currently not "
                    f"supported in ROCm.")
439
            if current_platform.is_tpu(
440
441
442
443
            ) and self.quantization not in tpu_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in TPU Backend.")
444
            if self.quantization not in optimized_quantization_methods:
445
                logger.warning(
446
                    "%s quantization is not fully "
447
                    "optimized yet. The speed can be slower than "
448
                    "non-quantized models.", self.quantization)
449
            if (self.quantization == "awq" and current_platform.is_rocm()
450
451
452
453
454
                    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
455
            if current_platform.is_neuron(
456
457
458
459
            ) and self.quantization not in neuron_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in Neuron Backend.")
460

461
    def _verify_cuda_graph(self) -> None:
462
463
464
465
        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)
466

467
468
    def _verify_bnb_config(self) -> None:
        """
469
        The current version of bitsandbytes (0.44.0) with 8-bit models does not
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
        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

489
490
491
492
493
494
495
496
497
498
499
500
    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

501
502
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
503
        if device_config.device_type not in ("cuda", "tpu", "xpu", "hpu"):
504
            logger.warning(
505
506
                "Async output processing is only supported for CUDA, TPU, XPU "
                "and HPU."
507
                "Disabling it for other platforms.")
508
509
510
511
512
513
514
515
516
            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

517
518
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
519
        if device_config.device_type == "cuda" and self.enforce_eager:
520
521
522
523
524
525
526
527
528
            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
529
        if self.task == "embedding":
530
531
            self.use_async_output_proc = False

532
533
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
534
535
536
537
538
        if speculative_config:
            logger.warning("Async output processing is not supported with"
                           " speculative decoding currently.")
            self.use_async_output_proc = False

539
540
541
542
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
543
544
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
545
546
547
548
549
550
551
552
        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
553
554
555
556
557
558
559
560
561
562
563
        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
564

565
566
    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], List[Optional[int]]]:
Woosuk Kwon's avatar
Woosuk Kwon committed
567
        """Get the sliding window size, or None if disabled."""
568
569
570
571

        # 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.
572
573
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
574
            return None
575
        return getattr(self.hf_text_config, "sliding_window", None)
576

577
    def get_sliding_window(self) -> Optional[Union[int, List[Optional[int]]]]:
578
579
580
581
582
583
584
585
        """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()

586
    def get_vocab_size(self) -> int:
587
        return self.hf_text_config.vocab_size
588

589
    def get_hidden_size(self) -> int:
590
        return self.hf_text_config.hidden_size
591
592

    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
593
594
595
596
597
598
        # 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
599
600
601
602

        if self.is_attention_free:
            return 0

603
604
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
605
        # FIXME(woosuk): This may not be true for all models.
606
607
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
608

609
610
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
611
        # For GPTBigCode & Falcon:
612
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
613
614
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
615
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
616
        new_decoder_arch_falcon = (
617
            self.hf_config.model_type in falcon_model_types
618
            and getattr(self.hf_config, "new_decoder_architecture", False))
619
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
620
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
621
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
622
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
623
            return 1
624

625
        # For DBRX and MPT
626
627
628
629
630
        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":
631
632
633
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

634
635
636
        if self.is_attention_free:
            return 0

637
638
639
640
641
642
643
644
645
646
        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:
647
            num_kv_heads = getattr(self.hf_text_config, attr, None)
648
649
650
651
652
            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.
653
        return self.hf_text_config.num_attention_heads
654
655
656
657
658
659
660
661
662
663

    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)
664

665
666
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
667
668
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
669

670
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
671
        from vllm.distributed.utils import get_pp_indices
Mor Zusman's avatar
Mor Zusman committed
672
673
        total_num_hidden_layers = getattr(self.hf_text_config,
                                          "num_hidden_layers", 0)
674
675
676
677
        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
678

679
680
681
682
    def get_num_attention_layers(self,
                                 parallel_config: "ParallelConfig") -> int:
        if self.is_attention_free:
            return 0
Mor Zusman's avatar
Mor Zusman committed
683
684
685

        num_layers = self.get_num_layers(parallel_config)

686
687
688
689
        # 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
690

691
692
693
694
695
696
697
698
699
700
701
702
    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

703
    @property
704
    def is_encoder_decoder(self) -> bool:
705
        """Extract the HF encoder/decoder model flag."""
706
707
708
709
710
        return is_encoder_decoder(self.hf_config)

    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)
711

712
713
714
715
    @property
    def is_multimodal_model(self) -> bool:
        return self.multimodal_config is not None

716
717
718
719
720
    @property
    def is_cross_encoder(self) -> bool:
        architectures = getattr(self.hf_config, "architectures", [])
        return ModelRegistry.is_cross_encoder_model(architectures)

721
722

class CacheConfig:
723
724
725
726
727
    """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
728
            vLLM execution.
729
        swap_space: Size of the CPU swap space per GPU (in GiB).
730
        cache_dtype: Data type for kv cache storage.
731
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
732
            profiled num_gpu_blocks if specified. Does nothing if None.
733
    """
734

735
736
737
738
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
739
        swap_space: float,
740
        cache_dtype: str,
741
        is_attention_free: bool = False,
742
        num_gpu_blocks_override: Optional[int] = None,
743
        sliding_window: Optional[int] = None,
744
        enable_prefix_caching: bool = False,
745
        cpu_offload_gb: float = 0,
746
747
748
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
749
        self.swap_space_bytes = swap_space * GiB_bytes
750
        self.num_gpu_blocks_override = num_gpu_blocks_override
751
        self.cache_dtype = cache_dtype
752
        self.is_attention_free = is_attention_free
753
        self.sliding_window = sliding_window
754
        self.enable_prefix_caching = enable_prefix_caching
755
        self.cpu_offload_gb = cpu_offload_gb
756

757
        self._verify_args()
758
        self._verify_cache_dtype()
759
        self._verify_prefix_caching()
760
761

        # Will be set after profiling.
762
763
        self.num_gpu_blocks: Optional[int] = None
        self.num_cpu_blocks: Optional[int] = None
764

765
    def metrics_info(self):
766
767
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
768
769
        return {key: str(value) for key, value in self.__dict__.items()}

770
771
772
773
774
775
    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}.")

776
777
778
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
779
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
780
            logger.info(
781
782
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
783
784
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
785
786
787
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

788
789
790
791
792
793
794
795
796
    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.")

797
798
799
800
801
802
803
804
805
806
    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

807
808
809
        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.")
810
811
812
        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:
813
            logger.warning("Possibly too large swap space. %s", msg)
814

815

816
817
818
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
819

820
821
822
823
824
825
826
827
    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
828
    pool_type: Union[str, Type["BaseTokenizerGroup"]]
829
830
831
    extra_config: dict

    def __post_init__(self):
832
833
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
834
835
836
837
838
839
            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(
840
841
        cls, tokenizer_pool_size: int,
        tokenizer_pool_type: Union[str, Type["BaseTokenizerGroup"]],
842
843
844
        tokenizer_pool_extra_config: Optional[Union[str, dict]]
    ) -> Optional["TokenizerPoolConfig"]:
        """Create a TokenizerPoolConfig from the given parameters.
845

846
        If tokenizer_pool_size is 0, return None.
847

848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
        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


870
871
872
873
874
875
876
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
877
    SHARDED_STATE = "sharded_state"
878
    GGUF = "gguf"
879
    BITSANDBYTES = "bitsandbytes"
880
    MISTRAL = "mistral"
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899


@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.
900
            "bitsandbytes" will load nf4 type weights.
901
        ignore_patterns: The list of patterns to ignore when loading the model.
902
            Default to "original/**/*" to avoid repeated loading of llama's
903
            checkpoints.
904
905
906
907
908
909
    """

    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)
910
    ignore_patterns: Optional[Union[List[str], str]] = None
911
912
913
914
915
916
917
918

    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()

919
920
921
922
923
924
925
        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/**/*"]

926
927
928
929
930
931
932
933
    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] = []
934
935
        if current_platform.is_rocm(
        ) and load_format in rocm_not_supported_load_format:
936
937
938
939
940
941
942
943
944
945
            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}")


946
@dataclass
947
class ParallelConfig:
948
    """Configuration for the distributed execution."""
949

950
951
    pipeline_parallel_size: int = 1  # Number of pipeline parallel groups.
    tensor_parallel_size: int = 1  # Number of tensor parallel groups.
952

953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
    # Deprecated, use distributed_executor_backend instead.
    worker_use_ray: Optional[bool] = None

    # Maximum number of multiple batches
    # when load model sequentially. To avoid RAM OOM when using tensor
    # parallel and large models.
    max_parallel_loading_workers: Optional[int] = None

    # Disable the custom all-reduce kernel and fall back to NCCL.
    disable_custom_all_reduce: bool = False

    # Config for the tokenizer pool. If None, will use synchronous tokenization.
    tokenizer_pool_config: Optional[TokenizerPoolConfig] = None

    # Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler.
    ray_workers_use_nsight: bool = False

    # ray distributed model workers placement group.
    placement_group: Optional["PlacementGroup"] = None

    # Backend to use for distributed model
    # workers, either "ray" or "mp" (multiprocessing). If the product
    # of pipeline_parallel_size and tensor_parallel_size is less than
    # or equal to the number of GPUs available, "mp" will be used to
    # keep processing on a single host. Otherwise, this will default
    # to "ray" if Ray is installed and fail otherwise. Note that tpu
    # and hpu only support Ray for distributed inference.
    distributed_executor_backend: Optional[Union[str,
                                                 Type["ExecutorBase"]]] = None

    # the full name of the worker class to use. If "auto", the worker class
    # will be determined based on the platform.
    worker_cls: str = "auto"

    world_size: int = field(init=False)

    rank: int = 0

    def __post_init__(self) -> None:
        self.world_size = self.pipeline_parallel_size * \
            self.tensor_parallel_size

        if self.worker_use_ray:
996
997
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
998
            elif not self.use_ray:
999
1000
1001
                raise ValueError(f"worker-use-ray can't be used with "
                                 f"distributed executor backend "
                                 f"'{self.distributed_executor_backend}'.")
1002
1003
1004
        ray_only_devices = ["tpu", "hpu"]
        if (current_platform.device_type in ray_only_devices
                and self.world_size > 1):
1005
1006
1007
1008
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
            if self.distributed_executor_backend != "ray":
                raise ValueError(
1009
1010
                    f"{current_platform.device_type.upper()} backend only "
                    "supports Ray for distributed inference.")
1011

1012
        if self.distributed_executor_backend is None and self.world_size > 1:
1013
1014
1015
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

1016
            from vllm.executor import ray_utils
1017
            backend = "mp"
1018
            ray_found = ray_utils.ray_is_available()
1019
            if (current_platform.is_cuda()
1020
                    and cuda_device_count_stateless() < self.world_size):
1021
1022
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
1023
1024
1025
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
1026
1027
                backend = "ray"
            elif ray_found:
1028
                if self.placement_group:
1029
                    backend = "ray"
1030
1031
1032
1033
1034
1035
                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"
1036
1037
1038
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
1039

1040
1041
        self._verify_args()

1042
1043
1044
1045
1046
1047
    @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)

1048
    def _verify_args(self) -> None:
1049
1050
1051
1052
1053
1054
1055
        # 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)):
1056
            raise ValueError(
1057
1058
1059
1060
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
                "values are 'ray', 'mp' or custom ExecutorBase subclass.")
        if self.use_ray:
1061
1062
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
1063
        if current_platform.is_rocm():
1064
1065
1066
1067
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
1068
        if self.ray_workers_use_nsight and not self.use_ray:
1069
1070
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
1071

1072

1073
@dataclass
1074
class SchedulerConfig:
1075
    """Scheduler configuration."""
1076

1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
    task: str = "generate"  # The task to use the model for.

    # Maximum number of tokens to be processed in a single iteration.
    max_num_batched_tokens: int = field(default=None)  # type: ignore

    # Maximum number of sequences to be processed in a single iteration.
    max_num_seqs: int = 128

    # Maximum length of a sequence (including prompt and generated text).
    max_model_len: int = 8192

    # 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.
    num_lookahead_slots: int = 0

    # Apply a delay (of delay factor multiplied by previous
    # prompt latency) before scheduling next prompt.
    delay_factor: float = 0.0

    # If True, prefill requests can be chunked based
    # on the remaining max_num_batched_tokens.
    enable_chunked_prefill: bool = False

    is_multimodal_model: bool = False
1103

1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
    # Whether to perform preemption by swapping or
    # 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.
    preemption_mode: Optional[str] = None

    num_scheduler_steps: int = 1

    multi_step_stream_outputs: bool = False

    # 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
    send_delta_data: bool = False

    # The scheduling policy to use. "fcfs" (default) or "priority".
    policy: str = "fcfs"

    chunked_prefill_enabled: bool = field(init=False)

    def __post_init__(self) -> None:
        if self.max_num_batched_tokens is None:
            if self.enable_chunked_prefill:
                if self.num_scheduler_steps > 1:
1131
1132
1133
1134
                    # 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.
1135
                    self.max_num_batched_tokens = max(self.max_model_len, 2048)
1136
                else:
1137
1138
1139
                    # This value is chosen to have a balance between ITL
                    # and TTFT. Note it is not optimized for throughput.
                    self.max_num_batched_tokens = 2048
1140
1141
1142
            else:
                # If max_model_len is too short, use 2048 as the default value
                # for higher throughput.
1143
                self.max_num_batched_tokens = max(self.max_model_len, 2048)
1144

1145
            if self.task == "embedding":
1146
                # For embedding, choose specific value for higher throughput
1147
1148
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1149
1150
                    _EMBEDDING_MODEL_MAX_NUM_BATCHED_TOKENS,
                )
1151
            if self.is_multimodal_model:
1152
                # The value needs to be at least the number of multimodal tokens
1153
1154
                self.max_num_batched_tokens = max(
                    self.max_num_batched_tokens,
1155
1156
1157
                    _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                )

1158
        if self.enable_chunked_prefill:
1159
1160
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
1161
                self.max_num_batched_tokens)
1162

1163
        self.chunked_prefill_enabled = self.enable_chunked_prefill
1164
1165
1166
        self._verify_args()

    def _verify_args(self) -> None:
1167
1168
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
1169
1170
1171
1172
1173
1174
1175
            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.")
1176

1177
1178
1179
1180
1181
        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}).")
1182

1183
1184
1185
1186
1187
1188
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
        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

1199

1200
class DeviceConfig:
1201
    device: Optional[torch.device]
1202
    device_type: str
1203

1204
1205
1206
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
1207
            self.device_type = current_platform.device_type
1208
            if not self.device_type:
1209
                raise RuntimeError("Failed to infer device type")
1210
1211
1212
1213
1214
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
1215
        if self.device_type in ["neuron", "openvino"]:
1216
            self.device = torch.device("cpu")
1217
1218
        elif self.device_type in ["tpu"]:
            self.device = None
1219
1220
1221
1222
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

1223

1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
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],
1237
        speculative_model_quantization: Optional[str],
1238
        speculative_draft_tensor_parallel_size: Optional[int],
1239
        num_speculative_tokens: Optional[int],
1240
        speculative_disable_mqa_scorer: Optional[bool],
1241
1242
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
1243
        disable_log_stats: bool,
1244
        speculative_disable_by_batch_size: Optional[int],
1245
1246
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1247
1248
1249
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
1250
        disable_logprobs: Optional[bool],
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
    ) -> 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.
1266
1267
1268
            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.
1269
1270
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
1271
            num_speculative_tokens (Optional[int]): The number of speculative
1272
1273
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
1274
1275
1276
            speculative_disable_mqa_scorer (Optional[bool]): Disable the MQA
                scorer for the speculative model and fall back to batch
                expansion for scoring.
1277
1278
1279
1280
1281
1282
            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.
1283
1284
1285
            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.
1286
1287
1288
1289
            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.
1290
1291
1292
1293
1294
1295
1296
1297
            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
1298
                accepted. This threshold is used only when we use the
1299
1300
1301
1302
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1303
1304
1305
1306
1307
            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.
1308

1309
1310
1311
1312
1313
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """

1314
1315
1316
1317
        if speculative_model is None:
            if num_speculative_tokens is not None:
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative_model.")
1318
1319
            return None

1320
1321
1322
1323
1324
1325
        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=}")

1326
1327
        # TODO: The user should be able to specify revision/max model len
        # for the draft model. It is not currently supported.
1328
1329
        draft_revision = None
        draft_code_revision = None
1330
        draft_quantization = speculative_model_quantization
1331

1332
1333
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1334
1335
1336
1337
1338
1339
1340
1341
                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=}")
1342

1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
            # 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,
1353
                task="draft",
1354
1355
1356
                tokenizer=target_model_config.tokenizer,
                tokenizer_mode=target_model_config.tokenizer_mode,
                trust_remote_code=target_model_config.trust_remote_code,
1357
1358
                allowed_local_media_path=target_model_config.
                allowed_local_media_path,
1359
1360
1361
1362
1363
1364
                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,
1365
                spec_target_max_model_len=target_model_config.max_model_len,
1366
1367
                quantization=draft_quantization,
                enforce_eager=target_model_config.enforce_eager,
1368
1369
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1370
1371
1372
                max_logprobs=target_model_config.max_logprobs,
            )

1373
            draft_hf_config = draft_model_config.hf_config
1374

1375
1376
1377
1378
1379
            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)
1380
1381
1382
1383
1384
1385
1386
1387
            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(
1388
1389
1390
                        "This speculative model supports a maximum of "
                        f"num_speculative_tokens={n_predict}, but "
                        f"{num_speculative_tokens=} was provided.")
1391

1392
1393
1394
1395
1396
1397
            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.")

1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
            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.")

1415
1416
1417
1418
1419
1420
1421
1422
1423
            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(
1424
                    target_parallel_config,
1425
                    speculative_draft_tensor_parallel_size, draft_hf_config))
1426

1427
1428
1429
1430
1431
1432
        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.")

1433
1434
1435
1436
        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
1437
1438
        if disable_logprobs is None:
            disable_logprobs = True
1439

1440
1441
1442
1443
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
1444
            speculative_disable_mqa_scorer,
1445
            speculative_disable_by_batch_size,
1446
1447
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
1448
1449
1450
1451
1452
            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,
1453
1454
            disable_logprobs=disable_logprobs,
            disable_log_stats=disable_log_stats,
1455
1456
        )

1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
    @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,
        )

1492
    @staticmethod
1493
1494
1495
1496
1497
1498
1499
    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.
1500
        """
1501
1502
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
1503
        if speculative_draft_tensor_parallel_size is None:
1504
1505
1506
1507
1508
1509
1510
1511
1512
            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
1513
1514
        elif speculative_draft_tensor_parallel_size not in (
                1, target_parallel_config.tensor_parallel_size):
1515
            raise ValueError(
1516
                f"{speculative_draft_tensor_parallel_size=} cannot be "
1517
                f"other value than 1 or target model tensor_parallel_size")
1518
        return speculative_draft_tensor_parallel_size
1519

1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
    @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.
        """
1530
1531
1532
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
1533
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
1534
1535
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
            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,
1553
        speculative_disable_mqa_scorer: Optional[bool],
1554
1555
1556
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1557
1558
1559
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
1560
        disable_logprobs: bool,
1561
        disable_log_stats: bool,
1562
1563
1564
1565
1566
1567
1568
1569
    ):
        """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.
1570
1571
1572
1573
1574
            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.
1575
1576
1577
1578
1579
1580
1581
1582
            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
1583
                accepted. This threshold is used only when we use the
1584
1585
1586
1587
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1588
            disable_logprobs: If set to True, token log probabilities will not
1589
                be returned even if requested by sampling parameters. This
1590
1591
1592
1593
                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.
1594
1595
            disable_log_stats: Whether to disable periodic printing of stage
                times in speculative decoding.
1596
1597
1598
1599
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
1600
        self.speculative_disable_mqa_scorer = speculative_disable_mqa_scorer
1601
1602
1603
1604
        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
1605
1606
1607
1608
1609
        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
1610
        self.disable_logprobs = disable_logprobs
1611
        self.disable_log_stats = disable_log_stats
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622

        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)
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
            # 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}")
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659

    @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:
1660
1661
1662
1663
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
1664
1665
1666
1667
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


1668
1669
1670
1671
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
1672
    fully_sharded_loras: bool = False
1673
    max_cpu_loras: Optional[int] = None
1674
    lora_dtype: Optional[Union[torch.dtype, str]] = None
1675
1676
1677
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
1678
    long_lora_scaling_factors: Optional[Tuple[float]] = None
1679
    bias_enabled: bool = False
1680
1681

    def __post_init__(self):
1682
1683
1684
        # 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)
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
        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
1701
                f"max_loras ({self.max_loras})")
1702
1703
1704
1705
1706
1707

    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)
1708
1709
1710
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
1711
            # TODO support marlin
1712
1713
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
1714
1715

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
1716
1717
        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
        # If the feature combo become valid
1718
1719
        if scheduler_config.chunked_prefill_enabled:
            raise ValueError("LoRA is not supported with chunked prefill yet.")
1720
1721


1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
@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)


1747
@dataclass
1748
class MultiModalConfig:
1749
1750
    """Controls the behavior of multimodal models."""

1751
    limit_per_prompt: Mapping[str, int] = field(default_factory=dict)
1752
1753
1754
1755
1756
    """
    The maximum number of multi-modal input instances allowed per prompt
    for each :class:`~vllm.multimodal.MultiModalPlugin`.
    """

1757
    # TODO: Add configs to init vision tower or not.
1758

1759

1760
1761
@dataclass
class PoolerConfig:
1762
    """Controls the behavior of output pooling in embedding models."""
1763
1764

    pooling_type: Optional[str] = None
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
    """
    The pooling method of the embedding model. This should be a key in
    :class:`vllm.model_executor.layers.pooler.PoolingType`.
    """

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

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

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

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

    @staticmethod
    def from_json(json_str: str) -> "PoolerConfig":
        return PoolerConfig(**json.loads(json_str))
1799
1800


1801
1802
1803
1804
1805
1806
1807
1808
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

1809
_ROCM_NOT_SUPPORTED_DTYPE: List[str] = []  #
1810

1811
1812
1813

def _get_and_verify_dtype(
    config: PretrainedConfig,
1814
    dtype: Union[str, torch.dtype],
1815
1816
1817
1818
1819
1820
1821
) -> 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

1822
1823
1824
1825
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
Woosuk Kwon's avatar
Woosuk Kwon committed
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
                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
1836
1837
            else:
                torch_dtype = config_dtype
1838
1839
1840
1841
1842
1843
1844

            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
1845
        else:
1846
1847
1848
1849
1850
            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
1851
    else:
1852
        raise ValueError(f"Unknown dtype: {dtype}")
1853
1854
1855
1856
1857

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
1858
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
1859
1860
1861
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
1862
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
1863
1864
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
1865
            # Casting between float16 and bfloat16 is allowed with a warning.
1866
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
1867
1868

    return torch_dtype
1869
1870
1871
1872
1873


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
1874
    disable_sliding_window: bool,
1875
    sliding_window_len: Optional[Union[int, List[Optional[int]]]],
1876
    spec_target_max_model_len: Optional[int] = None,
1877
    encoder_config: Optional[Any] = None,
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
) -> 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",
1888
1889
        # ChatGLM2
        "seq_length",
1890
1891
        # Command-R
        "model_max_length",
1892
1893
1894
1895
1896
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
1897
    # Choose the smallest "max_length" from the possible keys.
1898
    max_len_key = None
1899
    for key in possible_keys:
1900
1901
1902
1903
1904
        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)
1905
1906
1907
1908

    # 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:
1909
1910

        sliding_window_len_min = get_min_sliding_window(sliding_window_len)
1911
        max_len_key = "sliding_window" \
1912
1913
1914
            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)
1915
1916
1917

    # If none of the keys were found in the config, use a default and
    # log a warning.
1918
    if derived_max_model_len == float("inf"):
1919
1920
1921
1922
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

1923
1924
1925
1926
1927
        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

1928
1929
1930
1931
        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: "
1932
            "%s. Assuming the model's maximum length is %d.", possible_keys,
1933
            default_max_len)
1934
        derived_max_model_len = default_max_len
1935

1936
    rope_scaling = getattr(hf_config, "rope_scaling", None)
1937
    if rope_scaling is not None:
1938
1939
1940
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950

        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.")

1951
1952
1953
1954
            # 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)

1955
1956
1957
1958
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
1959

1960
1961
1962
    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

1963
1964
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
1965
    if max_model_len is None:
1966
        max_model_len = int(derived_max_model_len)
1967
    elif max_model_len > derived_max_model_len:
1968
1969
1970
1971
1972
        # 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:
1973
1974
1975
1976
1977
1978
1979
            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.")
1980
        else:
1981
            msg = (
1982
                f"User-specified max_model_len ({max_model_len}) is greater "
1983
1984
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
1985
                f"{model_max_length} in model's config.json). This may lead "
1986
1987
1988
1989
1990
1991
1992
1993
1994
                "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")
1995
    return int(max_model_len)
1996
1997


1998
1999
2000
2001
2002
2003
2004
2005
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


2006
2007
2008
def get_served_model_name(model: str,
                          served_model_name: Optional[Union[str, List[str]]]):
    """
2009
2010
2011
2012
    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
2013
2014
2015
2016
2017
2018
2019
2020
2021
    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


2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
@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}")


2037
2038
2039
2040
2041
@dataclass
class ObservabilityConfig:
    """Configuration for observability."""
    otlp_traces_endpoint: Optional[str] = None

2042
2043
2044
2045
2046
2047
2048
2049
    # 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

2050
    def __post_init__(self):
2051
2052
2053
2054
2055
        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}")
2056
2057


2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
class CompilationLevel:
    # constants for the levels of the compilation process
    NO_COMPILATION = 0
    DYNAMO_AS_IS = 1
    DYNAMO_ONCE = 2
    PIECEWISE = 3


class CompilationConfig(BaseModel):
    """
    Configuration for compilation.
    It has three parts:
    - Top-level Compilation control:
        - level: the level of compilation.
            - 0: no compilation.
            - 1: dynamo as is.
            - 2: dynamo once.
            - 3: piecewise compilation.
2076
2077
2078
2079
2080
2081
2082
        - backend: the backend for compilation. It needs to be a string.
            - "" (empty string): use the default backend.
            - "eager"/"openxla"/...: use the specified backend registered in PyTorch.
            - "full.module.name": a qualified name which can be used to import the backend function.
            We use string to avoid serialization issues when using compilation in a distributed setting.
            When the compilation level is 1 or 2, the backend is used for the compilation directly (it sees the whole graph).
            When the compilation level is 3, the backend is used for the piecewise compilation (it sees a part of the graph).
2083
2084
2085
2086
2087
2088
2089
2090
2091
        - custom_ops: fine-grained control over which custom ops to enable/disable.
            Use 'all' to enable all, 'none' to disable all.
            Also specify a list of custom op names to enable (prefixed with a '+'),
            or disable (prefixed with a '-').
            Examples:
                - 'all,-op1' to enable all except op1
                - 'none,+op1,+op2' to enable only op1 and op2
            By default, all custom ops are enabled when running without Inductor
                and disabled when running with Inductor (compile_level >= Inductor).
2092
        - splitting_ops: a list of ops to split the full graph into subgraphs, used in piecewise compilation.
2093
2094
2095
2096
    - CudaGraph capture:
        - use_cudagraph: whether to use cudagraph inside compilation.
            - False: cudagraph inside compilation is not used.
            - True: cudagraph inside compilation is used. It requires
2097
2098
2099
2100
                that all input buffers have fixed addresses, and all
                splitting ops write their outputs to input buffers.
            Note that this is orthogonal to the cudagraph capture logic
            outside of compilation.
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
            TODO: move outside cudagraph logic into compilation.
            torch.compile will handle cudagraph capture logic in the future.
        - cudagraph_capture_sizes: sizes to capture cudagraph.
            - None: capture sizes are inferred from compilation context.
            - List[int]: capture sizes are specified.
        - cudagraph_num_of_warmups: number of warmup runs for cudagraph.
            It means the first several runs will be treated as warmup runs.
            Only after that, the execution will be recorded, and the recorded
            cudagraph will be used for subsequent runs.
        - cudagraph_copy_inputs: whether to copy input tensors for
            cudagraph. If the caller can guarantee that the same input buffers
            are always used, it can set this to False. Otherwise, it should
            set this to True, and the compiler will copy the input to an
            internally managed buffer. Default is False.
    - Inductor compilation:
        - use_inductor: whether to use inductor compilation.
            - False: inductor compilation is not used. graph runs in eager.
            - True: inductor compilation is used. one graph for symbolic shape
                is compiled. In addition, compile for different sizes specified
                in inductor_compile_sizes, using configurations
                in inductor_compile_config.
        - inductor_compile_sizes: sizes to compile for inductor.
        - inductor_specialize_for_cudagraph_no_more_than: an optional integer
            to specialize inductor for cudagraph sizes no more than the
            specified size. It is useful when we want to specialize inductor
            with a subset of cudagraph sizes.
        - inductor_compile_config: additional configurations for inductor.
            - None: use default configurations.
        - inductor_passes: additional passes for inductor. It is a dictionary
            from pass name to pass function qualified name. We use function
            name because the config uses json format. If we pass the config
            from Python, functions can also be passed directly via Python object
            constructor, e.g. `CompilationConfig(inductor_passes={"a": func})`
2134
        - custom inductor passes: see PassConfig for more details
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
    
    Why we have different sizes for cudagraph and inductor:
    - cudagraph: a cudagraph captured for a specific size can only be used
        for the same size. We need to capture all the sizes we want to use.
    - inductor: a graph compiled by inductor for a general shape can be used
        for different sizes. Inductor can also compile for specific sizes,
        where it can have more information to optimize the graph with fully
        static shapes. However, we find the general shape compilation is
        sufficient for most cases. It might be beneficial to compile for
        certain small batchsizes, where inductor is good at optimizing.
    """ # noqa
    level: int = 0
2147
    backend: str = ""
2148
    custom_ops: List[str] = Field(default_factory=list)
2149
    splitting_ops: List[str] = Field(default_factory=lambda: [
2150
        "vllm.unified_attention",
2151
2152
        "vllm.unified_v1_flash_attention",
    ])
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164

    use_inductor: bool = True
    inductor_specialize_for_cudagraph_no_more_than: Optional[int] = None
    inductor_compile_sizes: Optional[List[int]] = Field(default_factory=dict)
    inductor_compile_config: Dict = Field(default_factory=dict)
    inductor_passes: Dict[str, str] = Field(default_factory=dict)

    use_cudagraph: bool = False
    cudagraph_num_of_warmups: int = 0
    cudagraph_capture_sizes: Optional[List[int]] = None
    cudagraph_copy_inputs: bool = False

2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
    class PassConfig(BaseModel):
        """
        Configuration for custom Inductor passes.
        This is separate from general CompilationConfig so that inductor passes
        don't all have access to full configuration - that would create a cycle
        as the PassManager is set as a property of config.
        - dump_graph_stages: list of stages for which we want to dump the graph.
            Each pass defines its own stages (before, after, maybe in-between).
        - dump_graph_dir: directory to dump the graphs. Default is .
        - enable_fusion: whether to enable the custom fusion pass.
        - enable_reshape: whether to enable the custom reshape elimination pass.
            TODO better pass enabling system.
        """
        dump_graph_stages: List[str] = Field(default_factory=list)
        dump_graph_dir: Path = Field(default=Path("."))
        enable_fusion: bool = True
        enable_reshape: bool = True

        def uuid(self):
            """
            Produces a hash unique to the pass configuration.
            Any new fields that affect compilation should be added to the hash.
            Do not include dump_graph_* in the hash - they don't affect
            compilation.
            """
            dict_ = self.model_dump(
                include={"enable_fusion", "enable_reshape"})
            encoded = json.dumps(dict_, sort_keys=True).encode("utf-8")
            return hashlib.sha256(encoded).digest()

        def model_post_init(self, __context: Any) -> None:
            if not self.enable_reshape and self.enable_fusion:
                print_warning_once(
                    "Fusion enabled but reshape elimination disabled."
                    "RMSNorm + quant (fp8) fusion might not work")

    pass_config: PassConfig = Field(default_factory=PassConfig)
2202
2203
2204
2205
2206

    # not configurable, computed after init
    compile_sizes: List[int] = PrivateAttr
    capture_sizes: List[int] = PrivateAttr

2207
2208
2209
2210
    # keep track of enabled and disabled custom ops
    enabled_custom_ops: Counter[str] = PrivateAttr
    disabled_custom_ops: Counter[str] = PrivateAttr

2211
2212
2213
2214
2215
    # Per-model forward context
    # Mainly used to store attention cls
    # Map from layer name to the attention cls
    static_forward_context: Dict[str, Any] = PrivateAttr

2216
2217
2218
2219
2220
2221
2222
    @classmethod
    def from_cli(cls, cli_value: str) -> "CompilationConfig":
        """Parse the CLI value for the compilation config."""
        if cli_value in ["0", "1", "2", "3"]:
            return cls(level=int(cli_value))
        return CompilationConfig.model_validate_json(cli_value)

2223
2224
2225
2226
2227
2228
2229
2230
2231
    def model_post_init(self, __context: Any) -> None:

        count_none = self.custom_ops.count("none")
        count_all = self.custom_ops.count("all")
        assert count_none + count_all <= 1, "Can only specify 'none' or 'all'"

        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
                assert callable(v), (
2232
2233
2234
                    f"pass {k} should be callable or a qualified name")
                self.inductor_compile_config[k] = v if isinstance(
                    v, InductorPass) else CallableInductorPass(v)
2235
2236
2237
2238
2239
2240
2241
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
2242
2243
            self.inductor_compile_config[k] = func if isinstance(
                func, InductorPass) else CallableInductorPass(func)
2244

2245
2246
        self.enabled_custom_ops = Counter()
        self.disabled_custom_ops = Counter()
2247
        self.static_forward_context = {}
2248

2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
    def init_backend(self) -> Union[str, Callable]:
        if self.level == CompilationLevel.NO_COMPILATION:
            raise ValueError("No compilation level is set.")

        from torch._dynamo.backends.registry import list_backends
        torch_backends = list_backends(exclude_tags=tuple())
        if self.level in [
                CompilationLevel.DYNAMO_AS_IS, CompilationLevel.DYNAMO_ONCE
        ]:
            if self.backend == "":
                return "eager"
            if self.backend in torch_backends:
                return self.backend
            return resolve_obj_by_qualname(self.backend)

        # TODO: pass user-specified backend to piecewise compilation
        # merge with the config use_inductor
        assert self.level == CompilationLevel.PIECEWISE
        from vllm.compilation.backends import VllmBackend
        return VllmBackend(self)

2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
    def init_during_runtime(self):
        """To complete the initialization of config,
        we need to know the compile context, which is only available
        during the first run of the model.
        """
        from vllm.compilation.compile_context import get_compile_context
        context = get_compile_context()
        context = copy.deepcopy(context) if context is not None else []
        sizes_to_specialize: List[int] = context
        if self.cudagraph_capture_sizes is None:
            self.capture_sizes = sizes_to_specialize
        else:
            self.capture_sizes = self.cudagraph_capture_sizes
            logger.info(("cudagraph sizes specified by model runner"
                         " %s is overridden by config %s"),
                        sizes_to_specialize, self.cudagraph_capture_sizes)
        if self.inductor_specialize_for_cudagraph_no_more_than is not None:
            assert self.inductor_compile_sizes is None, (
                "inductor_compile_sizes should be None when "
                "inductor_specialize_for_cudagraph_no_more_than is not None")
            self.compile_sizes = [
                x for x in self.capture_sizes
                if x <= self.inductor_specialize_for_cudagraph_no_more_than
            ]
        else:
            assert self.inductor_compile_sizes is not None, (
                "inductor_compile_sizes should not be None when "
                "inductor_specialize_for_cudagraph_no_more_than is None")
            self.compile_sizes = self.inductor_compile_sizes


2301
2302
2303
@dataclass
class VllmConfig:
    """Dataclass which contains all vllm-related configuration. This
2304
2305
2306
    simplifies passing around the distinct configurations in the codebase.
    """

2307
2308
    model_config: ModelConfig = field(default=None, init=True)  # type: ignore
    cache_config: CacheConfig = field(default=None, init=True)  # type: ignore
2309
2310
2311
2312
    parallel_config: ParallelConfig = field(default_factory=ParallelConfig,
                                            init=True)
    scheduler_config: SchedulerConfig = field(default_factory=SchedulerConfig,
                                              init=True)
2313
2314
2315
    device_config: DeviceConfig = field(default=None,
                                        init=True)  # type: ignore
    load_config: LoadConfig = field(default=None, init=True)  # type: ignore
2316
2317
2318
2319
2320
    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
2321
    quant_config: Optional[QuantizationConfig] = None
2322
2323
    compilation_config: CompilationConfig = field(default=None,
                                                  init=True)  # type: ignore
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351

    @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
2352

2353
2354
2355
2356
2357
2358
    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)

2359
2360
2361
    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
2362
2363
2364
2365
2366
2367
2368
2369
        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)
2370
2371
2372
2373
2374

        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
2375
2376
2377
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
2378
2379
2380
2381
2382

        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)
2383

2384
        if self.compilation_config is None:
2385
            self.compilation_config = CompilationConfig()
2386
        if envs.VLLM_USE_V1 and not self.model_config.enforce_eager:
2387
2388
2389
2390
2391
2392
2393
            # NOTE(woosuk): Currently, we use inductor because the piecewise
            # CUDA graphs do not work properly with the custom CUDA kernels.
            # FIXME(woosuk): Disable inductor to reduce the compilation time
            # and avoid any potential issues with the inductor.
            self.compilation_config.custom_ops = ["none"]
            self.compilation_config.use_cudagraph = True
            self.compilation_config.use_inductor = True
2394
2395
            self.compilation_config.pass_config.enable_fusion = False
            self.compilation_config.pass_config.enable_reshape = False
2396
            self.compilation_config.level = CompilationLevel.PIECEWISE
2397

2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
        if self.cache_config is not None and \
            self.cache_config.cpu_offload_gb > 0 and \
            self.compilation_config.level != CompilationLevel.NO_COMPILATION:
            logger.warning(
                "CPU offload is not supported with `torch.compile` yet."
                " Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

        if self.lora_config is not None and self.compilation_config.level !=\
             CompilationLevel.NO_COMPILATION:
            logger.warning("LoRA is not supported with `torch.compile` yet. "
                           "Disabling `torch.compile`.")
            self.compilation_config.level = CompilationLevel.NO_COMPILATION

2412
2413
        current_platform.check_and_update_config(self)

2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
    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,
2453
        self.model_config.mm_processor_kwargs)
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503


_current_vllm_config: Optional[VllmConfig] = None


@contextmanager
def set_current_vllm_config(vllm_config: VllmConfig):
    """
    Temporarily set the current VLLM config.
    Used during model initialization.
    We save the current VLLM config in a global variable,
    so that all modules can access it, e.g. custom ops
    can access the VLLM config to determine how to dispatch.
    """
    global _current_vllm_config
    old_vllm_config = _current_vllm_config
    from vllm.compilation.counter import compilation_counter
    num_models_seen = compilation_counter.num_models_seen
    try:
        _current_vllm_config = vllm_config
        yield
    finally:
        logger.debug("enabled custom ops: %s",
                     vllm_config.compilation_config.enabled_custom_ops)
        logger.debug("disabled custom ops: %s",
                     vllm_config.compilation_config.disabled_custom_ops)
        if vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
            and compilation_counter.num_models_seen == num_models_seen:
            # If the model supports compilation,
            # compilation_counter.num_models_seen should be increased
            # by at least 1.
            # If it is not increased, it means the model does not support
            # compilation (does not have @support_torch_compile decorator).
            logger.warning(
                "`torch.compile` is turned on, but the model %s"
                " does not support it. Please open an issue on GitHub"
                "if you want it to be supported.",
                vllm_config.model_config.model)
        _current_vllm_config = old_vllm_config


def get_current_vllm_config() -> VllmConfig:
    if _current_vllm_config is None:
        # in ci, usually when we test custom ops/modules directly,
        # we don't set the vllm config. In that case, we set a default
        # config.
        logger.warning("Current VLLM config is not set.")
        from vllm.config import VllmConfig
        return VllmConfig()
    return _current_vllm_config