model.py 85.6 KB
Newer Older
1
2
3
4
5
6
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import hashlib
import json
import warnings
7
from collections.abc import Callable
8
9
from dataclasses import InitVar, field
from importlib.util import find_spec
10
from typing import TYPE_CHECKING, Any, Literal, cast, get_args
11
12

import torch
13
from pydantic import ConfigDict, SkipValidation, field_validator, model_validator
14
15
16
17
from pydantic.dataclasses import dataclass
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE

import vllm.envs as envs
18
from vllm.config.multimodal import MMCacheType, MMEncoderTPMode, MultiModalConfig
19
from vllm.config.pooler import PoolerConfig
20
from vllm.config.scheduler import RunnerType
21
from vllm.config.utils import assert_hashable, config, getattr_iter
22
23
24
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.transformers_utils.config import (
25
26
27
28
29
30
31
    ConfigFormat,
    get_config,
    get_hf_image_processor_config,
    get_hf_text_config,
    get_pooling_config,
    get_sentence_transformer_tokenizer_config,
    is_encoder_decoder,
32
    try_get_dense_modules,
33
34
35
36
37
38
    try_get_generation_config,
    try_get_safetensors_metadata,
    try_get_tokenizer_config,
    uses_mrope,
)
from vllm.transformers_utils.runai_utils import ObjectStorageModel, is_runai_obj_uri
39
from vllm.transformers_utils.utils import maybe_model_redirect
40
from vllm.utils.import_utils import LazyLoader
41
from vllm.utils.torch_utils import common_broadcastable_dtype
42
43
44
45
46
47

if TYPE_CHECKING:
    from transformers import PretrainedConfig

    import vllm.model_executor.layers.quantization as me_quant
    import vllm.model_executor.models as me_models
48
    from vllm.attention.backends.registry import _Backend
49
50
51
52
53
54
55
    from vllm.config.load import LoadConfig
    from vllm.config.parallel import ParallelConfig
    from vllm.model_executor.layers.quantization import QuantizationMethods
    from vllm.v1.sample.logits_processor import LogitsProcessor
else:
    PretrainedConfig = Any

56
    _Backend = Any
57
58
59
60
    me_quant = LazyLoader(
        "model_executor", globals(), "vllm.model_executor.layers.quantization"
    )
    me_models = LazyLoader("model_executor", globals(), "vllm.model_executor.models")
61
62
63
64
65
66
67
    LoadConfig = Any
    ParallelConfig = Any
    QuantizationMethods = Any
    LogitsProcessor = Any

logger = init_logger(__name__)

68
RunnerOption = Literal["auto", RunnerType]
69
70
ConvertType = Literal["none", "embed", "classify", "reward"]
ConvertOption = Literal["auto", ConvertType]
71
72
73
74
75
76
77
78
79
80
81
TaskOption = Literal[
    "auto",
    "generate",
    "embedding",
    "embed",
    "classify",
    "score",
    "reward",
    "transcription",
    "draft",
]
82
83
TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
84
85
86
LogprobsMode = Literal[
    "raw_logits", "raw_logprobs", "processed_logits", "processed_logprobs"
]
87
HfOverrides = dict[str, Any] | Callable[[PretrainedConfig], PretrainedConfig]
88
ModelImpl = Literal["auto", "vllm", "transformers", "terratorch"]
89
LayerBlockType = Literal["attention", "linear_attention", "mamba"]
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119

_RUNNER_TASKS: dict[RunnerType, list[TaskOption]] = {
    "generate": ["generate", "transcription"],
    "pooling": ["embedding", "embed", "classify", "score", "reward"],
    "draft": ["draft"],
}

_RUNNER_CONVERTS: dict[RunnerType, list[ConvertType]] = {
    "generate": [],
    "pooling": ["embed", "classify", "reward"],
    "draft": [],
}


@config
@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
class ModelConfig:
    """Configuration for the model."""

    model: str = "Qwen/Qwen3-0.6B"
    """Name or path of the Hugging Face model to use. It is also used as the
    content for `model_name` tag in metrics output when `served_model_name` is
    not specified."""
    runner: RunnerOption = "auto"
    """The type of model runner to use. Each vLLM instance only supports one
    model runner, even if the same model can be used for multiple types."""
    convert: ConvertOption = "auto"
    """Convert the model using adapters defined in
    [vllm.model_executor.models.adapters][]. The most common use case is to
    adapt a text generation model to be used for pooling tasks."""
120
    task: TaskOption | None = None
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
    """[DEPRECATED] The task to use the model for. If the model supports more
    than one model runner, this is used to select which model runner to run.

    Note that the model may support other tasks using the same model runner.
    """
    tokenizer: SkipValidation[str] = None  # type: ignore
    """Name or path of the Hugging Face tokenizer to use. If unspecified, model
    name or path will be used."""
    tokenizer_mode: TokenizerMode = "auto"
    """Tokenizer mode:\n
    - "auto" will use the fast tokenizer if available.\n
    - "slow" will always use the slow tokenizer.\n
    - "mistral" will always use the tokenizer from `mistral_common`.\n
    - "custom" will use --tokenizer to select the preregistered tokenizer."""
    trust_remote_code: bool = False
    """Trust remote code (e.g., from HuggingFace) when downloading the model
    and tokenizer."""
138
    dtype: ModelDType | torch.dtype = "auto"
139
140
141
142
143
144
145
146
    """Data type for model weights and activations:\n
    - "auto" will use FP16 precision for FP32 and FP16 models, and BF16
    precision for BF16 models.\n
    - "half" for FP16. Recommended for AWQ quantization.\n
    - "float16" is the same as "half".\n
    - "bfloat16" for a balance between precision and range.\n
    - "float" is shorthand for FP32 precision.\n
    - "float32" for FP32 precision."""
147
    seed: int | None = None
148
149
    """Random seed for reproducibility. Initialized to None in V0, but
    initialized to 0 in V1."""
150
151
152
153
    hf_config: PretrainedConfig = field(init=False)
    """The Hugging Face config of the model."""
    hf_text_config: PretrainedConfig = field(init=False)
    """The Hugging Face config of the text model (same as hf_config for text models)."""
154
    hf_config_path: str | None = None
155
156
157
158
159
160
    """Name or path of the Hugging Face config to use. If unspecified, model
    name or path will be used."""
    allowed_local_media_path: str = ""
    """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."""
161
    allowed_media_domains: list[str] | None = None
162
    """If set, only media URLs that belong to this domain can be used for
163
    multi-modal inputs. """
164
    revision: str | None = None
165
166
    """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."""
167
    code_revision: str | None = None
168
169
170
171
172
173
    """The specific revision to use for the model code on the Hugging Face Hub.
    It can be a branch name, a tag name, or a commit id. If unspecified, will
    use the default version."""
    rope_scaling: dict[str, Any] = field(default_factory=dict)
    """RoPE scaling configuration. For example,
    `{"rope_type":"dynamic","factor":2.0}`."""
174
    rope_theta: float | None = None
175
176
    """RoPE theta. Use with `rope_scaling`. In some cases, changing the RoPE
    theta improves the performance of the scaled model."""
177
    tokenizer_revision: str | None = None
178
179
180
181
182
183
184
185
186
187
188
189
    """The specific revision to use for the tokenizer on the Hugging Face Hub.
    It can be a branch name, a tag name, or a commit id. If unspecified, will
    use the default version."""
    max_model_len: SkipValidation[int] = None  # type: ignore
    """Model context length (prompt and output). If unspecified, will be
    automatically derived from the model config.

    When passing via `--max-model-len`, supports k/m/g/K/M/G in human-readable
    format. Examples:\n
    - 1k -> 1000\n
    - 1K -> 1024\n
    - 25.6k -> 25,600"""
190
    spec_target_max_model_len: int | None = None
191
    """Specify the maximum length for spec decoding draft models."""
192
    quantization: SkipValidation[QuantizationMethods | None] = None
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
    """Method used to quantize the weights. If `None`, we first check the
    `quantization_config` attribute in the model config file. If that is
    `None`, we assume the model weights are not quantized and use `dtype` to
    determine the data type of the weights."""
    enforce_eager: bool = False
    """Whether to always use eager-mode PyTorch. 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 for maximal performance and
    flexibility."""
    max_logprobs: int = 20
    """Maximum number of log probabilities to return when `logprobs` is
    specified in `SamplingParams`. The default value comes the default for the
    OpenAI Chat Completions API. -1 means no cap, i.e. all (output_length *
    vocab_size) logprobs are allowed to be returned and it may cause OOM."""
    logprobs_mode: LogprobsMode = "raw_logprobs"
    """Indicates the content returned in the logprobs and prompt_logprobs.
    Supported mode:
    1) raw_logprobs, 2) processed_logprobs, 3) raw_logits, 4) processed_logits.
    Raw means the values before applying any logit processors, like bad words.
    Processed means the values after applying all processors, including
    temperature and top_k/top_p.
    """
    disable_sliding_window: bool = False
    """Whether to disable sliding window. If True, we will disable the sliding
    window functionality of the model, capping to sliding window size. If the
    model does not support sliding window, this argument is ignored."""
    disable_cascade_attn: bool = False
    """Disable cascade attention for V1. While cascade attention does not
    change the mathematical correctness, disabling it could be useful for
    preventing potential numerical issues. Note that even if this is set to
    False, cascade attention will be only used when the heuristic tells that
    it's beneficial."""
    skip_tokenizer_init: bool = False
    """Skip initialization of tokenizer and detokenizer. Expects valid
    `prompt_token_ids` and `None` for prompt from the input. The generated
    output will contain token ids."""
    enable_prompt_embeds: bool = False
    """If `True`, enables passing text embeddings as inputs via the
231
232
233
234
    `prompt_embeds` key.

    WARNING: The vLLM engine may crash if incorrect shape of embeddings is passed.
    Only enable this flag for trusted users!"""
235
    served_model_name: str | list[str] | None = None
236
237
238
239
240
241
242
    """The model name(s) used in the API. If multiple names are provided, the
    server will respond to any of the provided names. The model name in the
    model field of a response will be the first name in this list. If not
    specified, the model name will be the same as the `--model` argument. Noted
    that this name(s) will also be used in `model_name` tag content of
    prometheus metrics, if multiple names provided, metrics tag will take the
    first one."""
243
    config_format: str | ConfigFormat = "auto"
244
245
246
247
248
    """The format of the model config to load:\n
    - "auto" will try to load the config in hf format if available else it
    will try to load in mistral format.\n
    - "hf" will load the config in hf format.\n
    - "mistral" will load the config in mistral format."""
249
    hf_token: bool | str | None = None
250
251
252
253
254
255
    """The token to use as HTTP bearer authorization for remote files . If
    `True`, will use the token generated when running `huggingface-cli login`
    (stored in `~/.huggingface`)."""
    hf_overrides: HfOverrides = field(default_factory=dict)
    """If a dictionary, contains arguments to be forwarded to the Hugging Face
    config. If a callable, it is called to update the HuggingFace config."""
256
    logits_processor_pattern: str | None = None
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
    """Optional regex pattern specifying valid logits processor qualified names
    that can be passed with the `logits_processors` extra completion argument.
    Defaults to `None`, which allows no processors."""
    generation_config: str = "auto"
    """The folder path to the generation config. Defaults to `"auto"`, the
    generation config will be loaded from model path. If set to `"vllm"`, no
    generation config is loaded, vLLM defaults will be used. If set to a folder
    path, the generation config will be loaded from the specified folder path.
    If `max_new_tokens` is specified in generation config, then it sets a
    server-wide limit on the number of output tokens for all requests."""
    override_generation_config: dict[str, Any] = field(default_factory=dict)
    """Overrides or sets generation config. e.g. `{"temperature": 0.5}`. If
    used with `--generation-config auto`, the override parameters will be
    merged with the default config from the model. If used with
    `--generation-config vllm`, only the override parameters are used."""
    enable_sleep_mode: bool = False
    """Enable sleep mode for the engine (only cuda platform is supported)."""
274
    model_impl: str | ModelImpl = "auto"
275
276
277
278
279
280
281
282
    """Which implementation of the model to use:\n
    - "auto" will try to use the vLLM implementation, if it exists, and fall
    back to the Transformers implementation if no vLLM implementation is
    available.\n
    - "vllm" will use the vLLM model implementation.\n
    - "transformers" will use the Transformers model implementation.\n
    - "terratorch" will use the TerraTorch model implementation.
    """
283
    override_attention_dtype: str | None = None
284
    """Override dtype for attention"""
285
    logits_processors: list[str | type[LogitsProcessor]] | None = None
286
287
    """One or more logits processors' fully-qualified class names or class
    definitions"""
288
    io_processor_plugin: str | None = None
289
290
291
    """IOProcessor plugin name to load at model startup"""

    # Pooler config
292
    pooler_config: PoolerConfig | None = None
293
294
    """Pooler config which controls the behaviour of output pooling in pooling
    models."""
295
    override_pooler_config: dict | PoolerConfig | None = None
296
297
298
299
    """[DEPRECATED] Use `pooler_config` instead. This field will be removed in
    v0.12.0 or v1.0.0, whichever is sooner."""

    # Multimodal config and init vars
300
    multimodal_config: MultiModalConfig | None = None
301
302
    """Configuration for multimodal model. If `None`, this will be inferred
    from the architecture of `self.model`."""
303
    limit_mm_per_prompt: InitVar[dict[str, int | dict[str, int]] | None] = None
304
    enable_mm_embeds: InitVar[bool | None] = None
305
306
307
308
309
310
    media_io_kwargs: InitVar[dict[str, dict[str, Any]] | None] = None
    mm_processor_kwargs: InitVar[dict[str, Any] | None] = None
    mm_processor_cache_gb: InitVar[float | None] = None
    mm_processor_cache_type: InitVar[MMCacheType | None] = None
    mm_shm_cache_max_object_size_mb: InitVar[int | None] = None
    mm_encoder_tp_mode: InitVar[MMEncoderTPMode | None] = None
311
    mm_encoder_attn_backend: InitVar[_Backend | str | None] = None
312
313
314
    interleave_mm_strings: InitVar[bool | None] = None
    skip_mm_profiling: InitVar[bool | None] = None
    video_pruning_rate: InitVar[float | None] = None
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342

    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: list[Any] = []
        factors.append(self.model)
        factors.append(self.dtype)
        factors.append(self.quantization)
        factors.append(self.revision)
        factors.append(self.code_revision)
        factors.append(self.max_model_len)
        factors.append(self.max_logprobs)
        factors.append(self.disable_sliding_window)
        factors.append(self.trust_remote_code)
        factors.append(self.generation_config)
        factors.append(self.model_impl)
        factors.append(self.override_generation_config)
        factors.append(self.rope_scaling)
        factors.append(self.rope_theta)
343
        factors.append(self.video_pruning_rate)
344
        factors.append(self.enable_prompt_embeds)
345

346
        # hf_config can control how the model looks!
347
348
349
350
351
352
353
354
        try:
            hf_config_json = self.hf_config.to_json_string(use_diff=False)
        except TypeError:
            from transformers import PretrainedConfig

            from vllm.utils.jsontree import json_map_leaves

            # Handle nested HF configs with unserializable values gracefully
355
356
357
358
359
360
361
362
363
364
365
366
367
            hf_config_json = (
                json.dumps(
                    json_map_leaves(
                        lambda v: v.to_dict()
                        if isinstance(v, PretrainedConfig)
                        else str(v),
                        self.hf_config.to_dict(),
                    ),
                    indent=2,
                    sort_keys=True,
                )
                + "\n"
            )
368
369
370

        factors.append(hf_config_json)

371
372
373
374
        str_factors = str(factors)
        assert_hashable(str_factors)
        return hashlib.sha256(str(factors).encode()).hexdigest()

375
376
    def _update_nested(
        self,
377
        target: PretrainedConfig | dict[str, Any],
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
        updates: dict[str, Any],
    ) -> None:
        """Recursively updates a config or dict with nested updates."""
        for key, value in updates.items():
            if isinstance(value, dict):
                # Get the nested target
                if isinstance(target, dict):
                    nested_target = target.get(key)
                else:
                    nested_target = getattr(target, key, None)

                # If nested target exists and can be updated recursively
                if nested_target is not None and (
                    isinstance(nested_target, dict)
                    or hasattr(nested_target, "__dict__")
                ):
                    self._update_nested(nested_target, value)
                    continue

            # Set the value (base case)
            if isinstance(target, dict):
                target[key] = value
            else:
                setattr(target, key, value)

    def _apply_dict_overrides(
        self,
405
        config: PretrainedConfig,
406
407
408
409
410
411
412
413
414
415
416
417
418
419
        overrides: dict[str, Any],
    ) -> None:
        """Apply dict overrides, handling both nested configs and dict values."""
        from transformers import PretrainedConfig

        for key, value in overrides.items():
            attr = getattr(config, key, None)
            if attr is not None and isinstance(attr, PretrainedConfig):
                # It's a nested config - recursively update it
                self._update_nested(attr, value)
            else:
                # It's a dict-valued parameter - set it directly
                setattr(config, key, value)

420
    def __post_init__(
421
422
        self,
        # Multimodal config init vars
423
        limit_mm_per_prompt: dict[str, int] | None,
424
        enable_mm_embeds: bool | None,
425
426
427
428
429
430
        media_io_kwargs: dict[str, dict[str, Any]] | None,
        mm_processor_kwargs: dict[str, Any] | None,
        mm_processor_cache_gb: float | None,
        mm_processor_cache_type: MMCacheType | None,
        mm_shm_cache_max_object_size_mb: int | None,
        mm_encoder_tp_mode: MMEncoderTPMode | None,
431
        mm_encoder_attn_backend: _Backend | str | None,
432
433
434
        interleave_mm_strings: bool | None,
        skip_mm_profiling: bool | None,
        video_pruning_rate: float | None,
435
    ) -> None:
436
        # Set the default seed to 0 in V1.
437
        # NOTE(woosuk): In V1, we use separate processes for workers (unless
438
439
440
441
        # VLLM_ENABLE_V1_MULTIPROCESSING=0), so setting a seed here
        # doesn't affect the user process. However, without a consistent seed,
        # different tensor parallel workers would sample different tokens,
        # leading to inconsistent results.
442
        if self.seed is None:
443
444
445
446
447
448
            self.seed = 0
            if not envs.VLLM_ENABLE_V1_MULTIPROCESSING:
                logger.warning(
                    "The global random seed is set to %d. Since "
                    "VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may "
                    "affect the random state of the Python process that "
449
450
451
                    "launched vLLM.",
                    self.seed,
                )
452
453

        # Keep set served_model_name before maybe_model_redirect(self.model)
454
455
456
        self.served_model_name = get_served_model_name(
            self.model, self.served_model_name
        )
457
458
459
460
461
462
463
464
465
466
467
468
469
470
        self.model = maybe_model_redirect(self.model)
        # The tokenizer is consistent with the model by default.
        if self.tokenizer is None:
            self.tokenizer = self.model
        if self.tokenizer_revision is None:
            self.tokenizer_revision = self.revision
        self.tokenizer = maybe_model_redirect(self.tokenizer)

        if isinstance(self.hf_config_path, str):
            self.hf_config_path = maybe_model_redirect(self.hf_config_path)

        if callable(self.hf_overrides):
            hf_overrides_kw = {}
            hf_overrides_fn = self.hf_overrides
471
            dict_overrides: dict[str, Any] = {}
472
        else:
473
474
475
476
477
478
479
480
481
            # Separate dict overrides from flat ones
            # We'll determine how to apply dict overrides after loading the config
            hf_overrides_kw = {}
            dict_overrides = {}
            for key, value in self.hf_overrides.items():
                if isinstance(value, dict):
                    dict_overrides[key] = value
                else:
                    hf_overrides_kw[key] = value
482
483
484
485
486
487
488
489
            hf_overrides_fn = None

        if self.rope_scaling:
            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
            hf_overrides_kw.update(hf_override)
            hf_overrides_str = json.dumps(hf_overrides_kw)
            msg = (
                "`--rope-scaling` will be removed in a future release. "
490
491
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`"
            )
492
493
494
495
496
497
498
            warnings.warn(DeprecationWarning(msg), stacklevel=2)
        if self.rope_theta is not None:
            hf_override = {"rope_theta": self.rope_theta}
            hf_overrides_kw.update(hf_override)
            hf_overrides_str = json.dumps(hf_overrides_kw)
            msg = (
                "`--rope-theta` will be removed in a future release. "
499
500
                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`"
            )
501
502
503
504
            warnings.warn(DeprecationWarning(msg), stacklevel=2)

        self.maybe_pull_model_tokenizer_for_runai(self.model, self.tokenizer)

505
506
507
508
509
        if (
            (backend := envs.VLLM_ATTENTION_BACKEND)
            and backend == "FLASHINFER"
            and find_spec("flashinfer") is None
        ):
510
511
512
513
            raise ValueError(
                "VLLM_ATTENTION_BACKEND is set to FLASHINFER, but flashinfer "
                "module was not found. See "
                "https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile "  # noqa: E501
514
515
                "for instructions on how to install it."
            )
516
517
518

        from vllm.platforms import current_platform

519
        if self.override_attention_dtype is not None and not current_platform.is_rocm():
520
521
            warnings.warn(
                "override-attention-dtype is set but not using ROCm platform",
522
523
                stacklevel=2,
            )
524

525
526
527
528
529
530
531
532
533
534
535
536
        if self.enable_sleep_mode and not current_platform.is_sleep_mode_available():
            raise ValueError("Sleep mode is not supported on current platform.")

        hf_config = get_config(
            self.hf_config_path or self.model,
            self.trust_remote_code,
            self.revision,
            self.code_revision,
            self.config_format,
            hf_overrides_kw=hf_overrides_kw,
            hf_overrides_fn=hf_overrides_fn,
        )
537
538

        self.hf_config = hf_config
539
540
        if dict_overrides:
            self._apply_dict_overrides(hf_config, dict_overrides)
541
        self.hf_text_config = get_hf_text_config(self.hf_config)
542
543
544
        self.attention_chunk_size = getattr(
            self.hf_text_config, "attention_chunk_size", None
        )
545
546
        self.encoder_config = self._get_encoder_config()
        self.hf_image_processor_config = get_hf_image_processor_config(
547
548
            self.model, hf_token=self.hf_token, revision=self.revision
        )
549
550
551

        architectures = self.architectures
        registry = self.registry
552
        is_generative_model = registry.is_text_generation_model(architectures, self)
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
        is_pooling_model = registry.is_pooling_model(architectures, self)

        def _task_to_convert(task: TaskOption) -> ConvertType:
            if task == "embedding" or task == "embed":
                return "embed"
            if task == "classify":
                return "classify"
            if task == "reward":
                return "reward"
            if task == "score":
                new_task = self._get_default_pooling_task(architectures)
                return "classify" if new_task == "classify" else "embed"

            return "none"

        if self.task is not None:
            runner: RunnerOption = "auto"
            convert: ConvertOption = "auto"
571
572
573
574
            msg_prefix = (
                "The 'task' option has been deprecated and will be "
                "removed in v0.13.0 or v1.0, whichever comes first."
            )
575
576
577
578
579
580
581
582
583
            msg_hint = "Please remove this option."

            is_generative_task = self.task in _RUNNER_TASKS["generate"]
            is_pooling_task = self.task in _RUNNER_TASKS["pooling"]

            if is_generative_model and is_pooling_model:
                if is_generative_task:
                    runner = "generate"
                    convert = "auto"
584
585
586
587
588
                    msg_hint = (
                        "Please replace this option with `--runner "
                        "generate` to continue using this model "
                        "as a generative model."
                    )
589
590
591
                elif is_pooling_task:
                    runner = "pooling"
                    convert = "auto"
592
593
594
595
596
                    msg_hint = (
                        "Please replace this option with `--runner "
                        "pooling` to continue using this model "
                        "as a pooling model."
                    )
597
598
599
600
601
602
603
604
605
606
                else:  # task == "auto"
                    pass
            elif is_generative_model or is_pooling_model:
                if is_generative_task:
                    runner = "generate"
                    convert = "auto"
                    msg_hint = "Please remove this option"
                elif is_pooling_task:
                    runner = "pooling"
                    convert = _task_to_convert(self.task)
607
608
609
610
611
                    msg_hint = (
                        "Please replace this option with `--convert "
                        f"{convert}` to continue using this model "
                        "as a pooling model."
                    )
612
613
614
                else:  # task == "auto"
                    pass
            else:
615
616
617
618
619
                debug_info = {
                    "architectures": architectures,
                    "is_generative_model": is_generative_model,
                    "is_pooling_model": is_pooling_model,
                }
620
621
622
623
624
                raise AssertionError(
                    "The model should be a generative or "
                    "pooling model when task is set to "
                    f"{self.task!r}. Found: {debug_info}"
                )
625
626
627
628
629
630
631
632

            self.runner = runner
            self.convert = convert

            msg = f"{msg_prefix} {msg_hint}"
            warnings.warn(msg, DeprecationWarning, stacklevel=2)

        self.runner_type = self._get_runner_type(architectures, self.runner)
633
634
635
        self.convert_type = self._get_convert_type(
            architectures, self.runner_type, self.convert
        )
636
637
638
639
640

        if self.runner_type == "generate" and not is_generative_model:
            generate_converts = _RUNNER_CONVERTS["generate"]
            if self.convert_type not in generate_converts:
                # Currently we don't have any converters for generative models
641
                raise ValueError("This model does not support `--runner generate`.")
642
643
644
645
646
647
648
        if self.runner_type == "pooling" and not is_pooling_model:
            pooling_converts = _RUNNER_CONVERTS["pooling"]
            if self.convert_type not in pooling_converts:
                convert_option = "<" + "|".join(pooling_converts) + ">"
                raise ValueError(
                    "This model does not support `--runner pooling`. "
                    f"You can pass `--convert {convert_option} to adapt "
649
650
                    "it into a pooling model."
                )
651
652
653
654
655
656
657
658
659
660
661
662
663
664

        # Note: Initialize these attributes early because transformers fallback
        # may fail to load dynamic modules in child processes
        model_info, arch = registry.inspect_model_cls(architectures, self)
        self._model_info = model_info
        self._architecture = arch
        logger.info("Resolved architecture: %s", arch)

        # Init pooler config if needed
        if self.runner_type == "pooling":
            if self.override_pooler_config is not None:
                logger.warning_once(
                    "`override_pooler_config` is deprecated and will be "
                    "removed in v0.12.0 or v1.0.0, whichever is sooner. "
665
666
                    "Please use `pooler_config` instead."
                )
667
668

                if isinstance(self.override_pooler_config, dict):
669
                    self.pooler_config = PoolerConfig(**self.override_pooler_config)
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
                else:
                    self.pooler_config = self.override_pooler_config

            if self.pooler_config is None:
                self.pooler_config = 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(self.pooler_config, k) is None:
                        setattr(self.pooler_config, k, v)

            default_pooling_type = self._model_info.default_pooling_type
            if self.pooler_config.pooling_type is None:
                self.pooler_config.pooling_type = default_pooling_type

        self.dtype: torch.dtype = _get_and_verify_dtype(
            self.model,
            self.hf_config,
            self.dtype,
            is_pooling_model=self.runner_type == "pooling",
            revision=self.revision,
        )

        self.original_max_model_len = self.max_model_len
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
        # Init multimodal config if needed
        if self._model_info.supports_multimodal:
699
700
701
702
            if (
                mm_encoder_tp_mode == "data"
                and not self._model_info.supports_multimodal_encoder_tp_data
            ):
703
704
                logger.warning_once(
                    "This model does not support `--mm-encoder-tp-mode data`. "
705
706
                    "Falling back to `--mm-encoder-tp-mode weights`."
                )
707
708
709
710
                mm_encoder_tp_mode = "weights"

            mm_config_kwargs = dict(
                limit_per_prompt=limit_mm_per_prompt,
711
                enable_mm_embeds=enable_mm_embeds,
712
713
714
715
716
717
                media_io_kwargs=media_io_kwargs,
                mm_processor_kwargs=mm_processor_kwargs,
                mm_processor_cache_gb=mm_processor_cache_gb,
                mm_processor_cache_type=mm_processor_cache_type,
                mm_shm_cache_max_object_size_mb=mm_shm_cache_max_object_size_mb,
                mm_encoder_tp_mode=mm_encoder_tp_mode,
718
                mm_encoder_attn_backend=mm_encoder_attn_backend,
719
720
                interleave_mm_strings=interleave_mm_strings,
                skip_mm_profiling=skip_mm_profiling,
721
                video_pruning_rate=video_pruning_rate,
722
723
724
            )

            mm_config_kwargs = {
725
                k: v for k, v in mm_config_kwargs.items() if v is not None
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
            }

            self.multimodal_config = MultiModalConfig(**mm_config_kwargs)

        if self.disable_sliding_window:
            # Set after get_and_verify_max_len to ensure that max_model_len
            # can be correctly capped to sliding window size
            self.hf_text_config.sliding_window = None

        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()

        # Avoid running try_verify_and_update_config multiple times
        self.config_updated = False

        self._verify_quantization()
        self._verify_cuda_graph()
        self._verify_bnb_config()

    @field_validator("quantization", mode="before")
    @classmethod
    def validate_quantization_before(cls, value: Any) -> Any:
        if isinstance(value, str):
            return value.lower()
        return value

    @model_validator(mode="after")
    def validate_model_config_after(self: "ModelConfig") -> "ModelConfig":
        if not isinstance(self.tokenizer, str):
            raise ValueError("tokenizer must be a string after __post_init__.")
        if not isinstance(self.max_model_len, int):
757
            raise ValueError("max_model_len must be an integer after __post_init__.")
758
759
760
761
762
        return self

    def _get_transformers_backend_cls(self) -> str:
        """Determine which Transformers backend class will be used if
        `model_impl` is set to `transformers` or `auto`."""
763
764
765
766
        cls = "Transformers"
        # If 'hf_config != hf_text_config' it's a nested config, i.e. multimodal
        cls += "MultiModal" if self.hf_config != self.hf_text_config else ""
        cls += "MoE" if self.get_num_experts() > 1 else ""
767
768
        # Check if the architecture we're wrapping has defaults
        runner = None
769
        task = None
770
        if defaults := try_match_architecture_defaults(self.architectures[0]):
771
772
            _, (runner, task) = defaults
        # User specified value take precedence
773
774
        if self.runner != "auto":
            runner = self.runner
775
776
777
778
779
780
781
782
        # Only consider Transformers backend pooling classes if we're wrapping an
        # architecture that defaults to pooling. Otherwise, we return the LM class
        # and use adapters.
        if runner == "pooling" and task in {"embed", "classify"}:
            if task == "embed":
                cls += "EmbeddingModel"
            elif task == "classify":
                cls += "ForSequenceClassification"
783
784
785
        else:
            cls += "ForCausalLM"
        return cls
786
787
788

    def using_transformers_backend(self) -> bool:
        """Check if the model is using the Transformers backend class."""
789
790
791
        used_cls = self._model_info.architecture
        transformers_backend_cls = self._get_transformers_backend_cls()
        return used_cls == transformers_backend_cls
792
793
794
795
796
797
798
799
800
801
802
803
804
805

    @property
    def registry(self):
        return me_models.ModelRegistry

    @property
    def architectures(self) -> list[str]:
        return getattr(self.hf_config, "architectures", [])

    @property
    def architecture(self) -> str:
        """The architecture vllm actually used."""
        return self._architecture

806
    def maybe_pull_model_tokenizer_for_runai(self, model: str, tokenizer: str) -> None:
807
808
809
810
811
812
813
        """Pull model/tokenizer from Object Storage to temporary
        directory when needed.

        Args:
            model: Model name or path
            tokenizer: Tokenizer name or path
        """
814

815
816
817
818
        if not (is_runai_obj_uri(model) or is_runai_obj_uri(tokenizer)):
            return

        if is_runai_obj_uri(model):
819
            object_storage_model = ObjectStorageModel(url=model)
820
            object_storage_model.pull_files(
821
822
                model, allow_pattern=["*.model", "*.py", "*.json"]
            )
823
824
825
826
827
            self.model_weights = model
            self.model = object_storage_model.dir

            # If tokenizer is same as model, download to same directory
            if model == tokenizer:
828
829
830
831
832
833
834
835
836
837
                object_storage_model.pull_files(
                    model,
                    ignore_pattern=[
                        "*.pt",
                        "*.safetensors",
                        "*.bin",
                        "*.tensors",
                        "*.pth",
                    ],
                )
838
839
840
841
842
                self.tokenizer = object_storage_model.dir
                return

        # Only download tokenizer if needed and not already handled
        if is_runai_obj_uri(tokenizer):
843
            object_storage_tokenizer = ObjectStorageModel(url=tokenizer)
844
845
846
847
            object_storage_tokenizer.pull_files(
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors", "*.pth"],
            )
848
849
850
            self.tokenizer = object_storage_tokenizer.dir

    def _get_encoder_config(self):
851
        return get_sentence_transformer_tokenizer_config(self.model, self.revision)
852
853
854
855
856
857

    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = cast(TokenizerMode, self.tokenizer_mode.lower())
        if tokenizer_mode not in get_args(TokenizerMode):
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
858
859
                f"one of {get_args(TokenizerMode)}."
            )
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
        self.tokenizer_mode = tokenizer_mode

    def _get_default_runner_type(
        self,
        architectures: list[str],
    ) -> RunnerType:
        registry = self.registry

        # Some Sentence Transformers models use *ForCausalLM archs
        if get_pooling_config(self.model, self.revision):
            return "pooling"

        for arch in architectures:
            if arch in registry.get_supported_archs():
                if registry.is_pooling_model(architectures, self):
                    return "pooling"
                if registry.is_text_generation_model(architectures, self):
                    return "generate"

            match = try_match_architecture_defaults(arch)
            if match:
                _, (runner_type, _) = match
                return runner_type

        return "generate"

    def _get_runner_type(
        self,
        architectures: list[str],
        runner: RunnerOption,
    ) -> RunnerType:
        if runner != "auto":
            return runner

        runner_type = self._get_default_runner_type(architectures)

        # Don't log the most common case
        if runner_type != "generate":
            logger.info(
                "Resolved `--runner auto` to `--runner %s`. "
                "Pass the value explicitly to silence this message.",
901
902
                runner_type,
            )
903
904
905
906
907
908
909
910
911
912
913
914

        return runner_type

    def _get_default_convert_type(
        self,
        architectures: list[str],
        runner_type: RunnerType,
    ) -> ConvertType:
        registry = self.registry

        for arch in architectures:
            if arch in registry.get_supported_archs():
915
916
917
                if runner_type == "generate" and registry.is_text_generation_model(
                    architectures, self
                ):
918
                    return "none"
919
920
921
                if runner_type == "pooling" and registry.is_pooling_model(
                    architectures, self
                ):
922
923
                    return "none"

924
            match = try_match_architecture_defaults(arch, runner_type=runner_type)
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
            if match:
                _, (_, convert_type) = match
                return convert_type

        # This is to handle Sentence Transformers models that use *ForCausalLM
        # and also multi-modal pooling models which are not defined as
        # Sentence Transformers models
        if runner_type == "pooling":
            return "embed"

        return "none"

    def _get_convert_type(
        self,
        architectures: list[str],
        runner_type: RunnerType,
        convert: ConvertOption,
    ) -> ConvertType:
        if convert != "auto":
            return convert

946
        convert_type = self._get_default_convert_type(architectures, runner_type)
947
948
949
950
951
952

        # Don't log the most common case
        if convert_type != "none":
            logger.info(
                "Resolved `--convert auto` to `--convert %s`. "
                "Pass the value explicitly to silence this message.",
953
954
                convert_type,
            )
955
956
957
958
959
960
961
962
963
964
965

        return convert_type

    def _get_default_pooling_task(
        self,
        architectures: list[str],
    ) -> Literal["embed", "classify", "reward"]:
        if self.registry.is_cross_encoder_model(architectures, self):
            return "classify"

        for arch in architectures:
966
            match = try_match_architecture_defaults(arch, runner_type="pooling")
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
            if match:
                _, (_, convert_type) = match
                assert convert_type != "none"
                return convert_type

        return "embed"

    def _parse_quant_hf_config(self, hf_config: PretrainedConfig):
        quant_cfg = getattr(hf_config, "quantization_config", None)
        if quant_cfg is None:
            # compressed-tensors uses a "compression_config" key
            quant_cfg = getattr(hf_config, "compression_config", None)

        else:
            # Set quant_method for ModelOpt models.
            producer_name = quant_cfg.get("producer", {}).get("name")
            if producer_name == "modelopt":
984
                quant_algo = quant_cfg.get("quantization", {}).get("quant_algo")
985
986
987
988
989
                if quant_algo == "FP8":
                    quant_cfg["quant_method"] = "modelopt"
                elif quant_algo == "NVFP4":
                    quant_cfg["quant_method"] = "modelopt_fp4"
                elif quant_algo is not None:
990
                    raise ValueError(f"Unknown ModelOpt quant algo: {quant_algo}")
991
992
993
994
995
996

        return quant_cfg

    def _verify_quantization(self) -> None:
        supported_quantization = me_quant.QUANTIZATION_METHODS
        if self.quantization is not None:
997
            self.quantization = cast(me_quant.QuantizationMethods, self.quantization)
998
999
1000

        # Parse quantization method from the HF model config, if available.
        quant_cfg = self._parse_quant_hf_config(self.hf_config)
1001
1002
1003
        if quant_cfg is None and (
            text_config := getattr(self.hf_config, "text_config", None)
        ):
1004
1005
1006
1007
1008
1009
1010
1011
            # Check the text config as well for multi-modal models.
            quant_cfg = self._parse_quant_hf_config(text_config)

        if quant_cfg is not None:
            # Use the community standard 'quant_method'
            quant_method = quant_cfg.get("quant_method", "").lower()

            # Normalize library names
1012
1013
1014
            quant_method = quant_method.replace(
                "compressed_tensors", "compressed-tensors"
            )
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031

            quant_cfg["quant_method"] = quant_method

            # Quantization methods which are overrides (i.e. they have a
            # `override_quantization_method` method) must be checked in order
            # of preference (this is particularly important for GPTQ).
            overrides = [
                "bitblas",
                "gptq_marlin_24",
                "gptq_marlin",
                "gptq_bitblas",
                "awq_marlin",
                "ipex",
                "moe_wna16",
                "modelopt",
                "modelopt_fp4",
                "petit_nvfp4",
1032
1033
1034
                # Ensure heavy backends are probed last to avoid unnecessary
                # imports during override detection (e.g., MXFP4 imports Triton)
                "mxfp4",
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
            ]
            quantization_methods = [
                q for q in supported_quantization if q not in overrides
            ]
            # Any custom overrides will be in quantization_methods so we place
            # them at the start of the list so custom overrides have preference
            # over the built-in ones.
            quantization_methods = quantization_methods + overrides

            # Detect which checkpoint is it
            for name in quantization_methods:
                method = me_quant.get_quantization_config(name)
                quantization_override = method.override_quantization_method(
1048
1049
                    quant_cfg, self.quantization
                )
1050
1051
1052
1053
                if quantization_override is not None:
                    # Raise error if the override is not custom (custom would
                    # be in QUANTIZATION_METHODS but not QuantizationMethods)
                    # and hasn't been added to the overrides list.
1054
1055
1056
1057
                    if (
                        name in get_args(me_quant.QuantizationMethods)
                        and name not in overrides
                    ):
1058
1059
1060
1061
                        raise ValueError(
                            f"Quantization method {name} is an override but "
                            "is has not been added to the `overrides` list "
                            "above. This is necessary to ensure that the "
1062
1063
                            "overrides are checked in order of preference."
                        )
1064
1065
1066
1067
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break

1068
            quant_method = quant_method if quant_method != "" else None
1069
1070
1071
1072
1073
1074
1075
1076
            # Verify quantization configurations.
            if self.quantization is None:
                self.quantization = quant_method
            elif self.quantization != quant_method:
                raise ValueError(
                    "Quantization method specified in the model config "
                    f"({quant_method}) does not match the quantization "
                    f"method specified in the `quantization` argument "
1077
1078
                    f"({self.quantization})."
                )
1079
1080
1081
1082
1083

        if self.quantization is not None:
            if self.quantization not in supported_quantization:
                raise ValueError(
                    f"Unknown quantization method: {self.quantization}. Must "
1084
1085
                    f"be one of {supported_quantization}."
                )
1086
            from vllm.platforms import current_platform
1087

1088
1089
1090
1091
1092
            current_platform.verify_quantization(self.quantization)

    def _verify_cuda_graph(self) -> None:
        # CUDAGraph capture not supported for encoder-decoder models on ROCm
        unsupported_rocm = self.is_encoder_decoder
1093
        if unsupported_rocm and not self.enforce_eager and current_platform.is_rocm():
1094
1095
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1096
1097
1098
                "to eager mode.",
                self.hf_config.model_type,
            )
1099
1100
1101
1102
1103
1104
1105
1106
1107
            self.enforce_eager = True

    def _verify_bnb_config(self) -> None:
        """
        The current version of bitsandbytes (0.46.1) with 8-bit models does not
        yet support CUDA graph.
        # TODO Remove this when bitsandbytes supports.
        """
        is_bitsandbytes = self.quantization == "bitsandbytes"
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
        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(
            [
1118
1119
1120
1121
                is_bitsandbytes,
                has_quantization_config,
                is_8bit,
                not self.enforce_eager,
1122
1123
            ]
        ):
1124
1125
            logger.warning(
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1126
1127
                "fallback to the eager mode."
            )
1128
1129
1130
1131

            self.enforce_eager = True

    def _verify_with_expert_parallelism(self) -> None:
1132
        num_experts = self.get_num_experts()
1133
1134
1135
        if num_experts < 1:
            raise ValueError(
                "Number of experts in the model must be greater than 0 "
1136
1137
                "when expert parallelism is enabled."
            )
1138
1139
1140
1141
1142
1143
1144
1145

    def verify_dual_chunk_attention_config(
        self,
        load_config: LoadConfig,
    ) -> None:
        if hasattr(self.hf_config, "dual_chunk_attention_config"):
            # Try loading the sparse attention config
            from vllm.model_executor.model_loader.weight_utils import (
1146
1147
1148
                get_sparse_attention_config,
            )

1149
1150
1151
            sparse_attn_config = get_sparse_attention_config(self, load_config)
            if sparse_attn_config:
                self.hf_config.dual_chunk_attention_config[
1152
1153
1154
1155
1156
1157
                    "sparse_attention_config"
                ] = sparse_attn_config
                if (
                    "sparse_attention_enabled"
                    not in self.hf_config.dual_chunk_attention_config
                ):
1158
                    self.hf_config.dual_chunk_attention_config[
1159
1160
                        "sparse_attention_enabled"
                    ] = True
1161
1162
1163
1164
1165
1166
1167
1168

    def verify_with_parallel_config(
        self,
        parallel_config: ParallelConfig,
    ) -> None:
        if parallel_config.distributed_executor_backend == "external_launcher":
            assert self.seed is not None, (
                "Seed must be set when using external launcher backend to "
1169
1170
                "make sure sampling results are the same across workers."
            )
1171

1172
1173
1174
        total_num_attention_heads = getattr(
            self.hf_text_config, "num_attention_heads", 0
        )
1175
1176
1177
1178
1179
        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 "
1180
1181
                f"({tensor_parallel_size})."
            )
1182
1183
1184
1185
1186

        if parallel_config.enable_expert_parallel:
            self._verify_with_expert_parallelism()

        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1187
1188
1189
        if pipeline_parallel_size > 1 and not self.registry.is_pp_supported_model(
            self.architectures, self
        ):
1190
1191
            raise NotImplementedError(
                "Pipeline parallelism is not supported for this model. "
1192
1193
                "Supported models implement the `SupportsPP` interface."
            )
1194

1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
        decode_context_parallel_size = parallel_config.decode_context_parallel_size
        if decode_context_parallel_size > 1 and not self.use_mla:
            total_num_kv_heads = self.get_total_num_kv_heads()
            assert tensor_parallel_size > total_num_kv_heads, (
                f"tensor parallel size {tensor_parallel_size} must be greater "
                f"than total num kv heads {total_num_kv_heads} when enable "
                f"decode context parallel for GQA/MQA"
            )

            max_dcp_size = tensor_parallel_size // total_num_kv_heads
            assert decode_context_parallel_size <= max_dcp_size, (
                f"decode context parallel size must less than or equal to "
                f"(tensor parallel size {tensor_parallel_size} // total "
                f"num kv heads {total_num_kv_heads}) = {max_dcp_size}, "
                f"but got {decode_context_parallel_size}"
            )

1212
    def get_sliding_window(self) -> int | None:
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
        """Get the sliding window size from the HF text config if present."""
        return getattr(self.hf_text_config, "sliding_window", None)

    def get_vocab_size(self) -> int:
        return getattr(self.hf_text_config, "vocab_size", 0)

    def get_hidden_size(self) -> int:
        return getattr(self.hf_text_config, "hidden_size", 0)

    @property
    def is_deepseek_mla(self) -> bool:
        if not hasattr(self.hf_text_config, "model_type"):
            return False
1226
1227
1228
1229
1230
1231
        elif self.hf_text_config.model_type in (
            "deepseek_v2",
            "deepseek_v3",
            "deepseek_v32",
            "deepseek_mtp",
            "kimi_k2",
1232
            "kimi_linear",
1233
1234
            "longcat_flash",
        ):
1235
            return self.hf_text_config.kv_lora_rank is not None
1236
        elif self.hf_text_config.model_type == "eagle":
1237
1238
            # if the model is an EAGLE module, check for the
            # underlying architecture
1239
1240
1241
            return (
                self.hf_text_config.model.model_type
                in ("deepseek_v2", "deepseek_v3", "deepseek_v32")
1242
                and self.hf_text_config.kv_lora_rank is not None
1243
            )
1244
1245
1246
1247
1248
        return False

    def get_head_size(self) -> int:
        # TODO remove hard code
        if self.is_deepseek_mla:
1249
            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim", 0)
1250
1251
1252
            if self.use_mla:
                return self.hf_text_config.kv_lora_rank + qk_rope_head_dim
            else:
1253
                qk_nope_head_dim = getattr(self.hf_text_config, "qk_nope_head_dim", 0)
1254
1255
1256
                if qk_rope_head_dim and qk_nope_head_dim:
                    return qk_rope_head_dim + qk_nope_head_dim

1257
1258
1259
        if hasattr(self.hf_text_config, "model_type") and (
            self.hf_text_config.model_type == "zamba2"
        ):
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
            return self.hf_text_config.attention_head_dim

        if self.is_attention_free:
            return 0

        # NOTE: Some configs may set head_dim=None in the config
        if getattr(self.hf_text_config, "head_dim", None) is not None:
            return self.hf_text_config.head_dim

        # NOTE: Some models (such as PLaMo2.1) use `hidden_size_per_head`
1270
        if getattr(self.hf_text_config, "hidden_size_per_head", None) is not None:
1271
1272
1273
            return self.hf_text_config.hidden_size_per_head

        # FIXME(woosuk): This may not be true for all models.
1274
1275
1276
        return (
            self.hf_text_config.hidden_size // self.hf_text_config.num_attention_heads
        )
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286

    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
        # For GPTBigCode & Falcon:
        # NOTE: for falcon, when new_decoder_architecture is True, the
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
        new_decoder_arch_falcon = (
            self.hf_config.model_type in falcon_model_types
1287
1288
1289
1290
1291
            and getattr(self.hf_config, "new_decoder_architecture", False)
        )
        if not new_decoder_arch_falcon and getattr(
            self.hf_text_config, "multi_query", False
        ):
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
            # Multi-query attention, only one KV head.
            # Currently, tensor parallelism is not supported in this case.
            return 1

        # For DBRX and MPT
        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":
1302
1303
1304
1305
1306
            return getattr(
                self.hf_config.attn_config,
                "kv_n_heads",
                self.hf_config.num_attention_heads,
            )
1307
1308
1309
1310

        if self.hf_config.model_type == "nemotron-nas":
            for block in self.hf_config.block_configs:
                if not block.attention.no_op:
1311
1312
                    return (
                        self.hf_config.num_attention_heads
1313
                        // block.attention.n_heads_in_group
1314
                    )
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349

            raise RuntimeError("Couldn't determine number of kv heads")

        if self.is_attention_free:
            return 0

        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:
            num_kv_heads = getattr(self.hf_text_config, attr, None)
            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.
        return self.hf_text_config.num_attention_heads

    def get_num_kv_heads(self, parallel_config: ParallelConfig) -> int:
        """Returns the number of KV heads per GPU."""
        if self.use_mla:
            # When using MLA during decode it becomes MQA
            return 1

        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.
1350
        return max(1, total_num_kv_heads // parallel_config.tensor_parallel_size)
1351
1352
1353
1354
1355

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

1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
    def get_num_experts(self) -> int:
        """Returns the number of experts in the model."""
        num_expert_names = [
            "num_experts",  # Jamba
            "moe_num_experts",  # Dbrx
            "n_routed_experts",  # DeepSeek
            "num_local_experts",  # Mixtral
        ]
        num_experts = getattr_iter(self.hf_text_config, num_expert_names, 0)
        if isinstance(num_experts, list):
            # Ernie VL's remote code uses list[int]...
            # The values are always the same so we just take the first one.
            return num_experts[0]
        return num_experts

1371
    def get_layers_start_end_indices(
1372
1373
        self, parallel_config: ParallelConfig
    ) -> tuple[int, int]:
1374
        from vllm.distributed.utils import get_pp_indices
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389

        if (
            self.hf_text_config.model_type == "deepseek_mtp"
            or self.hf_config.model_type == "mimo_mtp"
            or self.hf_config.model_type == "glm4_moe_mtp"
            or self.hf_config.model_type == "ernie_mtp"
            or self.hf_config.model_type == "qwen3_next_mtp"
        ):
            total_num_hidden_layers = getattr(
                self.hf_text_config, "num_nextn_predict_layers", 0
            )
        elif self.hf_config.model_type == "longcat_flash_mtp":
            total_num_hidden_layers = getattr(
                self.hf_text_config, "num_nextn_predict_layers", 1
            )
1390
        else:
1391
1392
1393
            total_num_hidden_layers = getattr(
                self.hf_text_config, "num_hidden_layers", 0
            )
1394
        # the layout order is: DP x PP x TP
1395
1396
1397
        pp_rank = (
            parallel_config.rank // parallel_config.tensor_parallel_size
        ) % parallel_config.pipeline_parallel_size
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
        return start, end

    def get_num_layers(self, parallel_config: ParallelConfig) -> int:
        start, end = self.get_layers_start_end_indices(parallel_config)
        return end - start

    def get_num_layers_by_block_type(
        self,
        parallel_config: ParallelConfig,
1409
        block_type: LayerBlockType = "attention",
1410
1411
1412
    ) -> int:
        # This function relies on 'layers_block_type' in hf_config,
        # for w/o this attribute, we will need to have workarounds like so
1413
        attn_block_type = block_type == "attention"
1414
1415
1416
        is_transformer = (
            not self.is_hybrid and not self.has_noops and not self.is_attention_free
        )
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
        start, end = self.get_layers_start_end_indices(parallel_config)

        if is_transformer:
            # Handle the basic case first
            return end - start if attn_block_type else 0
        elif self.is_attention_free:
            # Attention free
            # Note that this code assumes there
            # is only one type of attention-free block type.
            return 0 if attn_block_type else end - start
        elif self.has_noops:
            block_configs = self.hf_config.block_configs
1429
            return sum(not bc.attention.no_op for bc in block_configs[start:end])
1430
1431
        else:
            # Hybrid model Jamba
1432
1433
1434
            layers_block_type_value = getattr(
                self.hf_text_config, "layers_block_type", None
            )
1435
            if layers_block_type_value is not None:
1436
1437
1438
                if hasattr(self.hf_text_config, "model_type") and (
                    self.hf_text_config.model_type == "zamba2"
                ):
1439
                    if attn_block_type:
1440
1441
1442
                        return sum(
                            t == "hybrid" for t in layers_block_type_value[start:end]
                        )
1443
1444
                    else:
                        return self.get_num_layers(parallel_config)
1445
                return sum(t == block_type for t in layers_block_type_value[start:end])
1446
1447
1448
1449
1450
1451
1452
1453
1454

            # Hybrid model Minimax
            attn_type_list = getattr(self.hf_config, "attn_type_list", None)
            if attn_type_list:
                return sum(t == 1 for t in attn_type_list[start:end])

            # Hybrid model Qwen3Next
            layer_types_value = getattr(self.hf_config, "layer_types", None)
            if layer_types_value is not None:
1455
                if block_type == "attention":
1456
1457
1458
                    return sum(
                        t == "full_attention" for t in layer_types_value[start:end]
                    )
1459
                elif block_type == "linear_attention":
1460
1461
1462
                    return sum(
                        t == "linear_attention" for t in layer_types_value[start:end]
                    )
1463
                else:
1464
                    return sum(t == block_type for t in layer_types_value[start:end])
1465
1466
1467
1468
1469
1470

            if (
                layers_block_type_value is None
                and attn_type_list is None
                and layer_types_value is None
            ):
1471
                raise ValueError(
1472
1473
1474
                    "The model is an hybrid without a layers_block_type or an "
                    "attn_type_list, or a layer_types in the hf_config, "
                    f"cannot determine the num of {block_type} layers"
1475
                )
1476

1477
    def get_mamba_chunk_size(self) -> int | None:
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
        """
        Returns the mamba chunk size if it exists
        """
        # used by e.g. Bamba, FalconH1, Granite, PLaMo2
        chunk_size = getattr(self.hf_text_config, "mamba_chunk_size", None)
        if chunk_size is None:
            # used by e.g. Mamba2, NemotronH, Zamba
            chunk_size = getattr(self.hf_text_config, "chunk_size", None)
        return chunk_size

    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

    def try_get_generation_config(self) -> dict[str, Any]:
        """
        This method attempts to retrieve the non-default values of the
        generation config for this model.

        The generation config can contain information about special tokens, as
        well as sampling parameters. Which is why this method exists separately
        to `get_diff_sampling_param`.

        Returns:
            A dictionary containing the non-default generation config.
        """
        if self.generation_config in {"auto", "vllm"}:
            config = try_get_generation_config(
                self.hf_config_path or self.model,
                trust_remote_code=self.trust_remote_code,
                revision=self.revision,
1517
                config_format=self.config_format,
1518
1519
1520
1521
1522
            )
        else:
            config = try_get_generation_config(
                self.generation_config,
                trust_remote_code=self.trust_remote_code,
1523
                config_format=self.config_format,
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
            )

        if config is None:
            return {}

        return config.to_diff_dict()

    def get_diff_sampling_param(self) -> dict[str, Any]:
        """
        This method returns a dictionary containing the non-default sampling
        parameters with `override_generation_config` applied.

        The default sampling parameters are:

        - vLLM's neutral defaults if `self.generation_config="vllm"`
        - the model's defaults if `self.generation_config="auto"`
        - as defined in `generation_config.json` if
            `self.generation_config="path/to/generation_config/dir"`

        Returns:
            A dictionary containing the non-default sampling parameters.
        """
        if self.generation_config == "vllm":
            config = {}
        else:
            config = self.try_get_generation_config()

        # Overriding with given generation config
        config.update(self.override_generation_config)

        available_params = [
            "repetition_penalty",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
            "max_new_tokens",
        ]
        if any(p in config for p in available_params):
            diff_sampling_param = {
1564
                p: config.get(p) for p in available_params if config.get(p) is not None
1565
1566
1567
1568
1569
            }
            # Huggingface definition of max_new_tokens is equivalent
            # to vLLM's max_tokens
            if "max_new_tokens" in diff_sampling_param:
                diff_sampling_param["max_tokens"] = diff_sampling_param.pop(
1570
1571
                    "max_new_tokens"
                )
1572
1573
1574
1575
1576
1577
1578
1579
        else:
            diff_sampling_param = {}

        if diff_sampling_param:
            logger.warning_once(
                "Default sampling parameters have been overridden by the "
                "model's Hugging Face generation config recommended from the "
                "model creator. If this is not intended, please relaunch "
1580
1581
                "vLLM instance with `--generation-config vllm`."
            )
1582
1583
1584
1585
1586
1587
1588
        return diff_sampling_param

    @property
    def is_encoder_decoder(self) -> bool:
        """Extract the HF encoder/decoder model flag."""
        return is_encoder_decoder(self.hf_config)

1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
    @property
    def uses_alibi(self) -> bool:
        cfg = self.hf_text_config

        return (
            getattr(cfg, "alibi", False)  # Falcon
            or "BloomForCausalLM" in self.architectures  # Bloom
            or getattr(cfg, "position_encoding_type", "") == "alibi"  # codellm_1b_alibi
            or (
                hasattr(cfg, "attn_config")  # MPT
                and (
                    (
                        isinstance(cfg.attn_config, dict)
                        and cfg.attn_config.get("alibi", False)
                    )
                    or (
                        not isinstance(cfg.attn_config, dict)
                        and getattr(cfg.attn_config, "alibi", False)
                    )
                )
            )
        )

1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)

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

    @property
    def is_multimodal_raw_input_only_model(self) -> bool:
        return self._model_info.supports_multimodal_raw_input_only

    @property
    def is_cross_encoder(self) -> bool:
1626
1627
1628
        return (
            self._model_info.supports_cross_encoding or self.convert_type == "classify"
        )
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649

    @property
    def is_pp_supported(self) -> bool:
        return self._model_info.supports_pp

    @property
    def is_attention_free(self) -> bool:
        return self._model_info.is_attention_free

    @property
    def is_hybrid(self) -> bool:
        return self._model_info.is_hybrid

    @property
    def has_noops(self) -> bool:
        return self._model_info.has_noops

    @property
    def has_inner_state(self):
        return self._model_info.has_inner_state

1650
1651
1652
1653
    @property
    def supports_mamba_prefix_caching(self) -> bool:
        return self._model_info.supports_mamba_prefix_caching

1654
1655
1656
1657
1658
1659
    @property
    def use_mla(self) -> bool:
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE

    @property
    def is_matryoshka(self) -> bool:
1660
1661
1662
        return bool(getattr(self.hf_config, "matryoshka_dimensions", None)) or getattr(
            self.hf_config, "is_matryoshka", False
        )
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685

    @property
    def matryoshka_dimensions(self):
        return getattr(self.hf_config, "matryoshka_dimensions", None)

    @property
    def use_pad_token(self) -> bool:
        # cross_encoder models defaults to using pad_token.
        # `llm as reranker` models defaults to not using pad_token.
        return getattr(self.hf_config, "use_pad_token", True)

    @property
    def head_dtype(self) -> torch.dtype:
        """
        "head" refers to the last Linear layer(s) of an LLM,
        such as the lm_head in a generation model,
        or the score or classifier in a classification model.

        `head_dtype` currently only supports pooling models.\n
        - The pooling model defaults to using fp32 head,
        you can use --hf-overrides '{"head_dtype": "model"}' to disable it.
        """

1686
1687
1688
        head_dtype = _get_head_dtype(
            config=self.hf_config, dtype=self.dtype, runner_type=self.runner_type
        )
1689
1690
1691
1692

        if self.runner_type != "pooling" and head_dtype != self.dtype:
            logger.warning_once(
                "`head_dtype` currently only supports pooling models."
1693
1694
1695
                "fallback to model dtype [%s].",
                self.dtype,
            )
1696
1697
1698
1699
1700
            return self.dtype

        if head_dtype not in current_platform.supported_dtypes:
            logger.warning_once(
                "The current platform does not support [%s] head dtype, "
1701
1702
1703
1704
                "fallback to model dtype [%s].",
                head_dtype,
                self.dtype,
            )
1705
1706
1707
1708
1709
            return self.dtype

        logger.debug_once("head dtype: %s", head_dtype)
        return head_dtype

1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
    @property
    def hidden_size(self):
        if hasattr(self.hf_config, "hidden_size"):
            return self.hf_config.hidden_size
        text_config = self.hf_config.get_text_config()
        return text_config.hidden_size

    @property
    def embedding_size(self):
        dense_modules = try_get_dense_modules(self.model, revision=self.revision)
        if dense_modules is not None:
            return dense_modules[-1]["out_features"]
        return self.hidden_size

1724
1725
1726
1727
    def get_and_verify_max_len(self, max_model_len: int):
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
        tokenizer_config = None
1728
1729
1730
1731
        if (
            self.runner_type == "pooling"
            and getattr(self.hf_config, "position_embedding_type", "") == "absolute"
        ):
1732
1733
1734
            tokenizer_config = try_get_tokenizer_config(
                self.tokenizer,
                trust_remote_code=self.trust_remote_code,
1735
1736
                revision=self.tokenizer_revision,
            )
1737
1738
1739
1740
1741
1742
1743
        max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
            tokenizer_config=tokenizer_config,
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
            sliding_window=self.get_sliding_window(),
            spec_target_max_model_len=self.spec_target_max_model_len,
1744
1745
            encoder_config=self.encoder_config,
        )
1746
1747
1748
1749
        logger.info("Using max model len %s", max_model_len)
        return max_model_len


1750
def get_served_model_name(model: str, served_model_name: str | list[str] | None):
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
    """
    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
    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


# Some model suffixes are based on auto classes from Transformers:
# https://huggingface.co/docs/transformers/en/model_doc/auto
# NOTE: Items higher on this list priority over lower ones
_SUFFIX_TO_DEFAULTS: list[tuple[str, tuple[RunnerType, ConvertType]]] = [
    ("ForCausalLM", ("generate", "none")),
    ("ForConditionalGeneration", ("generate", "none")),
    ("ChatModel", ("generate", "none")),
    ("LMHeadModel", ("generate", "none")),
    ("ForTextEncoding", ("pooling", "embed")),
    ("EmbeddingModel", ("pooling", "embed")),
    ("ForSequenceClassification", ("pooling", "classify")),
    ("ForAudioClassification", ("pooling", "classify")),
    ("ForImageClassification", ("pooling", "classify")),
    ("ForVideoClassification", ("pooling", "classify")),
    ("ClassificationModel", ("pooling", "classify")),
    ("ForRewardModeling", ("pooling", "reward")),
    ("RewardModel", ("pooling", "reward")),
    # Let other `*Model`s take priority
    ("Model", ("pooling", "embed")),
]


def iter_architecture_defaults():
    yield from _SUFFIX_TO_DEFAULTS


def try_match_architecture_defaults(
    architecture: str,
    *,
1794
1795
1796
    runner_type: RunnerType | None = None,
    convert_type: ConvertType | None = None,
) -> tuple[str, tuple[RunnerType, ConvertType]] | None:
1797
1798
1799
1800
1801
1802
1803
1804
1805
    for suffix, (
        default_runner_type,
        default_convert_type,
    ) in iter_architecture_defaults():
        if (
            (runner_type is None or runner_type == default_runner_type)
            and (convert_type is None or convert_type == default_convert_type)
            and architecture.endswith(suffix)
        ):
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
            return suffix, (default_runner_type, default_convert_type)

    return None


_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

# model_type -> reason
_FLOAT16_NOT_SUPPORTED_MODELS = {
    "gemma2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "gemma3": "Numerical instability. Please use bfloat16 or float32 instead.",
1823
    "gemma3_text": "Numerical instability. Please use bfloat16 or float32 instead.",
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
    "plamo2": "Numerical instability. Please use bfloat16 or float32 instead.",
    "glm4": "Numerical instability. Please use bfloat16 or float32 instead.",
}


def _is_valid_dtype(model_type: str, dtype: torch.dtype):
    if model_type in _FLOAT16_NOT_SUPPORTED_MODELS and dtype == torch.float16:  # noqa: E501, SIM103
        return False

    return True


def _check_valid_dtype(model_type: str, dtype: torch.dtype):
    if model_type in _FLOAT16_NOT_SUPPORTED_MODELS and dtype == torch.float16:
        reason = _FLOAT16_NOT_SUPPORTED_MODELS[model_type]
1839
1840
1841
        raise ValueError(
            f"The model type {model_type!r} does not support float16. Reason: {reason}"
        )
1842
1843
1844
1845
1846
1847
1848
1849

    return True


def _find_dtype(
    model_id: str,
    config: PretrainedConfig,
    *,
1850
    revision: str | None,
1851
):
1852
1853
1854
    # NOTE: getattr(config, "dtype", torch.float32) is not correct
    # because config.dtype can be None.
    config_dtype = getattr(config, "dtype", None)
1855
1856

    # Fallbacks for multi-modal models if the root config
1857
    # does not define dtype
1858
    if config_dtype is None:
1859
        config_dtype = getattr(config.get_text_config(), "dtype", None)
1860
    if config_dtype is None and hasattr(config, "vision_config"):
1861
        config_dtype = getattr(config.vision_config, "dtype", None)
1862
    if config_dtype is None and hasattr(config, "encoder_config"):
1863
        config_dtype = getattr(config.encoder_config, "dtype", None)
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894

    # Try to read the dtype of the weights if they are in safetensors format
    if config_dtype is None:
        repo_mt = try_get_safetensors_metadata(model_id, revision=revision)

        if repo_mt and (files_mt := repo_mt.files_metadata):
            param_dtypes: set[torch.dtype] = {
                _SAFETENSORS_TO_TORCH_DTYPE[dtype_str]
                for file_mt in files_mt.values()
                for dtype_str in file_mt.parameter_count
                if dtype_str in _SAFETENSORS_TO_TORCH_DTYPE
            }

            if param_dtypes:
                return common_broadcastable_dtype(param_dtypes)

    if config_dtype is None:
        config_dtype = torch.float32

    return config_dtype


def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    from vllm.platforms import current_platform

    supported_dtypes = [
1895
1896
        dtype
        for dtype in current_platform.supported_dtypes
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
        if _is_valid_dtype(model_type, dtype)
    ]

    if is_pooling_model and torch.float16 in supported_dtypes:
        preferred_dtype = torch.float16
    else:
        preferred_dtype = supported_dtypes[0]

    # Downcast for float32 models
    if config_dtype == torch.float32:
        config_dtype = preferred_dtype

    if config_dtype in supported_dtypes:
        return config_dtype

    # Ensure device compatibility
    device_name = current_platform.get_device_name()
    device_capability = current_platform.get_device_capability()

    if device_capability is None:
        device_str = f"{device_name!r}"
    else:
        version_str = device_capability.as_version_str()
        device_str = f"{device_name!r} (with compute capability {version_str})"

    logger.warning(
1923
        "Your device %s doesn't support %s. Falling back to %s for compatibility.",
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
        device_str,
        config_dtype,
        preferred_dtype,
    )

    return preferred_dtype


def _get_and_verify_dtype(
    model_id: str,
    config: PretrainedConfig,
1935
    dtype: str | torch.dtype,
1936
1937
    *,
    is_pooling_model: bool,
1938
    revision: str | None = None,
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
) -> torch.dtype:
    config_dtype = _find_dtype(model_id, config, revision=revision)
    model_type = config.model_type

    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            # Set default dtype from model config
            torch_dtype = _resolve_auto_dtype(
                model_type,
                config_dtype,
                is_pooling_model=is_pooling_model,
            )
        else:
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
                raise ValueError(f"Unknown dtype: {dtype!r}")
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
    else:
        raise ValueError(f"Unknown dtype: {dtype}")

    _check_valid_dtype(model_type, torch_dtype)

    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
        else:
            # Casting between float16 and bfloat16 is allowed with a warning.
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)

    return torch_dtype


1977
1978
1979
def _get_head_dtype(
    config: PretrainedConfig, dtype: torch.dtype, runner_type: str
) -> torch.dtype:
1980
    head_dtype: str | torch.dtype | None = getattr(config, "head_dtype", None)
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002

    if head_dtype == "model":
        return dtype
    elif isinstance(head_dtype, str):
        head_dtype = head_dtype.lower()
        if head_dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
            raise ValueError(f"Unknown dtype: {head_dtype!r}")
        return _STR_DTYPE_TO_TORCH_DTYPE[head_dtype]
    elif isinstance(head_dtype, torch.dtype):
        return head_dtype
    elif head_dtype is None:
        if torch.float32 not in current_platform.supported_dtypes:
            return dtype
        if runner_type == "pooling":
            return torch.float32
        return dtype
    else:
        raise ValueError(f"Unknown dtype: {head_dtype}")


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
2003
2004
    tokenizer_config: dict | None,
    max_model_len: int | None,
2005
    disable_sliding_window: bool,
2006
2007
2008
    sliding_window: int | None,
    spec_target_max_model_len: int | None = None,
    encoder_config: Any | None = None,
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
) -> 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",
        # ChatGLM2
        "seq_length",
        # Command-R
        "model_max_length",
        # Whisper
        "max_target_positions",
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
    # Choose the smallest "max_length" from the possible keys
    max_len_key = None
    for key in possible_keys:
        max_len = getattr(hf_config, key, None)
        if max_len is not None:
2035
            max_len_key = key if max_len < derived_max_model_len else max_len_key
2036
2037
2038
2039
2040
2041
2042
2043
            derived_max_model_len = min(derived_max_model_len, max_len)
    # For Command-R / Cohere, Cohere2 / Aya Vision models
    if tmp_max_len := getattr(hf_config, "model_max_length", None):
        max_len_key = "model_max_length"
        derived_max_model_len = tmp_max_len

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
2044
2045
2046
2047
2048
    if (
        disable_sliding_window
        and sliding_window is not None
        and sliding_window < derived_max_model_len
    ):
2049
2050
2051
2052
2053
2054
        max_len_key = "sliding_window"
        derived_max_model_len = sliding_window

    # Consider model_max_length in tokenizer_config
    if tokenizer_config:
        tokenizer_model_max_length = tokenizer_config.get(
2055
2056
2057
            "model_max_length", derived_max_model_len
        )
        derived_max_model_len = min(derived_max_model_len, tokenizer_model_max_length)
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074

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

        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

        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: "
2075
2076
2077
2078
            "%s. Assuming the model's maximum length is %d.",
            possible_keys,
            default_max_len,
        )
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
        derived_max_model_len = default_max_len

    rope_scaling = getattr(hf_config, "rope_scaling", None)
    # NOTE(woosuk): Gemma3's max_model_len (128K) is already scaled by RoPE
    # scaling, so we skip applying the scaling factor again.
    if rope_scaling is not None and "gemma3" not in hf_config.model_type:
        # No need to consider "type" key because of patch_rope_scaling when
        # loading HF config
        rope_type = rope_scaling["rope_type"]

        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 "
2096
2097
                    "investigate."
                )
2098
2099
2100
2101
2102
2103

            # 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)

            if rope_type == "yarn":
2104
                derived_max_model_len = rope_scaling["original_max_position_embeddings"]
2105
2106
2107
2108
2109
            derived_max_model_len *= scaling_factor

    if encoder_config and "max_seq_length" in encoder_config:
        derived_max_model_len = encoder_config["max_seq_length"]

2110
2111
    # If the user didn't specify `max_model_len`, then use that derived from
    # the model config as a default value.
2112
    if max_model_len is None:
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
        # For LongRoPE, default to original_max_position_embeddings to avoid
        # performance degradation for shorter sequences
        if rope_scaling is not None and rope_scaling["rope_type"] == "longrope":
            max_model_len = int(
                getattr(
                    hf_config, "original_max_position_embeddings", derived_max_model_len
                )
            )
        else:
            max_model_len = int(derived_max_model_len)
2123
        max_model_len = current_platform.check_max_model_len(max_model_len)
2124

2125
2126
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
    elif max_model_len > derived_max_model_len:
        # 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:
            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 "
2139
2140
                    "so we can investigate."
                )
2141
2142
2143
2144
2145
        else:
            msg = (
                f"User-specified max_model_len ({max_model_len}) is greater "
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
2146
2147
                f"{model_max_length} in model's config.json)."
            )
2148
2149
2150
2151
2152
2153
            warning = (
                "VLLM_ALLOW_LONG_MAX_MODEL_LEN must be used with extreme "
                "caution. If the model uses relative position encoding (RoPE), "
                "positions exceeding derived_max_model_len lead to nan. If the "
                "model uses absolute position encoding, positions exceeding "
                "derived_max_model_len will cause a CUDA array out-of-bounds "
2154
2155
                "error."
            )
2156
2157
2158
2159
2160
            if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
                logger.warning_once("%s %s", msg, warning)
            else:
                raise ValueError(
                    f"{msg} To allow overriding this maximum, set "
2161
2162
                    f"the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1. {warning}"
                )
2163
    return int(max_model_len)