model.py 86.4 KB
Newer Older
1
2
3
4
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import warnings
5
from collections.abc import Callable
6
from dataclasses import InitVar, field
7
from functools import cached_property
8
from typing import TYPE_CHECKING, Any, Literal, cast, get_args
9
10

import torch
11
from pydantic import ConfigDict, Field, field_validator, model_validator
12
13

import vllm.envs as envs
14
15
16
from vllm.config.model_arch import (
    ModelArchitectureConfig,
)
17
18
19
20
21
22
from vllm.config.multimodal import (
    MMCacheType,
    MMEncoderTPMode,
    MMTensorIPC,
    MultiModalConfig,
)
23
from vllm.config.pooler import PoolerConfig
24
from vllm.config.quantization import OnlineQuantizationConfigArgs
25
from vllm.config.scheduler import RunnerType
26
from vllm.config.utils import config, getattr_iter
27
28
from vllm.logger import init_logger
from vllm.platforms import current_platform
29
from vllm.tasks import PoolingTask, ScoreType, SupportedTask
30
from vllm.transformers_utils.config import (
31
32
33
34
35
36
37
    ConfigFormat,
    get_config,
    get_hf_image_processor_config,
    get_hf_text_config,
    get_pooling_config,
    get_sentence_transformer_tokenizer_config,
    is_encoder_decoder,
38
    is_rope_parameters_nested,
39
    try_get_dense_modules,
40
41
42
    try_get_generation_config,
    try_get_tokenizer_config,
    uses_mrope,
43
    uses_xdrope_dim,
44
)
45
from vllm.transformers_utils.gguf_utils import (
46
    is_gguf,
47
    is_remote_gguf,
48
    maybe_patch_hf_config_from_gguf,
49
50
    split_remote_gguf,
)
51
52
53
54
from vllm.transformers_utils.model_arch_config_convertor import (
    MODEL_ARCH_CONFIG_CONVERTORS,
    ModelArchConfigConvertorBase,
)
55
56
from vllm.transformers_utils.runai_utils import ObjectStorageModel, is_runai_obj_uri
from vllm.transformers_utils.utils import maybe_model_redirect
57
from vllm.utils.import_utils import LazyLoader
58
from vllm.v1.attention.backends.registry import AttentionBackendEnum
59
60
61
62
63
64
65
66
67
68
69
70
71

if TYPE_CHECKING:
    from transformers import PretrainedConfig

    import vllm.model_executor.layers.quantization as me_quant
    import vllm.model_executor.models as me_models
    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

72
73
74
75
    me_quant = LazyLoader(
        "model_executor", globals(), "vllm.model_executor.layers.quantization"
    )
    me_models = LazyLoader("model_executor", globals(), "vllm.model_executor.models")
76
77
78
79
80
81
82
    LoadConfig = Any
    ParallelConfig = Any
    QuantizationMethods = Any
    LogitsProcessor = Any

logger = init_logger(__name__)

83
RunnerOption = Literal["auto", RunnerType]
84
ConvertType = Literal["none", "embed", "classify"]
85
ConvertOption = Literal["auto", ConvertType]
86
TokenizerMode = Literal["auto", "hf", "slow", "mistral", "deepseek_v32"]
87
ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
88
89
90
LogprobsMode = Literal[
    "raw_logits", "raw_logprobs", "processed_logits", "processed_logprobs"
]
91
HfOverrides = dict[str, Any] | Callable[[PretrainedConfig], PretrainedConfig]
92
ModelImpl = Literal["auto", "vllm", "transformers", "terratorch"]
93
LayerBlockType = Literal["attention", "linear_attention", "mamba"]
94
95
96

_RUNNER_CONVERTS: dict[RunnerType, list[ConvertType]] = {
    "generate": [],
97
    "pooling": ["embed", "classify"],
98
99
100
    "draft": [],
}

101
102
103
104
AttnTypeStr = Literal[
    "decoder", "encoder", "encoder_only", "encoder_decoder", "attention_free", "hybrid"
]

105

106
107
@config(config=ConfigDict(arbitrary_types_allowed=True))
class ModelConfig:
108
109
110
111
112
113
    """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."""
114
    model_weights: str = ""
115
116
    """Original model weights path. Used when the model is pulled from object
    storage (e.g., RunAI) to preserve the original URI while `model` points to
117
    the local directory."""
118
119
120
121
122
123
124
    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."""
125
    tokenizer: str = None  # type: ignore[assignment]
126
127
128
    """Name or path of the Hugging Face tokenizer to use. If unspecified, model
    name or path will be used."""
    tokenizer_mode: TokenizerMode | str = "auto"
129
130
    """Tokenizer mode:

131
    - "auto" will use the tokenizer from `mistral_common` for Mistral models
132
133
134
135
136
137
      if available, otherwise it will use the "hf" tokenizer.
    - "hf" will use the fast tokenizer if available.
    - "slow" will always use the slow tokenizer.
    - "mistral" will always use the tokenizer from `mistral_common`.
    - "deepseek_v32" will always use the tokenizer from `deepseek_v32`.
    - "qwen_vl" will always use the tokenizer from `qwen_vl`.
138
    - Other custom values can be supported via plugins."""
139
140
141
    trust_remote_code: bool = False
    """Trust remote code (e.g., from HuggingFace) when downloading the model
    and tokenizer."""
142
    dtype: ModelDType | torch.dtype = "auto"
143
144
    """Data type for model weights and activations:

145
    - "auto" will use FP16 precision for FP32 and FP16 models, and BF16
146
147
148
149
150
      precision for BF16 models.
    - "half" for FP16. Recommended for AWQ quantization.
    - "float16" is the same as "half".
    - "bfloat16" for a balance between precision and range.
    - "float" is shorthand for FP32 precision.
151
    - "float32" for FP32 precision."""
152
153
154
155
156
157
    seed: int = 0
    """Random seed for reproducibility.

    We must set the global seed because otherwise,
    different tensor parallel workers would sample different tokens,
    leading to inconsistent results."""
158
159
160
161
    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)."""
162
    hf_config_path: str | None = None
163
164
    """Name or path of the Hugging Face config to use. If unspecified, model
    name or path will be used."""
165
166
167
168
169
170
171
    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."""
    allowed_media_domains: list[str] | None = None
    """If set, only media URLs that belong to this domain can be used for
    multi-modal inputs. """
172
    revision: str | None = None
173
174
    """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."""
175
    code_revision: str | None = None
176
177
178
    """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."""
179
180
181
182
    tokenizer_revision: str | None = None
    """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."""
183
    max_model_len: int = Field(default=None, ge=-1)  # type: ignore[assignment]
184
185
186
187
    """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
188
189
190
191
192
    format. Examples:

    - 1k -> 1000
    - 1K -> 1024
    - 25.6k -> 25,600
193
    - -1 or 'auto' -> Automatically choose the maximum model length that fits in
194
195
      GPU memory. This will use the model's maximum context length if it fits,
      otherwise it will find the largest length that can be accommodated."""
196
    spec_target_max_model_len: int | None = None
197
    """Specify the maximum length for spec decoding draft models."""
198
    quantization: QuantizationMethods | str | None = None
199
200
201
202
    """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."""
203
204
205
206
    quantization_config: dict[str, Any] | OnlineQuantizationConfigArgs | None = None
    """Arguments for online quantization.
    Auto-created when `quantization` equals to one of the string values of
    the `OnlineQuantScheme` enum."""
207
208
    allow_deprecated_quantization: bool = False
    """Whether to allow deprecated quantization methods."""
209
210
211
212
213
    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."""
214
215
    enable_return_routed_experts: bool = False
    """Whether to return routed experts."""
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
    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."""
233
    disable_cascade_attn: bool = True
234
235
    """Disable cascade attention for V1. While cascade attention does not
    change the mathematical correctness, disabling it could be useful for
236
237
238
239
    preventing potential numerical issues. This defaults to True, so users
    must opt in to cascade attention by setting this to False. Even when this
    is set to False, cascade attention will only be used when the heuristic
    tells that it's beneficial."""
240
241
242
243
    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."""
244
245
    enable_prompt_embeds: bool = False
    """If `True`, enables passing text embeddings as inputs via the
246
247
248
249
    `prompt_embeds` key.

    WARNING: The vLLM engine may crash if incorrect shape of embeddings is passed.
    Only enable this flag for trusted users!"""
250
    served_model_name: str | list[str] | None = None
251
252
253
254
255
256
257
    """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."""
258
    config_format: str | ConfigFormat = "auto"
259
260
    """The format of the model config to load:

261
    - "auto" will try to load the config in hf format if available after trying
262
263
      to load in mistral format.
    - "hf" will load the config in hf format.
264
    - "mistral" will load the config in mistral format."""
265
    hf_token: bool | str | None = None
266
    """The token to use as HTTP bearer authorization for remote files . If
267
268
    `True`, will use the token generated when running `hf auth login`
    (stored in `~/.cache/huggingface/token`)."""
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
    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."""
    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
285
286
    """Enable sleep mode for the engine (only cuda and
    hip platforms are supported)."""
287
    model_impl: str | ModelImpl = "auto"
288
289
290
291
292
293
    """Which implementation of the model to use:

    - "auto" will try to use the vLLM implementation, if it exists, and fall back to the
      Transformers implementation if no vLLM implementation is available.
    - "vllm" will use the vLLM model implementation.
    - "transformers" will use the Transformers model implementation.
294
295
    - "terratorch" will use the TerraTorch model implementation.
    """
296
    override_attention_dtype: str | None = None
297
    """Override dtype for attention"""
298
    logits_processors: list[str | type[LogitsProcessor]] | None = None
299
300
    """One or more logits processors' fully-qualified class names or class
    definitions"""
301
302
    io_processor_plugin: str | None = None
    """IOProcessor plugin name to load at model startup"""
303
304
305
306
    renderer_num_workers: int = 1
    """Number of worker threads in the renderer thread pool. This pool
    handles async tokenization, chat template rendering, and multimodal
    preprocessing."""
307
308

    # Pooler config
309
    pooler_config: PoolerConfig | None = None
310
311
312
313
    """Pooler config which controls the behaviour of output pooling in pooling
    models."""

    # Multimodal config and init vars
314
    multimodal_config: MultiModalConfig | None = None
315
316
    """Configuration for multimodal model. If `None`, this will be inferred
    from the architecture of `self.model`."""
317
    language_model_only: InitVar[bool] = False
318
    limit_mm_per_prompt: InitVar[dict[str, int | dict[str, int]] | None] = None
319
    enable_mm_embeds: InitVar[bool | None] = None
320
    media_io_kwargs: InitVar[dict[str, dict[str, Any]] | None] = None
321
322
323
324
    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
325
    mm_encoder_only: InitVar[bool | None] = None
326
    mm_encoder_tp_mode: InitVar[MMEncoderTPMode | None] = None
327
    mm_encoder_attn_backend: InitVar[AttentionBackendEnum | str | None] = None
328
329
330
    interleave_mm_strings: InitVar[bool | None] = None
    skip_mm_profiling: InitVar[bool | None] = None
    video_pruning_rate: InitVar[float | None] = None
331
    mm_tensor_ipc: InitVar[MMTensorIPC] = None
332
333
334
335
336
337
338
339
340
341
342
343
344

    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.
        """
345
346
        ignored_factors = {
            "convert",
347
348
            "tokenizer",
            "tokenizer_mode",
349
350
            "seed",
            "hf_config_path",
351
352
353
            "allowed_local_media_path",
            "allowed_media_domains",
            "tokenizer_revision",
354
355
356
357
            "spec_target_max_model_len",
            "enforce_eager",
            "logprobs_mode",
            "disable_cascade_attn",
358
            "skip_tokenizer_init",
359
360
361
362
363
364
            "served_model_name",
            "config_format",
            "hf_token",
            "hf_overrides",
            "override_attention_dtype",
            "logits_processors",
365
            "io_processor_plugin",
366
367
368
            "pooler_config",
            "multimodal_config",
            "limit_mm_per_prompt",
369
            "media_io_kwargs",
370
371
372
373
374
375
376
377
            "mm_processor_kwargs",
            "mm_processor_cache_gb",
            "mm_processor_cache_type",
            "mm_shm_cache_max_object_size_mb",
            "mm_encoder_tp_mode",
            "interleave_mm_strings",
            "skip_mm_profiling",
        }
378

379
        from vllm.config.utils import get_hash_factors, hash_factors
380

381
        factors = get_hash_factors(self, ignored_factors)
382
383
384
385
386
387

        # NOTE: For some models (e.g, Qwen3-VL), whether the MM code path is enabled
        # affects the computation graph of the language model, therefore we add it
        # here early.
        if self.multimodal_config:
            factors["language_model_only"] = self.multimodal_config.language_model_only
388
        return hash_factors(factors)
389

390
391
    def _update_nested(
        self,
392
        target: PretrainedConfig | dict[str, Any],
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
        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,
420
        config: PretrainedConfig,
421
422
423
424
425
426
427
428
429
430
431
432
433
434
        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)

435
    def __post_init__(
436
437
        self,
        # Multimodal config init vars
438
        language_model_only: bool,
439
        limit_mm_per_prompt: dict[str, int | dict[str, int]] | None,
440
        enable_mm_embeds: bool | None,
441
        media_io_kwargs: dict[str, dict[str, Any]] | None,
442
443
444
445
        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,
446
        mm_encoder_only: bool | None,
447
        mm_encoder_tp_mode: MMEncoderTPMode | None,
448
        mm_encoder_attn_backend: AttentionBackendEnum | str | None,
449
450
451
        interleave_mm_strings: bool | None,
        skip_mm_profiling: bool | None,
        video_pruning_rate: float | None,
452
        mm_tensor_ipc: MMTensorIPC,
453
    ) -> None:
454
        # Keep set served_model_name before maybe_model_redirect(self.model)
455
456
457
        self.served_model_name = get_served_model_name(
            self.model, self.served_model_name
        )
458
459
460
461
462
463
464
        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)
465
466
467
468
469

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

        if callable(self.hf_overrides):
470
            hf_overrides_kw: dict[str, Any] = {}
471
            hf_overrides_fn = self.hf_overrides
472
            dict_overrides: dict[str, Any] = {}
473
        else:
474
475
476
477
478
479
480
481
482
            # 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
483
484
            hf_overrides_fn = None

485
        self.maybe_pull_model_tokenizer_for_runai(self.model, self.tokenizer)
486

487
        if self.override_attention_dtype is not None and not current_platform.is_rocm():
488
489
            warnings.warn(
                "override-attention-dtype is set but not using ROCm platform",
490
491
                stacklevel=2,
            )
492

493
494
495
496
497
498
499
500
501
502
503
        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,
504
            token=self.hf_token,
505
        )
506
507
508
509
        hf_config = maybe_patch_hf_config_from_gguf(
            self.model,
            hf_config,
        )
510
511

        self.hf_config = hf_config
512
513
        if dict_overrides:
            self._apply_dict_overrides(hf_config, dict_overrides)
514
        self.hf_text_config = get_hf_text_config(self.hf_config)
515
516
517
        self.attention_chunk_size = getattr(
            self.hf_text_config, "attention_chunk_size", None
        )
518
519
        self.encoder_config = self._get_encoder_config()
        self.hf_image_processor_config = get_hf_image_processor_config(
520
521
            self.model, hf_token=self.hf_token, revision=self.revision
        )
522
        self.model_arch_config = self.get_model_arch_config()
523
524
525

        architectures = self.architectures
        registry = self.registry
526
        is_generative_model = registry.is_text_generation_model(architectures, self)
527
528
529
        is_pooling_model = registry.is_pooling_model(architectures, self)

        self.runner_type = self._get_runner_type(architectures, self.runner)
530
531
532
        self.convert_type = self._get_convert_type(
            architectures, self.runner_type, self.convert
        )
533
534
535
536
537

        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
538
                raise ValueError("This model does not support `--runner generate`.")
539
540
541
542
543
544
545
        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 "
546
547
                    "it into a pooling model."
                )
548
549
550
551
552
553
554
555

        # 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)

556
557
        # Set default tokenizer modes based on model architecture
        if self.tokenizer_mode == "auto":
558
559
560
            if self.model_impl == "terratorch":
                self.tokenizer_mode = "terratorch"
            elif arch == "Grok1ForCausalLM":
561
562
563
564
565
                self.tokenizer_mode = "grok2"
            elif arch == "MoonshotKimiaForCausalLM":
                self.tokenizer_mode = "kimi_audio"
            elif arch == "QwenVLForConditionalGeneration":
                self.tokenizer_mode = "qwen_vl"
566
567
            elif arch == "DeepseekV32ForCausalLM":
                self.tokenizer_mode = "deepseek_v32"
568
569
570
571
572
573
574
575

            if self.tokenizer_mode != "auto":
                logger.info(
                    "Defaulting to tokenizer_mode=%r for %s",
                    self.tokenizer_mode,
                    arch,
                )

576
577
578
579
580
581
582
583
584
585
586
587
        # Init pooler config if needed
        if self.runner_type == "pooling":
            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)

588
589
590
591
592
593
            default_seq_pooling_type = self._model_info.default_seq_pooling_type
            if self.pooler_config.seq_pooling_type is None:
                self.pooler_config.seq_pooling_type = default_seq_pooling_type
            default_tok_pooling_type = self._model_info.default_tok_pooling_type
            if self.pooler_config.tok_pooling_type is None:
                self.pooler_config.tok_pooling_type = default_tok_pooling_type
594
595
596
597
598
599
600

        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,
601
            config_format=self.config_format,
602
603
        )

604
605
606
607
608
609
610
611
        # Some checkpoints set sliding_window to 0 to indicate that sliding window is
        # disabled, but vLLM uses None for that. Convert 0 to None to avoid errors.
        # Set before get_and_verify_max_len to ensure that max_model_len does not get
        # capped to 0.
        if self.get_sliding_window() == 0:
            self.disable_sliding_window = True
            self.hf_text_config.sliding_window = None

612
        self.original_max_model_len = self.max_model_len
613
        self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
614
615

        if self.is_encoder_decoder:
616
            mm_processor_cache_gb = 0
617
618
            logger.info("Encoder-decoder model detected, disabling mm processor cache.")

619
620
        # Init multimodal config if needed
        if self._model_info.supports_multimodal:
621
622
623
624
            if (
                mm_encoder_tp_mode == "data"
                and not self._model_info.supports_multimodal_encoder_tp_data
            ):
625
626
                logger.warning_once(
                    "This model does not support `--mm-encoder-tp-mode data`. "
627
628
                    "Falling back to `--mm-encoder-tp-mode weights`."
                )
629
630
631
                mm_encoder_tp_mode = "weights"

            mm_config_kwargs = dict(
632
                language_model_only=language_model_only,
633
                limit_per_prompt=limit_mm_per_prompt,
634
                enable_mm_embeds=enable_mm_embeds,
635
                media_io_kwargs=media_io_kwargs,
636
637
638
639
                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,
640
                mm_encoder_only=mm_encoder_only,
641
                mm_encoder_tp_mode=mm_encoder_tp_mode,
642
                mm_encoder_attn_backend=mm_encoder_attn_backend,
643
644
                interleave_mm_strings=interleave_mm_strings,
                skip_mm_profiling=skip_mm_profiling,
645
                video_pruning_rate=video_pruning_rate,
646
                mm_tensor_ipc=mm_tensor_ipc,
647
648
649
            )

            mm_config_kwargs = {
650
                k: v for k, v in mm_config_kwargs.items() if v is not None
651
652
            }

653
            self.multimodal_config = MultiModalConfig(**mm_config_kwargs)  # type: ignore[arg-type]
654

655
656
657
658
659
660
661
662
663
664
665
666
667
            if (
                self.renderer_num_workers > 1
                and self.multimodal_config.mm_processor_cache_gb > 0
            ):
                raise ValueError(
                    "Cannot use --renderer-num-workers > 1 with the "
                    "multimodal processor cache enabled. The cache is "
                    "not thread-safe and does not support concurrent "
                    "renderer workers. Please set "
                    "--renderer-num-workers 1 (the default), or "
                    "disable the cache with --mm-processor-cache-gb 0."
                )

668
669
670
671
672
673
674
675
        # Multimodal GGUF models must use original repo for mm processing
        if is_gguf(self.tokenizer) and self.is_multimodal_model:
            raise ValueError(
                "Loading a multimodal GGUF model needs to use original "
                "tokenizer. Please specify the unquantized hf model's "
                "repo name or path using the --tokenizer argument."
            )

676
        if self.disable_sliding_window:
677
            # Set after get_and_verify_max_len to ensure that max_model_len
678
679
680
681
682
            # can be correctly capped to sliding window size
            self.hf_text_config.sliding_window = None

        # Avoid running try_verify_and_update_config multiple times
        self.config_updated = False
683
        self._try_verify_and_update_model_config()
684
685
686
687
        self._verify_quantization()
        self._verify_cuda_graph()
        self._verify_bnb_config()

688
689
690
691
692
693
694
695
696
    def get_model_arch_config(
        self,
    ) -> ModelArchitectureConfig:
        convertor_cls = MODEL_ARCH_CONFIG_CONVERTORS.get(
            self.hf_config.model_type, ModelArchConfigConvertorBase
        )
        convertor = convertor_cls(self.hf_config, self.hf_text_config)
        return convertor.convert()

697
698
699
700
701
702
703
704
    @field_validator("tokenizer", "max_model_len", mode="wrap")
    @classmethod
    def _skip_none_validation(cls, value: Any, handler: Callable) -> Any:
        """Skip validation if the value is `None` when initialisation is delayed."""
        if value is None:
            return value
        return handler(value)

705
706
707
708
    @field_validator("tokenizer_mode", mode="after")
    def _lowercase_tokenizer_mode(cls, tokenizer_mode: str) -> str:
        return tokenizer_mode.lower()

709
710
711
712
713
714
715
716
717
    @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":
718
        """Called after __post_init__"""
719
        if not isinstance(self.tokenizer, str):
720
721
722
723
724
            raise ValueError(
                f"tokenizer must be a string, got "
                f"{type(self.tokenizer).__name__}: {self.tokenizer!r}. "
                "Please provide a valid tokenizer path or HuggingFace model ID."
            )
725
        if not isinstance(self.max_model_len, int):
726
727
728
729
730
            raise ValueError(
                f"max_model_len must be a positive integer, "
                f"got {type(self.max_model_len).__name__}: {self.max_model_len!r}. "
                "Example: max_model_len=2048"
            )
731
732
733
        return self

    def _get_transformers_backend_cls(self) -> str:
734
        """Determine which Transformers modeling backend class will be used if
735
        `model_impl` is set to `transformers` or `auto`."""
736
737
738
        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 ""
739
        cls += "MoE" if self.is_moe else ""
740
741
        # Check if the architecture we're wrapping has defaults
        runner = None
742
        task = None
743
        if defaults := try_match_architecture_defaults(self.architectures[0]):
744
745
            _, (runner, task) = defaults
        # User specified value take precedence
746
747
        if self.runner != "auto":
            runner = self.runner
748
749
        # Only consider Transformers modeling backend pooling classes if we're wrapping
        # an architecture that defaults to pooling. Otherwise, we return the LM class
750
751
752
753
754
755
        # and use adapters.
        if runner == "pooling" and task in {"embed", "classify"}:
            if task == "embed":
                cls += "EmbeddingModel"
            elif task == "classify":
                cls += "ForSequenceClassification"
756
757
758
        else:
            cls += "ForCausalLM"
        return cls
759
760

    def using_transformers_backend(self) -> bool:
761
        """Check if the model is using the Transformers modeling backend class."""
762
763
764
        used_cls = self._model_info.architecture
        transformers_backend_cls = self._get_transformers_backend_cls()
        return used_cls == transformers_backend_cls
765
766
767
768
769
770
771

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

    @property
    def architectures(self) -> list[str]:
772
        return self.model_arch_config.architectures
773
774
775
776
777
778

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

779
780
781
782
783
784
785
786
787
    def maybe_pull_model_tokenizer_for_runai(self, model: str, tokenizer: str) -> None:
        """Pull model/tokenizer from Object Storage to temporary
        directory when needed.

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

788
789
790
791
        # Skip if model_weights is already set (model already pulled)
        if self.model_weights:
            return

792
        if not (is_runai_obj_uri(model) or is_runai_obj_uri(tokenizer)):
793
794
            return

795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
        if is_runai_obj_uri(model):
            object_storage_model = ObjectStorageModel(url=model)
            object_storage_model.pull_files(
                model, allow_pattern=["*.model", "*.py", "*.json"]
            )
            self.model_weights = model
            self.model = object_storage_model.dir

            # If tokenizer is same as model, download to same directory
            if model == tokenizer:
                object_storage_model.pull_files(
                    model,
                    ignore_pattern=[
                        "*.pt",
                        "*.safetensors",
                        "*.bin",
                        "*.tensors",
                        "*.pth",
                    ],
                )
                self.tokenizer = object_storage_model.dir
                return

        # Only download tokenizer if needed and not already handled
        if is_runai_obj_uri(tokenizer):
            object_storage_tokenizer = ObjectStorageModel(url=tokenizer)
            object_storage_tokenizer.pull_files(
                model,
                ignore_pattern=["*.pt", "*.safetensors", "*.bin", "*.tensors", "*.pth"],
            )
            self.tokenizer = object_storage_tokenizer.dir
826

827
    def _get_encoder_config(self) -> dict[str, Any] | None:
828
829
830
831
        model = self.model
        if is_remote_gguf(model):
            model, _ = split_remote_gguf(model)
        return get_sentence_transformer_tokenizer_config(model, self.revision)
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871

    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.",
872
873
                runner_type,
            )
874
875
876
877
878
879
880
881
882
883
884
885

        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():
886
887
888
                if runner_type == "generate" and registry.is_text_generation_model(
                    architectures, self
                ):
889
                    return "none"
890
891
892
                if runner_type == "pooling" and registry.is_pooling_model(
                    architectures, self
                ):
893
894
                    return "none"

895
            match = try_match_architecture_defaults(arch, runner_type=runner_type)
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
            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

917
        convert_type = self._get_default_convert_type(architectures, runner_type)
918
919
920
921
922
923

        # 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.",
924
925
                convert_type,
            )
926
927
928
929
930
931

        return convert_type

    def _verify_quantization(self) -> None:
        supported_quantization = me_quant.QUANTIZATION_METHODS
        if self.quantization is not None:
932
            self.quantization = cast(me_quant.QuantizationMethods, self.quantization)
933
934

        # Parse quantization method from the HF model config, if available.
935
        quant_cfg = self.model_arch_config.quantization_config
936
937

        if quant_cfg is not None:
938
            quant_method = quant_cfg["quant_method"]
939
940
941
942
943
944
            # 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 = [
                "gptq_marlin",
                "awq_marlin",
945
                "inc",
946
947
948
                "moe_wna16",
                "modelopt",
                "modelopt_fp4",
949
                "modelopt_mxfp8",
950
                "modelopt_mixed",
951
952
953
                # Ensure heavy backends are probed last to avoid unnecessary
                # imports during override detection (e.g., MXFP4 imports Triton)
                "mxfp4",
954
                "gpt_oss_mxfp4",
Li, Jiang's avatar
Li, Jiang committed
955
                "cpu_awq",
956
                "humming",
957
                "gguf",
958
            ]
959
960
961
            # if the user specifies humming, we should always use humming
            if self.quantization == "humming":
                overrides = ["humming"] + overrides
962
963
964
965
966
967
968
969
970
971
972
973
            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(
974
                    quant_cfg, self.quantization, hf_config=self.hf_config
975
                )
976
977
978
979
                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.
980
981
982
983
                    if (
                        name in get_args(me_quant.QuantizationMethods)
                        and name not in overrides
                    ):
984
985
986
987
                        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 "
988
989
                            "overrides are checked in order of preference."
                        )
990
991
992
993
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break

994
            quant_method = quant_method if quant_method != "" else None
995
996
997
998
999
1000
1001
1002
            # 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 "
1003
1004
                    f"({self.quantization})."
                )
1005
1006
1007
1008
1009

        if self.quantization is not None:
            if self.quantization not in supported_quantization:
                raise ValueError(
                    f"Unknown quantization method: {self.quantization}. Must "
1010
1011
                    f"be one of {supported_quantization}."
                )
1012
1013
            current_platform.verify_quantization(self.quantization)

1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
        if self.quantization in me_quant.DEPRECATED_QUANTIZATION_METHODS:
            if self.allow_deprecated_quantization:
                logger.warning(
                    "The quantization method %s is deprecated "
                    "and will be removed in future versions of vLLM.",
                    self.quantization,
                )
            else:
                raise ValueError(
                    "The quantization method %s is deprecated "
                    "and will be removed in future versions of vLLM. To bypass, "
                    "set `--allow-deprecated-quantization`.",
                    self.quantization,
                )

1029
1030
1031
    def _verify_cuda_graph(self) -> None:
        # CUDAGraph capture not supported for encoder-decoder models on ROCm
        unsupported_rocm = self.is_encoder_decoder
1032
        if unsupported_rocm and not self.enforce_eager and current_platform.is_rocm():
1033
1034
            logger.warning(
                "CUDA graph is not supported for %s on ROCm yet, fallback "
1035
                "to eager mode.",
1036
                self.model_arch_config.model_type,
1037
            )
1038
1039
1040
1041
1042
1043
1044
1045
1046
            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"
1047
        has_quantization_config = self.model_arch_config.quantization_config is not None
1048
        is_8bit = (
1049
            self.model_arch_config.quantization_config.get("load_in_8bit", False)  # type: ignore[union-attr]
1050
1051
1052
1053
1054
            if has_quantization_config
            else False
        )
        if all(
            [
1055
1056
1057
1058
                is_bitsandbytes,
                has_quantization_config,
                is_8bit,
                not self.enforce_eager,
1059
1060
            ]
        ):
1061
1062
            logger.warning(
                "CUDA graph is not supported on BitsAndBytes 8bit yet, "
1063
1064
                "fallback to the eager mode."
            )
1065
1066
1067
1068

            self.enforce_eager = True

    def _verify_with_expert_parallelism(self) -> None:
1069
        if not self.is_moe:
1070
1071
            raise ValueError(
                "Number of experts in the model must be greater than 0 "
1072
1073
                "when expert parallelism is enabled."
            )
1074

1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
    def _try_verify_and_update_model_config(self):
        # Avoid running try_verify_and_update_config multiple times
        if getattr(self, "config_updated", False):
            return

        architecture = self.architecture
        if architecture is None:
            return

        from vllm.model_executor.models.config import (
            MODELS_CONFIG_MAP,
        )

        cls = MODELS_CONFIG_MAP.get(architecture, None)
        if cls is not None:
            cls.verify_and_update_model_config(self)

1092
1093
1094
1095
1096
1097
1098
    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 (
1099
1100
1101
                get_sparse_attention_config,
            )

1102
1103
1104
            sparse_attn_config = get_sparse_attention_config(self, load_config)
            if sparse_attn_config:
                self.hf_config.dual_chunk_attention_config[
1105
1106
1107
1108
1109
1110
                    "sparse_attention_config"
                ] = sparse_attn_config
                if (
                    "sparse_attention_enabled"
                    not in self.hf_config.dual_chunk_attention_config
                ):
1111
                    self.hf_config.dual_chunk_attention_config[
1112
1113
                        "sparse_attention_enabled"
                    ] = True
1114
1115
1116
1117
1118

    def verify_with_parallel_config(
        self,
        parallel_config: ParallelConfig,
    ) -> None:
1119
        total_num_attention_heads = self.model_arch_config.total_num_attention_heads
1120
1121
1122
1123
1124
        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 "
1125
1126
                f"({tensor_parallel_size})."
            )
1127
1128
1129
1130
1131

        if parallel_config.enable_expert_parallel:
            self._verify_with_expert_parallelism()

        pipeline_parallel_size = parallel_config.pipeline_parallel_size
1132
1133
1134
        if pipeline_parallel_size > 1 and not self.registry.is_pp_supported_model(
            self.architectures, self
        ):
1135
1136
            raise NotImplementedError(
                "Pipeline parallelism is not supported for this model. "
1137
1138
                "Supported models implement the `SupportsPP` interface."
            )
1139

1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
        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}"
            )

1157
1158
1159
1160
1161
1162
1163
1164
            num_q_per_kv = total_num_attention_heads // total_num_kv_heads
            assert num_q_per_kv % decode_context_parallel_size == 0, (
                f"Total number of q per kv attn heads ({num_q_per_kv})"
                " must be divisible by dcp world size when enable "
                "decode context parallel for GQA "
                f"({parallel_config.decode_context_parallel_size})."
            )

1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
        # torch_shm uses a single IPC queue to rank 0; DP>1 is
        # incompatible because API servers can't know which
        # CoreEngine the scheduler will assign work to. TP>1 is
        # also not supported because this requires broadcasting
        # MM tensors between all TP ranks.
        if (
            self.multimodal_config is not None
            and self.multimodal_config.mm_tensor_ipc == "torch_shm"
            and parallel_config.world_size_across_dp > 1
        ):
            raise ValueError(
                "mm_tensor_ipc='torch_shm' is not supported with "
                "data_parallel_size > 1 or tensor_parallel_size > 1 "
                "or pipeline_parallel_size > 1."
            )

1181
    def get_sliding_window(self) -> int | None:
1182
1183
1184
1185
        """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:
1186
        return self.model_arch_config.vocab_size
1187
1188

    def get_hidden_size(self) -> int:
1189
        return self.model_arch_config.hidden_size
1190

1191
1192
1193
1194
    def get_inputs_embeds_size(self) -> int:
        # The size of inputs_embeds is usually identical to the size
        # of the hidden states, however there are exceptions, such as
        # embedding models like CLIP and SigLIP
1195
1196
1197
1198
        names = ("projection_dim", "projection_size")
        return getattr_iter(
            self.hf_text_config, names, default_factory=self.get_hidden_size
        )
1199

1200
1201
    @property
    def is_deepseek_mla(self) -> bool:
1202
        return self.model_arch_config.is_deepseek_mla
1203

1204
    @property
1205
    def is_mm_prefix_lm(self) -> bool:
1206
        return self.model_arch_config.is_mm_prefix_lm
1207

1208
    def get_head_size(self) -> int:
1209
        return self.model_arch_config.head_size
1210
1211
1212

    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
1213
        return self.model_arch_config.total_num_kv_heads
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225

    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.
1226
        return max(1, total_num_kv_heads // parallel_config.tensor_parallel_size)
1227
1228

    def get_num_attention_heads(self, parallel_config: ParallelConfig) -> int:
1229
        num_heads = self.model_arch_config.total_num_attention_heads
1230
1231
        return num_heads // parallel_config.tensor_parallel_size

1232
    def get_num_experts(self) -> int:
1233
        return self.model_arch_config.num_experts
1234

1235
    def get_total_num_hidden_layers(self) -> int:
1236
        return self.model_arch_config.total_num_hidden_layers
1237
1238
1239
1240
1241
1242
1243
1244

    def get_layers_start_end_indices(
        self, parallel_config: ParallelConfig
    ) -> tuple[int, int]:
        from vllm.distributed.utils import get_pp_indices

        total_num_hidden_layers = self.get_total_num_hidden_layers()

1245
        # the layout order is: DP x PP x TP
1246
1247
1248
        pp_rank = (
            parallel_config.rank // parallel_config.tensor_parallel_size
        ) % parallel_config.pipeline_parallel_size
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
        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,
1260
        block_type: LayerBlockType = "attention",
1261
1262
1263
    ) -> int:
        # This function relies on 'layers_block_type' in hf_config,
        # for w/o this attribute, we will need to have workarounds like so
1264
        attn_block_type = block_type == "attention"
1265
1266
1267
        is_transformer = (
            not self.is_hybrid and not self.has_noops and not self.is_attention_free
        )
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
        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
1280
            return sum(not bc.attention.no_op for bc in block_configs[start:end])
1281
1282
        else:
            # Hybrid model Jamba
1283
1284
1285
            layers_block_type_value = getattr(
                self.hf_text_config, "layers_block_type", None
            )
1286
            if layers_block_type_value is not None:
1287
                if self.model_arch_config.text_model_type == "zamba2":
1288
                    if attn_block_type:
1289
1290
1291
                        return sum(
                            t == "hybrid" for t in layers_block_type_value[start:end]
                        )
1292
1293
                    else:
                        return self.get_num_layers(parallel_config)
1294
                return sum(t == block_type for t in layers_block_type_value[start:end])
1295
1296
1297
1298
1299
1300

            # 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])

1301
1302
            # Hybrid model Qwen3Next Qwen3.5 Series
            layer_types_value = getattr(self.hf_text_config, "layer_types", None)
1303
            if layer_types_value is not None:
1304
                if block_type == "attention":
1305
1306
1307
                    return sum(
                        t == "full_attention" for t in layer_types_value[start:end]
                    )
1308
                elif block_type == "linear_attention":
1309
1310
1311
                    return sum(
                        t == "linear_attention" for t in layer_types_value[start:end]
                    )
1312
                else:
1313
                    return sum(t == block_type for t in layer_types_value[start:end])
1314
1315
1316
1317
1318
1319

            if (
                layers_block_type_value is None
                and attn_type_list is None
                and layer_types_value is None
            ):
1320
                raise ValueError(
1321
1322
1323
                    "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"
1324
                )
1325
            raise AssertionError(f"Unsupported block type: {block_type}")
1326

1327
    def get_mamba_chunk_size(self) -> int | None:
1328
1329
1330
1331
1332
1333
1334
1335
        """
        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)
1336
1337
1338
1339
1340
1341

        # Since Mamba1 does not have a chunk notion
        # we use a default chunk size of 1024.
        if chunk_size is None:
            chunk_size = 2048

1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
        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,
1373
                config_format=self.config_format,
1374
                hf_token=self.hf_token,
1375
1376
1377
1378
1379
            )
        else:
            config = try_get_generation_config(
                self.generation_config,
                trust_remote_code=self.trust_remote_code,
1380
                config_format=self.config_format,
1381
                hf_token=self.hf_token,
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
            )

        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.
        """
1404
1405
1406
        src = self.generation_config

        config = {} if src == "vllm" else self.try_get_generation_config()
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420

        # 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 = {
1421
                p: config.get(p) for p in available_params if config.get(p) is not None
1422
1423
1424
1425
1426
            }
            # 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(
1427
1428
                    "max_new_tokens"
                )
1429
1430
1431
        else:
            diff_sampling_param = {}

1432
        if diff_sampling_param and src != "vllm":
1433
            logger.warning_once(
1434
1435
1436
1437
1438
1439
                "Default vLLM sampling parameters have been overridden by %s: `%s`. "
                "If this is not intended, please relaunch vLLM instance "
                "with `--generation-config vllm`.",
                "the model's `generation_config.json`" if src == "auto" else src,
                str(diff_sampling_param),
                scope="local",
1440
            )
1441

1442
1443
        return diff_sampling_param

1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
    def get_pooling_task(
        self, supported_tasks: tuple[SupportedTask, ...]
    ) -> PoolingTask | None:
        if self.pooler_config is None:
            return None

        pooling_task = self.pooler_config.task

        if pooling_task is not None:
            if self.pooler_config.task in supported_tasks:
                return self.pooler_config.task
            else:
                raise RuntimeError(
                    f"Unsupported task: {pooling_task!r} "
                    f"Supported tasks: {supported_tasks}"
                )

        if "token_classify" in supported_tasks:
            for architecture in self.architectures:
                if "ForTokenClassification" in architecture:
                    return "token_classify"

        priority: list[PoolingTask] = [
            "embed&token_classify",
            "embed",
            "classify",
            "token_embed",
            "token_classify",
            "plugin",
        ]
        for task in priority:
            if task in supported_tasks:
                return task
        return None

1479
    @cached_property
1480
1481
1482
1483
    def is_encoder_decoder(self) -> bool:
        """Extract the HF encoder/decoder model flag."""
        return is_encoder_decoder(self.hf_config)

1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
    @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)
                    )
                )
            )
        )

1507
1508
1509
1510
    @property
    def uses_mrope(self) -> bool:
        return uses_mrope(self.hf_config)

1511
1512
1513
1514
    @property
    def uses_xdrope_dim(self) -> int:
        return uses_xdrope_dim(self.hf_config)

1515
1516
1517
1518
1519
1520
1521
1522
    @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

Patrick von Platen's avatar
Patrick von Platen committed
1523
1524
1525
1526
    @property
    def requires_raw_input_tokens(self) -> bool:
        return self._model_info.requires_raw_input_tokens

1527
    @property
1528
1529
    def score_type(self) -> ScoreType:
        """
1530
1531
1532
1533
1534
        Scoring API handles score/rerank for:

        - "classify" task (score_type: cross-encoder models)
        - "embed" task (score_type: bi-encoder models)
        - "token_embed" task (score_type: late interaction models)
1535
1536
1537
1538
1539
        """
        # fixme: self._model_info.score_type is the score type before
        #  as_seq_cls_model, which is "bi-encoder", rather than the
        #  score type after as_seq_cls_model, which is "cross-encoder".
        #  Therefore, the following logic is required.
1540
        return (
1541
1542
1543
            "cross-encoder"
            if self.convert_type == "classify"
            else self._model_info.score_type
1544
        )
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555

    @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:
1556
1557
        if not self._model_info.is_hybrid:
            return False
1558
1559
1560
        # Handle granite-4.0-micro case which uses hybrid config but does not
        # actually contain any non-attention layers.
        layer_types = getattr(self.hf_config, "layer_types", None)
1561
        return layer_types is None or not all(
1562
            layer == "attention" for layer in layer_types
1563
        )
1564
1565
1566
1567
1568
1569
1570
1571
1572

    @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

1573
1574
1575
1576
    @property
    def supports_mamba_prefix_caching(self) -> bool:
        return self._model_info.supports_mamba_prefix_caching

1577
1578
1579
1580
1581
1582
    @property
    def use_mla(self) -> bool:
        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE

    @property
    def is_matryoshka(self) -> bool:
1583
1584
1585
        return bool(getattr(self.hf_config, "matryoshka_dimensions", None)) or getattr(
            self.hf_config, "is_matryoshka", False
        )
1586
1587
1588
1589
1590
1591

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

    @property
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
    def use_sep_token(self) -> bool:
        # cross_encoder models defaults to using separating token.
        # `llm as reranker` defaults to not using separating token.

        use_pad_token = getattr(self.hf_config, "use_pad_token", None)
        if use_pad_token is not None:
            logger.warning_once(
                "use_pad_token has been deprecated; please use use_sep_token instead."
            )
            return use_pad_token

        return getattr(self.hf_config, "use_sep_token", True)
1604
1605
1606
1607
1608
1609
1610
1611

    @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.

1612
1613
1614
1615
        `head_dtype` currently only supports pooling models.

        - The pooling model defaults to using fp32 head, you can use
          --hf-overrides '{"head_dtype": "model"}' to disable it.
1616
1617
        """

1618
1619
1620
        head_dtype = _get_head_dtype(
            config=self.hf_config, dtype=self.dtype, runner_type=self.runner_type
        )
1621
1622
1623

        if self.runner_type != "pooling" and head_dtype != self.dtype:
            logger.warning_once(
1624
                "`head_dtype` currently only supports pooling models, "
1625
1626
1627
                "fallback to model dtype [%s].",
                self.dtype,
            )
1628
1629
1630
1631
1632
            return self.dtype

        if head_dtype not in current_platform.supported_dtypes:
            logger.warning_once(
                "The current platform does not support [%s] head dtype, "
1633
1634
1635
1636
                "fallback to model dtype [%s].",
                head_dtype,
                self.dtype,
            )
1637
1638
1639
1640
1641
            return self.dtype

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

1642
1643
    @property
    def embedding_size(self):
chengchengpei's avatar
chengchengpei committed
1644
1645
1646
1647
        # Check for embedding_size set by model config (e.g., Voyage models)
        override = getattr(self.hf_config, "embedding_size", None)
        if override is not None:
            return override
1648
1649
1650
        dense_modules = try_get_dense_modules(self.model, revision=self.revision)
        if dense_modules is not None:
            return dense_modules[-1]["out_features"]
1651
        return self.get_hidden_size()
1652

1653
    def get_and_verify_max_len(self, max_model_len: int):
1654
1655
1656
        # Consider max_model_len in tokenizer_config only when
        # pooling models use absolute position_embedding.
        tokenizer_config = None
1657
1658
1659
1660
        if (
            self.runner_type == "pooling"
            and getattr(self.hf_config, "position_embedding_type", "") == "absolute"
        ):
1661
            tokenizer_config = try_get_tokenizer_config(
1662
                self.tokenizer,
1663
                trust_remote_code=self.trust_remote_code,
1664
                revision=self.tokenizer_revision,
1665
            )
1666
        max_model_len = _get_and_verify_max_len(
1667
            hf_config=self.hf_text_config,
1668
            model_arch_config=self.model_arch_config,
1669
            tokenizer_config=tokenizer_config,
1670
            max_model_len=max_model_len,
1671
1672
1673
            disable_sliding_window=self.disable_sliding_window,
            sliding_window=self.get_sliding_window(),
            spec_target_max_model_len=self.spec_target_max_model_len,
1674
1675
            encoder_config=self.encoder_config,
        )
1676
1677
        logger.info("Using max model len %s", max_model_len)
        return max_model_len
1678

1679
1680
    @property
    def attn_type(self) -> AttnTypeStr:
1681
        """Determine the attention type based on model configuration."""
1682
        if self.pooler_config is not None:
1683
1684
            seq_pooling_type = self._model_info.default_seq_pooling_type
            if seq_pooling_type == "CLS":
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
                return "encoder_only"
            else:
                is_causal = getattr(self.hf_config, "is_causal", True)
                return "encoder_only" if not is_causal else self._model_info.attn_type
        elif self.is_hybrid:
            return "hybrid"
        elif self.is_attention_free:
            return "attention_free"
        elif self.is_encoder_decoder:
            return "encoder_decoder"
        else:
            return "decoder"

    @property
    def is_chunked_prefill_supported(self) -> bool:
        attn_type = self.attn_type
1701
1702

        if pooler_config := self.pooler_config:
1703
1704
1705
            # for pooling models
            if attn_type == "encoder_only":
                logger.debug(
1706
1707
                    "Pooling models with bidirectional attn "
                    "do not support chunked prefill."
1708
1709
                )
                return False
1710
1711
1712
1713
1714
1715

            if attn_type == "decoder":
                if (
                    pooler_config.seq_pooling_type in ("MEAN", "CLS")
                    or pooler_config.tok_pooling_type == "STEP"
                ):
1716
                    logger.debug(
1717
1718
1719
1720
                        "Pooling models with causal attn and %s/%s pooling "
                        "do not support chunked prefill.",
                        pooler_config.seq_pooling_type,
                        pooler_config.tok_pooling_type,
1721
1722
                    )
                    return False
1723
                else:
1724
                    logger.debug(
1725
1726
1727
1728
                        "Pooling models with causal attn and %s/%s pooling "
                        "support chunked prefill.",
                        pooler_config.seq_pooling_type,
                        pooler_config.tok_pooling_type,
1729
1730
                    )
                    return True
1731

1732
1733
1734
1735
            # vllm currently does not have pooling models using hybrid,
            # attention_free or encoder_decoder attn types.
            return attn_type != "encoder_decoder"
        else:
1736
            # for generative models
1737
            if attn_type == "encoder_decoder":
1738
                logger.debug("Encoder decoder models do not support chunked prefill.")
1739
                return False
1740

1741
1742
1743
1744
1745
1746
            logger.debug("Generative models support chunked prefill.")
            return True

    @property
    def is_prefix_caching_supported(self) -> bool:
        attn_type = self.attn_type
1747
1748

        if pooler_config := self.pooler_config:
1749
1750
1751
            # for pooling models
            if attn_type == "encoder_only":
                logger.debug(
1752
1753
                    "Pooling models with bidirectional attn "
                    "do not support prefix caching."
1754
1755
                )
                return False
1756
1757
1758
1759
1760
1761

            if attn_type == "decoder":
                if (
                    pooler_config.seq_pooling_type in ("MEAN", "CLS")
                    or pooler_config.tok_pooling_type == "STEP"
                ):
1762
                    logger.debug(
1763
1764
1765
1766
                        "Pooling models with causal attn and %s/%s pooling "
                        "do not support prefix caching.",
                        pooler_config.seq_pooling_type,
                        pooler_config.tok_pooling_type,
1767
1768
                    )
                    return False
1769
                else:
1770
                    logger.debug(
1771
1772
1773
1774
                        "Pooling models with causal attn and %s/%s pooling "
                        "support prefix caching.",
                        pooler_config.seq_pooling_type,
                        pooler_config.tok_pooling_type,
1775
1776
                    )
                    return True
1777

1778
1779
1780
1781
            # vllm currently does not have pooling models using hybrid,
            # attention_free or encoder_decoder attn types.
            return False
        else:
1782
            # for generative models
1783
1784
            if attn_type == "hybrid":
                logger.debug(
1785
                    "Hybrid models do not support prefix caching since the feature "
1786
1787
1788
1789
1790
                    "is still experimental."
                )
                return False
            elif attn_type == "attention_free":
                logger.debug(
1791
                    "Attention free models do not support prefix caching since the "
1792
1793
1794
1795
                    "feature is still experimental."
                )
                return False
            elif attn_type == "encoder_decoder":
1796
                logger.debug("Encoder decoder models do not support prefix caching.")
1797
1798
1799
1800
1801
                return False
            else:  # attn_type == "decoder"
                logger.debug("Generative models support prefix caching.")
                return True

1802
1803
1804
    @property
    def is_moe(self) -> bool:
        return self.get_num_experts() > 0
1805

1806
    @property
1807
1808
1809
    def is_quantized(self) -> bool:
        return getattr(self.hf_config, "quantization_config", None) is not None

1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
    def is_nvfp4_quantized(self) -> bool:
        # ModelOpt NVFP4 checkpoints resolve to modelopt_fp4 quantization method
        if self.quantization in ("modelopt_fp4",):
            return True

        # For Compressed Tensors we look for `"format": "nvfp4-pack-quantized"`
        # in the quantization config
        quant_config = self.model_arch_config.quantization_config
        return (
            self.quantization == "compressed-tensors"
            and quant_config is not None
            and "nvfp4" in quant_config.get("format", "").lower()
        )

1824

1825
def get_served_model_name(model: str, served_model_name: str | list[str] | None):
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
    """
    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")),
1851
    ("ForTokenClassification", ("pooling", "classify")),
1852
1853
1854
1855
    ("ForAudioClassification", ("pooling", "classify")),
    ("ForImageClassification", ("pooling", "classify")),
    ("ForVideoClassification", ("pooling", "classify")),
    ("ClassificationModel", ("pooling", "classify")),
1856
1857
    ("ForRewardModeling", ("pooling", "embed")),
    ("RewardModel", ("pooling", "embed")),
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
    # 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,
    *,
1870
1871
1872
    runner_type: RunnerType | None = None,
    convert_type: ConvertType | None = None,
) -> tuple[str, tuple[RunnerType, ConvertType]] | None:
1873
1874
1875
1876
1877
1878
1879
1880
1881
    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)
        ):
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
            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,
}

1895
1896
1897
1898
1899

def str_dtype_to_torch_dtype(type: str):
    return _STR_DTYPE_TO_TORCH_DTYPE.get(type)


1900
1901
1902
1903
# 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.",
1904
    "gemma3_text": "Numerical instability. Please use bfloat16 or float32 instead.",
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
    "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]
1920
1921
1922
        raise ValueError(
            f"The model type {model_type!r} does not support float16. Reason: {reason}"
        )
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933

    return True


def _resolve_auto_dtype(
    model_type: str,
    config_dtype: torch.dtype,
    *,
    is_pooling_model: bool,
):
    supported_dtypes = [
1934
1935
        dtype
        for dtype in current_platform.supported_dtypes
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
        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(
1962
        "Your device %s doesn't support %s. Falling back to %s for compatibility.",
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
        device_str,
        config_dtype,
        preferred_dtype,
    )

    return preferred_dtype


def _get_and_verify_dtype(
    model_id: str,
    config: PretrainedConfig,
1974
    dtype: str | torch.dtype,
1975
1976
    *,
    is_pooling_model: bool,
1977
    revision: str | None = None,
1978
    config_format: str | ConfigFormat = "hf",
1979
) -> torch.dtype:
1980
    config_dtype = ModelArchConfigConvertorBase.get_torch_dtype(
1981
        config, model_id, revision=revision, config_format=config_format
1982
    )
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
    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


2019
2020
2021
def _get_head_dtype(
    config: PretrainedConfig, dtype: torch.dtype, runner_type: str
) -> torch.dtype:
2022
    head_dtype: str | torch.dtype | None = getattr(config, "head_dtype", None)
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044

    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,
2045
    model_arch_config: ModelArchitectureConfig,
2046
2047
    tokenizer_config: dict | None,
    max_model_len: int | None,
2048
    disable_sliding_window: bool,
2049
2050
    sliding_window: int | None,
    spec_target_max_model_len: int | None = None,
2051
    encoder_config: dict[str, Any] | None = None,
2052
2053
) -> int:
    """Get and verify the model's maximum length."""
2054
2055
2056
    (derived_max_model_len, max_len_key) = (
        model_arch_config.derived_max_model_len_and_key
    )
2057
2058
2059

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
2060
2061
2062
2063
2064
    if (
        disable_sliding_window
        and sliding_window is not None
        and sliding_window < derived_max_model_len
    ):
2065
2066
2067
2068
2069
2070
        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(
2071
2072
2073
            "model_max_length", derived_max_model_len
        )
        derived_max_model_len = min(derived_max_model_len, tokenizer_model_max_length)
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088

    # 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(
2089
2090
2091
            "The model's config.json does not contain any of the keys "
            "to determine the original maximum length of the model. "
            "Assuming the model's maximum length is %d.",
2092
2093
            default_max_len,
        )
2094
2095
        derived_max_model_len = default_max_len

2096
2097
2098
    # In Transformers v5 rope_parameters could be TypedDict or dict[str, TypedDict].
    # To simplify the verification, we convert it to dict[str, TypedDict].
    rope_parameters = getattr(hf_config, "rope_parameters", None)
2099
    if rope_parameters and not is_rope_parameters_nested(rope_parameters):
2100
2101
        rope_parameters = {"": rope_parameters}

2102
2103
    # NOTE(woosuk): Gemma3's max_model_len (128K) is already scaled by RoPE
    # scaling, so we skip applying the scaling factor again.
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
    if rope_parameters is not None and "gemma3" not in hf_config.model_type:
        scaling_factor = 1.0
        for rp in rope_parameters.values():
            # No need to consider "type" key because of patch_rope_parameters when
            # loading HF config
            rope_type = rp["rope_type"]

            if rope_type not in ("su", "longrope", "llama3"):
                # NOTE: rope_type == "default" does not define factor https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/modeling_rope_utils.py
                # NOTE: This assumes all layer types have the same scaling factor.
                scaling_factor = rp.get("factor", scaling_factor)

                if rope_type == "yarn":
                    derived_max_model_len = rp["original_max_position_embeddings"]
2118
2119
2120
2121
2122
2123
2124
2125
2126
        if scaling_factor is None:
            # Fallback the factor to 1.0 if a user assigned `null`
            logger.warning_once(
                "The model's RoPE configuration has a null scaling "
                "factor which is unexpected. This likely indicates a bug "
                "in the model's HuggingFace config.json. Please notify the "
                "model vendor. Falling back the value to 1.0. "
            )
            scaling_factor = 1.0
2127
2128
        # Do this outside loop since all layer types should have the same scaling
        derived_max_model_len *= scaling_factor
2129
2130
2131
2132

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

2133
2134
2135
2136
    # If the user didn't specify `max_model_len` or specified -1 (auto-fit),
    # then use that derived from the model config as a default value.
    # When -1 is specified, the engine will later auto-fit to available memory.
    if max_model_len is None or max_model_len == -1:
2137
2138
        # For LongRoPE, default to original_max_position_embeddings to avoid
        # performance degradation for shorter sequences
2139
2140
2141
        if rope_parameters is not None and any(
            rp["rope_type"] == "longrope" for rp in rope_parameters.values()
        ):
2142
2143
2144
2145
2146
2147
2148
            max_model_len = int(
                getattr(
                    hf_config, "original_max_position_embeddings", derived_max_model_len
                )
            )
        else:
            max_model_len = int(derived_max_model_len)
2149
        max_model_len = current_platform.check_max_model_len(max_model_len)
2150

2151
2152
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
2153
2154
2155
2156
2157
    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)
2158
        if model_max_length is None or max_model_len > model_max_length:
2159
2160
2161
2162
            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="
2163
2164
                f"{model_max_length} in model's config.json)."
            )
2165
2166
2167
2168
2169
2170
            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 "
2171
2172
                "error."
            )
2173
2174
2175
2176
2177
            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 "
2178
2179
                    f"the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1. {warning}"
                )
2180
    return int(max_model_len)