config.py 41.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import os
5
from collections.abc import Callable
6
from dataclasses import asdict
7
from functools import cache, partial
8
from importlib.metadata import version
9
from pathlib import Path
10
from typing import Any, Literal, TypeAlias
Jasmond L's avatar
Jasmond L committed
11

Joe Runde's avatar
Joe Runde committed
12
import huggingface_hub
13
14
15
from huggingface_hub import (
    get_safetensors_metadata,
)
16
from packaging.version import Version
17
from transformers import GenerationConfig, PretrainedConfig
18
from transformers.models.auto.image_processing_auto import get_image_processor_config
19
20
21
22
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_MAPPING_NAMES,
)
23
from transformers.models.auto.tokenization_auto import get_tokenizer_config
24
from transformers.utils import CONFIG_NAME as HF_CONFIG_NAME
25

26
from vllm import envs
27
from vllm.logger import init_logger
28
from vllm.transformers_utils.utils import parse_safetensors_file_metadata
29
30

from .config_parser_base import ConfigParserBase
31
32
33
34
35
36
from .gguf_utils import (
    check_gguf_file,
    is_gguf,
    is_remote_gguf,
    split_remote_gguf,
)
37
from .repo_utils import (
38
39
40
41
42
43
44
    _get_hf_token,
    file_or_path_exists,
    get_hf_file_to_dict,
    list_repo_files,
    try_get_local_file,
    with_retry,
)
45

46
47
48
49
50
51
52
53
54
55
try:
    # Transformers v5
    from transformers.configuration_utils import ALLOWED_ATTENTION_LAYER_TYPES
except ImportError:
    # Transformers v4
    from transformers.configuration_utils import (
        ALLOWED_LAYER_TYPES as ALLOWED_ATTENTION_LAYER_TYPES,
    )


56
if envs.VLLM_USE_MODELSCOPE:
57
58
59
    from modelscope import AutoConfig
else:
    from transformers import AutoConfig
60

61
62
MISTRAL_CONFIG_NAME = "params.json"

63
64
logger = init_logger(__name__)

65

66
67
class LazyConfigDict(dict):
    def __getitem__(self, key):
68
69
70
        if isinstance(value := super().__getitem__(key), type):
            return value

71
        import vllm.transformers_utils.configs as configs
72

73
        return getattr(configs, value)
74
75
76


_CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
77
    afmoe="AfmoeConfig",
78
    bagel="BagelConfig",
79
80
    chatglm="ChatGLMConfig",
    deepseek_vl_v2="DeepseekVLV2Config",
81
    deepseek_v32="DeepseekV3Config",
82
    flex_olmo="FlexOlmoConfig",
83
    hunyuan_vl="HunYuanVLConfig",
84
    kimi_linear="KimiLinearConfig",
85
86
87
88
89
90
    kimi_vl="KimiVLConfig",
    RefinedWeb="RWConfig",  # For tiiuae/falcon-40b(-instruct)
    RefinedWebModel="RWConfig",  # For tiiuae/falcon-7b(-instruct)
    jais="JAISConfig",
    mlp_speculator="MLPSpeculatorConfig",
    medusa="MedusaConfig",
91
    midashenglm="MiDashengLMConfig",
92
93
94
    eagle="EAGLEConfig",
    speculators="SpeculatorsConfig",
    nemotron="NemotronConfig",
95
    olmo3="Olmo3Config",
96
97
98
99
    ovis="OvisConfig",
    ultravox="UltravoxConfig",
    step3_vl="Step3VLConfig",
    step3_text="Step3TextConfig",
100
    qwen3_next="Qwen3NextConfig",
Paul Pak's avatar
Paul Pak committed
101
    lfm2_moe="Lfm2MoeConfig",
102
    tarsier2="Tarsier2Config",
103
)
104

105
106
107
108
_CONFIG_ATTRS_MAPPING: dict[str, str] = {
    "llm_config": "text_config",
}

109
_AUTO_CONFIG_KWARGS_OVERRIDES: dict[str, dict[str, Any]] = {
110
    "internvl_chat": {"has_no_defaults_at_init": True},
111
    "Llama_Nemotron_Nano_VL": {"attn_implementation": "eager"},
112
    "NVLM_D": {"has_no_defaults_at_init": True},
113
114
}

115

116
117
118
119
120
121
122
123
def is_rope_parameters_nested(rope_parameters: dict[str, Any]) -> bool:
    """Check if rope_parameters is nested by layer types."""
    # Cannot be nested if rope_parameters is empty
    if not rope_parameters:
        return False
    return set(rope_parameters.keys()).issubset(ALLOWED_ATTENTION_LAYER_TYPES)


124
class HFConfigParser(ConfigParserBase):
125
126
    def parse(
        self,
127
        model: str | Path,
128
        trust_remote_code: bool,
129
130
        revision: str | None = None,
        code_revision: str | None = None,
131
132
        **kwargs,
    ) -> tuple[dict, PretrainedConfig]:
133
134
135
136
137
138
139
140
141
142
143
        kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
        config_dict, _ = PretrainedConfig.get_config_dict(
            model,
            revision=revision,
            code_revision=code_revision,
            token=_get_hf_token(),
            **kwargs,
        )
        # Use custom model class if it's in our registry
        model_type = config_dict.get("model_type")
        if model_type is None:
144
145
146
147
148
            model_type = (
                "speculators"
                if config_dict.get("speculators_config") is not None
                else model_type
            )
149
150
151
        # Allow hf_overrides to override model_type before checking _CONFIG_REGISTRY
        if (hf_overrides := kwargs.pop("hf_overrides", None)) is not None:
            model_type = hf_overrides.get("model_type", model_type)
152
153
154
155
156
157
158
159
160
161
162
163

        if model_type in _CONFIG_REGISTRY:
            config_class = _CONFIG_REGISTRY[model_type]
            config = config_class.from_pretrained(
                model,
                revision=revision,
                code_revision=code_revision,
                token=_get_hf_token(),
                **kwargs,
            )
        else:
            try:
164
                kwargs = _maybe_update_auto_config_kwargs(kwargs, model_type=model_type)
165
166
167
168
169
170
171
172
173
                config = AutoConfig.from_pretrained(
                    model,
                    trust_remote_code=trust_remote_code,
                    revision=revision,
                    code_revision=code_revision,
                    token=_get_hf_token(),
                    **kwargs,
                )
            except ValueError as e:
174
175
176
177
                if (
                    not trust_remote_code
                    and "requires you to execute the configuration file" in str(e)
                ):
178
179
180
181
182
                    err_msg = (
                        "Failed to load the model config. If the model "
                        "is a custom model not yet available in the "
                        "HuggingFace transformers library, consider setting "
                        "`trust_remote_code=True` in LLM or using the "
183
184
                        "`--trust-remote-code` flag in the CLI."
                    )
185
186
187
188
189
190
191
192
                    raise RuntimeError(err_msg) from e
                else:
                    raise e
        config = _maybe_remap_hf_config_attrs(config)
        return config_dict, config


class MistralConfigParser(ConfigParserBase):
193
194
    def parse(
        self,
195
        model: str | Path,
196
        trust_remote_code: bool,
197
198
        revision: str | None = None,
        code_revision: str | None = None,
199
200
        **kwargs,
    ) -> tuple[dict, PretrainedConfig]:
201
202
203
        # This function loads a params.json config which
        # should be used when loading models in mistral format
        config_dict = _download_mistral_config_file(model, revision)
204
205
206
        if (
            max_position_embeddings := config_dict.get("max_position_embeddings")
        ) is None:
207
            max_position_embeddings = _maybe_retrieve_max_pos_from_hf(
208
209
                model, revision, **kwargs
            )
210
211
212
213
            config_dict["max_position_embeddings"] = max_position_embeddings

        from vllm.transformers_utils.configs.mistral import adapt_config_dict

214
215
216
217
218
219
220
221
222
223
224
225
226
        # Get missing fields from HF config if available
        try:
            hf_config_dict, _ = PretrainedConfig.get_config_dict(
                model,
                revision=revision,
                code_revision=code_revision,
                token=_get_hf_token(),
                **kwargs,
            )
        except OSError:  # Not found
            hf_config_dict = {}

        config = adapt_config_dict(config_dict, defaults=hf_config_dict)
227
228
229

        # Mistral configs may define sliding_window as list[int]. Convert it
        # to int and add the layer_types list[str] to make it HF compatible
230
231
232
        if (sliding_window := getattr(config, "sliding_window", None)) and isinstance(
            sliding_window, list
        ):
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
            pattern_repeats = config.num_hidden_layers // len(sliding_window)
            layer_types = sliding_window * pattern_repeats
            config.layer_types = [
                "full_attention" if layer_type is None else "sliding_attention"
                for layer_type in layer_types
            ]
            config.sliding_window = next(filter(None, sliding_window), None)

        return config_dict, config


_CONFIG_FORMAT_TO_CONFIG_PARSER: dict[str, type[ConfigParserBase]] = {
    "hf": HFConfigParser,
    "mistral": MistralConfigParser,
}

ConfigFormat = Literal[
    "auto",
    "hf",
    "mistral",
]


def get_config_parser(config_format: str) -> ConfigParserBase:
    """Get the config parser for a given config format."""
    if config_format not in _CONFIG_FORMAT_TO_CONFIG_PARSER:
        raise ValueError(f"Unknown config format `{config_format}`.")
    return _CONFIG_FORMAT_TO_CONFIG_PARSER[config_format]()


def register_config_parser(config_format: str):
    """Register a customized vllm config parser.
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
     When a config format is not supported by vllm, you can register a customized
    config parser to support it.
     Args:
         config_format (str): The config parser format name.
     Examples:

         >>> from vllm.transformers_utils.config import (get_config_parser,
                                                         register_config_parser)
         >>> from vllm.transformers_utils.config_parser_base import ConfigParserBase
         >>>
         >>> @register_config_parser("custom_config_parser")
         ... class CustomConfigParser(ConfigParserBase):
         ...     def parse(
         ...         self,
         ...         model: Union[str, Path],
         ...         trust_remote_code: bool,
281
282
         ...         revision: str | None = None,
         ...         code_revision: str | None = None,
283
284
285
286
287
288
         ...         **kwargs,
         ...     ) -> tuple[dict, PretrainedConfig]:
         ...         raise NotImplementedError
         >>>
         >>> type(get_config_parser("custom_config_parser"))
         <class 'CustomConfigParser'>
289
290
291
292
293
294
    """  # noqa: E501

    def _wrapper(config_parser_cls):
        if config_format in _CONFIG_FORMAT_TO_CONFIG_PARSER:
            logger.warning(
                "Config format `%s` is already registered, and will be "
295
296
297
298
                "overwritten by the new parser class `%s`.",
                config_format,
                config_parser_cls,
            )
299
        if not issubclass(config_parser_cls, ConfigParserBase):
300
301
302
            raise ValueError(
                "The config parser must be a subclass of `ConfigParserBase`."
            )
303
        _CONFIG_FORMAT_TO_CONFIG_PARSER[config_format] = config_parser_cls
304
305
306
307
308
        logger.info(
            "Registered config parser `%s` with config format `%s`",
            config_parser_cls,
            config_format,
        )
309
310
311
        return config_parser_cls

    return _wrapper
312
313


314
315
316
317
318
319
320
321
322
323
def set_default_rope_theta(config: PretrainedConfig, default_theta: float) -> None:
    """Some models may have no rope_theta in their config but still use RoPE.
    This function sets a default rope_theta if it's missing."""
    if getattr(config, "rope_parameters", None) is None:
        config.rope_parameters = {"rope_type": "default"}
    if "rope_theta" not in config.rope_parameters:
        config.rope_parameters["rope_theta"] = default_theta


def patch_rope_parameters(config: PretrainedConfig) -> None:
324
    """Provide backwards compatibility for RoPE."""
325
326
    from vllm.config.utils import getattr_iter

327
328
329
330
331
332
333
    # Older custom models may use non-standard field names
    # which need patching for both Transformers v4 and v5.
    names = ["rope_theta", "rotary_emb_base"]
    rope_theta = getattr_iter(config, names, None, warn=True)
    names = ["partial_rotary_factor", "rotary_pct", "rotary_emb_fraction"]
    partial_rotary_factor = getattr_iter(config, names, None, warn=True)

334
335
    if Version(version("transformers")) < Version("5.0.0.dev0"):
        # Transformers v4 installed, legacy config fields may be present
336
337
        if (rope_scaling := getattr(config, "rope_scaling", None)) is not None:
            config.rope_parameters = rope_scaling
338
339
340
341
        if (
            rope_theta is not None or partial_rotary_factor is not None
        ) and not getattr(config, "rope_parameters", None):
            config.rope_parameters = {"rope_type": "default"}
342
        if rope_theta is not None:
343
            config.rope_parameters["rope_theta"] = rope_theta
344
345
        if partial_rotary_factor is not None:
            config.rope_parameters["partial_rotary_factor"] = partial_rotary_factor
346
    elif rope_theta is not None or getattr(config, "rope_parameters", None):
347
        # Transformers v5 installed
348
349
350
351
352
353
        # Patch these fields in case they used non-standard names
        if rope_theta is not None:
            config.rope_theta = rope_theta
        if partial_rotary_factor is not None:
            config.partial_rotary_factor = partial_rotary_factor
        # Standardize and validate RoPE parameters
354
355
        config.standardize_rope_params()
        config.validate_rope()
356
357

    # No RoPE parameters to patch
358
    if getattr(config, "rope_parameters", None) is None:
359
360
        return

361
362
363
364
    # Add original_max_position_embeddings if present
    if ompe := getattr(config, "original_max_position_embeddings", None):
        config.rope_parameters["original_max_position_embeddings"] = ompe

365
    # Handle nested rope_parameters in interleaved sliding attention models
366
    if is_rope_parameters_nested(config.rope_parameters):
367
        for rope_parameters_layer_type in config.rope_parameters.values():
368
369
            patch_rope_parameters_dict(rope_parameters_layer_type)
    else:
370
        patch_rope_parameters_dict(config.rope_parameters)
371
372


373
374
375
376
def patch_rope_parameters_dict(rope_parameters: dict[str, Any]) -> None:
    if "rope_type" in rope_parameters and "type" in rope_parameters:
        rope_type = rope_parameters["rope_type"]
        rope_type_legacy = rope_parameters["type"]
377
378
379
380
381
        if (rope_type_legacy == "su" and rope_type == "longrope") or (
            rope_type_legacy == "mrope" and rope_type == "default"
        ):
            pass  # No action needed
        elif rope_type != rope_type_legacy:
382
383
384
            raise ValueError(
                f"Found conflicts between 'rope_type={rope_type}' (modern "
                f"field) and 'type={rope_type_legacy}' (legacy field). "
385
386
                "You should only specify one of them."
            )
387

388
389
    if "rope_type" not in rope_parameters and "type" in rope_parameters:
        rope_parameters["rope_type"] = rope_parameters["type"]
390
391
        logger.info("Replacing legacy 'type' key with 'rope_type'")

392
393
    if "rope_type" not in rope_parameters:
        raise ValueError("rope_parameters should have a 'rope_type' key")
394

395
396
    if rope_parameters["rope_type"] == "su":
        rope_parameters["rope_type"] = "longrope"
397
        logger.warning("Replacing legacy rope_type 'su' with 'longrope'")
398
    elif rope_parameters["rope_type"] == "mrope":
399
400
401
402
        if "mrope_section" not in rope_parameters:
            raise ValueError(
                "Legacy rope_type 'mrope' requires 'mrope_section' in rope_parameters"
            )
403
        rope_parameters["rope_type"] = "default"
404
405
406
        logger.warning("Replacing legacy rope_type 'mrope' with 'default'")


407
def _uses_mrope(config: PretrainedConfig) -> bool:
408
409
    rope_parameters = getattr(config, "rope_parameters", None)
    if rope_parameters is None:
410
411
        return False

412
    return "mrope_section" in rope_parameters
413
414


415
416
def uses_mrope(config: PretrainedConfig) -> bool:
    """Detect if the model with this config uses M-ROPE."""
417
418
419
420
421
    return (
        _uses_mrope(config)
        or _uses_mrope(config.get_text_config())
        or thinker_uses_mrope(config)
    )
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436


def thinker_uses_mrope(config: PretrainedConfig) -> bool:
    """Detect if the model contains a thinker config and it uses M-ROPE."""
    thinker_config = getattr(config, "thinker_config", None)
    if thinker_config is None:
        return False

    thinker_text_config = getattr(thinker_config, "text_config", None)
    if thinker_text_config is None:
        return False

    return uses_mrope(thinker_text_config)


437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
def uses_xdrope_dim(config: PretrainedConfig) -> int:
    """Detect if the model with this config uses XD-ROPE."""
    xdrope_section = getattr(config, "xdrope_section", None)
    if xdrope_section is not None and isinstance(xdrope_section, list):
        return len(xdrope_section)
    rope_scaling = getattr(config, "rope_scaling", None)
    if rope_scaling is None:
        return 0

    if isinstance(rope_scaling, dict) and "xdrope_section" in rope_scaling:
        xdrope_section = rope_scaling["xdrope_section"]
        if xdrope_section is not None and isinstance(xdrope_section, list):
            return len(xdrope_section)

    return 0


454
455
456
def is_encoder_decoder(config: PretrainedConfig) -> bool:
    """Detect if the model with this config is used as an encoder/decoder."""

457
458
459
    def _is_encoder_decoder(config: PretrainedConfig) -> bool:
        return getattr(config, "is_encoder_decoder", False)

460
    return _is_encoder_decoder(config) or _is_encoder_decoder(config.get_text_config())
461
462


463
464
465
466
467
468
def is_interleaved(config: PretrainedConfig) -> bool:
    """
    Detect if the model with this config is used with interleaved attention.
    """
    text_config = config.get_text_config()
    if layer_types := getattr(text_config, "layer_types", None):
469
        return len(set(layer_types)) > 1
470
471
472
    return False


473
474
475
476
477
478
479
480
481
def _maybe_update_auto_config_kwargs(kwargs: dict[str, Any], model_type: str):
    """
    Update kwargs for AutoConfig initialization based on model_type
    """
    if model_type in _AUTO_CONFIG_KWARGS_OVERRIDES:
        kwargs.update(_AUTO_CONFIG_KWARGS_OVERRIDES[model_type])
    return kwargs


482
483
484
485
486
487
def _maybe_remap_hf_config_attrs(config: PretrainedConfig) -> PretrainedConfig:
    """Remap config attributes to match the expected names."""
    for old_attr, new_attr in _CONFIG_ATTRS_MAPPING.items():
        if hasattr(config, old_attr):
            if not hasattr(config, new_attr):
                config.update({new_attr: getattr(config, old_attr)})
488
            logger.debug("Remapped config attribute '%s' to '%s'", old_attr, new_attr)
489
490
491
    return config


492
def maybe_override_with_speculators(
493
    model: str,
494
    tokenizer: str | None,
495
    trust_remote_code: bool,
496
497
    revision: str | None = None,
    vllm_speculative_config: dict[str, Any] | None = None,
498
    **kwargs,
499
) -> tuple[str, str | None, dict[str, Any] | None]:
500
    """
501
502
503
504
505
506
507
508
509
510
511
512
513
514
    Resolve model configuration when speculators are detected.

    Checks if the provided model is a speculators model and if so, extracts
    the target model configuration and builds the speculative config.

    Args:
        model: Model name or path
        tokenizer: Tokenizer name or path
        trust_remote_code: Whether to trust remote code
        revision: Model revision
        vllm_speculative_config: Existing vLLM speculative config

    Returns:
        Tuple of (resolved_model, resolved_tokenizer, speculative_config)
515
    """
516
    if check_gguf_file(model):
517
518
        kwargs["gguf_file"] = Path(model).name
        gguf_model_repo = Path(model).parent
519
520
521
    elif is_remote_gguf(model):
        repo_id, _ = split_remote_gguf(model)
        gguf_model_repo = Path(repo_id)
522
523
    else:
        gguf_model_repo = None
524
    kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
525
    config_dict, _ = PretrainedConfig.get_config_dict(
526
        model if gguf_model_repo is None else gguf_model_repo,
527
528
529
        revision=revision,
        trust_remote_code=trust_remote_code,
        token=_get_hf_token(),
530
        **kwargs,
531
    )
532
533
534
535
536
537
538
    speculators_config = config_dict.get("speculators_config")

    if speculators_config is None:
        # No speculators config found, return original values
        return model, tokenizer, vllm_speculative_config

    # Speculators format detected - process overrides
539
    from vllm.transformers_utils.configs.speculators.base import SpeculatorsConfig
540

541
    speculative_config = SpeculatorsConfig.extract_vllm_speculative_config(
542
543
        config_dict=config_dict
    )
544
545

    # Set the draft model to the speculators model
546
    speculative_config["model"] = model
547
548
549
550
551

    # Override model and tokenizer with the verifier model from config
    verifier_model = speculators_config["verifier"]["name_or_path"]
    model = tokenizer = verifier_model

552
    return model, tokenizer, speculative_config
553
554


555
def get_config(
556
    model: str | Path,
557
    trust_remote_code: bool,
558
559
560
561
562
    revision: str | None = None,
    code_revision: str | None = None,
    config_format: str | ConfigFormat = "auto",
    hf_overrides_kw: dict[str, Any] | None = None,
    hf_overrides_fn: Callable[[PretrainedConfig], PretrainedConfig] | None = None,
563
564
565
    **kwargs,
) -> PretrainedConfig:
    # Separate model folder from file path for GGUF models
566

567
568
569
570
571
572
573
574
575
576
577
578
    _is_gguf = is_gguf(model)
    _is_remote_gguf = is_remote_gguf(model)
    if _is_gguf:
        if check_gguf_file(model):
            # Local GGUF file
            kwargs["gguf_file"] = Path(model).name
            model = Path(model).parent
        elif _is_remote_gguf:
            # Remote GGUF - extract repo_id from repo_id:quant_type format
            # The actual GGUF file will be downloaded later by GGUFModelLoader
            # Keep model as repo_id:quant_type for download, but use repo_id for config
            model, _ = split_remote_gguf(model)
579

580
    if config_format == "auto":
581
        try:
582
583
584
            # First check for Mistral to avoid defaulting to
            # Transformers implementation.
            if file_or_path_exists(model, MISTRAL_CONFIG_NAME, revision=revision):
585
                config_format = "mistral"
586
            elif (_is_gguf and not _is_remote_gguf) or file_or_path_exists(
587
588
589
                model, HF_CONFIG_NAME, revision=revision
            ):
                config_format = "hf"
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
            # Remote GGUF models must have config.json in repo,
            # otherwise the config can't be parsed correctly.
            # FIXME(Isotr0py): Support remote GGUF repos without config.json
            elif _is_remote_gguf and not file_or_path_exists(
                model, HF_CONFIG_NAME, revision=revision
            ):
                err_msg = (
                    "Could not find config.json for remote GGUF model repo. "
                    "To load remote GGUF model through `<repo_id>:<quant_type>`, "
                    "ensure your model has config.json (HF format) file. "
                    "Otherwise please specify --hf-config-path <original_repo> "
                    "in engine args to fetch config from unquantized hf model."
                )
                logger.error(err_msg)
                raise ValueError(err_msg)
605
606
607
            else:
                raise ValueError(
                    "Could not detect config format for no config file found. "
608
609
610
                    "With config_format 'auto', ensure your model has either "
                    "config.json (HF format) or params.json (Mistral format). "
                    "Otherwise please specify your_custom_config_format "
611
612
                    "in engine args for customized config parser."
                )
613
614
615
616
617
618
619
620
621
622
623

        except Exception as e:
            error_message = (
                "Invalid repository ID or local directory specified:"
                " '{model}'.\nPlease verify the following requirements:\n"
                "1. Provide a valid Hugging Face repository ID.\n"
                "2. Specify a local directory that contains a recognized "
                "configuration file.\n"
                "   - For Hugging Face models: ensure the presence of a "
                "'config.json'.\n"
                "   - For Mistral models: ensure the presence of a "
624
                "'params.json'.\n"
625
            ).format(model=model)
626
627

            raise ValueError(error_message) from e
628

629
630
631
632
633
634
    config_parser = get_config_parser(config_format)
    config_dict, config = config_parser.parse(
        model,
        trust_remote_code=trust_remote_code,
        revision=revision,
        code_revision=code_revision,
635
        hf_overrides=hf_overrides_kw,
636
637
        **kwargs,
    )
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659

    # Patching defaults for GGUF models
    if _is_gguf:
        # Some models have different default values between GGUF and HF.
        def apply_gguf_default(key: str, gguf_default: Any):
            """
            Apply GGUF defaults unless explicitly configured.

            This function reads/writes external `config` and `config_dict`.
            If the specified `key` is not in `config_dict` (i.e. not explicitly
            configured and the default HF value is used), it updates the
            corresponding `config` value to `gguf_default`.
            """
            if key not in config_dict:
                config.update({key: gguf_default})

        # Apply architecture-specific GGUF defaults.
        if config.model_type in {"qwen3_moe"}:
            # Qwen3 MoE: norm_topk_prob is always true.
            # Note that, this parameter is always false (HF default) on Qwen2 MoE.
            apply_gguf_default("norm_topk_prob", True)

660
    # Special architecture mapping check for GGUF models
661
    if _is_gguf:
662
        if config.model_type not in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
663
            raise RuntimeError(f"Can't get gguf config for {config.model_type}.")
664
665
666
        model_type = MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[config.model_type]
        config.update({"architectures": [model_type]})

667
668
669
    # Architecture mapping for models without explicit architectures field
    if not config.architectures:
        if config.model_type not in MODEL_MAPPING_NAMES:
670
671
672
673
674
675
676
677
            logger.warning(
                "Model config does not have a top-level 'architectures' field: "
                "expecting `hf_overrides={'architectures': ['...']}` to be passed "
                "in engine args."
            )
        else:
            model_type = MODEL_MAPPING_NAMES[config.model_type]
            config.update({"architectures": [model_type]})
678

679
680
681
682
683
684
    # ModelOpt 0.31.0 and after saves the quantization config in the model
    # config file.
    quantization_config = config_dict.get("quantization_config", None)

    # ModelOpt 0.29.0 and before saves the quantization config in a separate
    # "hf_quant_config.json" in the same directory as the model config file.
685
686
687
688
689
690
    if quantization_config is None and file_or_path_exists(
        model, "hf_quant_config.json", revision
    ):
        quantization_config = get_hf_file_to_dict(
            "hf_quant_config.json", model, revision
        )
691
692
693

    if quantization_config is not None:
        config.quantization_config = quantization_config
694
        # auto-enable DeepGEMM UE8M0 if model config requests it
695
        scale_fmt = quantization_config.get("scale_fmt", None)
696
        if scale_fmt in ("ue8m0",):
697
698
            if not envs.is_set("VLLM_USE_DEEP_GEMM_E8M0"):
                os.environ["VLLM_USE_DEEP_GEMM_E8M0"] = "1"
699
                logger.info_once(
700
701
                    (
                        "Detected quantization_config.scale_fmt=%s; "
702
                        "enabling UE8M0 for DeepGEMM."
703
                    ),
704
705
                    scale_fmt,
                )
706
            elif not envs.VLLM_USE_DEEP_GEMM_E8M0:
707
                logger.warning_once(
708
709
710
                    (
                        "Model config requests UE8M0 "
                        "(quantization_config.scale_fmt=%s), but "
711
712
                        "VLLM_USE_DEEP_GEMM_E8M0=0 is set; "
                        "UE8M0 for DeepGEMM disabled."
713
                    ),
714
715
                    scale_fmt,
                )
716

717
718
719
720
721
722
723
    if hf_overrides_kw:
        logger.debug("Overriding HF config with %s", hf_overrides_kw)
        config.update(hf_overrides_kw)
    if hf_overrides_fn:
        logger.debug("Overriding HF config with %s", hf_overrides_fn)
        config = hf_overrides_fn(config)

724
725
726
727
728
729
730
731
    # Exhaustively patch RoPE parameters everywhere they might be
    patch_rope_parameters(config)
    patch_rope_parameters(config.get_text_config())
    SubConfigs: TypeAlias = dict[str, PretrainedConfig]
    sub_configs: SubConfigs | None = getattr(config, "sub_configs", None)
    if sub_configs:
        for sub_config in sub_configs:
            patch_rope_parameters(getattr(config, sub_config))
732

733
734
735
    if trust_remote_code:
        maybe_register_config_serialize_by_value()

736
    return config
737
738


739
@cache
740
def get_pooling_config(model: str, revision: str | None = "main") -> dict | None:
741
    """
742
743
744
    This function gets the pooling and normalize
    config from the model - only applies to
    sentence-transformers models.
745
746

    Args:
747
        model: The name of the Hugging Face model.
748
        revision: The specific version of the model to use.
749
            Defaults to 'main'.
750
751

    Returns:
752
        A dictionary containing the pooling type and whether
753
            normalization is used, or None if no pooling configuration is found.
754
    """
755
756
    if is_remote_gguf(model):
        model, _ = split_remote_gguf(model)
757
758

    modules_file_name = "modules.json"
759
760

    modules_dict = None
761
762
763
    if file_or_path_exists(
        model=model, config_name=modules_file_name, revision=revision
    ):
764
        modules_dict = get_hf_file_to_dict(modules_file_name, model, revision)
765
766
767
768

    if modules_dict is None:
        return None

769
770
    logger.info("Found sentence-transformers modules configuration.")

771
772
773
774
775
776
777
778
    pooling = next(
        (
            item
            for item in modules_dict
            if item["type"] == "sentence_transformers.models.Pooling"
        ),
        None,
    )
779
    normalize = bool(
780
781
782
783
784
785
786
787
788
        next(
            (
                item
                for item in modules_dict
                if item["type"] == "sentence_transformers.models.Normalize"
            ),
            False,
        )
    )
789
790
791

    if pooling:
        pooling_file_name = "{}/config.json".format(pooling["path"])
792
        pooling_dict = get_hf_file_to_dict(pooling_file_name, model, revision)
793
        pooling_type_name = next(
794
795
            (item for item, val in pooling_dict.items() if val is True), None
        )
796
797
798
799

        if pooling_type_name is not None:
            pooling_type_name = get_pooling_config_name(pooling_type_name)

800
        logger.info("Found pooling configuration.")
801
802
803
804
805
        return {"pooling_type": pooling_type_name, "normalize": normalize}

    return None


806
def get_pooling_config_name(pooling_name: str) -> str | None:
807
808
809
810
811
812
813
814
815
    if "pooling_mode_" in pooling_name:
        pooling_name = pooling_name.replace("pooling_mode_", "")

    if "_" in pooling_name:
        pooling_name = pooling_name.split("_")[0]

    if "lasttoken" in pooling_name:
        pooling_name = "last"

816
    supported_pooling_types = ["LAST", "ALL", "CLS", "STEP", "MEAN"]
817
818
    pooling_type_name = pooling_name.upper()

819
820
821
    if pooling_type_name in supported_pooling_types:
        return pooling_type_name

822
    raise NotImplementedError(f"Pooling type {pooling_type_name} not supported")
823
824


825
@cache
826
def get_sentence_transformer_tokenizer_config(
827
    model: str | Path, revision: str | None = "main"
828
):
829
    """
830
    Returns the tokenization configuration dictionary for a
831
832
833
    given Sentence Transformer BERT model.

    Parameters:
834
    - model (str|Path): The name of the Sentence Transformer
835
836
837
838
839
    BERT model.
    - revision (str, optional): The revision of the m
    odel to use. Defaults to 'main'.

    Returns:
840
    - dict: A dictionary containing the configuration parameters
841
842
    for the Sentence Transformer BERT model.
    """
843
844
845
846
847
848
849
850
851
852
    sentence_transformer_config_files = [
        "sentence_bert_config.json",
        "sentence_roberta_config.json",
        "sentence_distilbert_config.json",
        "sentence_camembert_config.json",
        "sentence_albert_config.json",
        "sentence_xlm-roberta_config.json",
        "sentence_xlnet_config.json",
    ]
    encoder_dict = None
853
854

    for config_file in sentence_transformer_config_files:
855
856
857
858
        if (
            try_get_local_file(model=model, file_name=config_file, revision=revision)
            is not None
        ):
859
            encoder_dict = get_hf_file_to_dict(config_file, model, revision)
860
861
            if encoder_dict:
                break
862

863
    if not encoder_dict and not Path(model).is_absolute():
864
865
        try:
            # If model is on HuggingfaceHub, get the repo files
866
867
868
            repo_files = list_repo_files(
                model, revision=revision, token=_get_hf_token()
            )
869
        except Exception:
870
871
872
873
            repo_files = []

        for config_name in sentence_transformer_config_files:
            if config_name in repo_files:
874
                encoder_dict = get_hf_file_to_dict(config_name, model, revision)
875
876
877
                if encoder_dict:
                    break

878
879
880
    if not encoder_dict:
        return None

881
882
    logger.info("Found sentence-transformers tokenize configuration.")

883
884
885
886
887
    if all(k in encoder_dict for k in ("max_seq_length", "do_lower_case")):
        return encoder_dict
    return None


888
def maybe_register_config_serialize_by_value() -> None:
889
890
    """Try to register HF model configuration class to serialize by value

891
892
893
    If trust_remote_code is set, and the model's config file specifies an
    `AutoConfig` class, then the config class is typically an instance of
    a custom class imported from the HF modules cache.
894

895
    Examples:
896

897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
    >>> from transformers import AutoConfig
    >>> klass = AutoConfig.from_pretrained(
    ...     "meta-llama/Meta-Llama-3-8B", trust_remote_code=True
    ... )
    >>> klass.__class__  # transformers.models.llama.configuration_llama.LlamaConfig
    >>> import transformers_modules  # error, not initialized
    >>> klass = AutoConfig.from_pretrained(
    ...     "deepseek-ai/DeepSeek-V2.5", trust_remote_code=True
    ... )
    >>> import transformers_modules  # success, initialized
    >>> klass.__class__  # transformers_modules.deepseek-ai.DeepSeek-V2.5.98b11844770b2c3ffc18b175c758a803640f4e77.configuration_deepseek.DeepseekV2Config

    In the DeepSeek example, the config class is an instance of a custom
    class that is not serializable by default. This class will not be
    importable in spawned workers, and won't exist at all on
    other nodes, which breaks serialization of the config.

    In this function we tell the cloudpickle serialization library to pass
    instances of these generated classes by value instead of by reference,
    i.e. the class definition is serialized along with its data so that the
    class module does not need to be importable on the receiving end.

    See: https://github.com/cloudpipe/cloudpickle?tab=readme-ov-file#overriding-pickles-serialization-mechanism-for-importable-constructs
    """  # noqa
921
922
    try:
        import transformers_modules
923

924
        transformers_modules_available = True
925
    except ImportError:
926
        transformers_modules_available = False
927
928
929
930
931

    try:
        import multiprocessing
        import pickle

932
933
        import cloudpickle

934
        from vllm.config import VllmConfig
935

936
937
938
        # Register multiprocessing reducers to handle cross-process
        # serialization of VllmConfig objects that may contain custom configs
        # from transformers_modules
939
        def _reduce_config(config: VllmConfig):
940
            return (pickle.loads, (cloudpickle.dumps(config),))
941

942
        multiprocessing.reducer.register(VllmConfig, _reduce_config)
943

944
945
946
947
948
        # Register transformers_modules with cloudpickle if available
        if transformers_modules_available:
            cloudpickle.register_pickle_by_value(transformers_modules)

            # ray vendors its own version of cloudpickle
949
            from vllm.v1.executor.ray_utils import ray
950

951
952
953
            if ray:
                ray.cloudpickle.register_pickle_by_value(transformers_modules)

954
955
956
957
958
959
    except Exception as e:
        logger.warning(
            "Unable to register remote classes used by"
            " trust_remote_code with by-value serialization. This may"
            " lead to a later error. If remote code is not needed"
            " remove `--trust-remote-code`",
960
961
            exc_info=e,
        )
962
963


964
def get_hf_image_processor_config(
965
966
967
    model: str | Path,
    hf_token: bool | str | None = None,
    revision: str | None = None,
968
    **kwargs,
969
) -> dict[str, Any]:
970
    # ModelScope does not provide an interface for image_processor
971
    if envs.VLLM_USE_MODELSCOPE:
972
        return dict()
973
    # Separate model folder from file path for GGUF models
974
    if check_gguf_file(model):
975
        model = Path(model).parent
976
977
    elif is_remote_gguf(model):
        model, _ = split_remote_gguf(model)
978
979
980
    return get_image_processor_config(
        model, token=hf_token, revision=revision, **kwargs
    )
981
982


983
984
def get_hf_text_config(config: PretrainedConfig):
    """Get the "sub" config relevant to llm for multi modal models.
985
    No op for pure text models.
986
    """
987
988
    text_config = config.get_text_config()

989
990
991
992
993
994
995
    if text_config is not config and not hasattr(text_config, "num_attention_heads"):
        raise ValueError(
            "The text_config extracted from the model config does not have "
            "`num_attention_heads` attribute. This indicates a mismatch "
            "between the model config and vLLM's expectations. Please "
            "ensure that the model config is compatible with vLLM."
        )
996
997

    return text_config
998
999
1000
1001
1002


def try_get_generation_config(
    model: str,
    trust_remote_code: bool,
1003
1004
1005
    revision: str | None = None,
    config_format: str | ConfigFormat = "auto",
) -> GenerationConfig | None:
1006
1007
1008
1009
1010
1011
1012
    # GGUF files don't have generation_config.json - their config is embedded
    # in the file header. Skip all filesystem lookups to avoid re-reading the
    # memory-mapped file, which can hang in multi-process scenarios when the
    # EngineCore process already has the file mapped.
    if is_gguf(model):
        return None

1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
    try:
        return GenerationConfig.from_pretrained(
            model,
            revision=revision,
        )
    except OSError:  # Not found
        try:
            config = get_config(
                model,
                trust_remote_code=trust_remote_code,
                revision=revision,
1024
                config_format=config_format,
1025
1026
1027
1028
            )
            return GenerationConfig.from_model_config(config)
        except OSError:  # Not found
            return None
1029
1030


1031
1032
1033
def try_get_safetensors_metadata(
    model: str,
    *,
1034
    revision: str | None = None,
1035
1036
1037
1038
1039
):
    get_safetensors_metadata_partial = partial(
        get_safetensors_metadata,
        model,
        revision=revision,
1040
        token=_get_hf_token(),
1041
1042
1043
    )

    try:
1044
1045
1046
        return with_retry(
            get_safetensors_metadata_partial, "Error retrieving safetensors"
        )
1047
1048
    except Exception:
        return None
1049
1050
1051


def try_get_tokenizer_config(
1052
    pretrained_model_name_or_path: str | os.PathLike,
1053
    trust_remote_code: bool,
1054
1055
    revision: str | None = None,
) -> dict[str, Any] | None:
1056
1057
1058
1059
1060
1061
1062
1063
    try:
        return get_tokenizer_config(
            pretrained_model_name_or_path,
            trust_remote_code=trust_remote_code,
            revision=revision,
        )
    except Exception:
        return None
1064
1065


1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
@cache
def try_get_dense_modules(
    model: str | Path,
    revision: str | None = None,
) -> list[dict[str, Any]] | None:
    try:
        modules = get_hf_file_to_dict("modules.json", model, revision)
        if not modules:
            return None

        if isinstance(modules, dict):
            modules = modules.get("modules", [])

        dense_modules = [
            m for m in modules if m.get("type") == "sentence_transformers.models.Dense"
        ]
        if not dense_modules:
            return None

        layer_configs = []
        for module in dense_modules:
            folder = module.get("path", "")

            config_path = f"{folder}/config.json" if folder else "config.json"
            layer_config = get_hf_file_to_dict(config_path, model, revision)
            if not layer_config:
                continue
            layer_config["folder"] = folder
            layer_configs.append(layer_config)
        return layer_configs
    except Exception:
        return None


1100
1101
1102
def get_safetensors_params_metadata(
    model: str,
    *,
1103
    revision: str | None = None,
1104
1105
1106
1107
1108
1109
1110
1111
1112
) -> dict[str, Any]:
    """
    Get the safetensors metadata for remote model repository.
    """
    full_metadata = {}
    if (model_path := Path(model)).exists():
        safetensors_to_check = model_path.glob("*.safetensors")
        full_metadata = {
            param_name: info
1113
1114
1115
            for file_path in safetensors_to_check
            if file_path.is_file()
            for param_name, info in parse_safetensors_file_metadata(file_path).items()
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
        }
    else:
        repo_mt = try_get_safetensors_metadata(model, revision=revision)
        if repo_mt and (files_mt := repo_mt.files_metadata):
            full_metadata = {
                param_name: asdict(info)
                for file_mt in files_mt.values()
                for param_name, info in file_mt.tensors.items()
            }
    return full_metadata


1128
1129
1130
1131
1132
1133
1134
def _download_mistral_config_file(model, revision) -> dict:
    config_file_name = "params.json"
    config_dict = get_hf_file_to_dict(config_file_name, model, revision)
    if config_dict is None:
        raise ValueError(
            f"Failed to load mistral '{config_file_name}' config for model "
            f"{model}. Please check if the model is a mistral-format model "
1135
1136
            f"and if the config file exists."
        )
1137
1138
1139
1140
1141
1142
1143
1144
    assert isinstance(config_dict, dict)
    return config_dict


def _maybe_retrieve_max_pos_from_hf(model, revision, **kwargs) -> int:
    max_position_embeddings = 128_000
    try:
        trust_remote_code_val = kwargs.get("trust_remote_code", False)
1145
1146
1147
1148
1149
1150
        hf_config = get_config(
            model=model,
            trust_remote_code=trust_remote_code_val,
            revision=revision,
            config_format="hf",
        )
1151
1152
1153
1154
1155
1156
1157
        if hf_value := hf_config.get_text_config().max_position_embeddings:
            max_position_embeddings = hf_value
    except Exception as e:
        logger.warning(
            "The params.json file is missing 'max_position_embeddings'"
            " and could not get a value from the HF config."
            " Defaulting to 128000",
1158
1159
            exc_info=e,
        )
1160
1161

    return max_position_embeddings