config.py 39.2 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.configuration_utils import ALLOWED_LAYER_TYPES
19
from transformers.models.auto.image_processing_auto import get_image_processor_config
20
21
22
23
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_MAPPING_NAMES,
)
24
from transformers.models.auto.tokenization_auto import get_tokenizer_config
25
from transformers.utils import CONFIG_NAME as HF_CONFIG_NAME
26

27
from vllm import envs
28
from vllm.config.utils import getattr_iter
29
from vllm.logger import init_logger
30
from vllm.transformers_utils.utils import parse_safetensors_file_metadata
31
32

from .config_parser_base import ConfigParserBase
33
34
35
36
37
38
from .gguf_utils import (
    check_gguf_file,
    is_gguf,
    is_remote_gguf,
    split_remote_gguf,
)
39
from .repo_utils import (
40
41
42
43
44
45
46
    _get_hf_token,
    file_or_path_exists,
    get_hf_file_to_dict,
    list_repo_files,
    try_get_local_file,
    with_retry,
)
47

48
if envs.VLLM_USE_MODELSCOPE:
49
50
51
    from modelscope import AutoConfig
else:
    from transformers import AutoConfig
52

53
54
MISTRAL_CONFIG_NAME = "params.json"

55
56
logger = init_logger(__name__)

57

58
59
class LazyConfigDict(dict):
    def __getitem__(self, key):
60
61
62
        if isinstance(value := super().__getitem__(key), type):
            return value

63
        import vllm.transformers_utils.configs as configs
64

65
        return getattr(configs, value)
66
67
68


_CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
69
    afmoe="AfmoeConfig",
70
71
    chatglm="ChatGLMConfig",
    deepseek_vl_v2="DeepseekVLV2Config",
72
    deepseek_v32="DeepseekV3Config",
73
    flex_olmo="FlexOlmoConfig",
74
    hunyuan_vl="HunYuanVLConfig",
75
    kimi_linear="KimiLinearConfig",
76
77
78
79
80
81
    kimi_vl="KimiVLConfig",
    RefinedWeb="RWConfig",  # For tiiuae/falcon-40b(-instruct)
    RefinedWebModel="RWConfig",  # For tiiuae/falcon-7b(-instruct)
    jais="JAISConfig",
    mlp_speculator="MLPSpeculatorConfig",
    medusa="MedusaConfig",
82
    midashenglm="MiDashengLMConfig",
83
84
85
    eagle="EAGLEConfig",
    speculators="SpeculatorsConfig",
    nemotron="NemotronConfig",
86
    olmo3="Olmo3Config",
87
88
89
90
    ovis="OvisConfig",
    ultravox="UltravoxConfig",
    step3_vl="Step3VLConfig",
    step3_text="Step3TextConfig",
91
    qwen3_next="Qwen3NextConfig",
Paul Pak's avatar
Paul Pak committed
92
    lfm2_moe="Lfm2MoeConfig",
93
    tarsier2="Tarsier2Config",
94
)
95

96
97
98
99
_CONFIG_ATTRS_MAPPING: dict[str, str] = {
    "llm_config": "text_config",
}

100
_AUTO_CONFIG_KWARGS_OVERRIDES: dict[str, dict[str, Any]] = {
101
    "internvl_chat": {"has_no_defaults_at_init": True},
102
    "Llama_Nemotron_Nano_VL": {"attn_implementation": "eager"},
103
    "NVLM_D": {"has_no_defaults_at_init": True},
104
105
}

106

107
class HFConfigParser(ConfigParserBase):
108
109
    def parse(
        self,
110
        model: str | Path,
111
        trust_remote_code: bool,
112
113
        revision: str | None = None,
        code_revision: str | None = None,
114
115
        **kwargs,
    ) -> tuple[dict, PretrainedConfig]:
116
117
118
119
120
121
122
123
124
125
126
        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:
127
128
129
130
131
            model_type = (
                "speculators"
                if config_dict.get("speculators_config") is not None
                else model_type
            )
132
133
134
        # 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)
135
136
137
138
139
140
141
142
143
144
145
146

        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:
147
                kwargs = _maybe_update_auto_config_kwargs(kwargs, model_type=model_type)
148
149
150
151
152
153
154
155
156
                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:
157
158
159
160
                if (
                    not trust_remote_code
                    and "requires you to execute the configuration file" in str(e)
                ):
161
162
163
164
165
                    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 "
166
167
                        "`--trust-remote-code` flag in the CLI."
                    )
168
169
170
171
172
173
174
175
                    raise RuntimeError(err_msg) from e
                else:
                    raise e
        config = _maybe_remap_hf_config_attrs(config)
        return config_dict, config


class MistralConfigParser(ConfigParserBase):
176
177
    def parse(
        self,
178
        model: str | Path,
179
        trust_remote_code: bool,
180
181
        revision: str | None = None,
        code_revision: str | None = None,
182
183
        **kwargs,
    ) -> tuple[dict, PretrainedConfig]:
184
185
186
        # 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)
187
188
189
        if (
            max_position_embeddings := config_dict.get("max_position_embeddings")
        ) is None:
190
            max_position_embeddings = _maybe_retrieve_max_pos_from_hf(
191
192
                model, revision, **kwargs
            )
193
194
195
196
            config_dict["max_position_embeddings"] = max_position_embeddings

        from vllm.transformers_utils.configs.mistral import adapt_config_dict

197
198
199
200
201
202
203
204
205
206
207
208
209
        # 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)
210
211
212

        # 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
213
214
215
        if (sliding_window := getattr(config, "sliding_window", None)) and isinstance(
            sliding_window, list
        ):
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
            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.
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
     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,
264
265
         ...         revision: str | None = None,
         ...         code_revision: str | None = None,
266
267
268
269
270
271
         ...         **kwargs,
         ...     ) -> tuple[dict, PretrainedConfig]:
         ...         raise NotImplementedError
         >>>
         >>> type(get_config_parser("custom_config_parser"))
         <class 'CustomConfigParser'>
272
273
274
275
276
277
    """  # 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 "
278
279
280
281
                "overwritten by the new parser class `%s`.",
                config_format,
                config_parser_cls,
            )
282
        if not issubclass(config_parser_cls, ConfigParserBase):
283
284
285
            raise ValueError(
                "The config parser must be a subclass of `ConfigParserBase`."
            )
286
        _CONFIG_FORMAT_TO_CONFIG_PARSER[config_format] = config_parser_cls
287
288
289
290
291
        logger.info(
            "Registered config parser `%s` with config format `%s`",
            config_parser_cls,
            config_format,
        )
292
293
294
        return config_parser_cls

    return _wrapper
295
296


297
298
299
300
301
302
303
304
305
306
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:
307
    """Provide backwards compatibility for RoPE."""
308
309
    rope_theta_names = ("rope_theta", "rotary_emb_base")
    rope_theta = getattr_iter(config, rope_theta_names, None)
310
311
    if Version(version("transformers")) < Version("5.0.0.dev0"):
        # Transformers v4 installed, legacy config fields may be present
312
313
        if (rope_scaling := getattr(config, "rope_scaling", None)) is not None:
            config.rope_parameters = rope_scaling
314
        if rope_theta is not None:
315
316
317
            if not hasattr(config, "rope_parameters"):
                config.rope_parameters = {"rope_type": "default"}
            config.rope_parameters["rope_theta"] = rope_theta
318
319
320
321
322
323
        partial_rotary_factor_names = ("partial_rotary_factor", "rotary_pct")
        partial_rotary_factor = getattr_iter(config, partial_rotary_factor_names, None)
        if partial_rotary_factor is not None:
            if not hasattr(config, "rope_parameters"):
                config.rope_parameters = {"rope_type": "default"}
            config.rope_parameters["partial_rotary_factor"] = partial_rotary_factor
324
325
326
327
    elif rope_theta is not None or hasattr(config, "rope_parameters"):
        # Transformers v5 installed
        config.standardize_rope_params()
        config.validate_rope()
328
329

    # No RoPE parameters to patch
330
    if getattr(config, "rope_parameters", None) is None:
331
332
        return

333
334
335
336
    # Add original_max_position_embeddings if present
    if ompe := getattr(config, "original_max_position_embeddings", None):
        config.rope_parameters["original_max_position_embeddings"] = ompe

337
    # Handle nested rope_parameters in interleaved sliding attention models
338
339
    if set(config.rope_parameters.keys()).issubset(ALLOWED_LAYER_TYPES):
        for rope_parameters_layer_type in config.rope_parameters.values():
340
341
            patch_rope_parameters_dict(rope_parameters_layer_type)
    else:
342
        patch_rope_parameters_dict(config.rope_parameters)
343
344


345
346
347
348
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"]
349
350
351
352
353
        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:
354
355
356
            raise ValueError(
                f"Found conflicts between 'rope_type={rope_type}' (modern "
                f"field) and 'type={rope_type_legacy}' (legacy field). "
357
358
                "You should only specify one of them."
            )
359

360
361
    if "rope_type" not in rope_parameters and "type" in rope_parameters:
        rope_parameters["rope_type"] = rope_parameters["type"]
362
363
        logger.info("Replacing legacy 'type' key with 'rope_type'")

364
365
    if "rope_type" not in rope_parameters:
        raise ValueError("rope_parameters should have a 'rope_type' key")
366

367
368
    if rope_parameters["rope_type"] == "su":
        rope_parameters["rope_type"] = "longrope"
369
        logger.warning("Replacing legacy rope_type 'su' with 'longrope'")
370
    elif rope_parameters["rope_type"] == "mrope":
371
372
373
374
        if "mrope_section" not in rope_parameters:
            raise ValueError(
                "Legacy rope_type 'mrope' requires 'mrope_section' in rope_parameters"
            )
375
        rope_parameters["rope_type"] = "default"
376
377
378
        logger.warning("Replacing legacy rope_type 'mrope' with 'default'")


379
def _uses_mrope(config: PretrainedConfig) -> bool:
380
381
    rope_parameters = getattr(config, "rope_parameters", None)
    if rope_parameters is None:
382
383
        return False

384
    return "mrope_section" in rope_parameters
385
386


387
388
def uses_mrope(config: PretrainedConfig) -> bool:
    """Detect if the model with this config uses M-ROPE."""
389
390
391
392
393
    return (
        _uses_mrope(config)
        or _uses_mrope(config.get_text_config())
        or thinker_uses_mrope(config)
    )
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408


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)


409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
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


426
427
428
def is_encoder_decoder(config: PretrainedConfig) -> bool:
    """Detect if the model with this config is used as an encoder/decoder."""

429
430
431
    def _is_encoder_decoder(config: PretrainedConfig) -> bool:
        return getattr(config, "is_encoder_decoder", False)

432
    return _is_encoder_decoder(config) or _is_encoder_decoder(config.get_text_config())
433
434


435
436
437
438
439
440
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):
441
        return len(set(layer_types)) > 1
442
443
444
    return False


445
446
447
448
449
450
451
452
453
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


454
455
456
457
458
459
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)})
460
            logger.debug("Remapped config attribute '%s' to '%s'", old_attr, new_attr)
461
462
463
    return config


464
def maybe_override_with_speculators(
465
    model: str,
466
    tokenizer: str | None,
467
    trust_remote_code: bool,
468
469
    revision: str | None = None,
    vllm_speculative_config: dict[str, Any] | None = None,
470
    **kwargs,
471
) -> tuple[str, str | None, dict[str, Any] | None]:
472
    """
473
474
475
476
477
478
479
480
481
482
483
484
485
486
    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)
487
    """
488
    if check_gguf_file(model):
489
490
        kwargs["gguf_file"] = Path(model).name
        gguf_model_repo = Path(model).parent
491
492
493
    elif is_remote_gguf(model):
        repo_id, _ = split_remote_gguf(model)
        gguf_model_repo = Path(repo_id)
494
495
    else:
        gguf_model_repo = None
496
    kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
497
    config_dict, _ = PretrainedConfig.get_config_dict(
498
        model if gguf_model_repo is None else gguf_model_repo,
499
500
501
        revision=revision,
        trust_remote_code=trust_remote_code,
        token=_get_hf_token(),
502
        **kwargs,
503
    )
504
505
506
507
508
509
510
    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
511
    from vllm.transformers_utils.configs.speculators.base import SpeculatorsConfig
512

513
    speculative_config = SpeculatorsConfig.extract_vllm_speculative_config(
514
515
        config_dict=config_dict
    )
516
517

    # Set the draft model to the speculators model
518
    speculative_config["model"] = model
519
520
521
522
523

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

524
    return model, tokenizer, speculative_config
525
526


527
def get_config(
528
    model: str | Path,
529
    trust_remote_code: bool,
530
531
532
533
534
    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,
535
536
537
    **kwargs,
) -> PretrainedConfig:
    # Separate model folder from file path for GGUF models
538

539
540
541
542
543
544
545
546
547
548
549
550
    _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)
551

552
    if config_format == "auto":
553
        try:
554
555
556
            # First check for Mistral to avoid defaulting to
            # Transformers implementation.
            if file_or_path_exists(model, MISTRAL_CONFIG_NAME, revision=revision):
557
                config_format = "mistral"
558
            elif (_is_gguf and not _is_remote_gguf) or file_or_path_exists(
559
560
561
                model, HF_CONFIG_NAME, revision=revision
            ):
                config_format = "hf"
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
            # 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)
577
578
579
            else:
                raise ValueError(
                    "Could not detect config format for no config file found. "
580
581
582
                    "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 "
583
584
                    "in engine args for customized config parser."
                )
585
586
587
588
589
590
591
592
593
594
595

        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 "
596
                "'params.json'.\n"
597
            ).format(model=model)
598
599

            raise ValueError(error_message) from e
600

601
602
603
604
605
606
    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,
607
        hf_overrides=hf_overrides_kw,
608
609
        **kwargs,
    )
610
    # Special architecture mapping check for GGUF models
611
    if _is_gguf:
612
        if config.model_type not in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
613
            raise RuntimeError(f"Can't get gguf config for {config.model_type}.")
614
615
616
        model_type = MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[config.model_type]
        config.update({"architectures": [model_type]})

617
618
619
    # Architecture mapping for models without explicit architectures field
    if not config.architectures:
        if config.model_type not in MODEL_MAPPING_NAMES:
620
621
622
623
624
625
626
627
            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]})
628

629
630
631
632
633
634
    # 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.
635
636
637
638
639
640
    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
        )
641
642
643

    if quantization_config is not None:
        config.quantization_config = quantization_config
644
        # auto-enable DeepGEMM UE8M0 if model config requests it
645
        scale_fmt = quantization_config.get("scale_fmt", None)
646
        if scale_fmt in ("ue8m0",):
647
648
            if not envs.is_set("VLLM_USE_DEEP_GEMM_E8M0"):
                os.environ["VLLM_USE_DEEP_GEMM_E8M0"] = "1"
649
                logger.info_once(
650
651
                    (
                        "Detected quantization_config.scale_fmt=%s; "
652
                        "enabling UE8M0 for DeepGEMM."
653
                    ),
654
655
                    scale_fmt,
                )
656
            elif not envs.VLLM_USE_DEEP_GEMM_E8M0:
657
                logger.warning_once(
658
659
660
                    (
                        "Model config requests UE8M0 "
                        "(quantization_config.scale_fmt=%s), but "
661
662
                        "VLLM_USE_DEEP_GEMM_E8M0=0 is set; "
                        "UE8M0 for DeepGEMM disabled."
663
                    ),
664
665
                    scale_fmt,
                )
666

667
668
669
670
671
672
673
    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)

674
675
676
677
678
679
680
681
    # 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))
682

683
684
685
    if trust_remote_code:
        maybe_register_config_serialize_by_value()

686
    return config
687
688


689
@cache
690
def get_pooling_config(model: str, revision: str | None = "main") -> dict | None:
691
    """
692
693
694
    This function gets the pooling and normalize
    config from the model - only applies to
    sentence-transformers models.
695
696

    Args:
697
        model: The name of the Hugging Face model.
698
        revision: The specific version of the model to use.
699
            Defaults to 'main'.
700
701

    Returns:
702
        A dictionary containing the pooling type and whether
703
            normalization is used, or None if no pooling configuration is found.
704
    """
705
706
    if is_remote_gguf(model):
        model, _ = split_remote_gguf(model)
707
708

    modules_file_name = "modules.json"
709
710

    modules_dict = None
711
712
713
    if file_or_path_exists(
        model=model, config_name=modules_file_name, revision=revision
    ):
714
        modules_dict = get_hf_file_to_dict(modules_file_name, model, revision)
715
716
717
718

    if modules_dict is None:
        return None

719
720
    logger.info("Found sentence-transformers modules configuration.")

721
722
723
724
725
726
727
728
    pooling = next(
        (
            item
            for item in modules_dict
            if item["type"] == "sentence_transformers.models.Pooling"
        ),
        None,
    )
729
    normalize = bool(
730
731
732
733
734
735
736
737
738
        next(
            (
                item
                for item in modules_dict
                if item["type"] == "sentence_transformers.models.Normalize"
            ),
            False,
        )
    )
739
740
741

    if pooling:
        pooling_file_name = "{}/config.json".format(pooling["path"])
742
        pooling_dict = get_hf_file_to_dict(pooling_file_name, model, revision)
743
        pooling_type_name = next(
744
745
            (item for item, val in pooling_dict.items() if val is True), None
        )
746
747
748
749

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

750
        logger.info("Found pooling configuration.")
751
752
753
754
755
        return {"pooling_type": pooling_type_name, "normalize": normalize}

    return None


756
def get_pooling_config_name(pooling_name: str) -> str | None:
757
758
759
760
761
762
763
764
765
    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"

766
    supported_pooling_types = ["LAST", "ALL", "CLS", "STEP", "MEAN"]
767
768
    pooling_type_name = pooling_name.upper()

769
770
771
    if pooling_type_name in supported_pooling_types:
        return pooling_type_name

772
    raise NotImplementedError(f"Pooling type {pooling_type_name} not supported")
773
774


775
@cache
776
def get_sentence_transformer_tokenizer_config(
777
    model: str | Path, revision: str | None = "main"
778
):
779
    """
780
    Returns the tokenization configuration dictionary for a
781
782
783
    given Sentence Transformer BERT model.

    Parameters:
784
    - model (str|Path): The name of the Sentence Transformer
785
786
787
788
789
    BERT model.
    - revision (str, optional): The revision of the m
    odel to use. Defaults to 'main'.

    Returns:
790
    - dict: A dictionary containing the configuration parameters
791
792
    for the Sentence Transformer BERT model.
    """
793
794
795
796
797
798
799
800
801
802
    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
803
804

    for config_file in sentence_transformer_config_files:
805
806
807
808
        if (
            try_get_local_file(model=model, file_name=config_file, revision=revision)
            is not None
        ):
809
            encoder_dict = get_hf_file_to_dict(config_file, model, revision)
810
811
            if encoder_dict:
                break
812

813
    if not encoder_dict and not Path(model).is_absolute():
814
815
        try:
            # If model is on HuggingfaceHub, get the repo files
816
817
818
            repo_files = list_repo_files(
                model, revision=revision, token=_get_hf_token()
            )
819
        except Exception:
820
821
822
823
            repo_files = []

        for config_name in sentence_transformer_config_files:
            if config_name in repo_files:
824
                encoder_dict = get_hf_file_to_dict(config_name, model, revision)
825
826
827
                if encoder_dict:
                    break

828
829
830
    if not encoder_dict:
        return None

831
832
    logger.info("Found sentence-transformers tokenize configuration.")

833
834
835
836
837
    if all(k in encoder_dict for k in ("max_seq_length", "do_lower_case")):
        return encoder_dict
    return None


838
def maybe_register_config_serialize_by_value() -> None:
839
840
    """Try to register HF model configuration class to serialize by value

841
842
843
    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.
844

845
    Examples:
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
    >>> 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
871
872
    try:
        import transformers_modules
873

874
        transformers_modules_available = True
875
    except ImportError:
876
        transformers_modules_available = False
877
878
879
880
881

    try:
        import multiprocessing
        import pickle

882
883
        import cloudpickle

884
        from vllm.config import VllmConfig
885

886
887
888
        # Register multiprocessing reducers to handle cross-process
        # serialization of VllmConfig objects that may contain custom configs
        # from transformers_modules
889
        def _reduce_config(config: VllmConfig):
890
            return (pickle.loads, (cloudpickle.dumps(config),))
891

892
        multiprocessing.reducer.register(VllmConfig, _reduce_config)
893

894
895
896
897
898
        # 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
899
            from vllm.v1.executor.ray_utils import ray
900

901
902
903
            if ray:
                ray.cloudpickle.register_pickle_by_value(transformers_modules)

904
905
906
907
908
909
    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`",
910
911
            exc_info=e,
        )
912
913


914
def get_hf_image_processor_config(
915
916
917
    model: str | Path,
    hf_token: bool | str | None = None,
    revision: str | None = None,
918
    **kwargs,
919
) -> dict[str, Any]:
920
    # ModelScope does not provide an interface for image_processor
921
    if envs.VLLM_USE_MODELSCOPE:
922
        return dict()
923
    # Separate model folder from file path for GGUF models
924
    if check_gguf_file(model):
925
        model = Path(model).parent
926
927
    elif is_remote_gguf(model):
        model, _ = split_remote_gguf(model)
928
929
930
    return get_image_processor_config(
        model, token=hf_token, revision=revision, **kwargs
    )
931
932


933
934
def get_hf_text_config(config: PretrainedConfig):
    """Get the "sub" config relevant to llm for multi modal models.
935
    No op for pure text models.
936
    """
937
938
    text_config = config.get_text_config()

939
940
941
942
943
944
945
    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."
        )
946
947

    return text_config
948
949
950
951
952


def try_get_generation_config(
    model: str,
    trust_remote_code: bool,
953
954
955
    revision: str | None = None,
    config_format: str | ConfigFormat = "auto",
) -> GenerationConfig | None:
956
957
958
959
960
961
962
963
964
965
966
    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,
967
                config_format=config_format,
968
969
970
971
            )
            return GenerationConfig.from_model_config(config)
        except OSError:  # Not found
            return None
972
973


974
975
976
def try_get_safetensors_metadata(
    model: str,
    *,
977
    revision: str | None = None,
978
979
980
981
982
):
    get_safetensors_metadata_partial = partial(
        get_safetensors_metadata,
        model,
        revision=revision,
983
        token=_get_hf_token(),
984
985
986
    )

    try:
987
988
989
        return with_retry(
            get_safetensors_metadata_partial, "Error retrieving safetensors"
        )
990
991
    except Exception:
        return None
992
993
994


def try_get_tokenizer_config(
995
    pretrained_model_name_or_path: str | os.PathLike,
996
    trust_remote_code: bool,
997
998
    revision: str | None = None,
) -> dict[str, Any] | None:
999
1000
1001
1002
1003
1004
1005
1006
    try:
        return get_tokenizer_config(
            pretrained_model_name_or_path,
            trust_remote_code=trust_remote_code,
            revision=revision,
        )
    except Exception:
        return None
1007
1008


1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
@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


1043
1044
1045
def get_safetensors_params_metadata(
    model: str,
    *,
1046
    revision: str | None = None,
1047
1048
1049
1050
1051
1052
1053
1054
1055
) -> 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
1056
1057
1058
            for file_path in safetensors_to_check
            if file_path.is_file()
            for param_name, info in parse_safetensors_file_metadata(file_path).items()
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
        }
    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


1071
1072
1073
1074
1075
1076
1077
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 "
1078
1079
            f"and if the config file exists."
        )
1080
1081
1082
1083
1084
1085
1086
1087
    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)
1088
1089
1090
1091
1092
1093
        hf_config = get_config(
            model=model,
            trust_remote_code=trust_remote_code_val,
            revision=revision,
            config_format="hf",
        )
1094
1095
1096
1097
1098
1099
1100
        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",
1101
1102
            exc_info=e,
        )
1103
1104

    return max_position_embeddings