config.py 44 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import os
Julien Denize's avatar
Julien Denize committed
5
6
from collections.abc import Callable, Iterator
from contextlib import contextmanager
7
from dataclasses import asdict
8
from functools import cache, partial
9
from importlib.metadata import version
10
from pathlib import Path
11
from typing import Any, Literal, TypeAlias
Jasmond L's avatar
Jasmond L committed
12

Joe Runde's avatar
Joe Runde committed
13
import huggingface_hub
Julien Denize's avatar
Julien Denize committed
14
15
import torch
from huggingface_hub import constants, get_safetensors_metadata
16
from packaging.version import Version
Julien Denize's avatar
Julien Denize committed
17
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
18
from transformers import GenerationConfig, PretrainedConfig
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.logger import init_logger
29
from vllm.transformers_utils.repo_utils import is_mistral_model_repo
30
31
32
33
from vllm.transformers_utils.utils import (
    parse_safetensors_file_metadata,
    without_trust_remote_code,
)
Julien Denize's avatar
Julien Denize committed
34
from vllm.utils.torch_utils import common_broadcastable_dtype
35
36

from .config_parser_base import ConfigParserBase
37
38
39
40
41
42
from .gguf_utils import (
    check_gguf_file,
    is_gguf,
    is_remote_gguf,
    split_remote_gguf,
)
43
from .repo_utils import (
44
45
46
47
48
49
    file_or_path_exists,
    get_hf_file_to_dict,
    list_repo_files,
    try_get_local_file,
    with_retry,
)
50

51
52
53
54
55
56
57
58
59
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,
    )

60
if envs.VLLM_USE_MODELSCOPE:
61
62
63
    from modelscope import AutoConfig
else:
    from transformers import AutoConfig
64

65
66
MISTRAL_CONFIG_NAME = "params.json"

67
68
logger = init_logger(__name__)

69

70
71
class LazyConfigDict(dict):
    def __getitem__(self, key):
72
73
74
        if isinstance(value := super().__getitem__(key), type):
            return value

75
        import vllm.transformers_utils.configs as configs
76

77
        return getattr(configs, value)
78
79
80


_CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
81
    afmoe="AfmoeConfig",
82
    bagel="BagelConfig",
83
    chatglm="ChatGLMConfig",
84
    colmodernvbert="ColModernVBertConfig",
85
    colpali="ColPaliConfig",
86
87
    colqwen3="ColQwen3Config",
    ops_colqwen3="OpsColQwen3Config",
88
    qwen3_vl_nemotron_embed="Qwen3VLNemotronEmbedConfig",
89
    deepseek_vl_v2="DeepseekVLV2Config",
90
    deepseek_v32="DeepseekV3Config",
91
    flex_olmo="FlexOlmoConfig",
92
    funaudiochat="FunAudioChatConfig",
93
    hunyuan_vl="HunYuanVLConfig",
oscardev256's avatar
oscardev256 committed
94
    isaac="IsaacConfig",
95
    kimi_k2="DeepseekV3Config",  # Kimi K2 uses same architecture as DeepSeek V3
96
    kimi_linear="KimiLinearConfig",
97
    kimi_vl="KimiVLConfig",
Roger Wang's avatar
Roger Wang committed
98
    kimi_k25="KimiK25Config",
99
100
101
102
103
    RefinedWeb="RWConfig",  # For tiiuae/falcon-40b(-instruct)
    RefinedWebModel="RWConfig",  # For tiiuae/falcon-7b(-instruct)
    jais="JAISConfig",
    mlp_speculator="MLPSpeculatorConfig",
    medusa="MedusaConfig",
104
    midashenglm="MiDashengLMConfig",
105
106
107
    eagle="EAGLEConfig",
    speculators="SpeculatorsConfig",
    nemotron="NemotronConfig",
108
    olmo3="Olmo3Config",
109
    olmo_hybrid="OlmoHybridConfig",
110
111
112
113
    ovis="OvisConfig",
    ultravox="UltravoxConfig",
    step3_vl="Step3VLConfig",
    step3_text="Step3TextConfig",
csy0225's avatar
csy0225 committed
114
    step3p5="Step3p5Config",
Roger Wang's avatar
Roger Wang committed
115
    qwen3_asr="Qwen3ASRConfig",
116
    qwen3_next="Qwen3NextConfig",
117
118
    qwen3_5="Qwen3_5Config",
    qwen3_5_moe="Qwen3_5MoeConfig",
Paul Pak's avatar
Paul Pak committed
119
    lfm2_moe="Lfm2MoeConfig",
120
    tarsier2="Tarsier2Config",
121
)
122

123
124
125
126
_CONFIG_ATTRS_MAPPING: dict[str, str] = {
    "llm_config": "text_config",
}

127
_AUTO_CONFIG_KWARGS_OVERRIDES: dict[str, dict[str, Any]] = {
128
    "internvl_chat": {"has_no_defaults_at_init": True},
129
    "Llama_Nemotron_Nano_VL": {"attn_implementation": "eager"},
130
    "NVLM_D": {"has_no_defaults_at_init": True},
131
132
}

133

134
135
136
137
138
139
140
141
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)


Julien Denize's avatar
Julien Denize committed
142
143
144
145
146
147
148
149
150
151
152
153
154
@contextmanager
def _mistral_patch_hf_hub_constants() -> Iterator[None]:
    hf_safetensors_single_file = constants.SAFETENSORS_SINGLE_FILE
    hf_safetensors_index_file = constants.SAFETENSORS_INDEX_FILE
    constants.SAFETENSORS_SINGLE_FILE = "consolidated.safetensors"
    constants.SAFETENSORS_INDEX_FILE = "consolidated.safetensors.index.json"
    try:
        yield
    finally:
        constants.SAFETENSORS_SINGLE_FILE = hf_safetensors_single_file
        constants.SAFETENSORS_INDEX_FILE = hf_safetensors_index_file


155
class HFConfigParser(ConfigParserBase):
156
157
    def parse(
        self,
158
        model: str | Path,
159
        trust_remote_code: bool,
160
161
        revision: str | None = None,
        code_revision: str | None = None,
162
163
        **kwargs,
    ) -> tuple[dict, PretrainedConfig]:
164
        kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
165
166
        trust_remote_code |= kwargs.get("trust_remote_code", False)
        kwargs = without_trust_remote_code(kwargs)
167
168
169
170
171
172
173
174
175
        config_dict, _ = PretrainedConfig.get_config_dict(
            model,
            revision=revision,
            code_revision=code_revision,
            **kwargs,
        )
        # Use custom model class if it's in our registry
        model_type = config_dict.get("model_type")
        if model_type is None:
176
177
178
179
180
            model_type = (
                "speculators"
                if config_dict.get("speculators_config") is not None
                else model_type
            )
181
182
        # Allow hf_overrides to override model_type before checking _CONFIG_REGISTRY
        if (hf_overrides := kwargs.pop("hf_overrides", None)) is not None:
183
184
185
186
187
188
189
190
191
192
            if isinstance(hf_overrides, dict) and "model_type" in hf_overrides:
                model_type = hf_overrides["model_type"]
            elif callable(hf_overrides):
                # If hf_overrides doesn't modify model_type, it will be passed straight
                # through and remain unchanged by this elif block
                dummy_model_type = f"dummy_{model_type}"
                dummy_kwargs = dict(architectures=[""], model_type=dummy_model_type)
                dummy_config = PretrainedConfig(**dummy_kwargs)
                dummy_model_type = hf_overrides(dummy_config).model_type
                model_type = dummy_model_type.removeprefix("dummy_")
193
194
195
196
197
198
199

        if model_type in _CONFIG_REGISTRY:
            config_class = _CONFIG_REGISTRY[model_type]
            config = config_class.from_pretrained(
                model,
                revision=revision,
                code_revision=code_revision,
200
                trust_remote_code=trust_remote_code,
201
202
203
204
                **kwargs,
            )
        else:
            try:
205
                kwargs = _maybe_update_auto_config_kwargs(kwargs, model_type=model_type)
206
207
208
209
210
211
212
213
                config = AutoConfig.from_pretrained(
                    model,
                    trust_remote_code=trust_remote_code,
                    revision=revision,
                    code_revision=code_revision,
                    **kwargs,
                )
            except ValueError as e:
214
215
216
217
                if (
                    not trust_remote_code
                    and "requires you to execute the configuration file" in str(e)
                ):
218
219
220
221
222
                    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 "
223
224
                        "`--trust-remote-code` flag in the CLI."
                    )
225
226
227
228
229
230
231
232
                    raise RuntimeError(err_msg) from e
                else:
                    raise e
        config = _maybe_remap_hf_config_attrs(config)
        return config_dict, config


class MistralConfigParser(ConfigParserBase):
233
234
    def parse(
        self,
235
        model: str | Path,
236
        trust_remote_code: bool,
237
238
        revision: str | None = None,
        code_revision: str | None = None,
239
240
        **kwargs,
    ) -> tuple[dict, PretrainedConfig]:
241
242
243
        # 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)
244
245
246
        if (
            max_position_embeddings := config_dict.get("max_position_embeddings")
        ) is None:
247
            max_position_embeddings = _maybe_retrieve_max_pos_from_hf(
248
249
                model, revision, **kwargs
            )
250
251
252
253
            config_dict["max_position_embeddings"] = max_position_embeddings

        from vllm.transformers_utils.configs.mistral import adapt_config_dict

254
255
256
257
258
259
        # Get missing fields from HF config if available
        try:
            hf_config_dict, _ = PretrainedConfig.get_config_dict(
                model,
                revision=revision,
                code_revision=code_revision,
260
                **without_trust_remote_code(kwargs),
261
262
263
264
            )
        except OSError:  # Not found
            hf_config_dict = {}

Julien Denize's avatar
Julien Denize committed
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
        if config_dict.get("dtype") is None:
            with _mistral_patch_hf_hub_constants():
                model_str = model if isinstance(model, str) else model.as_posix()
                param_mt = get_safetensors_params_metadata(model_str, revision=revision)
            if param_mt:
                param_dtypes: set[torch.dtype] = {
                    _SAFETENSORS_TO_TORCH_DTYPE[dtype]
                    for info in param_mt.values()
                    if (dtype := info.get("dtype", None))
                    and dtype in _SAFETENSORS_TO_TORCH_DTYPE
                }

                if param_dtypes:
                    config_dict["dtype"] = common_broadcastable_dtype(param_dtypes)
                    logger.info_once(
                        "Inferred from consolidated*.safetensors files "
                        f"{config_dict['dtype']} dtype."
                    )

284
        config = adapt_config_dict(config_dict, defaults=hf_config_dict)
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309

        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.
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
     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,
326
327
         ...         revision: str | None = None,
         ...         code_revision: str | None = None,
328
329
330
331
332
333
         ...         **kwargs,
         ...     ) -> tuple[dict, PretrainedConfig]:
         ...         raise NotImplementedError
         >>>
         >>> type(get_config_parser("custom_config_parser"))
         <class 'CustomConfigParser'>
334
335
336
337
338
339
    """  # 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 "
340
341
342
343
                "overwritten by the new parser class `%s`.",
                config_format,
                config_parser_cls,
            )
344
        if not issubclass(config_parser_cls, ConfigParserBase):
345
346
347
            raise ValueError(
                "The config parser must be a subclass of `ConfigParserBase`."
            )
348
        _CONFIG_FORMAT_TO_CONFIG_PARSER[config_format] = config_parser_cls
349
350
351
352
353
        logger.info(
            "Registered config parser `%s` with config format `%s`",
            config_parser_cls,
            config_format,
        )
354
355
356
        return config_parser_cls

    return _wrapper
357
358


359
360
361
362
363
364
365
366
367
368
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:
369
    """Provide backwards compatibility for RoPE."""
370
371
    from vllm.config.utils import getattr_iter

372
373
374
375
376
377
    # 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)
378
    ompe = getattr(config, "original_max_position_embeddings", None)
379

380
    if Version(version("transformers")) < Version("5.0.0"):
381
        # Transformers v4 installed, legacy config fields may be present
382
383
        if (rope_scaling := getattr(config, "rope_scaling", None)) is not None:
            config.rope_parameters = rope_scaling
384
        if (
385
386
387
            rope_theta is not None
            or partial_rotary_factor is not None
            or ompe is not None
388
389
        ) and not getattr(config, "rope_parameters", None):
            config.rope_parameters = {"rope_type": "default"}
390
        # Patch legacy fields into rope_parameters
391
        if rope_theta is not None:
392
            config.rope_parameters["rope_theta"] = rope_theta
393
394
        if partial_rotary_factor is not None:
            config.rope_parameters["partial_rotary_factor"] = partial_rotary_factor
395
396
        if ompe is not None:
            config.rope_parameters["original_max_position_embeddings"] = ompe
397
    elif rope_theta is not None or getattr(config, "rope_parameters", None):
398
        # Transformers v5 installed
399
400
401
402
403
404
        # 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
405
406
        config.standardize_rope_params()
        config.validate_rope()
407
408

    # No RoPE parameters to patch
409
    if getattr(config, "rope_parameters", None) is None:
410
411
412
        return

    # Handle nested rope_parameters in interleaved sliding attention models
413
    if is_rope_parameters_nested(config.rope_parameters):
414
        for rope_parameters_layer_type in config.rope_parameters.values():
415
416
            patch_rope_parameters_dict(rope_parameters_layer_type)
    else:
417
        patch_rope_parameters_dict(config.rope_parameters)
418
419


420
421
422
423
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"]
424
425
426
427
428
        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:
429
430
431
            raise ValueError(
                f"Found conflicts between 'rope_type={rope_type}' (modern "
                f"field) and 'type={rope_type_legacy}' (legacy field). "
432
433
                "You should only specify one of them."
            )
434

435
436
    if "rope_type" not in rope_parameters and "type" in rope_parameters:
        rope_parameters["rope_type"] = rope_parameters["type"]
437
438
        logger.info("Replacing legacy 'type' key with 'rope_type'")

439
440
    if "rope_type" not in rope_parameters:
        raise ValueError("rope_parameters should have a 'rope_type' key")
441

442
443
    if rope_parameters["rope_type"] == "su":
        rope_parameters["rope_type"] = "longrope"
444
        logger.warning("Replacing legacy rope_type 'su' with 'longrope'")
445
    elif rope_parameters["rope_type"] == "mrope":
446
447
448
449
        if "mrope_section" not in rope_parameters:
            raise ValueError(
                "Legacy rope_type 'mrope' requires 'mrope_section' in rope_parameters"
            )
450
        rope_parameters["rope_type"] = "default"
451
452
453
        logger.warning("Replacing legacy rope_type 'mrope' with 'default'")


454
def _uses_mrope(config: PretrainedConfig) -> bool:
455
456
    rope_parameters = getattr(config, "rope_parameters", None)
    if rope_parameters is None:
457
458
        return False

459
    return "mrope_section" in rope_parameters
460
461


462
463
def uses_mrope(config: PretrainedConfig) -> bool:
    """Detect if the model with this config uses M-ROPE."""
464
465
466
467
468
    return (
        _uses_mrope(config)
        or _uses_mrope(config.get_text_config())
        or thinker_uses_mrope(config)
    )
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483


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)


484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
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


501
502
503
def is_encoder_decoder(config: PretrainedConfig) -> bool:
    """Detect if the model with this config is used as an encoder/decoder."""

504
505
506
    def _is_encoder_decoder(config: PretrainedConfig) -> bool:
        return getattr(config, "is_encoder_decoder", False)

507
    return _is_encoder_decoder(config) or _is_encoder_decoder(config.get_text_config())
508
509


510
511
512
513
514
515
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):
516
        return len(set(layer_types)) > 1
517
518
519
    return False


520
521
522
523
524
525
526
527
528
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


529
530
531
532
533
534
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)})
535
            logger.debug("Remapped config attribute '%s' to '%s'", old_attr, new_attr)
536
537
538
    return config


539
def maybe_override_with_speculators(
540
    model: str,
541
    tokenizer: str | None,
542
    trust_remote_code: bool,
543
544
    revision: str | None = None,
    vllm_speculative_config: dict[str, Any] | None = None,
545
    hf_token: bool | str | None = None,
546
    **kwargs,
547
) -> tuple[str, str | None, dict[str, Any] | None]:
548
    """
549
550
551
552
553
554
555
556
557
558
559
    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
560
        hf_token: HuggingFace token for authenticated model access
561
562
563

    Returns:
        Tuple of (resolved_model, resolved_tokenizer, speculative_config)
564
    """
565
    if check_gguf_file(model):
566
567
        kwargs["gguf_file"] = Path(model).name
        gguf_model_repo = Path(model).parent
568
569
570
    elif is_remote_gguf(model):
        repo_id, _ = split_remote_gguf(model)
        gguf_model_repo = Path(repo_id)
571
572
    else:
        gguf_model_repo = None
573
    kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
574
    config_dict, _ = PretrainedConfig.get_config_dict(
575
        model if gguf_model_repo is None else gguf_model_repo,
576
        revision=revision,
577
        token=hf_token,
578
        **without_trust_remote_code(kwargs),
579
    )
580
581
582
583
584
585
586
    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
587
    from vllm.transformers_utils.configs.speculators.base import SpeculatorsConfig
588

589
    speculative_config = SpeculatorsConfig.extract_vllm_speculative_config(
590
591
        config_dict=config_dict
    )
592
593

    # Set the draft model to the speculators model
594
    speculative_config["model"] = model
595
596
597
598
599

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

600
    return model, tokenizer, speculative_config
601
602


603
def get_config(
604
    model: str | Path,
605
    trust_remote_code: bool,
606
607
608
609
610
    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,
611
612
613
    **kwargs,
) -> PretrainedConfig:
    # Separate model folder from file path for GGUF models
614

615
616
617
618
619
620
621
622
623
624
625
626
    _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)
627

628
    if config_format == "auto":
629
        try:
630
631
            # First check for Mistral to avoid defaulting to
            # Transformers implementation.
632
633
634
635
636
            if is_mistral_model_repo(
                model_name_or_path=str(model), revision=revision
            ) and file_or_path_exists(
                model=model, config_name=MISTRAL_CONFIG_NAME, revision=revision
            ):
637
                config_format = "mistral"
638
            elif (_is_gguf and not _is_remote_gguf) or file_or_path_exists(
639
640
641
                model, HF_CONFIG_NAME, revision=revision
            ):
                config_format = "hf"
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
            # 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)
657
658
659
            else:
                raise ValueError(
                    "Could not detect config format for no config file found. "
660
661
662
                    "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 "
663
664
                    "in engine args for customized config parser."
                )
665
666
667
668
669
670
671
672
673
674
675

        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 "
676
                "'params.json'.\n"
677
            ).format(model=model)
678
679

            raise ValueError(error_message) from e
680

681
682
683
684
685
686
    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,
687
        hf_overrides=hf_overrides_kw or hf_overrides_fn,
688
689
        **kwargs,
    )
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711

    # 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)

712
    # Special architecture mapping check for GGUF models
713
    if _is_gguf:
714
        if config.model_type not in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
715
            raise RuntimeError(f"Can't get gguf config for {config.model_type}.")
716
717
718
        model_type = MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[config.model_type]
        config.update({"architectures": [model_type]})

719
720
721
    # Architecture mapping for models without explicit architectures field
    if not config.architectures:
        if config.model_type not in MODEL_MAPPING_NAMES:
722
723
724
725
726
727
728
729
            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]})
730

731
732
733
734
735
736
    # 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.
737
738
739
740
741
742
    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
        )
743
744
745

    if quantization_config is not None:
        config.quantization_config = quantization_config
746
        # auto-enable DeepGEMM UE8M0 if model config requests it
747
        scale_fmt = quantization_config.get("scale_fmt", None)
748
        if scale_fmt in ("ue8m0",):
749
750
            if not envs.is_set("VLLM_USE_DEEP_GEMM_E8M0"):
                os.environ["VLLM_USE_DEEP_GEMM_E8M0"] = "1"
751
                logger.info_once(
752
753
                    (
                        "Detected quantization_config.scale_fmt=%s; "
754
                        "enabling UE8M0 for DeepGEMM."
755
                    ),
756
757
                    scale_fmt,
                )
758
            elif not envs.VLLM_USE_DEEP_GEMM_E8M0:
759
                logger.warning_once(
760
761
762
                    (
                        "Model config requests UE8M0 "
                        "(quantization_config.scale_fmt=%s), but "
763
764
                        "VLLM_USE_DEEP_GEMM_E8M0=0 is set; "
                        "UE8M0 for DeepGEMM disabled."
765
                    ),
766
767
                    scale_fmt,
                )
768

769
770
771
772
773
774
775
    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)

776
777
778
779
780
781
782
783
    # 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))
784

785
786
787
    if trust_remote_code:
        maybe_register_config_serialize_by_value()

788
    return config
789
790


791
@cache
792
793
794
795
def get_pooling_config(
    model: str,
    revision: str | None = "main",
) -> dict[str, Any] | None:
796
    """
797
798
799
    This function gets the pooling and normalize
    config from the model - only applies to
    sentence-transformers models.
800
801

    Args:
802
        model: The name of the Hugging Face model.
803
        revision: The specific version of the model to use.
804
            Defaults to 'main'.
805
806

    Returns:
807
        A dictionary containing the pooling type and whether
808
            normalization is used, or None if no pooling configuration is found.
809
    """
810
811
    if is_remote_gguf(model):
        model, _ = split_remote_gguf(model)
812
813

    modules_file_name = "modules.json"
814
815

    modules_dict = None
816
817
818
    if file_or_path_exists(
        model=model, config_name=modules_file_name, revision=revision
    ):
819
        modules_dict = get_hf_file_to_dict(modules_file_name, model, revision)
820
821
822
823

    if modules_dict is None:
        return None

824
825
    logger.info("Found sentence-transformers modules configuration.")

826
827
828
829
830
831
832
833
    pooling = next(
        (
            item
            for item in modules_dict
            if item["type"] == "sentence_transformers.models.Pooling"
        ),
        None,
    )
834
    normalize = bool(
835
836
837
838
839
840
841
842
843
        next(
            (
                item
                for item in modules_dict
                if item["type"] == "sentence_transformers.models.Normalize"
            ),
            False,
        )
    )
844
845

    if pooling:
846
        from vllm.config.pooler import SEQ_POOLING_TYPES, TOK_POOLING_TYPES
847

848
849
        pooling_file_name = "{}/config.json".format(pooling["path"])
        pooling_dict = get_hf_file_to_dict(pooling_file_name, model, revision) or {}
850

851
        logger.info("Found pooling configuration.")
852

853
        config: dict[str, Any] = {"use_activation": normalize}
854
855
856
857
858
859
860
861
862
863
864
        for key, val in pooling_dict.items():
            if val is True:
                pooling_type = parse_pooling_type(key)
                if pooling_type in SEQ_POOLING_TYPES:
                    config["seq_pooling_type"] = pooling_type
                elif pooling_type in TOK_POOLING_TYPES:
                    config["tok_pooling_type"] = pooling_type
                else:
                    logger.debug("Skipping unrelated field: %r=%r", key, val)

        return config
865
866
867
868

    return None


869
def parse_pooling_type(pooling_name: str):
870
871
872
873
    if "pooling_mode_" in pooling_name:
        pooling_name = pooling_name.replace("pooling_mode_", "")

    if "_" in pooling_name:
874
        pooling_name = pooling_name.split("_", 1)[0]
875
876
877
878

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

879
    return pooling_name.upper()
880
881


882
@cache
883
def get_sentence_transformer_tokenizer_config(
884
    model: str | Path, revision: str | None = "main"
885
) -> dict[str, Any] | None:
886
    """
887
    Returns the tokenization configuration dictionary for a
888
889
890
    given Sentence Transformer BERT model.

    Parameters:
891
    - model (str|Path): The name of the Sentence Transformer
892
893
894
895
896
    BERT model.
    - revision (str, optional): The revision of the m
    odel to use. Defaults to 'main'.

    Returns:
897
    - dict: A dictionary containing the configuration parameters
898
899
    for the Sentence Transformer BERT model.
    """
900
901
902
903
904
905
906
907
908
909
    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
910
911

    for config_file in sentence_transformer_config_files:
912
913
914
915
        if (
            try_get_local_file(model=model, file_name=config_file, revision=revision)
            is not None
        ):
916
            encoder_dict = get_hf_file_to_dict(config_file, model, revision)
917
918
            if encoder_dict:
                break
919

920
    if not encoder_dict and not Path(model).is_absolute():
921
922
        try:
            # If model is on HuggingfaceHub, get the repo files
923
            repo_files = list_repo_files(model, revision=revision)
924
        except Exception:
925
926
927
928
            repo_files = []

        for config_name in sentence_transformer_config_files:
            if config_name in repo_files:
929
                encoder_dict = get_hf_file_to_dict(config_name, model, revision)
930
931
932
                if encoder_dict:
                    break

933
934
935
    if not encoder_dict:
        return None

936
937
    logger.info("Found sentence-transformers tokenize configuration.")

938
939
940
941
942
    if all(k in encoder_dict for k in ("max_seq_length", "do_lower_case")):
        return encoder_dict
    return None


943
def maybe_register_config_serialize_by_value() -> None:
944
945
    """Try to register HF model configuration class to serialize by value

946
947
948
    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.
949

950
    Examples:
951

952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
    >>> 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
976
977
    try:
        import transformers_modules
978

979
        transformers_modules_available = True
980
    except ImportError:
981
        transformers_modules_available = False
982
983
984
985
986

    try:
        import multiprocessing
        import pickle

987
988
        import cloudpickle

989
        from vllm.config import VllmConfig
990

991
992
993
        # Register multiprocessing reducers to handle cross-process
        # serialization of VllmConfig objects that may contain custom configs
        # from transformers_modules
994
        def _reduce_config(config: VllmConfig):
995
            return (pickle.loads, (cloudpickle.dumps(config),))
996

997
        multiprocessing.reducer.register(VllmConfig, _reduce_config)
998

999
1000
1001
1002
1003
        # 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
1004
            from vllm.v1.executor.ray_utils import ray
1005

1006
1007
1008
            if ray:
                ray.cloudpickle.register_pickle_by_value(transformers_modules)

1009
1010
1011
1012
1013
1014
    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`",
1015
1016
            exc_info=e,
        )
1017
1018


1019
def get_hf_image_processor_config(
1020
1021
1022
    model: str | Path,
    hf_token: bool | str | None = None,
    revision: str | None = None,
1023
    **kwargs,
1024
) -> dict[str, Any]:
1025
    # ModelScope does not provide an interface for image_processor
1026
    if envs.VLLM_USE_MODELSCOPE:
1027
        return dict()
1028
    # Separate model folder from file path for GGUF models
1029
    if check_gguf_file(model):
1030
        model = Path(model).parent
1031
1032
    elif is_remote_gguf(model):
        model, _ = split_remote_gguf(model)
1033
1034
1035
    return get_image_processor_config(
        model, token=hf_token, revision=revision, **kwargs
    )
1036
1037


1038
1039
def get_hf_text_config(config: PretrainedConfig):
    """Get the "sub" config relevant to llm for multi modal models.
1040
    No op for pure text models.
1041
    """
1042
1043
    text_config = config.get_text_config()

1044
1045
1046
1047
1048
1049
1050
    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."
        )
1051
1052

    return text_config
1053
1054
1055
1056
1057


def try_get_generation_config(
    model: str,
    trust_remote_code: bool,
1058
1059
    revision: str | None = None,
    config_format: str | ConfigFormat = "auto",
1060
    hf_token: bool | str | None = None,
1061
) -> GenerationConfig | None:
1062
1063
1064
1065
1066
1067
1068
    # 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

1069
1070
1071
1072
    try:
        return GenerationConfig.from_pretrained(
            model,
            revision=revision,
1073
            token=hf_token,
1074
1075
1076
1077
1078
1079
1080
        )
    except OSError:  # Not found
        try:
            config = get_config(
                model,
                trust_remote_code=trust_remote_code,
                revision=revision,
1081
                config_format=config_format,
1082
                token=hf_token,
1083
1084
1085
1086
            )
            return GenerationConfig.from_model_config(config)
        except OSError:  # Not found
            return None
1087
1088


1089
1090
1091
def try_get_safetensors_metadata(
    model: str,
    *,
1092
    revision: str | None = None,
1093
1094
):
    get_safetensors_metadata_partial = partial(
1095
        get_safetensors_metadata, model, revision=revision
1096
1097
1098
    )

    try:
1099
1100
1101
        return with_retry(
            get_safetensors_metadata_partial, "Error retrieving safetensors"
        )
1102
1103
    except Exception:
        return None
1104
1105
1106


def try_get_tokenizer_config(
1107
    pretrained_model_name_or_path: str | os.PathLike,
1108
    trust_remote_code: bool,
1109
1110
    revision: str | None = None,
) -> dict[str, Any] | None:
1111
1112
1113
1114
1115
1116
1117
1118
    try:
        return get_tokenizer_config(
            pretrained_model_name_or_path,
            trust_remote_code=trust_remote_code,
            revision=revision,
        )
    except Exception:
        return None
1119
1120


1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
@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", [])

1134
1135
1136
1137
1138
        _DENSE_MODULE_TYPES = {
            "sentence_transformers.models.Dense",
            "pylate.models.Dense.Dense",
        }
        dense_modules = [m for m in modules if m.get("type") in _DENSE_MODULE_TYPES]
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
        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


1157
1158
1159
def get_safetensors_params_metadata(
    model: str,
    *,
1160
    revision: str | None = None,
1161
1162
) -> dict[str, Any]:
    """
1163
    Get the safetensors parameters metadata for remote/local model repository.
1164
1165
1166
1167
1168
1169
    """
    full_metadata = {}
    if (model_path := Path(model)).exists():
        safetensors_to_check = model_path.glob("*.safetensors")
        full_metadata = {
            param_name: info
1170
1171
1172
            for file_path in safetensors_to_check
            if file_path.is_file()
            for param_name, info in parse_safetensors_file_metadata(file_path).items()
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
        }
    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


1185
1186
1187
1188
1189
1190
1191
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 "
1192
1193
            f"and if the config file exists."
        )
1194
1195
1196
1197
1198
1199
1200
1201
    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)
1202
1203
1204
1205
1206
1207
        hf_config = get_config(
            model=model,
            trust_remote_code=trust_remote_code_val,
            revision=revision,
            config_format="hf",
        )
1208
1209
1210
1211
1212
1213
1214
        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",
1215
1216
            exc_info=e,
        )
1217
1218

    return max_position_embeddings