config.py 32.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
import enum
import json
6
import os
7
import time
8
from functools import cache, partial
9
from pathlib import Path
10
from typing import Any, Callable, Literal, Optional, TypeVar, Union
Jasmond L's avatar
Jasmond L committed
11

Joe Runde's avatar
Joe Runde committed
12
import huggingface_hub
13
from huggingface_hub import get_safetensors_metadata, hf_hub_download
14
15
from huggingface_hub import list_repo_files as hf_list_repo_files
from huggingface_hub import try_to_load_from_cache
16
from huggingface_hub.utils import (EntryNotFoundError, HfHubHTTPError,
17
                                   HFValidationError, LocalEntryNotFoundError,
18
19
                                   RepositoryNotFoundError,
                                   RevisionNotFoundError)
20
from torch import nn
21
from transformers import GenerationConfig, PretrainedConfig
22
23
from transformers.models.auto.image_processing_auto import (
    get_image_processor_config)
24
25
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES)
26
from transformers.models.auto.tokenization_auto import get_tokenizer_config
27
from transformers.utils import CONFIG_NAME as HF_CONFIG_NAME
28

29
from vllm import envs
30
from vllm.logger import init_logger
31
32
# yapf conflicts with isort for this block
# yapf: disable
33
34
35
36
from vllm.transformers_utils.configs import (ChatGLMConfig, Cohere2Config,
                                             DbrxConfig, DeepseekVLV2Config,
                                             EAGLEConfig, ExaoneConfig,
                                             H2OVLChatConfig,
37
                                             InternVLChatConfig, JAISConfig,
38
                                             KimiVLConfig, MedusaConfig,
39
40
41
42
                                             MiniMaxText01Config,
                                             MiniMaxVL01Config, MllamaConfig,
                                             MLPSpeculatorConfig, MPTConfig,
                                             NemotronConfig, NVLM_D_Config,
43
44
45
                                             OvisConfig, RWConfig,
                                             SkyworkR1VChatConfig, SolarConfig,
                                             Telechat2Config, UltravoxConfig)
46
# yapf: enable
47
from vllm.transformers_utils.utils import check_gguf_file
48
from vllm.utils import resolve_obj_by_qualname
49

50
if envs.VLLM_USE_MODELSCOPE:
51
52
53
    from modelscope import AutoConfig
else:
    from transformers import AutoConfig
54

55
56
MISTRAL_CONFIG_NAME = "params.json"

57
58
logger = init_logger(__name__)

59
_CONFIG_REGISTRY_OVERRIDE_HF: dict[str, type[PretrainedConfig]] = {
60
61
62
    "mllama": MllamaConfig
}

63
_CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = {
GoHomeToMacDonal's avatar
GoHomeToMacDonal committed
64
    "chatglm": ChatGLMConfig,
65
    "cohere2": Cohere2Config,
66
    "dbrx": DbrxConfig,
67
    "deepseek_vl_v2": DeepseekVLV2Config,
68
    "kimi_vl": KimiVLConfig,
69
    "mpt": MPTConfig,
Zhuohan Li's avatar
Zhuohan Li committed
70
71
    "RefinedWeb": RWConfig,  # For tiiuae/falcon-40b(-instruct)
    "RefinedWebModel": RWConfig,  # For tiiuae/falcon-7b(-instruct)
72
    "jais": JAISConfig,
73
    "mlp_speculator": MLPSpeculatorConfig,
74
    "medusa": MedusaConfig,
75
    "eagle": EAGLEConfig,
76
    "exaone": ExaoneConfig,
77
    "h2ovl_chat": H2OVLChatConfig,
78
    "internvl_chat": InternVLChatConfig,
79
80
    "minimax_text_01": MiniMaxText01Config,
    "minimax_vl_01": MiniMaxVL01Config,
81
    "nemotron": NemotronConfig,
82
    "NVLM_D": NVLM_D_Config,
83
    "ovis": OvisConfig,
84
    "solar": SolarConfig,
85
    "skywork_chat": SkyworkR1VChatConfig,
86
    "telechat": Telechat2Config,
87
    "ultravox": UltravoxConfig,
88
    **_CONFIG_REGISTRY_OVERRIDE_HF
89
90
91
}


92
93
94
95
96
97
class ConfigFormat(str, enum.Enum):
    AUTO = "auto"
    HF = "hf"
    MISTRAL = "mistral"


98
99
100
101
102
103
104
105
106
_R = TypeVar("_R")


def with_retry(
    func: Callable[[], _R],
    log_msg: str,
    max_retries: int = 2,
    retry_delay: int = 2,
) -> _R:
107
108
109
110
111
112
113
114
115
116
117
118
    for attempt in range(max_retries):
        try:
            return func()
        except Exception as e:
            if attempt == max_retries - 1:
                logger.error("%s: %s", log_msg, e)
                raise
            logger.error("%s: %s, retrying %d of %d", log_msg, e, attempt + 1,
                         max_retries)
            time.sleep(retry_delay)
            retry_delay *= 2

119
120
    raise AssertionError("Should not be reached")

121
122
123
124
125
126
127
128
129
130
131

# @cache doesn't cache exceptions
@cache
def list_repo_files(
    repo_id: str,
    *,
    revision: Optional[str] = None,
    repo_type: Optional[str] = None,
    token: Union[str, bool, None] = None,
) -> list[str]:

132
133
134
135
136
137
138
139
    def lookup_files() -> list[str]:
        # directly list files if model is local
        if (local_path := Path(repo_id)).exists():
            return [
                str(file.relative_to(local_path))
                for file in local_path.rglob('*') if file.is_file()
            ]
        # if model is remote, use hf_hub api to list files
140
        try:
141
            if envs.VLLM_USE_MODELSCOPE:
142
143
144
145
                from vllm.transformers_utils.utils import (
                    modelscope_list_repo_files)
                return modelscope_list_repo_files(repo_id,
                                                  revision=revision,
146
147
148
                                                  token=os.getenv(
                                                      "MODELSCOPE_API_TOKEN",
                                                      None))
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
            return hf_list_repo_files(repo_id,
                                      revision=revision,
                                      repo_type=repo_type,
                                      token=token)
        except huggingface_hub.errors.OfflineModeIsEnabled:
            # Don't raise in offline mode,
            # all we know is that we don't have this
            # file cached.
            return []

    return with_retry(lookup_files, "Error retrieving file list")


def file_exists(
    repo_id: str,
    file_name: str,
    *,
    repo_type: Optional[str] = None,
    revision: Optional[str] = None,
    token: Union[str, bool, None] = None,
) -> bool:
    file_list = list_repo_files(repo_id,
                                repo_type=repo_type,
                                revision=revision,
                                token=token)
    return file_name in file_list


# In offline mode the result can be a false negative
178
179
def file_or_path_exists(model: Union[str, Path], config_name: str,
                        revision: Optional[str]) -> bool:
180
181
    if (local_path := Path(model)).exists():
        return (local_path / config_name).is_file()
182

Joe Runde's avatar
Joe Runde committed
183
184
185
186
187
188
189
190
191
192
    # Offline mode support: Check if config file is cached already
    cached_filepath = try_to_load_from_cache(repo_id=model,
                                             filename=config_name,
                                             revision=revision)
    if isinstance(cached_filepath, str):
        # The config file exists in cache- we can continue trying to load
        return True

    # NB: file_exists will only check for the existence of the config file on
    # hf_hub. This will fail in offline mode.
193
194

    # Call HF to check if the file exists
195
196
197
    return file_exists(str(model),
                       config_name,
                       revision=revision,
198
                       token=os.getenv('HF_TOKEN', None))
199
200


201
202
203
204
205
206
207
208
209
210
211
def patch_rope_scaling(config: PretrainedConfig) -> None:
    """Provide backwards compatibility for RoPE."""
    text_config = getattr(config, "text_config", None)
    if text_config is not None:
        patch_rope_scaling(text_config)

    rope_scaling = getattr(config, "rope_scaling", None)
    if rope_scaling is not None:
        patch_rope_scaling_dict(rope_scaling)


212
def patch_rope_scaling_dict(rope_scaling: dict[str, Any]) -> None:
213
214
215
216
217
218
219
220
221
    if "rope_type" in rope_scaling and "type" in rope_scaling:
        rope_type = rope_scaling["rope_type"]
        rope_type_legacy = rope_scaling["type"]
        if rope_type != rope_type_legacy:
            raise ValueError(
                f"Found conflicts between 'rope_type={rope_type}' (modern "
                f"field) and 'type={rope_type_legacy}' (legacy field). "
                "You should only specify one of them.")

222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
    if "rope_type" not in rope_scaling and "type" in rope_scaling:
        rope_scaling["rope_type"] = rope_scaling["type"]
        logger.info("Replacing legacy 'type' key with 'rope_type'")

    if "rope_type" not in rope_scaling:
        raise ValueError("rope_scaling should have a 'rope_type' key")

    if rope_scaling["rope_type"] == "su":
        rope_scaling["rope_type"] = "longrope"
        logger.warning("Replacing legacy rope_type 'su' with 'longrope'")
    elif rope_scaling["rope_type"] == "mrope":
        assert "mrope_section" in rope_scaling
        rope_scaling["rope_type"] = "default"
        logger.warning("Replacing legacy rope_type 'mrope' with 'default'")


238
def _uses_mrope(config: PretrainedConfig) -> bool:
239
240
241
242
243
244
245
    rope_scaling = getattr(config, "rope_scaling", None)
    if rope_scaling is None:
        return False

    return "mrope_section" in rope_scaling


246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
def uses_mrope(config: PretrainedConfig) -> bool:
    """Detect if the model with this config uses M-ROPE."""
    return _uses_mrope(config) or thinker_uses_mrope(config)


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)


264
265
266
267
268
269
270
271
272
def is_encoder_decoder(config: PretrainedConfig) -> bool:
    """Detect if the model with this config is used as an encoder/decoder."""
    text_config = getattr(config, "text_config", None)
    if text_config is not None:
        return is_encoder_decoder(text_config)

    return getattr(config, "is_encoder_decoder", False)


273
274
275
276
277
def get_config(
    model: Union[str, Path],
    trust_remote_code: bool,
    revision: Optional[str] = None,
    code_revision: Optional[str] = None,
278
    config_format: ConfigFormat = ConfigFormat.AUTO,
279
280
281
    **kwargs,
) -> PretrainedConfig:
    # Separate model folder from file path for GGUF models
282

283
    is_gguf = check_gguf_file(model)
284
285
286
287
    if is_gguf:
        kwargs["gguf_file"] = Path(model).name
        model = Path(model).parent

288
    if config_format == ConfigFormat.AUTO:
289
290
291
292
293
294
295
296
        try:
            if is_gguf or file_or_path_exists(
                    model, HF_CONFIG_NAME, revision=revision):
                config_format = ConfigFormat.HF
            elif file_or_path_exists(model,
                                     MISTRAL_CONFIG_NAME,
                                     revision=revision):
                config_format = ConfigFormat.MISTRAL
297
298
299
300
301
            else:
                raise ValueError(
                    "Could not detect config format for no config file found. "
                    "Ensure your model has either config.json (HF format) "
                    "or params.json (Mistral format).")
302
303
304
305
306
307
308
309
310
311
312

        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 "
313
314
315
316
                "'params.json'.\n"
                "3. For GGUF: pass the local path of the GGUF checkpoint.\n"
                "   Loading GGUF from a remote repo directly is not yet "
                "supported.\n").format(model=model)
317
318

            raise ValueError(error_message) from e
319
320
321

    if config_format == ConfigFormat.HF:
        config_dict, _ = PretrainedConfig.get_config_dict(
322
323
324
            model,
            revision=revision,
            code_revision=code_revision,
325
            token=os.getenv('HF_TOKEN', None),
326
327
            **kwargs,
        )
328
329
330
331
332

        # Use custom model class if it's in our registry
        model_type = config_dict.get("model_type")
        if model_type in _CONFIG_REGISTRY:
            config_class = _CONFIG_REGISTRY[model_type]
333
334
335
336
            config = config_class.from_pretrained(
                model,
                revision=revision,
                code_revision=code_revision,
337
                token=os.getenv('HF_TOKEN', None),
338
339
                **kwargs,
            )
340
341
342
343
344
345
346
        else:
            try:
                config = AutoConfig.from_pretrained(
                    model,
                    trust_remote_code=trust_remote_code,
                    revision=revision,
                    code_revision=code_revision,
347
                    token=os.getenv('HF_TOKEN', None),
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
                    **kwargs,
                )
            except ValueError as e:
                if (not trust_remote_code
                        and "requires you to execute the configuration file"
                        in str(e)):
                    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 "
                        "`--trust-remote-code` flag in the CLI.")
                    raise RuntimeError(err_msg) from e
                else:
                    raise e

    elif config_format == ConfigFormat.MISTRAL:
365
        config = load_params_config(model, revision, **kwargs)
366
    else:
367
368
369
370
371
372
373
374
        supported_formats = [
            fmt.value for fmt in ConfigFormat if fmt != ConfigFormat.AUTO
        ]
        raise ValueError(
            f"Unsupported config format: {config_format}. "
            f"Supported formats are: {', '.join(supported_formats)}. "
            f"Ensure your model uses one of these configuration formats "
            f"or specify the correct format explicitly.")
375
376
377
378
379
380
381
382
383

    # Special architecture mapping check for GGUF models
    if is_gguf:
        if config.model_type not in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
            raise RuntimeError(
                f"Can't get gguf config for {config.model_type}.")
        model_type = MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[config.model_type]
        config.update({"architectures": [model_type]})

384
385
    patch_rope_scaling(config)

386
387
388
    if trust_remote_code:
        maybe_register_config_serialize_by_value()

389
    return config
390
391


392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
def try_get_local_file(model: Union[str, Path],
                       file_name: str,
                       revision: Optional[str] = 'main') -> Optional[Path]:
    file_path = Path(model) / file_name
    if file_path.is_file():
        return file_path
    else:
        try:
            cached_filepath = try_to_load_from_cache(repo_id=model,
                                                     filename=file_name,
                                                     revision=revision)
            if isinstance(cached_filepath, str):
                return Path(cached_filepath)
        except HFValidationError:
            ...
    return None


410
411
def get_hf_file_to_dict(file_name: str,
                        model: Union[str, Path],
412
                        revision: Optional[str] = 'main'):
413
    """
414
    Downloads a file from the Hugging Face Hub and returns
415
416
417
418
419
    its contents as a dictionary.

    Parameters:
    - file_name (str): The name of the file to download.
    - model (str): The name of the model on the Hugging Face Hub.
420
    - revision (str): The specific version of the model.
421
422

    Returns:
423
    - config_dict (dict): A dictionary containing
424
425
426
    the contents of the downloaded file.
    """

427
428
429
430
    file_path = try_get_local_file(model=model,
                                   file_name=file_name,
                                   revision=revision)

431
    if file_path is None:
432
433
        try:
            hf_hub_file = hf_hub_download(model, file_name, revision=revision)
434
435
        except huggingface_hub.errors.OfflineModeIsEnabled:
            return None
436
437
438
439
440
441
442
443
444
445
446
447
448
449
        except (RepositoryNotFoundError, RevisionNotFoundError,
                EntryNotFoundError, LocalEntryNotFoundError) as e:
            logger.debug("File or repository not found in hf_hub_download", e)
            return None
        except HfHubHTTPError as e:
            logger.warning(
                "Cannot connect to Hugging Face Hub. Skipping file "
                "download for '%s':",
                file_name,
                exc_info=e)
            return None
        file_path = Path(hf_hub_file)

    if file_path is not None and file_path.is_file():
450
451
        with open(file_path) as file:
            return json.load(file)
452

453
454
455
    return None


456
@cache
457
def get_pooling_config(model: str, revision: Optional[str] = 'main'):
458
    """
459
460
461
    This function gets the pooling and normalize
    config from the model - only applies to
    sentence-transformers models.
462
463
464

    Args:
        model (str): The name of the Hugging Face model.
465
        revision (str, optional): The specific version
466
467
468
        of the model to use. Defaults to 'main'.

    Returns:
469
        dict: A dictionary containing the pooling
470
471
472
473
        type and whether normalization is used.
    """

    modules_file_name = "modules.json"
474
475
476
477
478
479

    modules_dict = None
    if file_or_path_exists(model=model,
                           config_name=modules_file_name,
                           revision=revision):
        modules_dict = get_hf_file_to_dict(modules_file_name, model, revision)
480
481
482
483

    if modules_dict is None:
        return None

484
485
    logger.info("Found sentence-transformers modules configuration.")

486
487
488
489
490
491
492
493
494
495
496
    pooling = next((item for item in modules_dict
                    if item["type"] == "sentence_transformers.models.Pooling"),
                   None)
    normalize = bool(
        next((item for item in modules_dict
              if item["type"] == "sentence_transformers.models.Normalize"),
             False))

    if pooling:

        pooling_file_name = "{}/config.json".format(pooling["path"])
497
        pooling_dict = get_hf_file_to_dict(pooling_file_name, model, revision)
498
499
500
501
502
503
        pooling_type_name = next(
            (item for item, val in pooling_dict.items() if val is True), None)

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

504
        logger.info("Found pooling configuration.")
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
        return {"pooling_type": pooling_type_name, "normalize": normalize}

    return None


def get_pooling_config_name(pooling_name: str) -> Union[str, None]:
    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"

    supported_pooling_types = ['LAST', 'ALL', 'CLS', 'STEP', 'MEAN']
    pooling_type_name = pooling_name.upper()

    try:
        if pooling_type_name in supported_pooling_types:
            return pooling_type_name
    except NotImplementedError as e:
        logger.debug("Pooling type not supported", e)
        return None
    return None


532
@cache
533
def get_sentence_transformer_tokenizer_config(model: str,
534
535
                                              revision: Optional[str] = 'main'
                                              ):
536
    """
537
    Returns the tokenization configuration dictionary for a
538
539
540
    given Sentence Transformer BERT model.

    Parameters:
541
    - model (str): The name of the Sentence Transformer
542
543
544
545
546
    BERT model.
    - revision (str, optional): The revision of the m
    odel to use. Defaults to 'main'.

    Returns:
547
    - dict: A dictionary containing the configuration parameters
548
549
    for the Sentence Transformer BERT model.
    """
550
551
552
553
554
555
556
557
558
559
    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
560
561
562
563
564
565

    for config_file in sentence_transformer_config_files:
        if try_get_local_file(model=model,
                              file_name=config_file,
                              revision=revision) is not None:
            encoder_dict = get_hf_file_to_dict(config_file, model, revision)
566
567
            if encoder_dict:
                break
568

569
    if not encoder_dict and not model.startswith("/"):
570
571
572
573
        try:
            # If model is on HuggingfaceHub, get the repo files
            repo_files = list_repo_files(model,
                                         revision=revision,
574
                                         token=os.getenv('HF_TOKEN', None))
575
        except Exception:
576
577
578
579
580
581
582
583
584
            repo_files = []

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

585
586
587
    if not encoder_dict:
        return None

588
589
    logger.info("Found sentence-transformers tokenize configuration.")

590
591
592
593
594
    if all(k in encoder_dict for k in ("max_seq_length", "do_lower_case")):
        return encoder_dict
    return None


595
def maybe_register_config_serialize_by_value() -> None:
596
597
    """Try to register HF model configuration class to serialize by value

598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
        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.

        Examples:

        >>> 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.
616
617
618
619

        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
620
        class module does not need to be importable on the receiving end.
621
622

        See: https://github.com/cloudpipe/cloudpickle?tab=readme-ov-file#overriding-pickles-serialization-mechanism-for-importable-constructs
623
    """ # noqa
624
625
626
    try:
        import transformers_modules
    except ImportError:
627
        # the config does not need trust_remote_code
628
629
630
631
632
633
634
635
636
637
638
639
        return

    try:
        import cloudpickle
        cloudpickle.register_pickle_by_value(transformers_modules)

        # ray vendors its own version of cloudpickle
        from vllm.executor.ray_utils import ray
        if ray:
            ray.cloudpickle.register_pickle_by_value(transformers_modules)

        # multiprocessing uses pickle to serialize arguments when using spawn
640
        # Here we get pickle to use cloudpickle to serialize config objects
641
642
643
644
645
646
        # that contain instances of the custom config class to avoid
        # serialization problems if the generated module (and model) has a `.`
        # in its name
        import multiprocessing
        import pickle

647
        from vllm.config import VllmConfig
648

649
650
        def _reduce_config(config: VllmConfig):
            return (pickle.loads, (cloudpickle.dumps(config), ))
651

652
        multiprocessing.reducer.register(VllmConfig, _reduce_config)
653
654
655
656
657
658
659
660
661
662

    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`",
            exc_info=e)


663
def load_params_config(model: Union[str, Path], revision: Optional[str],
664
                       **kwargs) -> PretrainedConfig:
665
666
667
668
669
    # This function loads a params.json config which
    # should be used when loading models in mistral format

    config_file_name = "params.json"

670
    config_dict = get_hf_file_to_dict(config_file_name, model, revision)
671
672
673
674
675
    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 "
            f"and if the config file exists.")
676
    assert isinstance(config_dict, dict)
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692

    config_mapping = {
        "dim": "hidden_size",
        "norm_eps": "rms_norm_eps",
        "n_kv_heads": "num_key_value_heads",
        "n_layers": "num_hidden_layers",
        "n_heads": "num_attention_heads",
        "hidden_dim": "intermediate_size",
    }

    def recurse_elems(elem: Any):
        if isinstance(elem, dict):
            config_dict = {}
            for key, value in elem.items():
                key = config_mapping.get(key, key)
                config_dict[key] = recurse_elems(value)
693
694

            return config_dict
695
696
697
698
699
700
701
        else:
            return elem

    config_dict["model_type"] = config_dict.get("model_type", "transformer")
    config_dict["hidden_act"] = config_dict.get("activation", "silu")
    config_dict["tie_word_embeddings"] = config_dict.get(
        "tie_embeddings", False)
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719

    if config_dict.get("max_position_embeddings") is None:
        max_position_embeddings = 128_000
        try:
            trust_remote_code_val = kwargs.get("trust_remote_code", False)
            hf_config = get_config(model=model,
                                   trust_remote_code=trust_remote_code_val,
                                   revision=revision,
                                   config_format=ConfigFormat.HF)
            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",
                exc_info=e)
        config_dict["max_position_embeddings"] = max_position_embeddings
720

721
722
723
724
725
726
727
728
    if config_dict.get("quantization") is not None:
        quantization = config_dict.get("quantization", {})
        if quantization.get("qformat_weight") == "fp8_e4m3":
            # This maps to the FP8 static per-tensor quantization scheme
            quantization_config = {
                "quant_method": "fp8",
                "activation_scheme": "static"
            }
729
730
731
        elif quantization.get("quant_method") == "compressed-tensors":
            # Pass through the quantization config to compressed-tensors
            quantization_config = quantization
732
733
734
735
736
737
738
739
740
741
        else:
            raise ValueError(
                f"Found unknown quantization='{quantization}' in config")

        config_dict["quantization_config"] = quantization_config

    config_type: Literal["text",
                         "multimodal"] = "multimodal" if config_dict.get(
                             "vision_encoder") is not None else "text"

Patrick von Platen's avatar
Patrick von Platen committed
742
743
744
745
746
    if config_dict.get("moe") is not None:
        config_dict["architectures"] = ["MixtralForCausalLM"]
    else:
        config_dict["architectures"] = ["MistralForCausalLM"]

747
    if config_type == "multimodal":
Patrick von Platen's avatar
Patrick von Platen committed
748
        multimodal_config = config_dict.pop("vision_encoder")
749
        quantization_config = config_dict.get("quantization_config", {})
750

Patrick von Platen's avatar
Patrick von Platen committed
751
752
753
754
755
756
        config_dict = {
            "text_config": config_dict,
            "vision_config": multimodal_config
        }
        config_dict["architectures"] = ["PixtralForConditionalGeneration"]
        config_dict["model_type"] = "pixtral"
757
758
        if quantization_config:
            config_dict["quantization_config"] = quantization_config
Patrick von Platen's avatar
Patrick von Platen committed
759

760
761
    config_dict.update(kwargs)

762
763
764
765
766
767
768
769
770
771
    config_dict = recurse_elems(config_dict)

    # transform to HF config format
    if config_type == "multimodal":
        config_dict["text_config"] = PretrainedConfig(
            **config_dict["text_config"])
        config_dict["vision_config"] = PretrainedConfig(
            **config_dict["vision_config"])

    return PretrainedConfig(**config_dict)
772
773


774
775
def get_hf_image_processor_config(
    model: Union[str, Path],
776
    hf_token: Optional[Union[bool, str]] = None,
777
778
    revision: Optional[str] = None,
    **kwargs,
779
) -> dict[str, Any]:
780
    # ModelScope does not provide an interface for image_processor
781
    if envs.VLLM_USE_MODELSCOPE:
782
        return dict()
783
    # Separate model folder from file path for GGUF models
784
    if check_gguf_file(model):
785
        model = Path(model).parent
786
787
788
789
    return get_image_processor_config(model,
                                      token=hf_token,
                                      revision=revision,
                                      **kwargs)
790
791


792
793
def get_hf_text_config(config: PretrainedConfig):
    """Get the "sub" config relevant to llm for multi modal models.
794
    No op for pure text models.
795
    """
796
797
    # This block should be unnecessary after https://github.com/huggingface/transformers/pull/37517
    if hasattr(config, "thinker_config"):
798
799
800
801
        # TODO(suyang.fy): Refactor code.
        #  For Qwen2.5-Omni, change hf_text_config to
        #  thinker_config.text_config.
        return config.thinker_config.text_config
802
803
804
805
806
807
808
809
810
811

    text_config = config.get_text_config()

    if text_config is not config:
        # The code operates under the assumption that text_config should have
        # `num_attention_heads` (among others). Assert here to fail early
        # if transformers config doesn't align with this assumption.
        assert hasattr(text_config, "num_attention_heads")

    return text_config
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833


def try_get_generation_config(
    model: str,
    trust_remote_code: bool,
    revision: Optional[str] = None,
) -> Optional[GenerationConfig]:
    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,
            )
            return GenerationConfig.from_model_config(config)
        except OSError:  # Not found
            return None
834
835
836
837


def get_cross_encoder_activation_function(config: PretrainedConfig):

838
839
840
841
842
843
844
    function_name: Optional[str] = None
    if hasattr(config, "sentence_transformers") and "activation_fn" in \
        config.sentence_transformers:
        function_name = config.sentence_transformers["activation_fn"]

    elif (hasattr(config, "sbert_ce_default_activation_function")
          and config.sbert_ce_default_activation_function is not None):
845
        function_name = config.sbert_ce_default_activation_function
846
847

    if function_name is not None:
848
849
850
851
852
853
        assert function_name.startswith("torch.nn.modules."), \
            "Loading of activation functions is restricted to " \
            "torch.nn.modules for security reasons"
        return resolve_obj_by_qualname(function_name)()
    else:
        return nn.Sigmoid() if config.num_labels == 1 else nn.Identity()
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872


def try_get_safetensors_metadata(
    model: str,
    *,
    revision: Optional[str] = None,
):
    get_safetensors_metadata_partial = partial(
        get_safetensors_metadata,
        model,
        revision=revision,
        token=os.getenv('HF_TOKEN', None),
    )

    try:
        return with_retry(get_safetensors_metadata_partial,
                          "Error retrieving safetensors")
    except Exception:
        return None
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887


def try_get_tokenizer_config(
    pretrained_model_name_or_path: Union[str, os.PathLike],
    trust_remote_code: bool,
    revision: Optional[str] = None,
) -> Optional[dict[str, Any]]:
    try:
        return get_tokenizer_config(
            pretrained_model_name_or_path,
            trust_remote_code=trust_remote_code,
            revision=revision,
        )
    except Exception:
        return None