Unrecognized configuration class <class 'transformers.models.qwen2_vl.configuration_qwen2_vl.Qwen2VLConfig'> for this kind of AutoModel: AutoModelForCausalLM.

[Jun-Howie]

已经从源码编译了 transformers

我的推理代码如下 def load(self): from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor from transformers.generation import GenerationConfig

    if self._check_tensorizer_integrity():
        self._model, self._tokenizer = self._load_tensorizer(
            code_revision=self.model_spec.model_revision
        )
        self._apply_lora()
        return

    device = self._pytorch_model_config.get("device", "auto")
    device = select_device(device)
    self._device = device
    # for multiple GPU, set back to auto to make multiple devices work
    device = "auto" if device == "cuda" else device

    self._tokenizer = AutoProcessor.from_pretrained(
        self.model_path,
        trust_remote_code=True
    )
    self._model = Qwen2VLForConditionalGeneration.from_pretrained(
        self.model_path,
        device_map=device,
        trust_remote_code=True
    ).eval()

报错信息如下

model = AutoModelForCausalLM.from_pretrained(

File "/miniconda/envs/comfyui/lib/python3.10/site-packages/transformers/models/auto/auto_factory.py", line 567, in from_pretrained raise ValueError( ValueError: [address=0.0.0.0:33177, pid=22511] Unrecognized configuration class <class 'transformers.models.qwen2_vl.configuration_qwen2_vl.Qwen2VLConfig'> for this kind of AutoModel: AutoModelForCausalLM. Model type should be one of BartConfig, BertConfig, BertGenerationConfig, BigBirdConfig, BigBirdPegasusConfig, BioGptConfig, BlenderbotConfig, BlenderbotSmallConfig, BloomConfig, CamembertConfig, LlamaConfig, CodeGenConfig, CohereConfig, CpmAntConfig, CTRLConfig, Data2VecTextConfig, DbrxConfig, ElectraConfig, ErnieConfig, FalconConfig, FalconMambaConfig, FuyuConfig, GemmaConfig, Gemma2Config, GitConfig, GPT2Config, GPT2Config, GPTBigCodeConfig, GPTNeoConfig, GPTNeoXConfig, GPTNeoXJapaneseConfig, GPTJConfig, GraniteConfig, JambaConfig, JetMoeConfig, LlamaConfig, MambaConfig, Mamba2Config, MarianConfig, MBartConfig, MegaConfig, MegatronBertConfig, MistralConfig, MixtralConfig, MptConfig, MusicgenConfig, MusicgenMelodyConfig, MvpConfig, NemotronConfig, OlmoConfig, OpenLlamaConfig, OpenAIGPTConfig, OPTConfig, PegasusConfig, PersimmonConfig, PhiConfig, Phi3Config, PLBartConfig, ProphetNetConfig, QDQBertConfig, Qwen2Config, Qwen2MoeConfig, RecurrentGemmaConfig, ReformerConfig, RemBertConfig, RobertaConfig, RobertaPreLayerNormConfig, RoCBertConfig, RoFormerConfig, RwkvConfig, Speech2Text2Config, StableLmConfig, Starcoder2Config, TransfoXLConfig, TrOCRConfig, WhisperConfig, XGLMConfig, XLMConfig, XLMProphetNetConfig, XLMRobertaConfig, XLMRobertaXLConfig, XLNetConfig, XmodConfig.

[simonJJJ]

what is the version of your transformers?

[Jun-Howie]

transformers 4.45.0.dev0

[JumpingRain]

Same question, transformers 4.45.0.dev0 accelerate-0.33.0

[ruizheliUOA]

When I tried code from the Qwen/Qwen2-VL-7B-Instruct huggingface website top right command:

# Load model directly
from transformers import AutoProcessor, AutoModelForSeq2SeqLM

processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
model = AutoModelForSeq2SeqLM.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")

I had a similar error for AutoModelForSeq2SeqLM: ValueError: Unrecognized configuration class <class 'transformers.models.qwen2_vl.configuration_qwen2_vl.Qwen2VLConfig'> for this kind of AutoModel: AutoModelForSeq2SeqLM. Model type should be one of BartConfig, BigBirdPegasusConfig, BlenderbotConfig, BlenderbotSmallConfig, EncoderDecoderConfig, FSMTConfig, GPTSanJapaneseConfig, LEDConfig, LongT5Config, M2M100Config, MarianConfig, MBartConfig, MT5Config, MvpConfig, NllbMoeConfig, PegasusConfig, PegasusXConfig, PLBartConfig, ProphetNetConfig, Qwen2AudioConfig, SeamlessM4TConfig, SeamlessM4Tv2Config, SwitchTransformersConfig, T5Config, UMT5Config, XLMProphetNetConfig.

My package version: transformers 4.45.0.dev0 accelerate 0.27.2

[kq-chen]

Qwen2VLForConditionalGeneration can't be load from AutoModelForCausalLM or AutoModelForSeq2SeqLM. Instead, it's under AutoModelForVision2Seq according to the transformers conventions.

you can use AutoModelForVision2Seq as follows:

from transformers import AutoModelForVision2Seq
model = AutoModelForVision2Seq.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")

Alternatively, you can directly use:

from transformers import Qwen2VLForConditionalGeneration
model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")

[kq-chen]

@Jun-Howie It seems that the error in your code occurs in AutoModelForCausalLM.from_pretrained, rather than loading the model from Qwen2VLForConditionalGeneration. If your framework always loads models from AutoModelForCausalLM (which is technically incorrect) and you cannot modify the framework's code, you can temporarily register Qwen2VLForConditionalGeneration under AutoModelForCausalLM:

# register before xxxx.from_pretrained
from transformers import AutoModelForCausalLM, Qwen2VLConfig, Qwen2VLForConditionalGeneration
AutoModelForCausalLM.register(config_class=Qwen2VLConfig, model_class=Qwen2VLForConditionalGeneration)

# now you can load Qwen2VLForConditionalGeneration from AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")

However, please note that this approach does not conform to standard practices and should only be used temporarily.

[kq-chen]

The example here is wrong. (I am not sure if this is a bug of transformers wibsite.