问题描述
model=AutoModelForCausalLM.from_pretrained(^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\AppData\Roaming\Python\Python312\site-packages\transformers\models\auto\auto_factory.py",line597,infrom_pretrainedreturnmodel_class.from_pretrained(^^^^^^^^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\AppData\Roaming\Python\Python312\site-packages\transformers\modeling_utils.py",line277,in_wrapperreturnfunc(*args,**kwargs)^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\AppData\Roaming\Python\Python312\site-packages\transformers\modeling_utils.py",line4971,infrom_pretrained model=cls(config,*model_args,**model_kwargs)^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\.cache\huggingface\modules\transformers_modules\Youtu_hyphen_VL_hyphen_4B_hyphen_Instruct\modeling_youtu_vl.py",line1173,in__init__ self.siglip2=Siglip2VisionModel._from_config(config.vision_config)^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\AppData\Roaming\Python\Python312\site-packages\transformers\modeling_utils.py",line277,in_wrapperreturnfunc(*args,**kwargs)^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\AppData\Roaming\Python\Python312\site-packages\transformers\modeling_utils.py",line2311,in_from_config model=cls(config,**kwargs)^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\.cache\huggingface\modules\transformers_modules\Youtu_hyphen_VL_hyphen_4B_hyphen_Instruct\modeling_siglip2.py",line1284,in__init__ self.vision_model=Siglip2VisionTransformer(config)^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\.cache\huggingface\modules\transformers_modules\Youtu_hyphen_VL_hyphen_4B_hyphen_Instruct\modeling_siglip2.py",line817,in__init__ self.encoder=Siglip2Encoder(config)^^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\.cache\huggingface\modules\transformers_modules\Youtu_hyphen_VL_hyphen_4B_hyphen_Instruct\modeling_siglip2.py",line604,in__init__ self.layers=nn.ModuleList([Siglip2EncoderLayer(config)for_inrange(config.num_hidden_layers)])^^^^^^^^^^^^^^^^^^^^^^^^^^^File"C:\Users\wh109\.cache\huggingface\modules\transformers_modules\Youtu_hyphen_VL_hyphen_4B_hyphen_Instruct\modeling_siglip2.py",line534,in__init__ self.self_attn=VISION_ATTENTION_CLASSES[config._attn_implementation](config=config)~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^KeyError:'sdpa'解决方法
将attn_implementation设置为eager，
attn_implementation="eager",# 关键参数代码如下：
fromtransformersimportAutoModelForCausalLM,AutoProcessor,AutoConfigimporttorch# 方法1：通过 config 设置config=AutoConfig.from_pretrained("Youtu-VL-4B-Instruct")config._attn_implementation="eager"# 使用标准注意力机制model=AutoModelForCausalLM.from_pretrained("Youtu-VL-4B-Instruct",config=config,torch_dtype=torch.bfloat16,# 或者 torch.float16device_map="auto")# 方法2：在 from_pretrained 中直接指定model=AutoModelForCausalLM.from_pretrained("Youtu-VL-4B-Instruct",torch_dtype=torch.bfloat16,attn_implementation="eager",# 关键参数device_map="auto")
