jinaai

/

jina-embeddings-v4

like 272

Follow

Jina AI 1.21k

Hello, I encountered an error. Can you help me?

from sentence_transformers import SentenceTransformer

# 1. Load a pretrained Sentence Transformer model
model = SentenceTransformer("/data/models/jina-embeddings-v4", trust_remote_code=True)

# The sentences to encode
sentences = [
    "The weather is lovely today.",
    "It's so sunny outside!",
    "He drove to the stadium.",
]

# 2. Calculate embeddings by calling model.encode()
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 384]

# 3. Calculate the embedding similarities
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.6660, 0.1046],
#         [0.6660, 1.0000, 0.1411],
#         [0.1046, 0.1411, 1.0000]])

When i run the above, it shows the following runtime:

You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`.
---------------------------------------------------------------------------
Exception                                 Traceback (most recent call last)
Cell In[1], line 4
      1 from sentence_transformers import SentenceTransformer
      3 # 1. Load a pretrained Sentence Transformer model
----> 4 model = SentenceTransformer("/data/models/jina-embeddings-v4", trust_remote_code=True)
      6 # The sentences to encode
      7 sentences = [
      8     "The weather is lovely today.",
      9     "It's so sunny outside!",
     10     "He drove to the stadium.",
     11 ]

File /opt/anaconda3/lib/python3.11/site-packages/sentence_transformers/SentenceTransformer.py:309, in SentenceTransformer.__init__(self, model_name_or_path, modules, device, prompts, default_prompt_name, similarity_fn_name, cache_folder, trust_remote_code, revision, local_files_only, token, use_auth_token, truncate_dim, model_kwargs, tokenizer_kwargs, config_kwargs, model_card_data, backend)
    300         model_name_or_path = __MODEL_HUB_ORGANIZATION__ + "/" + model_name_or_path
    302 if is_sentence_transformer_model(
    303     model_name_or_path,
    304     token,
   (...)
    307     local_files_only=local_files_only,
    308 ):
--> 309     modules, self.module_kwargs = self._load_sbert_model(
    310         model_name_or_path,
    311         token=token,
    312         cache_folder=cache_folder,
    313         revision=revision,
    314         trust_remote_code=trust_remote_code,
    315         local_files_only=local_files_only,
    316         model_kwargs=model_kwargs,
    317         tokenizer_kwargs=tokenizer_kwargs,
    318         config_kwargs=config_kwargs,
    319     )
    320 else:
    321     modules = self._load_auto_model(
    322         model_name_or_path,
    323         token=token,
   (...)
    330         config_kwargs=config_kwargs,
    331     )

File /opt/anaconda3/lib/python3.11/site-packages/sentence_transformers/SentenceTransformer.py:1808, in SentenceTransformer._load_sbert_model(self, model_name_or_path, token, cache_folder, revision, trust_remote_code, local_files_only, model_kwargs, tokenizer_kwargs, config_kwargs)
   1805 # Try to initialize the module with a lot of kwargs, but only if the module supports them
   1806 # Otherwise we fall back to the load method
   1807 try:
-> 1808     module = module_class(model_name_or_path, cache_dir=cache_folder, backend=self.backend, **kwargs)
   1809 except TypeError:
   1810     module = module_class.load(model_name_or_path)

File /data/cache/huggingface/modules/transformers_modules/jina-embeddings-v4/custom_st.py:45, in Transformer.__init__(self, model_name_or_path, max_seq_length, config_args, model_args, tokenizer_args, cache_dir, backend, **kwargs)
     40 if self.default_task and self.default_task not in self.config.task_names:
     41     raise ValueError(
     42         f"Invalid task: {self.default_task}. Must be one of {self.config.task_names}."
     43     )
---> 45 self.model = AutoModel.from_pretrained(
     46     model_name_or_path, config=self.config, cache_dir=cache_dir, **model_kwargs
     47 )
     48 self.processor = AutoProcessor.from_pretrained(
     49     model_name_or_path,
     50     cache_dir=cache_dir,
     51     use_fast=True,
     52     **tokenizer_kwargs,
     53 )
     54 self.max_seq_length = max_seq_length or 8192

File /opt/anaconda3/lib/python3.11/site-packages/transformers/models/auto/auto_factory.py:564, in _BaseAutoModelClass.from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)
    562     cls.register(config.__class__, model_class, exist_ok=True)
    563     model_class = add_generation_mixin_to_remote_model(model_class)
--> 564     return model_class.from_pretrained(
    565         pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
    566     )
    567 elif type(config) in cls._model_mapping.keys():
    568     model_class = _get_model_class(config, cls._model_mapping)

File /data/cache/huggingface/modules/transformers_modules/jina-embeddings-v4/modeling_jina_embeddings_v4.py:565, in JinaEmbeddingsV4Model.from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs)
    562 if not is_flash_attn_2_available():
    563     kwargs["attn_implementation"] = "sdpa"
--> 565 base_model = super().from_pretrained(
    566     pretrained_model_name_or_path, *args, **kwargs
    567 )
    569 # Configure adapter directory
    570 if os.path.isdir(base_model.name_or_path):

File /opt/anaconda3/lib/python3.11/site-packages/transformers/modeling_utils.py:309, in restore_default_torch_dtype.<locals>._wrapper(*args, **kwargs)
    307 old_dtype = torch.get_default_dtype()
    308 try:
--> 309     return func(*args, **kwargs)
    310 finally:
    311     torch.set_default_dtype(old_dtype)

File /opt/anaconda3/lib/python3.11/site-packages/transformers/modeling_utils.py:4508, in PreTrainedModel.from_pretrained(cls, pretrained_model_name_or_path, config, cache_dir, ignore_mismatched_sizes, force_download, local_files_only, token, revision, use_safetensors, weights_only, *model_args, **kwargs)
   4499     config = cls._autoset_attn_implementation(
   4500         config,
   4501         use_flash_attention_2=use_flash_attention_2,
   4502         torch_dtype=torch_dtype,
   4503         device_map=device_map,
   4504     )
   4506 with ContextManagers(model_init_context):
   4507     # Let's make sure we don't run the init function of buffer modules
-> 4508     model = cls(config, *model_args, **model_kwargs)
   4510 # Make sure to tie the weights correctly
   4511 model.tie_weights()

File /data/cache/huggingface/modules/transformers_modules/jina-embeddings-v4/modeling_jina_embeddings_v4.py:145, in JinaEmbeddingsV4Model.__init__(self, config)
    143 self._init_projection_layer(config)
    144 self.post_init()
--> 145 self.processor = JinaEmbeddingsV4Processor.from_pretrained(
    146     self.name_or_path, trust_remote_code=True, use_fast=True
    147 )
    148 self.multi_vector_projector_dim = config.multi_vector_projector_dim
    149 self._task = None

File /opt/anaconda3/lib/python3.11/site-packages/transformers/processing_utils.py:1185, in ProcessorMixin.from_pretrained(cls, pretrained_model_name_or_path, cache_dir, force_download, local_files_only, token, revision, **kwargs)
   1182 if token is not None:
   1183     kwargs["token"] = token
-> 1185 args = cls._get_arguments_from_pretrained(pretrained_model_name_or_path, **kwargs)
   1186 processor_dict, kwargs = cls.get_processor_dict(pretrained_model_name_or_path, **kwargs)
   1187 return cls.from_args_and_dict(args, processor_dict, **kwargs)

File /opt/anaconda3/lib/python3.11/site-packages/transformers/processing_utils.py:1248, in ProcessorMixin._get_arguments_from_pretrained(cls, pretrained_model_name_or_path, **kwargs)
   1245     else:
   1246         attribute_class = cls.get_possibly_dynamic_module(class_name)
-> 1248     args.append(attribute_class.from_pretrained(pretrained_model_name_or_path, **kwargs))
   1249 return args

File /opt/anaconda3/lib/python3.11/site-packages/transformers/tokenization_utils_base.py:2025, in PreTrainedTokenizerBase.from_pretrained(cls, pretrained_model_name_or_path, cache_dir, force_download, local_files_only, token, revision, trust_remote_code, *init_inputs, **kwargs)
   2022     else:
   2023         logger.info(f"loading file {file_path} from cache at {resolved_vocab_files[file_id]}")
-> 2025 return cls._from_pretrained(
   2026     resolved_vocab_files,
   2027     pretrained_model_name_or_path,
   2028     init_configuration,
   2029     *init_inputs,
   2030     token=token,
   2031     cache_dir=cache_dir,
   2032     local_files_only=local_files_only,
   2033     _commit_hash=commit_hash,
   2034     _is_local=is_local,
   2035     trust_remote_code=trust_remote_code,
   2036     **kwargs,
   2037 )

File /opt/anaconda3/lib/python3.11/site-packages/transformers/tokenization_utils_base.py:2278, in PreTrainedTokenizerBase._from_pretrained(cls, resolved_vocab_files, pretrained_model_name_or_path, init_configuration, token, cache_dir, local_files_only, _commit_hash, _is_local, trust_remote_code, *init_inputs, **kwargs)
   2276 # Instantiate the tokenizer.
   2277 try:
-> 2278     tokenizer = cls(*init_inputs, **init_kwargs)
   2279 except import_protobuf_decode_error():
   2280     logger.info(
   2281         "Unable to load tokenizer model from SPM, loading from TikToken will be attempted instead."
   2282         "(Google protobuf error: Tried to load SPM model with non-SPM vocab file).",
   2283     )

File /opt/anaconda3/lib/python3.11/site-packages/transformers/models/qwen2/tokenization_qwen2_fast.py:120, in Qwen2TokenizerFast.__init__(self, vocab_file, merges_file, tokenizer_file, unk_token, bos_token, eos_token, pad_token, **kwargs)
    109 unk_token = (
    110     AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
    111     if isinstance(unk_token, str)
    112     else unk_token
    113 )
    114 pad_token = (
    115     AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
    116     if isinstance(pad_token, str)
    117     else pad_token
    118 )
--> 120 super().__init__(
    121     vocab_file=vocab_file,
    122     merges_file=merges_file,
    123     tokenizer_file=tokenizer_file,
    124     unk_token=unk_token,
    125     bos_token=bos_token,
    126     eos_token=eos_token,
    127     pad_token=pad_token,
    128     **kwargs,
    129 )

File /opt/anaconda3/lib/python3.11/site-packages/transformers/tokenization_utils_fast.py:117, in PreTrainedTokenizerFast.__init__(self, *args, **kwargs)
    114     fast_tokenizer = copy.deepcopy(tokenizer_object)
    115 elif fast_tokenizer_file is not None and not from_slow:
    116     # We have a serialization from tokenizers which let us directly build the backend
--> 117     fast_tokenizer = TokenizerFast.from_file(fast_tokenizer_file)
    118 elif slow_tokenizer:
    119     # We need to convert a slow tokenizer to build the backend
    120     fast_tokenizer = convert_slow_tokenizer(slow_tokenizer)

Exception: expected value at line 1 column 1

This error seems to be related to Flash Attention 2. However, jina-embeddings-v3 is normal. And this problem also happens with transformers as well. Looking forward to your reply.