microsoft

/

Florence-2-large-ft

like 379

Follow

Microsoft 18.3k

at the begining of the month i started working on finetuning florence 2 and with great success i managed, i had 2 notebooks used to finetune and to test inference.

sadly when i tried today it seems none of them work anymore because of this error
'Florence2ForConditionalGeneration' object has no attribute '_supports_sdpa'

happens always when i try this

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[5], line 8
      5 device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
      7 try:
----> 8     model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype='auto', trust_remote_code=True).eval().to(device)
      9     print("Model loaded successfully to MPS.")
     10     # Add a small dummy input to trigger generation (replace with actual input if needed)

File [/usr/local/lib/python3.10/dist-packages/transformers/models/auto/auto_factory.py:593](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/usr/local/lib/python3.10/dist-packages/transformers/models/auto/auto_factory.py#line=592), in _BaseAutoModelClass.from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)
    591         model_class.register_for_auto_class(auto_class=cls)
    592     model_class = add_generation_mixin_to_remote_model(model_class)
--> 593     return model_class.from_pretrained(
    594         pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
    595     )
    596 elif type(config) in cls._model_mapping.keys():
    597     model_class = _get_model_class(config, cls._model_mapping)

File [/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py:315](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py#line=314), in restore_default_torch_dtype.<locals>._wrapper(*args, **kwargs)
    313 old_dtype = torch.get_default_dtype()
    314 try:
--> 315     return func(*args, **kwargs)
    316 finally:
    317     torch.set_default_dtype(old_dtype)

File [/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py:4927](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py#line=4926), 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)
   4924 config = copy.deepcopy(config)  # We do not want to modify the config inplace in from_pretrained.
   4925 with ContextManagers(model_init_context):
   4926     # Let's make sure we don't run the init function of buffer modules
-> 4927     model = cls(config, *model_args, **model_kwargs)
   4929 if _torch_distributed_available and device_mesh is not None:
   4930     model = distribute_model(model, distributed_config, device_mesh, tp_size)

File [~/.cache/huggingface/modules/transformers_modules/microsoft/Florence-2-base-ft/9803f52844ec1ae5df004e6089262e9a23e527fd/modeling_florence2.py:2534](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/~/.cache/huggingface/modules/transformers_modules/microsoft/Florence-2-base-ft/9803f52844ec1ae5df004e6089262e9a23e527fd/modeling_florence2.py#line=2533), in Florence2ForConditionalGeneration.__init__(self, config)
   2533 def __init__(self, config: Florence2Config):
-> 2534     super().__init__(config)
   2535     assert config.vision_config.model_type == 'davit', 'only DaViT is supported for now'
   2536     self.vision_tower = DaViT.from_config(config=config.vision_config)

File [/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py:2190](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py#line=2189), in PreTrainedModel.__init__(self, config, *inputs, **kwargs)
   2186 self.config = config
   2188 # Check the attention implementation is supported, or set it if not yet set (on the internal attr, to avoid
   2189 # setting it recursively)
-> 2190 self.config._attn_implementation_internal = self._check_and_adjust_attn_implementation(
   2191     self.config._attn_implementation, is_init_check=True
   2192 )
   2194 # for initialization of the loss
   2195 loss_type = self.__class__.__name__

File [/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py:2738](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py#line=2737), in PreTrainedModel._check_and_adjust_attn_implementation(self, attn_implementation, is_init_check)
   2735 elif applicable_attn_implementation == "sdpa":
   2736     # Sdpa is the default, so we try it and fallback to eager otherwise when not possible
   2737     try:
-> 2738         self._sdpa_can_dispatch(is_init_check)
   2739     except (ValueError, ImportError) as e:
   2740         # In this case, sdpa was requested explicitly, but we can't use it, so let's raise
   2741         if attn_implementation == "sdpa":

File [/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py:2602](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py#line=2601), in PreTrainedModel._sdpa_can_dispatch(self, is_init_check)
   2591 def _sdpa_can_dispatch(self, is_init_check: bool = False) -> bool:
   2592     """
   2593     Check the availability of SDPA for a given model.
   2594 
   (...)
   2600             before instantiating the full models if we know that the model does not support the requested attention.
   2601     """
-> 2602     if not self._supports_sdpa:
   2603         raise ValueError(
   2604             f"{self.__class__.__name__} does not support an attention implementation through torch.nn.functional.scaled_dot_product_attention yet."
   2605             " Please request the support for this architecture: https://github.com/huggingface/transformers/issues/28005. If you believe"
   2606             ' this error is a bug, please open an issue in Transformers GitHub repository and load your model with the argument `attn_implementation="eager"` meanwhile. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="eager")`'
   2607         )
   2608     if not is_torch_sdpa_available():

File [/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py:1688](https://qheip44eaodzmj-8888.proxy.runpod.net/lab/tree/workspace/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py#line=1687), in Module.__getattr__(self, name)
   1686     if name in modules:
   1687         return modules[name]
-> 1688 raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")

AttributeError: 'Florence2ForConditionalGeneration' object has no attribute '_supports_sdpa'

or

AttributeError                            Traceback (most recent call last)
[/tmp/ipython-input-5-2911465734.py](https://localhost:8080/#) in <cell line: 0>()
      6 
      7 try:
----> 8     model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype='auto', trust_remote_code=True).eval().to(device)
      9     print("Model loaded successfully to MPS.")
     10     # Add a small dummy input to trigger generation (replace with actual input if needed)

7 frames
[/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in __getattr__(self, name)
   1938             if name in modules:
   1939                 return modules[name]
-> 1940         raise AttributeError(
   1941             f"'{type(self).__name__}' object has no attribute '{name}'"
   1942         )

AttributeError: 'Florence2ForConditionalGeneration' object has no attribute '_supports_sdpa'

happens even in the official inference example notebook