Allow sending directly torch/tf/np dtype for generating dummy inputs

[JingyaHuang]

What does this PR do?

Currently always via git push DTYPE_MAPPER:

optimum/optimum/utils/input_generators.py

Line 73 in 4a7cb29

class DTYPE_MAPPER:

It would be nice to directly accept the different dtypes of frameworks. Eg. I could load a model in torch.bfloat16 in this case, I would like to define the dummy inputs dtype with model.dtype instead of going through a reversed dict of the DTYPE_MAPPER.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[IlyasMoutawwakil]

Thanks ! This looks better and more flexible ! I left a couple of comments

[IlyasMoutawwakil]

tests are failing, seems like some string inputs are still being passed around

dtype = 'int64'

    @staticmethod
    def infer_framework_from_dtype(dtype):
        if is_torch_available() and isinstance(dtype, torch.dtype):
            framework = "pt"
        elif is_tf_available() and isinstance(dtype, tf.dtypes.DType):
            framework = "tf"
        elif isinstance(dtype, type):
            framework = "np"
        else:
>           raise ValueError(
                f"Unable to create a tensor/array with dtype({dtype}). Please ensure the corresponding framework is installed."
            )
E           ValueError: Unable to create a tensor/array with dtype(int64). Please ensure the corresponding framework is installed.