Cast inputs type to dictionary.

[ostrowskimarcin]

I run some tests of hf_clip model using torchbench.py (https://github.com/pytorch/pytorch/blob/main/benchmarks/dynamo/torchbench.py). This script at one point clones input data with custom function (https://github.com/pytorch/pytorch/blob/main/torch/_dynamo/utils.py, L1926):

def clone_inputs(example_inputs):
    res: Union[dict[Any, Any], list[Any]]
    if type(example_inputs) is dict:
        res = dict(example_inputs)
        for key, value in res.items():
            if isinstance(value, tuple):
                res[key] = clone_inputs(value)
            else:
                assert isinstance(value, torch.Tensor), type(value)
                res[key] = clone_input(value)
        return res

    res = list(example_inputs)
    for i in range(len(res)):
        if isinstance(res[i], torch.Tensor):
            res[i] = clone_input(res[i])
    return res

hf_clip model generates example data using 'CLIPProcessor' class imported from transformers repo. We can use this data as a dictionary, but officialy its type is BatchEncoding (<class 'transformers.tokenization_utils_base.BatchEncoding'>). When clone_inputs() compares input types it doesn't see it as a dict type. Hence, it casts the data to list and returns only its keys leading to error: AttributeError: 'str' object has no attribute 'shape'

I proposed simple casting to dict in model's init script. This fixed the issue and doesn't affect other test scripts.

[ostrowskimarcin]

@xuzhao9

[ostrowskimarcin]

@xuzhao9