Fix max_memory example on README

[MilesCranmer]

Running the current README example on max_memory results in this strange error:

File ~/miniforge3/envs/llms/lib/python3.9/site-packages/accelerate/utils/modeling.py:807, in get_balanced_memory(model, max_memory, no_split_module_classes, dtype, special_dtypes, low_zero)
    805 # Get default / clean up max_memory
    806 user_not_set_max_memory = max_memory is None
--> 807 max_memory = get_max_memory(max_memory)
    809 if not is_xpu_available():
    810     num_devices = len([d for d in max_memory if torch.device(d).type == "cuda" and max_memory[d] > 0])

File ~/miniforge3/envs/llms/lib/python3.9/site-packages/accelerate/utils/modeling.py:693, in get_max_memory(max_memory)
    690     return max_memory
    692 for key in max_memory:
--> 693     if isinstance(max_memory[key], str):
    694         max_memory[key] = convert_file_size_to_int(max_memory[key])
    696 # Need to sort the device by type to make sure that we allocate the gpu first.
    697 # As gpu/xpu are represented by int, we need to sort them first.

TypeError: string indices must be integers

This is because the new max_memory syntax in accelerate expects a dictionary: https://huggingface.co/docs/accelerate/v0.25.0/en/concept_guides/big_model_inference#designing-a-device-map.

This PR also:

• Updates the README example to account for multiple devices
• Fixes the model name not being enclosed in matching quotes

[TimDettmers]

Thank you, Miles! Accelerate updated the behavior of max_memory. @younesbelkada can you confirm that it is not required to pass all devices together with max_memory or what is simplest usage?

[younesbelkada]

LGTM thanks !

[Titus-von-Koeller]

Thanks a lot @MilesCranmer for raising this 🤗 Looking forward to further contributions, if you feel like it :)