exist_ok mds_kwargs not passable to dataframe_to_mds()

Hello.

I want dataframe_to_mds() to overwrite the output directory if the directory already exists.

The documentation for mds_kwargs states to "Refer to https://docs.mosaicml.com/projects/streaming/en/stable/api_reference/generated/streaming.MDSWriter.html"

When looking at the documentation for MDSWriter(), one of the accepted **kwargs is the "exist_ok" field. I tried setting this field to True, but when looking at my stacktrace, it showed that it was still defaulting to False.

In dataframe_to_mds.py line 286, CloudUploader.get() is called, but exist_ok is not passed through. Therefore, CloudUploader uses the default value of False.

In line 299 of dataframe_to_mds.py, the write_mds() function is applied. The exist_ok field is correctly passed in this function (which calls MDSWriter), but it is not passed in line 286 to CloudUploader.

To me, this seems like a simple fix of passing the exist_ok field from **kwargs in the CloudUploader() call in line 286.

That said, I am not too familiar with DataLoaders for distributed training so there may be some reasoning that exist_ok should always be false for CloudUploader() in this context?

Read dataframe_to_mds.py around lines 286 and 299, comparing the CloudUploader.get() call with write_mds() and how each receives kwargs. Confirm whether exist_ok should apply to CloudUploader in this context, then verify that an existing output directory can be overwritten without breaking the dataframe_to_mds() flow.