DirectML storage compatibility issues

[sqxccdy]

Feature request

I'm trying to run transformers models using DirectML, and it works fine in most cases. However, since Microsoft's DirectML does not support storage, this causes an error whenever an untyped_storage method is involved.

Package versions installed:

torch~=2.3.1
torch_directml=0.2.2.dev240614
transformers=4.42.3

Motivation

compatibility issues

Your contribution

I hope there is a way to provide a custom device context method without modifying the transformers source code. In my current use case, I can write compatible code for storage and skip this part.

[amyeroberts]

Hi @sqxccdy, thanks for opening this feature request.

Could you provide some more details? Without a reproducible code snippet, explanation of the observed and expected behaviour, and all the relevant information about the error e.g. full error traceback there's not much we can do.