Is it possible to use int8 for other task?

[haritsahm]

From your paper and blog post, the quantization was tested using transformers. Is it possible to use the library for objet detection models and will it experience performance degradation?

[TimDettmers]

I have not tested this, but it should be possible for vision transformers. I known already from data that vision transformers behave the very same in terms of outliers than language model transformers. As such, I think it will work.

However, it will not straightforwardly work for convolutional networks, since the current implementation just supports the replacement of torch.nn.Linear layers and not convolutional layers with the 8-bit equivalent.