> we implemented customized int8 kernels for matrix multiplications and attention I would be curious how this differs from [1] which is supported in Huggingfac

> we implemented customized int8 kernels for matrix multiplications and attention I would be curious how this differs from [1] which is supported in Huggingface’s transformers library. [1] https://arxiv.org/abs/2208.07339 https://arxiv.org/abs/2208.07339