Currently, TGI does not support FP8. We have raised [issue](https://github.com/huggingface/text-generation-inference/issues/2654) about it.