With the tensor subclass guide published in #391, FP6-LLM implementation, which is currently a Linear module replacement, should migrate to tensor subclass too.