Comment by cjbprime

Looks like its only for 16/32-bit ops, which are almost irrelevant to LLM inference, but are relevant to training.