Comment by mistymountains

Again, the problem is custom kernels in CUDA. It’s not straightforward for many applications (LLMs are probably the most straightforward).