3 ms·
Again, the problem is custom kernels in CUDA. It’s not straightforward for many applications (LLMs are probably the most straightforward).
by mistymountains 2y ago
Again, the problem is custom kernels in CUDA. It’s not straightforward for many applications (LLMs are probably the most straightforward).