AI_hacker··on Llama 3 implemented in pure NumPyHow does the performance of llama3.np compare to other implementations, especially considering it's a pure NumPy implementation?