a90eb94ca9
* CUDA: add fused rope * move k forward_expand up * create helper function instead of re-using params * make assert statement more in line with comment * rope_norm: coalesced writes to global mem