Performance
agilob
•
7mo ago
•
90%
LLaMA Now Goes Faster on CPUs
justine.lolMy kernels go 2x faster than MKL for matrices that fit in L2 cache, which makes them a work in progress, since the speedup works best for prompts having fewer than 1,000 tokens.
Comments 0